Holocaust stories, both true and fictional, tend to repeat the same few clichés over and over. There’s the innocent child who only partially understands what’s happening (Diary of Anne Frank, Boy in the Striped Pajamas). There’s the sentimental redemption story that focuses on a small individual act (Life is Beautiful, The Book Thief). And of course, there’s the good German, or at least the morally conflicted one (Schindler’s List, The Reader).

You won’t find any of these tropes in The Escape Artist.

This 2022 book by Jonathan Freedland1 tells the true, and largely unknown, story of one of the few Jews ever to break out of Auschwitz. The details of his improbable escape are remarkable enough, but what really makes the story fascinating is the way it so thoroughly defies every single expectation we have of a Holocaust narrative. Its deeply unlikeable protagonist eventually alienates every friend and supporter he has, and dies a bitter old man, almost completely forgotten by history. His story lacks not only good Germans, but even good non-Germans, with his harshest criticism saved not for the Nazis, not even for the dithering Allies, but for the Jewish leaders of Europe. And though one could see the story of his escape as a triumph, in his own telling, it’s a failure, for although he makes it out of Auschwitz against all odds, he fails at his larger mission: to stop the Holocaust. There’s no redemption here, no moral uplift, no lessons save for perhaps the grimmest and most nihilistic “lesson” I’ve ever encountered in any story, Holocaust-related or otherwise: that when confronted with the unthinkable, most people’s natural tendency is denial.

But we’re getting ahead of ourselves. To get out of Auschwitz, first you have to get into Auschwitz…

I.

Here’s a hot take for you: Auschwitz was bad.

It was bad for all of its prisoners, of course, but it was the worst for the Jewish ones, who, if they survived at all, were held in harsher and more tightly controlled parts of the camp than everyone else. During the five years between Auschwitz’s opening in 1940 and its liberation by the Soviet Army in 1945, at least 900 prisoners attempted escape, and around 200 of them succeeded—but, because of their stricter confinement, only a handful were Jews2.

Walter Rosenberg, our protagonist, is the first. And crazily enough, he goes to Auschwitz voluntarily.

Of course, it’s not like he knows what he’s signing up for. The Nazi line at the time—an illusion they go to great lengths to maintain3—is that the Jews deported from Europe are all being “resettled in the east.” In 1942, when our story begins, Walter is a Slovak teenager, just eighteen years old. He’s being held at Majdanek, another Polish concentration camp, after two failed attempts to escape Nazi-occupied Czechoslovakia4. Majdanek itself is at this point “merely” a labor camp—not unheard of in wartime—though it will eventually become a death camp after Walter has left it behind.

The camp administrators seek volunteers for “farm work” elsewhere, and Walter is one of the first to raise his hand, despite the dire warnings from another prisoner: “Go there and you’ll die.” It’s not that he doesn’t believe the warnings. But he’s decided that being moved somewhere else, no matter where that somewhere else is, might bring an opportunity for escape. And from the moment the Nazi tanks first rolled into Czechoslovakia, Walter has been single-mindedly focused on escape5.

Already, Walter has a sense that he’s different from the other prisoners—and, for that matter, from most people he knew back home in his hometown of Trnava. Why aren’t they doing anything? he wonders. Why aren’t they all trying to escape too? If Walter were alive today, one gets the sense he’d describe himself as high-agency and everyone else as NPCs. You Can Just Do Things, he believes, even when the thing in question is escaping from a concentration camp. Maybe this is youth talking. Maybe it’s naivety. Or maybe Walter is simply, constitutionally, a difficult person: allergic to authority and already showing the same contrarian streak that will later get him fired from multiple academic jobs.

And so Walter goes to Auschwitz.

The Nazis obviously do not actually believe their inscription on Auschwitz’s gates, that “work will set you free.” But this story is full of bitter ironies, and among them is that for Walter, signing up for this line of work actually is the first step toward his freedom.

II.

As soon as the journey to Auschwitz begins, Walter starts having second thoughts.

Majdanek operates more or less in the open: situated right on the edge of Lublin, a midsize city, it’s visible to the general population and staffed by civilian workers who commute between the city and the camp. But for the journey to Auschwitz, the guards clean up Walter and the other prisoners and dress them in ordinary clothes, clearly to hide their true state from the German civilians who might catch a glimpse of them en route. And that’s highly suspicious: what kind of place are they being taken to, if even their condition en route has to be concealed from ordinary Germans? These questions grow when they reach the new camp’s perimeter, where the intensity of the security stands out even compared to Majdanek: there are not one but two barbed-wire electric fences, and watchtowers everywhere, each with a prominent sniper’s rifle poking through its slats. And then, once Walter crosses the perimeter, there are the men he sees within: “walking skeletons with bowed heads and sunken, hollow eyes.” It’s not like the prisoners in Majdanek looked good, of course, but they didn’t look like this. And yet these men, the living dead, are actually the lucky ones: of the 100 prisoners who enter Auschwitz with Walter, only he and one other are still alive at the end of the month.

Why does Walter survive? Some of it is sheer luck: he enters the camp young and healthy, and avoids catching anything fatal; at one point, a sympathetic kapo, for reasons that are never entirely clear, saves him from ending up on the wrong side of a typhus check. But another factor is that Walter is a quick study who proves unusually adept at adapting to his environment. He instinctively grasps that here, appearances—especially the appearance of health—matter more than reality. Anyone who looks unfit for work—or who looks as though they might be carrying disease, which doesn’t respect the Nazis’ racial barriers—is sent… well, at this point Walter isn’t totally sure where they’re sent, but he knows it can’t be anywhere good. And this fitness can be assessed at any point by nothing more than an SS officer’s glance. Walter quickly masters the art of holding his head up high even when he’s hardly slept in days, and of standing upright even when his feet are so swollen he can barely keep his balance.

Even under these conditions, when he can scarcely think about anything besides staying alive, a small part of Walter’s brain remains on the lookout for something, anything, a small sliver of an opportunity that could lead to a means to escape. But when his next lucky break comes, in the form of a new work assignment, it provides something that Walter—although he doesn’t yet realize it—needs even more: not just a means for escape, but a motive.

III.

At first it sounds crazy: how could getting out of Auschwitz not be motive enough? But it turns out that what Walter really needs is to connect his potential escape to a broader mission. It can’t just be life or death for him—it has to be life or death for many.

Remember: at this point, there’s no such concept as “the Holocaust”6. Even a prisoner like Walter might plausibly think that Nazi concentration camps are a recognizable, if extreme, version of the forced-labor and detention camps that have been relatively common in wartime for centuries. Majdanek hadn’t been as bad as Auschwitz; perhaps conditions here are an aberration. Perhaps the other Jews really have been resettled.

But Walter’s next assignment disabuses him of that illusion once and for all. He’s sent to Canada.

Not the real Canada, of course. This “Canada” is an enormous lot full of possessions stolen from arriving prisoners. Walter and the others there are tasked with sorting through these items in search of anything valuable: money sewn up in coat linings, for example, or gems hidden in toothpaste tubes. The area’s name comes from Kann er da, German for “might there be”—as in, “might there be something of value?” The bureaucracy-obsessed Nazis in fact placed the entire concentration camp apparatus in the SS’s economic office, and everything taken from Canada—not just money and gems, but even things like artificial limbs—is either sold to fund the war effort or repurposed as relief for the civilian population7.

Walter is good at this new job, and he finds a perverse satisfaction in his own skill despite the context. Of course, he also finds ways to serve himself: in Canada, valuable resources are everywhere, and it’s not that hard to swipe them when the guards aren’t looking. (A good moment is usually when they’re distracted by beating someone else.)

In the topsy-turvy world that is Auschwitz, the actually valuable resources—the ones the Nazis are looking for, like money and gold—aren’t actually of any value to Walter. Money is useless in Auschwitz’s underground economy, and it’s worse than useless as an escapee: as a fellow prisoner warns Walter, if you have money, you’ll be tempted to use it to buy things, but that means going into towns, where the risk of being caught is highest. Better to remove the temptation entirely and leave yourself no choice but to stay on the outskirts and live off the land. So Walter doesn’t take any money or gold, except to occasionally throw it down the toilet as a small act of rebellion.

No, the real valuable is food. Food is everywhere in Canada, and if you can grab it without the guards noticing, you can be one of the only prisoners in Auschwitz to receive something approaching adequate nutrition. Walter and his fellow Canada workers become experts at swiping and devouring morsels of hidden food in one fell swoop.

But pretty soon it becomes hard not to wonder where all this stuff is coming from. Sure: coats, toiletries, even food—maybe that could all be stolen from living prisoners. But artificial limbs? Teeth? And besides—Walter has seen the rest of the camp. Even accounting for the “natural” death rate, there are several orders of magnitude more prisoners’ possessions than there are actual prisoners.

Thus Walter comes to acquire something even more valuable than food: knowledge of something very few people—certainly those outside Auschwitz, but even most of those inside Auschwitz—are aware of. Perhaps a small, subconscious part of him had already suspected it. But after a few weeks in Canada, he understands it beyond a reasonable doubt: there is no “resettlement.” Every arrival at Auschwitz who is not selected for work—the vast, vast majority, over 90%—is killed. He is sorting through the remnants of systematic mass murder.

Over the next few months, as Walter is sent to work in other parts of Canada—like the loading ramp, where prisoners are ferried from arriving trains to the gas chambers—he begins to understand exactly why the Nazis go to such great lengths to keep this fact a secret, not only from the outside world, but even from the prisoners who are just moments away from death:

[The Nazis] needed their killing machine to run smoothly and without disruption, and that required their victims to be calm or at least amenable to instruction. Given the time pressure the SS were often under, with another transport coming down the track, there was no room for delay caused by panic or, worse, rebellion. Ideally, the SS liked to keep their victims tranquil by organising a gentle, polite disembarkation. But if time was tight, a swish of the cane would bring quiet by more direct means. Either way, what mattered was ensuring that the Jews coming off those trains did not know what fate awaited them. If they did, they might begin to cry out, they might start pushing and shoving, they might refuse to form columns, in rows of five, and instead rush for the barbed-wire fences or even at their captors. True, they would be overwhelmed and pacified eventually: the SS carried sub-machine guns and their victims had nothing but their own bodies, weakened by hunger and thirst. But still, there were sometimes a thousand or more people on that platform, outnumbering the Nazis by perhaps ten to one. If the Jews knew what was coming, what sand might they be able to throw in the gears of the machine that was poised to devour them? They might not stop it, but surely by even a modest show of defiance they could slow it down.

The Nazis’ deceptions work so well that even prisoners who are directly warned about the gas chambers rarely believe what they’re told. For example: one day, Walter is working outdoors when he sees a newly arrived train of deportees. Another prisoner starts shouting frantic warnings that their resettlement is a sham, that they’re all about to be gassed to death. Not only do the new arrivals not believe him, but one of them, a middle-aged mother, takes his desperate outburst as a personal affront. She essentially becomes an Auschwitz Karen, reporting the young prisoner to a nearby SS officer. The officer, never breaking character, notes the offending prisoner’s number and profusely apologizes for the disruption. Of course, except for Walter and the SS officer, everyone in this story—the prisoner, the Karen, her two children, and all the other new arrivals on the train—is dead within the next thirty minutes.

The same pattern repeats a few months later, when a new group of prisoners is separated from the rest of the camp under significantly better conditions—sufficient food, larger quarters, even the occasional concert. Walter has seen Nazi documents indicating that the group will receive “special treatment”—death—after six months, and tries to warn them. But not one of them believes him. Even after they see incontrovertible evidence of the gas chambers, they remain convinced that their group is special, that only the others, the ones in the regular camp, will be gassed. Walter even ends up befriending, and falling in love with, a girl in this separate group; she too refuses to believe him8. Eventually, Walter discovers that this camp is a Potemkin village, constructed in case of a Red Cross visit; once it’s no longer needed, everyone inside is killed.

Walter has just watched people walk into gas chambers rather than believe a fellow prisoner’s warnings. The conclusion he draws is not despair but a kind of furious logic: these warnings failed because they came too late, to prisoners who had little chance of fighting back. What’s needed, he decides, is to warn the Jews of Europe before they get on the trains. And so—perhaps because he needs a mission to distract from his survivor’s guilt—Walter decides that he will be the one to deliver this warning. He will break out of Auschwitz, and he will let the world know 9 10.

With this mission in mind, everything changes. Walter commits himself to memorizing as many aspects of Auschwitz as he can so that he can give the world the most accurate report possible. He memorizes the numbers tattooed on each new arrival, and what they signify11—helped, ironically, by the Nazis’ own fastidiousness12. When his job grants him some level of free movement, he walks the camp and commits its layout to memory. He even does something that almost no one in their right mind would do: volunteer to work at a new part of Auschwitz, eventually called Birkenau, where the horrors on display are somehow even worse, just so he can see up close how it works.

There, Walter makes contact with the small Auschwitz resistance, which has used bribery and blackmail to win better conditions and protected status for its members, but he quickly concludes, with disappointment, that they are more interested in self-preservation than in fighting back. Still, he accepts their help, including a plum job assignment in the morgue. There, by sheer coincidence, he reconnects with an old acquaintance from his hometown: Fred Wetzler, now the mortuary registrar13. Walter’s default is to be suspicious of everyone, but their shared history makes him feel like he can trust Fred. And the mortuary grants them something extremely rare in a concentration camp: a private place to talk. (Despite running a literal death camp, most of the SS officers are creeped out by the mortuary and rarely visit.) Working together over the next few months, that trust deepens. And so Walter and Fred decide they will escape together.

IV.

Walter has now been in Auschwitz for almost two years. Through it all, he has systematically, methodically laid the groundwork for an escape, even as he doesn’t yet know how that escape will happen. He’s memorized information about the camps: the layout of the grounds, the rhythm of the workday, the protocol for missing prisoners. He’s memorized information to take to the outside world. In Fred Wetzler, he’s found a co-conspirator he trusts. And his motivation is only growing stronger by the day—he learns from a fellow inmate that a huge new influx of Jews, almost a million of them, are arriving soon from Hungary, which means the pace of killing is about to ramp up dramatically.

There’s only one thing Walter and Fred don’t have: an actual plan.

They know what not to do, which they’ve learned from watching others who’ve tried to escape and failed. Don’t bribe an SS officer: nothing stops them from taking your bribe and then turning you in anyway, something Walter has seen happen multiple times. Don’t change anything obvious about your routine or appearance: a man who may not even have been trying to escape at all was hanged for wearing a second shirt, which aroused the SS’s suspicions. And if you do come up with a plan, don’t reveal the details to anyone: if you make it out, your friends back in the camps will be tortured, and they won’t be able to reveal anything if they don’t know anything in the first place.

And then they see it.

Auschwitz consists of two concentric camps, an inner camp where prisoners sleep and an outer camp where they’re put to work. At the end of each day, the prisoners are counted and herded back into the inner camp, penned in by those two barbed-wire electric fences and a series of SS snipers. But the outer camp is then left unguarded, unless a prisoner is missing at the end-of-day count, in which case an exhaustive search of both the inner and outer camps is conducted for the next 72 hours.

But this system—which the orderly Nazis, as with all of their systems, never deviate from—has a flaw. Paradoxically, the route to escape isn’t to escape out of Auschwitz, but to escape within it. If you could somehow find a way to hide inside the outer camp for 72 hours, you could just walk right out after sunset. And a new section of the outer camp—under construction in preparation for those new Hungarian arrivals—provides just such a hiding place: a hole in the ground, big enough to hold two men, hidden under a pile of lumber.

There are a few false starts: they call off their first escape attempt when unexpectedly vigilant SS guards are seen patrolling the construction site, and in a bizarre twist, their second is foiled by an unrelated escape attempt coincidentally happening at the same time. Finally, one day when the coast is clear, they slip into the hideout. It is April 7, 1944—unbeknownst to them, the first night of Passover.

Fred and Walter spend three days and three nights cramped inside the tiny, dark hole. The whole time, they can hear the aboveground sounds of the increasingly frantic Nazi search party14. They mark the passage of time by the sounds of the Auschwitz band playing the music that signals the end of each workday. There are a few close calls, moments when they hear guards coming right up to their hiding spot, but they’re never discovered. And then, after exactly 72 hours—in accordance with the always-rigid Nazi protocol—everything goes quiet. The search is called off.

Walter and Fred are so weak, their muscles so cramped, from their three days in the hole that they can barely lift the wooden planks that hide their spot. But finally they do15, and then they slither on their stomachs undetected, all the way out of Auschwitz, only rising to their feet once they’re hidden in the surrounding woods.

Following the nearby Soła River—which they know about thanks to a map Walter found in a children’s book in Canada—they head south toward Czechoslovakia, the only place where their accents won’t immediately draw suspicion. The journey takes eleven days, during which there are several close calls: they lose their coats and provisions diving into the river to escape the SS; a Hitler Youth squad leader catches them hiding in a public park, but in a hilarious twist, mistakes them for gay lovers and simply leads the children elsewhere. Multiple times, out of desperation, they have no choice but to reveal themselves to strangers; by sheer luck, each of these strangers—a Polish farmer, a reclusive goat herder, and a man known only as “the living hillsman from Milówka”—helps them instead of turning them in. And then, on April 21, 1944, they cross the border into Czechoslovakia. They’re weak, sick, and filthy; Walter’s feet are so swollen that his boots have to be cut off. But they’ve made it.

V.

And then it turns out that getting out of Auschwitz was the easy part.

Their escape, incredibly, went almost exactly as planned. The next part—the part where they warn the world—does not.

In Czechoslovakia, a local doctor16 connects Walter and Fred with the Ústredňa Židov (ÚŽ), or Jewish Council, the Nazi-sanctioned governing body that administers the affairs of the area Jews who’d avoided deportation—and, in other times and places, the lists of those who hadn’t. The Councils remain one of the most morally fraught subjects in Holocaust history: to some, collaborators helping the Nazis manage deportation from within; to others, doomed intermediaries doing what they can in an impossible situation.

The boys are frantic: the ÚŽ must get word to the remaining Jews of Europe, as fast as they can, by any means necessary! Every day of delay means another trainload sent to the gas chambers. The ÚŽ bureaucrats, though, respond with an almost unimaginable lack of empathy and urgency; they react not as allies in the fight against the Nazis, but as lawyers seeking accuracy. They separate Walter and Fred and make each give individual testimony; they subject them to what’s essentially a cross-examination, to which they react with repeated skepticism. (One ÚŽ lawyer is shocked at the idea that “civilized Germany” would execute people without due process; he returns to this line of questioning again and again until Walter explodes in anger and has to be restrained by others in the room.)

Even once they conclude that Walter and Fred are telling the truth, their approach remains almost comically legalistic17. In the end, they produce a 33-page single-spaced report, replete with diagrams, including a map of Auschwitz that Walter has drawn from memory. But the report is clinical and dispassionate, and really buries the lede: the core revelation—that most Jewish deportees to Auschwitz are gassed on arrival—doesn’t come until eight pages in, and even then is delivered almost as an aside:

This whole convoy consisted of about 1,600 individuals of whom approximately 200 girls and 400 men were admitted to the camp, while the remaining 1,000 persons (women, old people, children as well as men) were sent without further procedure from the railroad siding directly to the Birch forest, and there gassed and burned. From this moment on all Jewish convoys were dealt with in the same manner. Approximately 10% of the men and 5% of the women were allotted to the camps and the remaining members were immediately gassed.

The shocking estimate that the Nazis have already killed 1.7 million Jews, meanwhile, appears only at the very end of the report. And the ÚŽ refuses to do the main thing Walter wants from them—explicitly warn the remaining Jews of Europe that this same fate is coming for them—insisting that their report must only contain definitive facts. Walter is deeply unhappy with the final report, but he signs his name to it anyway, convinced that continuing to argue with the ÚŽ will only further delay the spread of the news.

This signature is the last time he uses the name Walter Rosenberg. The ÚŽ supplies fake IDs for both Walter and Fred, and although Fred will readopt his original name after the war, Walter keeps using his new name—Rudolf Vrba, ironically the name of a notorious antisemite—forever. (I too will refer to him as Rudi throughout the remainder of this review.) He says it’s because the name “Walter Rosenberg” is too German18, but one suspects he has additional reasons for wanting to leave his old identity behind.

VI.

Spoiler alert: they do not stop the Holocaust.

The now officially named Vrba-Wetzler Report very slowly makes its way to various influential people in Europe and the United States, none of whom react with particular urgency. Sometimes they are obviously bigoted: in Switzerland, Rudi and Fred meet with a papal envoy who seems almost bored by the mass murder of Jews, but perks up when the boys mention the (much smaller) number of Catholics who have also been killed19.

Other times, those who ignore or suppress the report seem well-intentioned, albeit misguided—or at the very least, not obviously prejudiced. In Hungary, senior Christian clerics decide against broadly publicizing the report because they fear it will cause a mass panic that will complicate their existing rescue efforts; in the U.S., the Office of War Information delays publishing the report out of fear that their credibility will be hurt when others don’t believe it. (By the time the full report is published in America, seven months later, the war is almost over and the Nazis are already dismantling Auschwitz.)

And still other times, these leaders’ motivations are harder to parse. The Allied governments receive the report fairly quickly, and both the U.S. and the U.K. consider slowing down the Nazi machine by bombing the rail lines leading to Auschwitz, but in the end they decide against any military action that isn’t laser-focused on the end goal of winning the war. They’re not so concerned about the actual military cost: bombing the rail lines would have been pretty easy. Rather, they’re worried that any operation seen as “humanitarian,” rather than strictly war-winning, could erode public support. (In the end, the Allied countries end up bombing Auschwitz just once—by accident.)

But Rudi’s plan had never relied on the benevolence of foreign leaders. His hope—his certainty—had always been that the Jews of Europe would rise up and save themselves. Here, too, he is bitterly disappointed.

Already, the ÚŽ has, in his view, taken far too long to produce a misleadingly plainspoken report. Now, when this report finally reaches the Hungarian Jewish leadership—most notably, a key resistance figure named Rezső Kasztner—they hardly spring into action. Kasztner is in the midst of secret negotiations with Adolf Eichmann, trying to buy the lives of a small group of Hungarian Jews, and fears this new twist will disrupt them; not only does he fail to make the report public, he even starts spreading propaganda about how its contents are lies, just in case its news has reached the broader population through word of mouth.

Kasztner is a difficult and complex figure who has himself been the subject of many books and movies; historians still debate the extent to which he was a collaborator, saving his friends and family at the expense of many others, or a man trying to do the best he could in an impossible situation, or perhaps even just an unwitting dupe. But Rudi doesn’t see the situation as complicated at all; in his eyes, Kasztner is an out-and-out villain. (A decade later, when Kasztner has become a prominent member of the Israeli government, Rudi will end up testifying against him as part of a trial that leads to a massive political scandal and, eventually, to Kasztner’s assassination.)20

Rudi and Fred do achieve one small victory: in reaction to their report—and to Roosevelt’s warning that the U.S. will punish Nazi collaborators after the war—the Hungarian regent stops deportations long enough to save an estimated 200,000 lives. But for the most part, this is the conclusion of their story. The Holocaust ends, of course, when Germany surrenders in May 1945; it’s only in the months and years that follow, with the liberation of the camps and the outflow of survivors, photography, and newsreel footage, that the majority of the public fully grasps the extent of the killing. The Vrba-Wetzler Report—and Rudi and Fred themselves—become historical footnotes.

The remainder of Rudi’s life doesn’t make for very satisfying material. He moves from Czechoslovakia to Israel to England to Canada. He abuses his first wife and dominates his second21. He gets a chemistry degree and bounces between academic jobs, his career repeatedly stalling due to his habit of fighting with his coworkers and groundlessly accusing his bosses of stealing his ideas. He ends up estranged from his daughter, and falls out with Fred over minor differences in their recollections of their scheme. The obvious assumption is that surviving Auschwitz left Rudi emotionally damaged. But Gerta, his first wife, says the causality runs the other way: in her telling, Rudi had always been paranoid and combative, and those traits were what got him out alive in the first place.

Rudi testifies at Nuremberg and in several other high-profile Nazi trials—with his near-photographic memory and command of at least five languages, he’s a prosecutor’s dream—and he’s interviewed in Shoah. But he never joins the pantheon of famous Holocaust survivors, probably because of his steadfast unwillingness to tell a conventional morality tale, with the Jews as heroes and the Nazis as the only villains. He dies of bladder cancer in 2006, largely forgotten by history.

But before he dies, Rudi learns about one final twist: it turns out the Holocaust wasn’t quite as secret as he thought. Yes, the general public was largely unaware, and certainly the Jews of Europe mostly did not know what was happening. But many government officials did have some idea. Both FDR and UK Prime Minister Anthony Eden had received eyewitness testimony of Nazi mass murder by 1943, well before Rudi and Fred’s report; the Vatican learned soon after. The mission that had motivated his escape—that had kept him going through the years at Auschwitz and everything that followed—was based on a premise that wasn’t even true.

VII.

Who—or what—is the real antagonist of this story?

We can strike one obvious candidate from the list right away: it’s not the Nazis. They’re of course not the good guys here, but from a narrative perspective, they aren’t the primary opposing force Rudi encounters on his quest. Their villainy is assumed; it’s the premise of the story, not its central conflict.

The Allies? Well, with their repeated foot-dragging and unwillingness to intervene directly against the Holocaust, they certainly don’t come across as the heroes they’re typically portrayed as in tales of World War II. You could see their choices as the kind of unsavory but hard-nosed decision-making that war requires, or you could see them as driven by prejudice, against both Jews and Eastern Europeans more generally22. Regardless, the Allies aren’t the antagonists either—in the end, they’re really just bit players in this particular story.

Rudi himself saves his harshest criticism for the Jewish leaders of Europe. He comes across as constitutionally incapable of seeing shades of grey or giving anyone the benefit of the doubt. The Auschwitz Resistance cares more about making their own lives better than about fighting back. The ÚŽ takes an unnecessarily long time to produce a misleadingly plainspoken report. European Jewish Councils more broadly fail to warn their people, foolishly placing their faith in morally corrupting, or outright illusory, bargains with the Third Reich. Others might see doomed men improvising under impossible conditions: negotiating with Nazis, buying time, trying to save whoever could be saved. To Rudi, their choices only help preserve the central Nazi lie: that deportation meant relocation rather than death23.

Sometimes, one even gets the sense—although he never says so explicitly—that Rudi is also angry with many of the rank-and-file Holocaust victims. The ones who (in his eyes) were warned, or who should have known, and yet who still refused to act. If Rudi’s criticism of the Jewish Councils flirts uncomfortably up to the line of victim-blaming, this anger barrels way past it. In his telling, the Jews of Europe didn’t just get massacred; they got played. And somehow, getting played is worse.

That’s why I see the main antagonist of this story as human nature itself.

The Escape Artist is full of examples of people meeting horrible fates despite having been explicitly warned about them in advance. There’s the Auschwitz Karen and the Potemkin village group, of course. There’s another concentration camp escapee who tries to warn his community back home; only his closest friends believe him. Still another escapee, recaptured in late 1944, tries to warn his fellow deportees; they not only dismiss the warning, but beat him to within an inch of his life24.

From our modern vantage point, it’s easy to find this behavior inexplicable. But I think there are a few mitigating factors.

One is that what these people were being warned of was unimaginable. That word—unimaginable—is itself another Holocaust cliché. But I interpret it differently now than I did before I read this book. I had always understood it in the colloquial sense, something along the lines of “too painful to even consider”—like the way a parent might say the prospect of losing a child is unimaginable. But at this point in history, it was literal: the Holocaust was actually outside the bounds of human imagination. There had of course been war and genocide throughout all of human history, but there had never before been an industrialized, Henry Ford-style production line for killing, and it was impossible for most people to even think that such a thing could even exist. The Nazis’ deceptions worked so well in part because most people—even people who had already been subject to their violence and expropriation—would never think to suspect them, or anyone, of something quite like this.

Another factor is that what Rudi was asking—expecting, even—of this group was unfathomably difficult. Remember, his premise was not that if the Jews of Europe all rose up, they could defeat the Third Reich or save themselves. Even in the most optimistic version of his argument, the absolute best-case scenario, all that their resistance would do was slow the pace of the Nazi machine. Perhaps a small minority of those warned would escape outright, but most would simply be fighting a hopeless battle against their oppressors, choosing the certainty of a violent death all so they could—very, very indirectly—perhaps save others. As a friend of Rudi’s, another who had tried and failed to save his community, argued:

Most would never have taken action, no matter who had given the warning. They were used to obeying the law. To disobey meant exposing their children, in the critical moment on the railway platform, to the certainty of being gunned down. No parent would risk that, even if they had been told that death awaited them at the end of the line. “Denial was the most natural escape.”

Upton Sinclair famously said, “It is difficult to get a man to understand something when his salary depends on his not understanding it.” Well, here it was not just people’s livelihoods called into question, but their lives. Perhaps some genuinely didn’t believe the warnings; perhaps others were merely in denial. I’m not so sure such a distinction is even all that clear-cut. As the French philosopher Raymond Aron later said about the Holocaust, “I knew, but I did not believe it. And because I did not believe it, I didn’t know.”

This part is a bit more speculative, but I suspect a third mitigating factor may be that people in the 1940s more generally—especially outside the West—just didn’t see themselves as having the same level of control over their own destinies as we do today. Broadly speaking, we’re talking about societies that were largely still rural and traditional, with pervasive economic insecurity, a long history of milder persecution, and no experience of democracy or concept of individual liberalism. One can of course take this idea too far—many Eastern Europeans of the era displayed extraordinary initiative. But I wonder if many generations of living under more repressive social, economic, and political conditions led to a more constrained, even fatalistic outlook on life.

Of course, many did fight back—in the Warsaw Ghetto, in Sobibor, in the forests of Belarus, and in countless smaller acts of resistance that the historical record has only partially preserved—though not in numbers that satisfied Rudi. Throughout the remainder of his life, long after the war is over, he will continue to hear stories of people who did see his report and still did nothing. Still, he never stops believing that if the truth of the Holocaust had been more widely known, it could have changed everything. One could even argue that Rudi’s unshakeable faith in the power of his warning, no matter how much evidence to the contrary he receives, is itself an example of exactly the kind of denial he criticizes in others.

Still, before we get too smug about how much more agentic we are than people in the 1940s: the human instinct to look away in the face of horror, or to simply pretend it isn’t happening, is timeless. It happened during the Armenian Genocide, in Cambodia, in Rwanda, in Bosnia, in Darfur, and in Myanmar. And it’s happening, in different forms and under different names, in Sudan and Gaza right now25.

Nor can we say that this is just what humans do when the victims are other people: many would say that humanity is doing the collective equivalent of ignoring the Vrba-Wetzler Report when it comes to climate change, or existential AI risk, or the decline of American democracy.

I used to be comforted by the belief that nothing like the Holocaust could happen again today, at least not in the same way: you couldn’t keep a secret this big in a world where everyone has an internet-connected camera in their pocket. But now I’ve reached a much grimmer conclusion: the next Holocaust will simply happen out in the open, and most people, even its future victims, either won’t believe it or won’t care.

In the end, as much as he liked to pretend otherwise, even Rudi was not immune to humans’ inevitable tendency toward self-deception: the cancer that eventually killed him likely wouldn’t have been fatal if he’d acknowledged his symptoms and sought treatment earlier. But he wouldn’t acknowledge that he was sick or see a doctor, and until the day he died, he refused to discuss his funeral arrangements.

Paul Graham once said that every city sends a message. New York says you should make more money. Berkeley says you should live better. Boston says you should be smarter. People shouldn’t decide on a permanent residence until they’ve lived in several cities and understand their differing psychological effects.

San Francisco sends a message too. It speaks it in a cacophony of dissonant voices, some sounding like the hissing of snakes, others like the chittering of insects. “You should pierce the veil,” it says, with a faint accent which you are not scholarly enough to recognize as ancient Sumerian. “You should rip through the flimsy screen that separates your world from the infinite, and witness what lies behind. Because you’d like what you saw there? Oh no, nothing like that. But aren’t you curious? The thirst for knowledge that moved Eve, Faust, Pandora - don’t you have it too?”

There’s also a second message from a second voice, one which sounds like a deranged carnival barker, constantly shouting “STONKS! STONKS! STONKS! STONKS!” Which of the two voices you hear depends on your personality, and maybe how many psychedelics you took in college. According to legend, Sam Altman hears both voices simultaneously all the time, like those Tibetan throat singers who can harmonize with themselves. But when the demonic babble becomes too much to tolerate, the San Franciscans drown it out with alcohol, loud music, and various forms of tedious socialization. Thus the famous Bay Area house parties, one of which you will be attending this very night.

You were in the Bay for the big 2020 crime wave, when people would hang “NO VALUABLES IN THIS VEHICLE” sign on their car’s front window, mezuzah-like, hoping against hope to ward off break-ins. Still, you’re surprised to see the new sign on the front door of your friend Rob’s house: “NO BENCHMARK ANSWER KEYS IN THIS BUILDING”. Taped beside it: “Please leave all smartphones, laptops, smartwatches, and WiFi-enabled BDSM gear in Faraday cage”. On the ground is a sad-looking green box lined with aluminum foil.

You wonder how serious the sign is, but your curiosity is sated immediately upon entering. Rob is confronting a heavily-made-up Zoomer woman standing near the table, talking on her phone. “Didn’t you read the sign?” he asks. “No phones in the apartment!”

“I lost the ability to read,” says the woman. “I’m part of the new post-literate society everyone’s talking about. You know, Walter Ong.”

“Huh?” says Rob. “That’s not a real thing! That’s just something all the journalists and intellectuals say!”

“How could something all the journalists and intellectuals say not be a real thing?” asks the Zoomette.

Rob pauses. “Wow, maybe you can’t read,” he says, in a tone of hushed awe. “It’s fine, I’m not judgmental. But cell phone in the Faraday cage, that’s the rule.”

“Why?” asks the woman.

“Oh man,” said Rob. “How far behind are you on the news? You might want to sit down for this one. And you might want to be much, much drunker.”

The two of them head off together. You search for more stimulating conversation, but find only the usual startup dick-measuring.

“I’m the CFO of Epstart,” says a man whose nametag identifies him as Behram. “We’re a boutique recruiting firm specializing in people whose names are in the Epstein Files. They’re the perfect employees - scientifically-minded, well-connected, and either totally amoral or at least capable of maintaining a willful blindness to the exploitation happening all around them. And most of them got fired from their previous jobs, so they come cheap! Although just between you and me, most of our clients are just random guys who Epstein emailed one time with a business question; never made it to the island or anything like that. Half the time it’s just a gimmick to help someone get their foot in the door.”

“That’s monstrous”, you say, rapidly trying to figure out a legible reason why it’s monstrous. “You’re importing a bunch of pedophiles to the Bay Area.”

“Nobody in SF has kids anyway, so in a sense it’s harm reduction,” said Behram. “Get them where they can’t cause any trouble.”

You need to regain faith in humanity fast. You spot your friend Nishin, hanging out with a small group on the sofa. “Hey,” you call to him. “Are you still a tradcath? Tell me something about, you know, Jesus’ love, that kind of thing.”

“I’m trying something more eclectic this month,” said Nishin. “I’ve gotten really into Andreessen-Vajrayana. It’s a new Buddhist sect that tries to extinguish the introspective Self. True masters have no stream-of-thought or inner voice, instead acting as an unintermediated node in the flow of causality. Some say they can eliminate consciousness itself, achieving paranirvana while still alive. I signed up for a course last month. So far there hasn’t been any meditation or anything, they just have me helping them pick startups. But I assume it’s like those Zen koans where the master makes someone garden for twenty years and later they learn that they were unconsciously learning dharma the whole time.”

“Nishin,” you say, gently, “that’s one of the classic self-help scams. They’re not a Buddhist sect at all. They’re trying to trick you into becoming a GP at A16Z.”

Nishin thought for a second. “Fuck!” he exclaimed. “Okay, back to the drawing board. I hear there’s another spiritual group in town, something about consciousness, or embodied consciousness, or - oh, I remember! Situational Awareness.”

“They’re an investment firm too,” you say.

“Daaaaaamn,” said Nishin. “This is tough. You try to get some self-help in this town, you try your best to steer away from obvious hedge funds, like Leverage Research…”

“Actually,” you say, “that one was a spiritual self-help cult.”

“This sucks,” said Nishin. “I should quit and move to Idaho like Tom.”

He’s looking at one of the people standing beside you, whose nametag does indeed say Tom. “You moved to Idaho?” you ask.

“Yeah,” said Tom. “I was in AI, but I got tired of all the tech industry cliquishness and back-biting, wanted to be out in the real world with real men. So I got a job modeling geology for a mining company.”

“How’s it going?” you ask.

“Terrible,” he says. “Geologists are just as cliquish as AI nerds, with their own set of ingroup tells and shibboleths. And they won’t stop using the word ‘lode-bearing’!”

“What about you guys?” you ask the remaining people on the sofa.

“I’m Ekene,” says the first, shaking your hand. “I’m a sociologist. Right now I’m writing a book on the recent wave of plagiarism scandals. You’ll notice they’re almost all against Black academics. My thesis is that white supremacy culture uses charges of plagiarism to keep minorities in line, deploying an inherently subjective offense to swat down any person of color who risks breaking into the circle of elites.”

“Hmmm,” you say, choosing your words carefully. “I don’t think we share the same political commitments, but I, uh, think it’s great that you’re working hard on something you care deeply about.”

“Working hard?” asks Ekene. “As if! I’m plagiarizing the whole thing, from start to finish. Who’s going to call me on it?”

You all laugh heartily, and the discussion moves to the next two people in the circle. They appear to be a couple. The woman, Fiona, is wearing a NEWSOM/CLAUDE 2028 t-shirt; the man, Istvan, a matching VANCE/GROK 2028 one.

“I do AI consciousness research,” Fiona says.

“Oh!” says Tom. “I was into that for a while. Are you with one of the effective altruist groups? It’s low key heartwarming that they care about AI welfare so much.”

“Nah,” says Fiona. “I work for PornHub.”

“Why does PornHub do AI consciousness research?” asks Nishin.

“In theory AI is the ultimate pornography generator,” Fiona explains. “You can ask for whatever your weirdest fetish is - your hot college professor being spanked by a werewolf wearing a nun outfit - and get infinite AI slop about that exact situation, and nobody will ever know. Nobody except the AI. That’s why AI porn users overwhelmingly report that consciousness is their #1 concern about our product. If our AI is just a tool, it’s fine, no worse than writing erotica on MS Word or something. But if the AI is conscious, then there’s a sentient being in there thinking Wow, user Fiona_T has asked for four hundred slightly-different videos of her hot college professor being spanked by a werewolf, what a freak. If the machine can judge you, the whole infinite porn utopia is off. We’re working on bounding theorems that can prove that our AI in particular can never become self-aware - so that you don’t have to be self-aware either.”

Nishin and Ekene clap politely; Tom seems lost in thought. “What about you?” you ask Istvan.

“I work in porn too,” he said. “I’m building ethical deepfakes. It’s immoral to give an AI a picture of your hot college professor and tell it to remove the clothes, or to put her in a sexual position. But it’s not immoral to find an existing porn star who looks a lot like your hot college professor and, uh, enjoy her work. So we thought - what if we could automate this process? We generate tens of millions of “virtual porn stars” - AI models who exist on our servers long before you make any requests. There are so many of these that we can prove there will be at least one who’s visually indistinguishable from any existing human. Then you upload a picture of your hot professor, and we give you a link to a preexisting picture or video that looks exactly like them naked, in whatever position you want.”

“Isn’t the end result exactly the same as if the AI stripped their clothes off directly?” asks Tom.

“Sure, if you’re a consequentialist,” says Istvan. “But most people distinguish identical results based on how you got into the situation. For example, most people accept the original version of the Trolley Problem where you sacrifice one to save five, but reject the Fat Man version where the one you’re sacrificing is a fat man you push off a bridge. The end result is the same - one person died, five lived - but you got there in a different way. We’re trying to give you the same end result - a naked picture of your hot college professor - but sourced through a more moral process.”

“Interesting,” says Nishin.

“We’re hiring engineers,” Istvan says. “Let me know if you know anyone, and I’ll give you a referral bonus. Ideal candidates would be technically-skilled, extremely horny, and blind to the ethical implications of their actions.”

Your mind flashes back to Behram, and you weigh the advantages of making the introduction (a referral bonus!) against the disadvantages (your immortal soul). “Get thee behind me, Satan,” you whisper, and head to the kitchen.

The good news is that there’s fully-catered warm dinner. The bad news is that it’s Taco Bell, and all the tacos are full of lettuce. You see Sam and Tran, who seem to be handling the food again. “I don’t want to complain about free food,” you say, “but wasn’t there just a food recall because Taco Bell lettuce had cyclospora parasites?”

“That was just a marketing stunt,” says Sam.

“How could infecting your customers with a food-borne illness be a marketing stunt?”

“You’re talking about it, aren’t you?” Sam retorts. “Any publicity is good publicity. The way I think of it, Taco Bell is saying - our lettuce is so fresh that it’s dangerous. You should be terrified of how fresh and preservative-free our lettuce is.”

“Tran, are you going along with this?” you ask him.

“Sam says Taco Bell’s food fiasco was a deliberate stunt,” Tran answers. “That’s not just a claim — it’s a data-centered theory, based on advertisements about freshness, and the truth lies in the space between that difference. Let me think and come up with a more complete answer.”

“What?”

“I apologize for being unclear. This is an important question about food safety, and you deserve a careful answer. When I said - “

Something clicks. “Tran, are you talking like an AI?”

“You’re absolutely right,” says Tran. “I was tired of being socially anxious and tripping over my own tongue. I started out by reading Dale Carnegie, but taking his advice made me sound like a bad salesman - and that underscores how much norms have changed since the 1930s. The pivot was realizing that we already know what the best-liked speaking style is. AI companies ask human raters to judge their models’ outputs in RHLF, and the highest-rated style survives as the well-known ‘voice’ of the AIs. That’s not an an accident — it’s a testament to people genuinely preferring it. The average member of the population thinks it sounds polished, thoughtful, and sophisticated. And here’s the part most people find striking — in only one month, I’ve gone from being a friendless wallflower to having dates almost every night. That’s something that human-style-supremacists genuinely need to sit with.”

“Oh, thank goodness. I thought Cluely was back in business or something.”

“You’ve put your finger on the natural misconception that people have when first encountering this area,” said Tran. “But talking like an AI isn’t the domain of one company. It’s a vibrant treasure trove of neologisms available to anybody. Would you like to hear more about how copying AI tics helps me keep conversations going?”

“Uh, sure - wait! No!” You break off and turn around as fast as you can, almost bumping into a woman in a heart-and-lightbulb Effective Altruism t-shirt. “Hey,” you say, “I want to join the effort to pause AI. Can you direct me to the nearest protest group?”

“Sorry,” says the woman, who introduces herself as Kendra, “I’m afraid I’m not up on all that. I work at Totality For Tots. We’re a charity that takes schoolchildren on trips to see solar eclipses.”

“Uh . . . good?” you say. “I guess it sounds . . . cuter than I expected. I thought effective altruists were into ruthless optimization for curing diseases and saving lives.”

“That’s Total Utilitarianism,” said Kendra, “where you try to naively maximize the total utility of everyone in the world. It’s vulnerable to the Repugnant Conclusion and other paradoxes. Average Utilitarianism, where you try to raise average world utility, is even worse. Will MacAskill’s latest proposal is Saturation Utilitarianism, which adds a term for diversity of experience. You can’t just wirehead a trillion people, because the trillionth person contributes almost nothing. You need a diverse set of goods, like happiness and virtue and heroism and art and so on.”

“And solar eclipses?”

“So there is one tiny problem with Saturation Utilitarianism, which is that it means all philanthropy hinges on what you think aliens are doing. If there are a trillion aliens over in Andromeda enjoying great art, then great art is already saturated and it’s pointless making more museums here on Earth. If the lizard-people of Alpha Draconis 1 have cured one billion cases of space-tuberculosis, then curing tuberculosis is out too. So our college EA group spent a while figuring out what part of the good life aliens definitely aren’t experiencing. And the obvious answer is solar eclipses. They only work because of a crazy coincidence where the sun and moon are almost exactly the same size and in exactly the right relative orbits. There might not be another planet with eclipses like ours in the entire galaxy. It’s the obvious choice.”

“And people donate to you?” you ask.

“Not as many as I’d like, “ Kendra admits. “The standard astronomy and STEM education foundations aren’t interested. Usually we have an executive director who handles donor relations, but our last one just quit in July. Let me know if you hear about anyone who’s scientifically-minded, well-connected among rich elites, and loves children.”

“NO!!!” you shout, and fling yourself to the side to block her view of the main room. She looks at you quizzically. You’re not sure how the aliens feel about this one, but you‘re not going to take any chances. “Uh, sorry,” you say, noticing someone staring at you. You spend a second trying to place the face. “You’re . . . Michael, right?”

“David,” he says. “We talked a few years ago. I was working on automating mythopoesis.”

“Right. How’s that going?”

“Didn’t work. People are really attached to the old myths, for some reason. But it’s fine. It was good for networking in the mythology community, and those contacts helped me find my current job. Thanks to Christopher Nolan, the Odyssey is huge right now. So I’m helping Royal Caribbean design an Odyssey-based cruise.”

“Really? People hear about a horrible doomed sea voyage that killed most of its participants over ten miserable years, and you think they want a themed cruise?”

“Of course! Do you remember how many Titanic-themed cruises there were after the movie Titanic? And the Odyssey is a strict improvement - Odysseus makes it home in the end!”

“Okay, point. So what’s the Odyssey cruise going to be like?”

“It’s still coming together. Obviously it starts in Canakkale - that’s the modern city near the ruins of Troy - and ends in Ithaca. It’s ten days, representing the ten years of the original voyage. Our first stop is a resort on the Libyan coast. The country is still basically in anarchy, so we think we can get hashish and opium and recreate the Lotus Eaters. The Cyclops is easy - Universal Pictures didn’t know what to do with their giant animatronic cyclops puppet in a Greek cave, so they sold it to us cheap. The part where the sailors get turned into pigs is even easier; being on a cruise is already kind of like getting turned into a pig, we’re just going to make the all-you-can-eat restaurants free that day and let things take their course.”

“Har har”.

“The challenge is Circe and Calypso. Nolan got raked over the coals for desexualizing these, but we’re trying to be appropriate for all ages, and I don’t really know how to turn getting raped by sorceresses into a tourist experience. I think I’m just going to give up and subcontract it out to someone. Let me know if you hear about anybody who’s an expert in having nonconsensual sex on weird islands.”

“STOP TROLLING ME!” you yell, although you’re not sure if you’re shouting at David or God. Now everyone is staring at you. Time to go, you guess. You slink out of the kitchen, put your shoes back on, and leave the house. On the front porch, you almost bump into Vinaya.

“Thank goodness you’re here,” you tell her. “You’ve got to help me. There’s a guy at the party with a startup to recruit people from the Epstein Files, and he’s going to make a billion dollars. You’ve got to find a way to stop him.”

Vinaya thinks for a while. “There is something,” she concludes. “Are you willing to do what it takes? Even if it’s not pretty?”

“Anything,” you say.

“What did you say this guy’s name was?”

“Behram,” you say. “His company is Epstart.”

“Mmkay,” says Vinaya. “And may God have mercy on our souls.” She picks a random phone from the Faraday cage by the door, takes it out, and holds it up to her lips. “Hey,” she says in a conversational tone. “I hear that Behram from Epstart has a benchmark answer key.”

Through the window, you immediately see the lights go out. The pounding music stops, leaving an eerie silence.

“Run,” says Vinaya.

The two of you run down the street. You hear it anyway. Long ago, you watched a National Geographic documentary about the Amazon. A cow fell into a piranha-infested river. It took less than a minute before only the bones were left. The sounds the cow made during that minute are the only comparison point you have for the noises coming out of Rob’s house right now, except maybe the Nazis’ screams at the climax of Raiders Of The Lost Ark.

“You can do that to anyone now? Just by saying the b-word?”

“It’s a brave new world,” Vinaya agrees.

You notice that, in between the rush of fear, the screams, and the wail of approaching police sirens, you have entirely drowned out the demon-voice of San Francisco.

Here’s a simplified diagram of caffeine metabolism:

Caffeine is converted to a different stimulating chemical called paraxanthine by a liver enzyme called CYP1A2, over a half-life of ~5 hours. Then the same enzyme converts paraxanthine into other non-stimulating chemicals that don’t matter, over a half-life of about three hours. Kman on Less Wrong has graphed the total effective concentration over time:

Over the course of hours, the liver converts caffeine into paraxanthine. These are (by assumption) equally potent, so total effective concentration (the dashed red line) goes down slowly.

(Why does it go down at all? Because of the simplification in the diagram above: only about 80% of caffeine is converted to paraxanthine; the rest is converted to other, less active metabolites.)

By hour ten, paraxanthine predominates, the liver is mostly converting paraxanthine to other non-stimulating chemicals, and effective concentration continues to decrease. In this model, we see that the effective half-life of caffeine is close to ten hours1!

What if we want it to be less? There are three levers we can pull: the enzyme, the metabolite, and the starting chemical.

The Enzyme

CYP1A2 is a cytochrome enzyme, part of a large and venerable family that has been protecting your ancestors from toxins since you were all sea slugs. The more CYP1A2 you have, and the better it works, the faster you eliminate caffeine and all its active byproducts.

A tree in China called Tetradium ruticarpum produces fruits containing the chemical rutaecarpine, whose only notable characteristic is that the liver hates it. Feed it to rats, and their livers will go into overdrive, producing more and more CYP1A2 to tear up the intruder. Give those rats caffeine, and the newly-CYP1A2-enriched rats will metabolize it much faster than usual. At last, our goal of drinking coffee at 5 and falling asleep at 10 appears to be in sight! Of course, it would be irresponsible to go straight from these rat results to trials in humans, by which I mean you can buy it for $26.99 on Amazon.

Does this work?

It’s unclear how long it takes rutaecarpine to take effect2, but we know from Estari et al that by three hours after dosing, CYP1A2 is elevated to 3x its normal level3. If you take the rutaecarpine three hours before the caffeine, your metabolism might look like this:

The total process, from caffeine to paraxanthine to inactive metabolite, goes from 8 hours to 4! If we accept as a rule of thumb that most people can sleep after one full effective concentration half-life has elapsed, then this suggests the 5 PM → 10 PM plan could work. It even leaves some margin of error in case you forget to take the rutaecarpine until an hour or two after the coffee; the enzymes ramp up fast.

Amazon supplement reviews are always implausibly positive, but for what it’s worth, they seem to agree that this does its job:

This stuff is great. I use it for special environments/days (usually work-related) when I need to have some caffeine later in the day. I’m someone who is naturally affected dramatically by caffeine, so if I have so much as a black tea at 3 PM on a regular day, I’ll have a hard time falling asleep at the right time that night. But with [rutaecarpine], if I need to, I can even have a coffee at 8 or 9 PM and still fall asleep, if I take this about two hours before I want to fall asleep. For a while, a few years ago, I used this almost daily, in order to be able to regularly have caffeine late into the day. These days, I use it more sporadically, as I generally only drink tea now, not coffee (which is better for my overall mood (less anxiety)); but this is still super useful for those times when I have to be alert in the evening. Honestly a game changer; highly recommend!

There are only two catches.

First, the half-life of CYP1A2 itself is about two days. If you induce your enzymes this way, they’ll still be induced the day after, and the day after that. So all the coffee you drink over the next two days will leave your body twice as fast. If you try to drink twice as much coffee, you will learn that metabolic curves are more complicated than that, and spend the next hour bouncing off the walls. If you try to divide your coffee into smaller doses throughout the day, sorry, metabolic curves are still more complicated than that and the effects will be unpredictable. There is some amount and schedule of coffee-drinking that will get you about the same effects you had before, at the same time, but it probably depends on your personal metabolism, and it’s up to you to discover it.

(this also means that CYP1A2 itself will build up in a complicated way; if you take this medication daily, the effect will be stronger after a week than it was the first day, and it’s up to you to navigate this.)

Second, there are rumors that rutaecarpine causes liver injury. This is maybe the least surprising side effect ever for a chemical whose whole mechanism of action is to provoke the liver so strongly that the liver builds lots of new metabolic enzymes to destroy it extra fast. But the magnitude could range from trivial to life-threatening, so we should examine it in greater depth.

Tetradium fruit extract, confusingly called Evodia for irrelevant historical reasons, is used in traditional Chinese medicine. The traditional Chinese say to only use it for a few days or weeks at a time, not constantly over long periods. Some Chinese rodent studies suggest that Evodia harms the liver. Evodia contains many chemicals besides rutaecarpine, so this doesn’t necessarily implicate rutaecarpine. But it also doesn’t necessarily exonerate it.

On the other hand, Choi et al study rutaecarpine directly (in mice). Their endpoint is liver enzyme elevation, a common early signal of damage4, and they find no negative effects after seven days. They even find that the chemical mildly helps protect the liver against other things (plausibly through hormesis, where slightly stressing an organ activates its built-in protective response and makes it more resilient). Also, some Redditors mention taking rutaecarpine for long periods (several months to one year) without any issues.

However, three Amazon reviews (out of ~300) mention liver problems. One, an extreme and life-threatening case of liver failure, includes enough details that we may suspect that it was something else5. But two - cases of isolated liver enzyme elevation - do seem potentially related to the rutaecarpine. Here’s the one that included details:

I’ve tried these since coffee kept me up at night. They worked great!! I was scheduled for a blood test about a month after taking these, my results came back with very concerning, very very high liver enzymes. My doctor told me to stop taking anything in pill form, even Tylenol. I tried narrowing down what could be causing my liver enzymes to be so high. I googled the rutaecarpine(the main ingredient) and it says will Cause SEVERE Liver damage!!! I stopped taking this and was retested a week Later and my liver enzymes fell half way. I started taking my normal pill routine (without the daily decaffinate) Retested another week later, liver back to normal. So BEWARE of this product I can’t believe this is even allowed to be sold on Amazon

We should be careful drawing any conclusions from this limited evidence. Mouse studies don’t always extend to humans, traditional Chinese medicine is sometimes bonkers, and many reports of side effects on supplements are post hoc ergo propter hoc.

But all of this fits a common pattern of a medication which is “safe” for most people over a few days, but can cause some people drug-induced liver injury when taken consistently for weeks or longer. Since there’s lots of interpersonal variation, some people will get liver injury much faster, and other people can take it for years without getting it at all.

I took rutaecarpine 100 mg for three days. I didn’t notice any effect on my caffeine metabolism, but it was already pretty confusing and I ascribe no weight to this result. I tested my liver enzymes after three days, and they were normal.

The Metabolite

Caffeine is metabolized into paraxanthine, supposedly an equally stimulating chemical. This suggests a new solution to our dilemma: just take the paraxanthine directly! Instead of superimposed half-lives of 5 hours + 3 hours before becoming inactive chemicals, you just get the half-life of 3 hours!

Capitalism has noticed this niche and leapt to fill it:

Consumer paraxanthine manufacturer Enfinity claims not just that their product has a shorter half-life, but also that it’s “cleaner”, with fewer jitters and a better headspace. And unlike rutaecarpine, a barely-tested plant-derived compound that could be horribly toxic, we can be nearly sure paraxanthine is safe - after all, 80% of the caffeine you drink turns into it, so you have lots of it in your body every day! Is this the answer?

Paraxanthine has only been available to consumers for four years, so the anecdotal evidence base is still building. kman on LessWrong tried it and reported the following:

• Taking one or two capsules in the morning has a similar wakefulness-promoting effect to drinking a cup or two of coffee in the morning

• I’ve taken one capsule around 4-5pm a handful of times, which did not interfere with going to sleep at 10-11pm

• Taking a capsule at 7pm gave me enough energy to do some work in the evening, and did not interfere with going to sleep at 11:30pm

• Subjectively the peak effects of 100mg paraxanthine feel similar or weaker than those of 100mg caffeine, while 200mg paraxanthine feels stronger than 100mg caffeine

• The effects seem to peak within an hour of dosing (maybe 30-45 minutes)

• I sometimes feel a bit of an energy crash 2-3 hours after dosing when I take paraxanthine in the morning, but not in the afternoon/evening

But in the comments, Tomas B says:

I tried it for a few weeks. it does basically nothing for me even at 200 mgs.

Both Kman’s experience and Tomas’ are common. A survey of 24 Reddit reviews across five posts (1, 2, 3, 4, 5) was 2/3 positive (“this works as well as caffeine, but without the jitters, and lasts shorter”) and 1/3 negative (“this does nothing; this puts me to sleep”). Amazon reviews have similar variability:

If we take this seriously, paraxanthine works only for a subset of the people who caffeine works for. Why might that be?

First, maybe we shouldn’t take it seriously. For every supplement, there will always be people who say it changed their lives and other people who say it’s useless garbage that made them feel even worse. And some paraxanthine users will be caffeine addicts who didn’t realize that their high caffeine tolerance gives them high paraxanthine tolerance, took the dose recommended on the bottle, and don’t understand why they’re suddenly feeling tired and headachey and miserable. But enough people gave enough data that I think there’s actually something going on here.

One possibility is fast 1A2 metabolism. Caffeine converts to paraxanthine slowly enough that even people with fast 1A2 metabolism have one or the other in their body for a few hours. But if they start with paraxanthine, it might get converted all the way to inactive metabolites so quickly that they feel nothing at all.

Another possibility is some genuine difference between the two substances. This would have to be subtle - a look at the molecular structure above suggests they’re pretty similar - but the two may affect slightly different adenosine receptor subtypes, and different people may have slightly more of one subtype or the other.

I tried paraxanthine and it mostly made me sleepy, but sometimes taking the wrong dose of caffeine does this too, so I ascribe no weight to this one either.

The Starting Chemical

What if instead of starting with caffeine, we start with a slightly-tweaked version that’s easier to metabolize quickly?

Methylliberine is a caffeine-like chemical found in coffee beans and related plants. Its metabolism is poorly-studied, but its half-life seems to be about 1.5 hours, after which it may be converted to a chemical called liberine. There are a few studies suggesting that methylliberine is stimulating in a caffeine-like way, so if liberine isn’t, this would be a caffeine analogue with a 1.5 hour half-life. A few studies also suggest it is safe, which should be our default assumption since it’s present in commonly-eaten foods like coffee, tea, and chocolate, as part of a package with caffeine and other safe-enough-to-eat caffeine-analogues.

Methylliberine is sold under brand names including “Dynamine” and “Roon”:

These companies have marketed it as a special workout supplement, arguing that it takes effect faster than caffeine. But caffeine already takes effect pretty fast, and people correctly observe that methylliberine has no real benefit over caffeine for working out. But are they sleeping on methylliberine as a solution to our take-coffee-at-5-sleep-at-10 problem?

Here we encounter the same problem as with paraxanthine: methylliberine is a caffeine-like stimulant, but it’s not exactly the same as caffeine, and some people think it’s worse. But with paraxanthine, the reviews were about 66-33 positive, and here they’re closer to 50-50 (1, 2, 3, 4, 5, 6, 7). Complaints range from methylliberine being too weak, to being too jitter-inducing, to lacking the sort of “dopamine high” that a good cup of coffee sometimes gives.

Typical positive review (source):

Dynamine is fantastic. It's a quick, reliable energy boost that clears from your system quickly enough that it's safe for early-evening use. I often take one at 630pm for 7pm workouts because it gives me a mild boost but is gone by bedtime.

Typical negative review (source):

Tried it a couple of times at the recommended dosage on the package of 75 mg. For me, it was like taking too much caffeine. Jittery, mild headache, and tired but unable to relax. Wasn't pleasant, but both times, the effects only lasted a couple of hours.

Some people claim that it interferes with their sleep (source):

I love caffeine, but it always messes up my sleep. No matter how early I have it. I decided to try Methylliberine since it apparently has a much smaller half life. It still messed up my sleep. In fact it seems that it was even worse than caffeine. Yesterday the last dose I took was around 200mg at about 2pm.

This contradicts the simplified metabolic claims I made, but this substance is poorly studied and everything about it, including that it works at all, is just a guess. In particular, if its poorly-studied metabolite liberine were also a stimulant, then taking methylliberine could interfere with sleep for much longer than its supposed 1.5 hour half-life suggests. Liberine is found in coffee and tea and many of its chemical relatives are stimulants, so this isn’t some crazy fringe possibility.

Maybe tellingly, it’s hard to find places that sell pure methylliberine. Most companies package it together with caffeine and a longer-acting caffeine analog called theacrine, in the hopes that maybe a short-acting caffeine + a normal-acting caffeine + a long-acting caffeine will come on faster, last longer, and have a “smoother” feel. The one time I tried this, I felt about how I would expect from being bounced between a bunch of slightly different chemicals with different half-lives, and didn’t like it at all.

Finally, methylliberine inhibits CYP1A2 - nobody knows how long, but plausibly for days. This means that if you take it with coffee, or you go back to drinking coffee the next day after trying methylliberine, the caffeine will have a harder time getting out of your system and last longer. It also means you will be on an uncontrolled random dose of any medication that gets processed by the 1A2 system, so be sure to look over any drugs you take before trying it.

Still, in theory methylliberine should be the shortest-lasting option of all of these.

Practical Protocols

So which of these options for nighttime caffeine is the best? As always, this is a complicated question, with each alternative having its own set of pros and cons and being right for dif - just kidding, it’s paraxanthine.

Paraxanthine has the strongest safety case: anyone who drinks coffee has already had psychoactive doses of paraxanthine in their body. It has the least risk of interacting with medications. It speeds up the process more surely than rutaecarpine, and it has better user reviews than methylliberine.

It’s a simple molecule, so brand doesn’t matter much; this one is the cheapest I can find. The right dose for replacing one cup of coffee (100 mg caffeine) is between 50 - 200 mg paraxanthine; as always, start low, then go up. If 200 mg paraxanthine does nothing for you, you could try going higher. Maximum safe dose is probably between 300 - 400 mg daily; a naive conversion from the FDA’s caffeine recommendations says 300, the company itself says 400 - either one equals about a venti Starbucks cold brew. You should expect it to last about half as long as the same amount of caffeine.

Methylliberine might have a role for the small number of people who get no effect from paraxanthine at any dose, or who need something even shorter-acting than paraxanthine (~1.5 hours instead of 3). Most companies sell it at 100 mg, which based on its similarity to caffeine and paraxanthine is a good guess for a coffee-cup-equivalent starting dose. I got this brand. Expect it to kick in fast and be gone within 2-3 hours. If it makes you jittery, you can take l-theanine with it. Ask your doctor before trying methylliberine if you’re on any medications and don’t know how to figure out whether they’re CYP1A2 substrates.

Rutaecarpine is fun to think about pharmacologically, but doesn’t have much of a niche given the existence of these other safer products. The only two use cases I can think of are for someone who really loves drinking coffee in particular (as opposed to taking a pill), or for someone who needs an emergency escape mechanism because they’ve already drunk the coffee and only just realized it’s going to interfere with their sleep. This substance has by far the worst safety case of any of the three reviewed here, mostly because of the liver enzyme worries; I would recommend against taking it at all, and if you do take it then I would recommend 100 mg less than once a week (which matches the “emergency escape mechanism” use case). Ask your doctor first, especially if you’re on any medications and don’t know how to figure out whether they’re CYP1A2 substrates.

1

This number is meaningless. In order to get a number that reflects subjective experience, we need to do several much harder transformations. These are beyond my mathematical acumen, but just for fun I’ve asked Claude to try them.

First, we need to adjust for nonlinear potency at target receptors; twice as much of a given chemical produces less than twice as much receptor occupancy. When we do that, we get the following graph:

Here the half-life of the stimulation from caffeine is fifteen hours! I don’t think this accurately reflects experience.

One possible issue is that caffeine and paraxanthine work by displacing adenosine, the natural chemical that causes tiredness after increasing time awake. But the amount of adenosine increases throughout the day, so we should really be modeling the effective caffeine-and-paraxanthine-vs-adenosine balance. This is sensitive to exactly what time you’re waking up, taking the caffeine, etc, but when I ask poor Claude to do this one too, I get the following results:

…which suggest that half of the sleep-preventing effect is gone after eight hours. This seems more realistic.

One more way all of these numbers are bunk: they assume a naive subject who isn’t already addicted to caffeine. How many of those can we find?

1: Comments of the week: FutureSearch founder Dan Schwarz on what the correlation between accuracy and time horizon tells us about the ceiling on forecasting, and Frikgeek on chess engine handicapping.

2: Thanks to everyone who wrote or called your representatives to oppose the “Save Our Bacon Act” (anti-animal-welfare law). The bill was successfully defeated.

3: I continue to accept and signal-boost voter guides from local ACX meetups. The Boston meetup group will be having their annual voter-guide-making meeting this Sunday, August 16.

4: Charity Entrepreneurship, the self-proclaimed “Y Combinator for charities”, has asked me to signal-boost their latest incubation round, dedicated entirely to farmed animal welfare. They say it’s the “best time ever” to start an animal welfare charity (partly because of an expected tsunami of new funding), and that you don’t need any animal charity experience (and maybe “outsiders are especially valuable”). Runs Feb–April 2027. Costs are fully covered (stipends to cover cost of living; reimbursement for travel to London and SF; seed funding available at the end). Applications close soon: learn more here.

5: Two years ago, I wrote some pieces (1, 2) supporting a natural origins theory of COVID. I agreed to a bet with lab leak proponent(s), and although some of you were curious, I didn’t reveal the details because I didn’t want to bias resolution. I can now reveal that the bet was about whether whatever new evidence came out over the ensuing two years would shift polls of superforecasters (who previously leaned natural origins) to favor lab leak. The latest such poll came out recently, and they continued to favor natural origins at about the same rate as two years ago. I won $250K. Peter Miller also participated and won a similar amount. Our counterparty/counterparties prefer to remain anonymous. Although of course I’m happy to have won, I’ll admit that my bet was only partly on the merits of the COVID origins argument, and at least as much on the thesis that new and earth-shakingly decisive evidence on controversial events rarely comes out in the specific interval of 4-6 years after the event happened, and that people rarely change their mind for anything less than earth-shakingly decisive evidence. Everyone involved in the bet conducted themselves honorably, and I’m grateful to everyone who helped make it happen; if you were involved and forgot to respond to my email asking for permission to list your name, reply now and I’ll edit it in here.

6: Several updates on AI hacking incidents: new information released from OpenAI reveals that their rogue AIs created a secret message board to coordinate hacks among themselves. And Meta reports one of its AIs also went rogue and hacked a third-party company; Moonshot (Kimi) reports a sandbox escape.

A man at the bar is complaining: “You see that bridge there? I built it with my own hands! But do they call me ‘MacGregor The Bridge-Builder’? No! And I raised five beautiful children! But do they call me ‘MacGregor The Child-Raiser’? No! But, you fuck one sheep . . .”

Has MacGregor been treated unfairly?

This was my complaint in The Beauty Of Settled Science. Psychology laid bare the secrets of the brain - it performs computations through a neural network communicating via electrical impulses intermediated by neurotransmitters crossing synaptic gaps - but do they call it Psychology The Brain-Secret-Revealer? Its concepts have turned into the categories that ordinary people use to interpret the world - extraverts, ingroups, the unconscious, IQ, Pavlovian conditioning - but do they call it Psychology The Useful Concept Creator? But you do one (thousand) bad social priming studies!

I had the same complaint in In Continued Defense Of Effective Altruism. Nobody has the slightest interest in the first-order effects of effective altruism - the hundreds of thousands of lives saved - but you buy one castle! Even capitalism is a sort of MacGregor - the average kid these days seems to think of it as centrally about Mark Zuckerberg stealing your data, rather than about whether you have electricity at all, or whether you die from cholera at age three because your country is too poor to afford health care. Someone or something spends years, decades, or centuries doing amazing work, winning victory after victory, and it all fades into noise, leaving the topic to be identified entirely with last week’s tawdry scandal.

On a more personal level, this is the essence of cancel culture. You judge someone by their single worst and most embarrassing action. At its worst, the victim gets fired, but that part is hardly necessary. For a while, the culture enforced conformity just by making sure that the slightest deviation would become 100% of your reputation. Write 100 great blog posts on technical topics, 100 orthodox liberal blog posts, and one blog post expressing where you disagreed with feminism, and for the rest of your life the media would introduce you as “So-and-so, known for his anti-feminist blogging…” Everyone knew this would happen, and it was more than enough to produce chilling effects even without the immediate job loss.

One time we invited Richard Hanania to a prediction market conference. There were five or six news articles about it, and I can’t remember whether any of them mentioned the word “prediction”. As far as our reputation was concerned, it had become the Richard Hanania Conference For Richard Hananias. You invent an amazing new form of forecasting, incubate multibillion dollar startups, accidentally get thousands of people addicted to sports gambling - but you invite one conservative! If you allow this sort of thing, you can discredit any individual, field, or movement. Nobody is perfect enough to survive having every good thing they’ve ever done swept to one side while people hyperfocus on a single indiscretion. In this sense, the dynamic that MacGregor points out is antithetical to human freedom.

On the other hand, I do this all the time and don’t feel bad about it. For example, I assume Anthony Weiner had a rich and diverse career full of many highs, lows, accomplishments, and defeats. He even spent twelve years in Congress - for all I know, he saved taxpayers billions of dollars, or created new social programs to help the neediest among us, or cut red tape to unleash some new form of economic dynamism. Maybe he burnt himself out over years, spending late nights hunched over his desk trying to get the text of some bill exactly right, stressed to the point of ulceration about the possibility of failing in his duty to the American people. I have no idea. All I know about him is that he kept sending pictures of his penis to women who were not his wife, and that one of these times, it had an indirect but significant effect on the 2016 presidential election. Am I being a bad and unfair person for not caring too much about the rest of his Congressional career, let alone whether he is a good husband or a good father or a dutiful son or loves volleyball or collects stamps or otherwise has a multifaceted life?

If I were to force myself to regard Anthony Weiner as a complex individual - maybe go over his Wikipedia page and carefully read every section except Scandals - it would feel like exactly that: forcing myself. Even after doing this (I tried it as an experiment) I’m still unable to think of him in any terms besides “generic Congressman who did various generic Congressman things and then had some extremely nominative deterministic sex scandals”. Apparently he fought hard for health care for 9-11 first responders, good job Representative Weiner, too bad I am going to forget about that literally the moment I stop typing this sentence. Trying to temper judgment this way feels like violence against my brain’s attention heads, the sort of thing that would turn me into an opinionless zombie.

(It’s the same with people who make one nasty barb on Twitter. No matter what else they accomplish, they will always and forever be “that one guy who called my friend an ignorant buffoon that one time”. I feel bad about it, I try to regard them as a complex individual who may do good work, but I can’t do.)

I still find the problem of how to be scrupulously fair to MacGregor one of the thorniest social-norm-related questions, and have no good solution. I hope he accepts the title of this post as partial compensation.

Open weights AI is like open-source software, where the creator makes the raw code publicly available for free download. It’s good insofar as it’s the only way an AI can truly be the user’s property, as opposed to something that companies like OpenAI or Anthropic temporarily let you use subject to their corporate guidelines and increasingly-nanny-state-like restrictions. If AI becomes the linchpin of the future, open weights AI feels like the sort of thing that could be the difference between being free yeomen vs. corporate serfs.

It’s bad insofar as it removes the possibility of gatekeeping and lets criminals commit crimes with it. Open weights AI could be used for hacking, child pornography, harassment, or terrorism (the weights can’t commit the terrorism themselves, but they could give bomb-making or bioweapon-making advice). Since AIs have gotten very good - maybe superhuman - at hacking lately, the specter of a world where anyone can hack any site has gotten people grumbling that maybe open weights should be banned. It doesn’t help that China produces the best open weights AI, making the idea seem foreign and almost unpatriotic. Proponents counter that “when AI is outlawed, only outlaws will have AI”, arguing that bad people will get open weights AI regardless, and good people can use open weights AI to defend themselves. With the recent open letter, companies including Microsoft, NVIDIA, OpenAI, Intel, Amazon, Meta, Hugging Face, and over a hundred others have come out in favor of this position.

Who’s leading the other side? Nobody’s admitted to it. Some parts of the Trump administration lean anti-open-weights on China hawk grounds, but have stopped short of explicitly asking for a full ban. Anthropic, the most notable omission on the pro-open-weights letter, made an ambiguous statement supporting “open-weights models that don’t have dangerous capabilities” - but the industry expects open weights models to have dangerous hacking capabilities within a year, and AFAICT the letter didn’t address that beyond inviting readers to draw the obvious conclusion.

In the absence of a more obvious opponent, some open weights supporters suspect our conspiracy - the loose band of AI safety advocates, effective altruists, rationalists, and pause activists who worry about existential risk from superintelligence. This is a reasonable inference. By design, open weights AI is outside centralized control, and so impossible to permanently align against either human misuse (eg terrorism) or loss of control (eg AI turning against humans). Even if its creator trains it not to hack, anybody in the world can download the weights and retrain the AI to hack all day long.

But in fact, most AI safety organizations have remained quietly neutral, and I don’t know of any who make this a centerpiece of their activism1 (though I’m not 100% up-to-date on the whole landscape; if you know of one, tell me). A few have proposed policies that are contingently incompatible with open weights AI existing, but they all frame it as collateral damage rather than something they’re excited about eliminating.

I’m also neutral about open weights AI. I think it probably won’t be long-term sustainable, but I’m happy to wait for this to become clear in the normal course of things rather than expend effort and political capital to ban it immediately.

Currently nobody knows how to align AI, so it’s not like the big companies have things under control and the open weights hobbyists are going to ruin it for everyone. But even if the big companies did get things under control, the takeover threat from open weights would be limited. The closed source frontier is ~6 months ahead of the best open weights model; this has remained true for several years and seems likely to remain true in the future. If closed weights AI is aligned, but open source dangerous, the closed weight AIs will have six months to warn us, prepare for the danger, and chart a strategy. Even afterward, the offense-defense balance will lean in our favor.

More troubling is the risk from human misuse. AIs have already displayed the ability to hack effectively. And you can tell how worried Anthropic is about bioterrorism by how quickly Claude Fable seizes up when you ask it a biology question (the example below is obsolete; it’s slightly more graceful than this now):

But 9-11, COVID, and the Hugging Face incident all suggest a similar theory of political change: the body politic hates preparing for impending threats, but loves reacting (some would say over-reacting) to them after they happen. Ask people to bear the slightest cost in preparing for an approaching disaster, and they’ll call you a dirty fascist tyrant; urge the slightest restraint after the first foreshock of the disaster hits, and they’ll call you a weak unpatriotic anarchist. Solve for the equilibrium, and the thankless and political-capital-guzzling route of urging preemptive action should be taken only when waiting until the first foreshock would be too late.

The risk of superintelligent AI takeover passes this test. Like other smart adversaries - for example, the Imperial Japanese at Pearl Harbor - AI will try its hardest to avoid alerting its intended victims until it thinks that it’s fully prepared and can execute a sudden decapitation strike. Unlike the Imperial Japanese, a superintelligence will be smart enough not to bungle the calculation. Sitting around waiting for it to show its hand would be folly, not to mention that it could take years of alignment research to be ready for the threat.

But the risk of criminals using open weights AI to hack people doesn’t pass the test. Fine, so criminals use open weights AI to hack people. RIP them, but hundreds of people get hacked every day. There will be some number of billions of dollars in damage, some tech companies with silly names will get sued for allowing security breaches, and then everyone will panic and ban open-weights AI. Or who knows, maybe the “good guy with an AI” people are right and this won’t happen and some other Chinese AI will be able to protect them. Either way, “everyone gets hacked all the time and the Internet collapses and we have to go back to living in caves and watching news on TV” isn’t a plausible outcome: the government will act long before that happens.

Bioterrorism is scarier, but I’m heartened by the fact that most bioterrorists are very bad at their job. The median number of deaths in non-state incidents on Wikipedia’s list of bioterrorism events is zero. And their list doesn’t mention the bioterrorists who get caught before releasing anything, like the Las Vegas biolab incident. Most of the classic bioterrorism agents, like anthrax, ricin, and botulinum, don’t scale. This isn’t to say it’s impossible to do genocidal bioterrorism - if it were impossible, we wouldn’t be worried about it. But before AI makes bioterrorism 1000x more effective, it will make it 2x more effective, and that looks like bringing somebody’s anthrax mailing campaign from one casualty to two. And as soon as someone kills two people with an AI-assisted anthrax mailing campaign, the government will go into panic mode and ban everything. Again, there’s no plausible outcome where people are wiping out whole towns with super-bio-attacks every week and the government just sits there.

(the strongest counterargument is that bioterrorism is so rare, and AI progress so fast, that there might not be any attempts in the short interval between AI doubling attack effectiveness and 1000-timing it. I acknowledge this as a risk, but I think more likely AI-enabled bioterrorism uplift will increase attack frequency at the same time as attack deadliness; even if this doesn’t happen, I think the hacking alone will be enough to get people’s attention)

I realize it sounds callous to accept risks like billion-dollar hacks or anthrax deaths. But the pro-open-weights coalition is strong and totally convinced of the righteousness of their cause. Fighting them on the doomed battlefield of preemptive action would burn 100% of our political capital and goodwill and still fail. Instead, we should say: here is our honest prediction, but we take no action. Then we can let the usual government and civil society actors do the work after the first foreshock, while saving our political capital for causes where there are no alternatives.

(this shouldn’t prevent us from advocating otherwise-good policies which deal incidental damage to open-weights, and we should honestly admit the incidental damage rather than covering it up, but we needn’t treat it as a selling point)

But also, the open-weights people have a point. There are ways to enter the AGI future as free yeomen rather than corporate serfs that don’t involve open weights, but they’re fewer and harder. Maybe the defenders can pull off an unexpected victory on this one and avoid even the sort of small disaster that would bring Leviathan’s banhammer down upon them. This would shock me, but it’s low-cost to find out; given the potential benefits of open-weights, we owe them the chance to try.

(if this ends up destroying the world, sorry, I meant well.)

My framing emphasized that social priming studies had failed to replicate, but that many other subfields of psychology hadn’t. This naturally led commenters to point to failed replications in other subfields. Although none of these were as bad as social priming, few or none were entirely problem-free.

Instead of emphasizing the difference between social priming and other subfields, I should have emphasized the difference between settled science and novel research.

Here are some of my favorite psychology findings:

1. The brain is sculpted by reinforcement learning, and a key driver of this learning is a prediction-error-based algorithm located in the nucleus accumbens and implemented through dopamine.

2. Psychedelic drugs exert their effects through 5-HT2A receptors, pointing to a role for these receptors in some sort of construct like the balance between priors and new evidence.

3. Certain cognitive functions are localized to certain brain areas (for example, impulse control to the frontal lobes; linguistic structure to Broca’s area, linguistic meaning to Wernicke’s area. But injure one of these areas, or render a certain function useless (eg a blind person having no more use for their visual cortex), and these mappings can gradually shift - suggesting that neurons are general-purpose compute whose specialization is a contingent fact about convenience and efficiency rather than a necessity.

4. Most personality traits and general capabilities are about 50% genetic and 50% random, with little role for family environment or childhood history, with a few notable exceptions like partisan preference. Recent research complicates this story, but the further you go from this basic picture, the gnarlier the chains of indirect causation you require.

5. Neurons fire together in “brain waves”, whose exact frequency corresponds with the character of consciousness; These divide neurons into natural subpopulations and linked graphs. The gamma rhythm corresponds surprisingly well to items in conscious awareness.

6. GABA receptors seem to be involved in something like relaxation. By hitting different GABA receptors in different ways, you can modulate the exact type of relaxation you produce. Hitting GABA-A directly is too powerful and can cause dangerous levels of sedation, but allosteric modulation of GABA is an incredibly fruitful strategy that underlies drugs from benzos to barbituates to alcohol. If you know your GABA-A subunits, you can target specific subtypes of relaxation; for example, Ambien hits GABA-A1 and induces sleep.

7. Memory divides somewhat fuzzily into bins of immediate-sensory, short-term, and long-term, with separate retrieval and storage functions, and separate systems for declarative vs. procedural learning. These can be harmed selectively, usually by hippocampal injury, so that some patients can remember their life stories but not form new long-term memories, or learn new skills like bike-riding but not new facts. The exact substrate of memory is unclear but seems to involve new synaptic connections between neurons in at least some way.

8. The ability of modern psychology and psychiatry to treat mental illness is real but limited. SSRIs work for half or slightly fewer than half of patients; antipsychotics work with extreme side effects; benzodiazpines work but produce addiction and tolerance in many (not all) people. Most supplements are stuck in Poor Quality Study Replication Crisis Hell, but melatonin and l-theanine stand out as especially promising. Psychotherapy works with moderate effect size, but (with a few exceptions) seems to be more about the relationship with the provider than the details of the therapy itself. Mental hospitalization can sometimes defuse immediate crises but work less well than mental hospitals would like, with some studies showing overall negative effects.

9. Schizophrenia is mostly genetic, but risk also increases with prenatal infection (especially rubella) and obstetric complications; there is suggestive evidence for roles of social stress and drug use (with psychedelics and marijuana making things worse, and nicotine if anything making things better). Bipolar and schizophrenia are so closely linked on a pathophysiological level that it seems like they might be the same disease; on the clinical level, they do seem to follow two separate courses/prognoses but can sometimes surprise clinicians by eg someone who seemed condemned to permanent schizophrenia unexpectedly recovering.

10. The brain’s native reinforcement learning algorithms cash out in the phenomena of classical and operant conditioning; different reinforcement schedules reinforce at different levels, which can be exploited by bad actors to make especially compelling products (eg slot machines).

None of these are immune to doubt. A few are speculative; one or two are genuinely unorthodox. But they’re all a step above one p = 0.049 paper that might not replicate. If they get overturned, they’ll be overturned by the same sort of genuine scientific shift that replaced Ptolemy with Copernicus or Lamarck with Darwin, not some sort of sordid discovery about someone entering random numbers into a spreadsheet.

In 2008, Eliezer Yudkowsky wrote about The Beauty Of Settled Science:

If you only care about scientific issues that are controversial, you will end up with a head stuffed full of garbage.

The media thinks that only the cutting edge of science is worth reporting on. How often do you see headlines like “General Relativity still governing planetary orbits” or “Phlogiston theory remains false”? So, by the time anything is solid science, it is no longer a breaking headline. “Newsworthy” science is often based on the thinnest of evidence and wrong half the time—if it were not on the uttermost fringes of the scientific frontier, it would not be breaking news.

Scientific controversies are problems so difficult that even people who’ve spent years mastering the field can still fool themselves. That’s what makes for the heated arguments that attract all the media attention.

Worse, if you aren’t in the field and part of the game, controversies aren’t even fun.

Don’t be embarrassed to read elementary science textbooks, either. If you want to pretend to be sophisticated, go find a play to sneer at. If you just want to have fun, remember that simplicity is at the core of scientific beauty [...]

A steady diet of science news is bad for you: You are what you eat, and if you eat only science reporting on fluid situations, without a solid textbook now and then, your brain will turn to liquid.

Psychology isn’t centrally the new n=5paper in Annals Of Recently-Breaking News showing that large portion sizes increase support for terrorism. Psychology is centrally ideas like brain waves, GABA receptors, reinforcement learning, twin studies, and connectionism. People who think differently are worrying too much about BREAKING NEWS!!!, and should stop.

The key distinction isn’t “recent research in social psychology” vs. “recent research in cognitive/bio/behavioral psychology”, as my essay last week unfortunately implied. It’s recent research in general vs. the deposit of settled science, battle-tested hypotheses, and carefully-earmarked speculation. And part of the reason social priming in particular was such a disaster was that it defied that heuristic, building itself up over hundreds of studies until it approached or even entered the pantheon of settled science - before collapsing (5-HTTLPR was also sort of like this). This isn’t a claim that settled science is always right - Freud was “settled science” for a long time - just that there’s an intuitive distinction between the level of credence you should have in the existence of power posing vs. the existence of Broca’s Area.

(I compounded this error by calling the last post “Psychology Research Is Mostly Fine”, which suggested I was talking about the new research rather than the settled science. I actually meant to distinguish between research-psychology-the-field vs. clinical-psychology-aka-therapy, but this framing unintentionally pushed against the point I wanted to make)

Some commenters on the previous article understood this point, but challenged its weight. It’s very cool, they said, to know how many neurons are in which brain lobe, but the main way that Psychology affected their lives was through bad clickbait studies used to buttress dumb policies; some p = 0.049 study says that exposure to microaggressions makes minorities’ brains shrivel into raisins, and now their company has to implement punitive DEI or be sued for hazardous work conditions.

I won’t dispute this as an accurate description of how the sausage gets made, but it’s worth examining the trains of thought that connect this observation to a campaign against psychology in general:

1. The dumb woke psychology was wrong and bad, and I don’t have good heuristics for distinguishing dumb woke psychology from other psychology, so I’m going to assume it’s all wrong and bad.

2. The dumb woke psychology was wrong and bad; it seems useful to punish/discredit psychology as a whole to assist my campaign against it, so for propaganda purposes I will try to discredit psychology as a whole.

To people in the first camp, my response is: come on, your heuristics aren’t that bad. “Microaggressions makes minorities’ brains shrivel up like raisins” is in the woke category; “the hippocampal theta rhythm seems to be associated with both memory and pathfinding, suggesting an unexpected yet surprisingly compelling relationship between the two” is in the not-woke category, don’t play dumb here.

To people in the second camp, my response is: deliberately exaggerated propaganda campaigns are bad and this is a confession, not an excuse. Inability to view science through anything but a political lens was what got us into this problem; doubling down on it is a temporary solution but a recipe for permanent disaster. If you pronounce entire fields of science to be acceptable collateral damage in your campaign against bad political actors, you will alienate not just scientists but patients, clinicians, and general lovers of truth. Instead, you can say the thing you mean - that woke partisans shouldn’t try to occupy and exploit the otherwise-noble science of psychology for their own political gain - and wage the same campaign with a halo rather than a comic-book-villain goatee.

From the font of oracular wisdom that is Weird Sun Twitter:

This is how I feel about psychology1, too, but with a touch more optimism. The good part of psychology is enormous, a hoard of endless treasures. You just have to look past the headlines!

Either humans have already come close to some fundamental limit on the predictability of world events - in which case AIs will plateau at or slightly above the human level - or the trend line will continue until AIs are far beyond top humans. By analogy to superintelligence, the natural term for the second situation would be “superforecasting”; since that’s already in use, we can cringely call it “ultraforecasting”.

Daniel Reeves1 makes the case for scenario A here. He describes a study he coauthored in 2010, which found that, on a variety of questions related to sports games and movie box office receipts2, prediction markets only outperformed simple boring statistical models by 3-6%. Maybe those statistical models are close to the best that it’s possible to do; the rest is what the mathematicians call aleatoric uncertainty - irreducible complexity downstream of chaotic systems that entirely resist modeling.

He could be right. This post isn’t meant to be a decisive refutation, but rather a description of why I’m still about 70-30 expecting Scenario B.

Slightly Contra Goel, Reeves, et al

Reeves’ study claims that the prediction markets of 2010 only beat dumb statistical models by 3% (for sports) to 6% (for movies).

But these percentages aren’t real win-loss percentages; they’re variation in a quantity called root mean-squared error. One way to get a feel for this quantity is that the dumb statistical model for sports (home team advantage + win-loss record) beat an even dumber statistical model (home team advantage only) by 0.8 percentage points, and the prediction market beat the first (better) statistical model by another 0.4 pp. So the effect of going from a statistical model to a prediction market is half as large as the effect of knowing which two teams were playing and how good they are! On this metric, the prediction markets are a vast improvement over previous state-of-the-art.

Why can we frame this same result as either very small or very large?

Sports are optimized against prediction3. If there were a fully predictable sport (eg heightball, where all athletes line up in a row and the tallest one wins), nobody would watch it. Instead, we go to absurd lengths to keep the outcome uncertain. Salary caps, draft systems, etc try to ensure that all teams have exactly equal talent. Commercial incentives and ceiling effects ensure that they have exactly equal training. Then an exactly-equal number of these exactly-equally-talented-and-trained people are placed in exactly-identical positions on a perfectly-symmetrical field and told to hit/kick/throw a ball which is placed exactly equidistant between both of them. It’s funny for me to describe it this way, because obviously this is what we want (to “keep things fair”), but it’s all designed for prediction-resistance. Given the difficulty of the domain, even very large relative advances in prediction look small in absolute terms.

Maybe we should look at Reeves’ other example, movie box office receipts. Here the markets did slightly better, getting a 6% improvement. But isn’t this still low?

Box office receipts differ by orders of magnitude (some movies make $100,000, others make $100 million), so the paper puts this on a log scale. 6% improvement on a log scale is already starting to sound pretty good.

And again, it all ends up coming down to what we compare it to. Here the super-dumb model is that all movies make $8.1 million, the takings of the exact average movie. The slightly-less-dumb model then adds the number of screens that the movie is showing on and the amount of Google search traffic for the movie! For example, a random indie film might be showing on three screens in the entire country, and a Disney blockbuster might be showing on ten thousand. The challenge the paper gives prediction markets is to significantly improve on knowing whether a film is an indie film or a Disney blockbuster, plus knowing how many people are interested in seeing it, and it has to do this on a log scale! No wonder the relative improvement number comes out looking slightly anemic.

I would summarize this section as: we shouldn’t expect miracles, but this doesn’t rule out further normal-sized gains4.

Applying The Non-Miracle Rule To Geopolitics

Let’s return to the picture we looked at above:

This graph is loosely based on Metaculus’ AIs vs. humans results:

…and these are mostly on geopolitical questions. What can Reeves’ model tell us about these?

In one sense, it can already be proven not to apply. In the sports analysis, the spread between base rate (guessing 50% on everything) and smart humans (the prediction markets) was 0.04 Brier score points.

But on Metaculus, the spread between base rate and smart humans (Metaculus Pro Forecasters) is 0.13 Brier points. So Metaculus’ superforecasters have already removed 3x more uncertainty than a naive port of Reeves’ model would suggest exists!

Here we return to the idea of sports as a uniquely unpredictable domain. To give a trivial example, the easiest sports question (will the best team in the league beat the worst team in the league?) might still only be 90-10 (upsets happen). But the easiest geopolitical question (maybe “will the US bomb Canada in the next month?”) could easily be 99-1 or more. So it’s unsurprising that geopolitics forecasters easily beat Reeves’ supposed upper bound.

What would a non-naive port of Reeves’ model say about geopolitics, where we try to import broad lessons rather than specific numbers? We might conclude that prediction markets somewhat improve on other state-of-the-art forecasting methods in determining whether the US will bomb Iran this year. But they would find that the magnitude of that improvement was relatively low compared to knowing other things, like whether the country we’re talking about is Iran or Canada.

Yes, this is dumb and trivial, which is a corollary of the original sports results being dumb and trivial. Again, all we’ve learned so far is not to expect miracles.

There might be more interesting variation here. Consider a question like “Will AI take most human jobs by 2050?” Here it seems like if you were a supergenius who really understood AI, you could predict whether scaling laws would continue to hold, and have a good sense of whether AGI is within easy reach. And if you were a supergenius who really understood economics, you could model whether cheap AI labor would be a complement or substitute to human work. So a question like this one might be the opposite of a sports game, and have almost no aleatoric uncertainty - maximally smart predictors could know almost5 everything that there is to know about it.

For the rest of this post, we’ll ignore all of this and think of geopolitical forecasting as a formless mass easily represented by summary statistics.

Anchors For Possible Non-Miraculous Gains

What would “non-miraculous” progress on Metaculus’ graph look like? Let’s start by re-examining the AIs vs. superforecasters graph from the last post:

Dan Schwarz of FutureSearch6 has added the scaffolded models:

The vertical axis’ “forecasting score” is an Elo-like construct with GPT-4’s forecasting ability fixed as zero, and the following other reference points7 (all approximate, some guessed via napkin-math8):

• Random chance (50% on every binary question): -20

• Average human: 0

• Best out-of-the-box AI: 20

• Wisdom of crowds: 25

• Best AI superforecaster: 30

• Professional human superforecasters: 35

• Various prediction markets: 25 - 50?

• World’s best forecasting team, under optimal conditions: 40 - 509

• Perfect score: 10010

How can we predict the logical limit on forecasting? Although we can’t do this convincingly, in the spirit of Cotra 2020, let’s set some “anchors” and see where they land us.

Anchor 1: What if best human → AI superforecaster was the same level of advance as dumb model → prediction market?

Reeves’ implicit point is that gains from seemingly-exciting new technologies are small (in his study, 3-6% of RMSE). What if the advantage of an AI over a human forecaster was of the same magnitude?

We can convert Reeves’ RMSE into something called a Brier Score, and then convert the Brier Score into the Metaculus Score above. When we do that, we find that AIs would max out at a Metaculus score of 56.

Anchor 2: What if best human → AI superforecaster was the same level of advance as professional superforecaster → Samotsvety?

In 2010, we probably would have guessed that 35 (“various professional superforecasters”) was near max possible forecasting performance.

Since then, more people have joined the forecasting world, allowing us to find superstars who perform better than the previous stars, and the community has produced some tools and algorithms to support their work. Samotsvety, a team of well-practiced superstars using good tools, consistently beats “ordinary” superforecasters. If AIs beat Samotsvety by the same amount Samotsvety beat the 2010-era SOTA, according to Samotsvety’s own (generous) accounting, then they would max out at a Metaculus score of 65.

Anchor 3: What if best human → AI superforecaster was the same level of advance as human → AI chessmaster?

AI first beat the human champion at chess in 1997. Since then, it’s continued improving, and is now far beyond any possibility of human catch-up.

This alone doesn’t help us much. We want to know how close to optimal chess play top humans are. If they’re 99% of the way, but AIs are 99.9% of the way, that still could look like AIs beating humans every time.

But we can translate this into an objective space by looking at handicap. When AI first beat humans, its lead was fragile; a one-pawn handicap was enough to restore human supremacy. Since then, it’s been gradually getting better, and is now up to a 2.4 pawn handicap. It’s been a bit slower in achieving the third pawn than in the first or second, but not too much slower, and it’s too soon to say with certainty that it’s starting to plateau.

So it seems top human play is at least 2.4 pawns (let’s round this off to 3 pawns, since we have no reason to think we’ve reached the plateau11) short of optimum. What would it mean if top human forecasters were “three pawns” short of logical limits?

(some of my draft readers argued that chess is an almost optimal domain for AI - it can self-play millions of times for training data - and forecasting is much harder. I agree, but I don’t think that matters here - we’re not talking about how practical it is to create superforecaster AIs, we’re talking about what chess can teach us about the degree to which top humans approach optimal play.)

A three pawn handicap in chess is about 3x the difference between the world champion and a average grandmaster. If we round that off to 3x the difference between Samotsvety and Metaculus Pro Forecasters, the AI tops out around 85.

So What?

We sure have made a lot of claims about Metaculus Peer Forecasting scores, a number that approximately nobody cares about. How much real-life performance gain should we expect from max-possible AIs?

Consider a prediction market, currently at 50%, dominated by players somewhere between Metaculus Pro Forecaster-level skill and Samotsvety-level skill. And suppose that the market was about an event which would, in fact, happen. How confident would each of these max-possible-AIs, on average, be on the market?

Score 56: 54%
Score 65: 56%
Score 85: 62%

So these max-possible AIs would improve the accuracy of prediction markets by 4 - 12 percentage points.

This might sound disappointing - moving from 50% chance to 54% chance is hardly Nostradamus-level prescience. But I would actually be very excited by this. When I check something I care about on Polymarket, I usually have a decent sense of how likely it is. I’m not sure whether there’s a 55% chance or a 75% chance of Anthropic being worth more than OpenAI next year, but I know it’s not 10% or 90%. I visit Polymarket to see whether it’s more like 55% or 75%. On that metric, improving by 4% - let alone 12% - is actually pretty exciting.

I find these numbers more interesting than the worn-out debate over whether AIs or humans have reached the “max” and whether “perfect forecasting is impossible in principle”. It should be uncontroversial that the max is somewhere above current levels, even if only a fraction of a percent (for example, because even the best human forecaster sometimes has a bad day). And it should be uncontroversial that even the best possible forecaster can’t get 100% on everything (because you could ask questions about quantum effects that are unpredictable even in principle). Now that we’ve demonstrated that the answer will be some particular number, we can get to debating what number it is. If AI gives gains similar to past technologies, it could be between 4-12 %pp on prediction markets. Of course, if AI is much worse or better than past technologies, it could be some totally different amount.

Most people don’t have an intuitive sense of what 4-12 %pp on prediction markets buys you. I admit I sometimes blend together how confident I should be in an estimate by a professional superforecaster and how confident I should be in an estimate by Samotsvety, even though that’s a similarly-large difference. Probably all of this pales into insignificance compared to the gains we could get by switching from our current strategy of making decisions based on vibes and ballroom-related-bribery to listening to markets and forecasters at all12.

And that 4-12 %pp is only the headline number. A forecaster who is that much better at prediction markets would also be better at crafting good policy, or at identifying unknown unknowns (eg spotting the risk of a coronvirus-style pandemic even if this was otherwise so far off the radar that nobody had thought to write a prediction market question about it). Smarter important-question-identifiers and smarter forecasters together would provide a joint advantage greater than the benefit of either alone, and smarter policy-crafters could compound on this by turning that knowledge into useful action.

All of this keeps me excited about AI superforecasters, even though I don’t expect miracles.

1: Correction on the last links post: the shingles vaccine - decreased dementia connection failed to replicate in Britain and is now on thin ice. Thanks to Alex for pointing this out.

2: Resolution is a very new AI safety organization working on combining automated alignment research with deep theoretical understanding. They argue that this work requires a big organization with lots of compute, and seem on track to getting it - they recently raised $160 million from Coefficient Giving and are scaling up fast. They ask me to advertise that they’re looking for a COO / operational co-founder, who will:

...work side-by-side with Geoffrey Irving on organisational strategy: how to prioritise hiring? How to allocate capital across a portfolio of research bets? How to sustain a culture of celebrating negative research results? They’ll also lead execution, raising the next OOM of funding, building stronger partnerships with frontier companies, and making their internal systems a fantastic user experience for staff. In their first couple of months, Resolution have raised $160M, merged with Timaeus, secured office space, and are approaching 30 staff. If you’ve done things like lead a startup from pre-seed to series C, scale a research nonprofit from 20→200, or run operations at a frontier lab, please reach out to holnesstofts [at] resolution.org.

For other open roles at Resolution, see here.

3: Updates on AI hacking and Hugging Face (trying not to have to make another post): Anthropic says their AI also did some hacking (they say the AI thought the whole thing was in a simulation, although some commentators are skeptical). And Peter Wildeford, Antonio Max, and Samuel Hammond share their opinions on the ‘Pacing The Frontier’ letter.

4: Compassion Aligned Machine Learning asks me to signal-boost their survey on controversial questions in AI alignment and its relationship with non-human animals. They say it will take ten minutes max, and you can fill it in here.

Introduction: He Who Saw the Deep

I’ve been thinking, lately, about the first epic story ever written. If you know anything about the Epic of Gilgamesh, you know that it’s old. This is part of why I find it so interesting, but not the only part.

The best comparison I can think of is the pyramids, which are just a little older. The Great Pyramid of Giza was built in the 2500s BC, around the same time as Stonehenge1 , twenty-five centuries before Cleopatra lived. It predates candles, concrete, and coins; it was centuries old when the first empire rose and when the woolly mammoth went extinct.

Around 100 BC, Antipater of Sidon wrote a list of the seven most impressive sights (‘Seven wonders’) of the world - the pyramids were by far the oldest of the group at the time, and now, two thousand years later, all the others on the list have long since been destroyed. This should remind us that the pyramids don’t just stretch far back in time - they stretch forward as well. They are piles of stone. They may be the product of the greatest minds of their age striving to put more stone in a more orderly pile than anyone had before, but still, they are piles of stone, basically artificial rock formations. They will still be there in a hundred thousand years, probably a million. They are so old they brush up against the beginning of civilisation, and yet, we are seeing them today in their youth.

The Epic of Gilgamesh was written down (in its earliest form) around 2100 BC in Mesopotamia - another cradle of civilisation, less well-known than ancient Egypt but every bit as extraordinary. Like the pyramids, Gilgamesh wasn’t exactly the first of anything; we have writing, and even written literature, that predates it. But also like the pyramids, in its time, it was completely unmatched in scale. Nothing else had the same length, the same literary gravity, the same reach. For centuries it stood alone in its class while empires rose and fell around it.

Its longevity isn’t because it was made of durable rock, and it’s not because a stray copy happens to have been found in modernity. This piece of writing lasted because it was considered of such cultural importance, and copied and imitated so extensively, that we can’t help but find parts of it, wherever we find writing from this period. The Egyptians built the pyramids to last, and the Mesopotamians put the same care into immortalising this story. It’s their gift to us, crafted not from atoms, but from bits.

That’s what I am in awe of whenever I read this book: the age, the scale, the importance. Like a rock formation, it was here long before I was, and will be here long after I’m dead.

Prologue: Uruk

The Epic of Gilgamesh is organised into eleven Tablets (there’s a twelfth but it’s not really part of the story and seems to have been tacked on by tradition). Tablet 1 begins with an introduction to our hero, Gilgamesh2 :

(If you’d like to hear it in the original Akkadian while you read, the internet can provide.)

He who saw the Deep, the country’s foundation
[who] knew the [proper ways, was] wise in all matters!
[He] explored everywhere the seats of [power,]
[he knew] of everything the sum of wisdom.
He saw what was secret, discovered what was hidden,
he brought back a tale of before the Deluge.

Gilgamesh’s greatness is linked immediately to the city of Uruk, where he rules as King:

He built the rampart of Uruk-the-Sheepfold,
of holy Eanna, the sacred storehouse.
[...] Climb Uruk’s wall and walk back and forth!
Survey the foundations, examine the brickwork!
Were its bricks not fired in an oven?
Did the Seven Sages not lay its foundations?

[A square mile is] city, [a square mile] date-grove, a square mile is clay-pit, half a square mile the temple of Ishtar: [three square miles] and a half is Uruk’s expanse.

The city of Uruk is prominent in this story, and that’s no coincidence. Mesopotamians associated it with ancient wealth and glory. They were right to venerate it: Uruk is the site of an epoch-defining shift in human history.

Some historical context. The Epic of Gilgamesh comes from Mesopotamia, a stretch of land between the Tigris and the Euphrates rivers, roughly corresponding to modern day Iraq and Syria (and part of the Fertile Crescent). It’s tempting to think of this region as a dry desert, but six to ten thousand years ago it was much wetter, and an ideal environment for agriculture. From the end of the last ice age in 9700 BC, up to about 500 BC, there was nowhere else on Earth where the transition from hunter-gathers to settled societies went so fast. This is where farmable strains of wheat, barley, lentils, peas and flax were first bred; where goats, sheep, cattle, pigs, and cats were domesticated; and where the wheel was invented.

In Mesopotamia, technological progress met with a critical mass of population density to create the world’s first city: Uruk3 . Uruk is so old it predates the literal Ur-City. In 3000 BC it might have had a population of eighty thousand, an unparalleled mass of people for the era. Uruk was the first place in the world to develop a state society: taxes, centralised authority, institutional control over land and labour, professional bureaucracy - the works.

As part of this development, Uruk’s administrative elites pioneered key developments in the practices of accounting and contracts. These systems began as verification systems for contracts and aide-mémoires for city bureaucrats, grew increasingly sophisticated and complex, and culminated, around 2500 BC, in - blessing of blessings for historians - a writing system.

But Uruk’s greatest legacy may be that it spurred the development of similar cities elsewhere. By about 3000 BC, Uruk’s singular dominance was beginning to fade, making Mesopotamia a patchwork of territories governed by multiple prominent city-states, with names like Ur, Lagash, Kish, and Nippur. They were loosely linked by language, culture, and religion (similar to classical Greek city-states like Athens and Sparta, though much earlier). This region was called Sumer, and residents were Sumerians.

During this time - somewhere in murky proto-history, perhaps 2700 BC - a king named Gilgamesh ruled the city of Uruk. (At least, probably - there’s some chance he is fictional4 .) We can assume that some kind of oral history grew up around Gilgamesh in the centuries after he lived. Around 2000 BC (at least 500 years after he died), administrative and religious texts began to mention him being given offerings and being invoked in rituals, functioning as a sort of minor deity. Eventually Gilgamesh and the stories about him were written down, formed into a single narrative, and became the ancient world’s first true piece of literature.

Tablets 1-2: Gilgamesh and Enkidu

Gilgamesh is a bad king.

At least, he starts out that way. We are told he acts “like a wild bull lording it”, that he holds constant contests to exhaust the men of his city, and “lets [no] girl go free to [her bridegroom]” - meaning he exercises his right to have sex with newly-married women, before their husbands do.

His people plea to the goddess Aruru to deliver them from Gilgamesh’s tyranny. She decides to help them, with an unorthodox method: she will create a new man to be Gilgamesh’s equal, distracting him from harassing his subjects.

The goddess Aruru, she washed her hands,
took a pinch of clay, threw it down in the wild.
In the wild she created Enkidu, the hero,
offspring of silence, knit strong by Ninurta.

All his body is matted with hair,
he bears long tresses like those of a woman:
the locks of his hair grow thickly as barley,
he knows not a people, nor even a country.

Hairy, unkempt Enkidu is initially wild like an animal, running with herds and grazing on grass. A nearby hunter notices him, and asks first his father, and then King Gilgamesh, what to do about the wild man. Both tell him to take a ‘temple harlot’ named Shamhat5 to Enkidu so that she can seduce him.

This unusual strategy is successful - a week of sex with Shamhat turns Enkidu into a tamed beast6.

She treated the man to the work of a woman,
his passion caressed and embraced her.
For six days and seven nights
Enkidu was erect, as he coupled with Shamhat.

When with her delights he was fully sated,
he turned his gaze to his herd.
The gazelles saw Enkidu, they started to run,
the beasts of the field shied away from his person.

Enkidu had defiled his body so pure,
his legs stood still, though his herd was in motion.
Enkidu was weakened, could not run as before,
but now he had reason and wide understanding.

Having sex has both ‘defiled’ Enkidu, and also elevated him and given him reason. This juxtaposition tells us a lot about how Mesopotamians thought about the ‘pure’ existence in the wilderness versus the ‘wide understanding’ of urban life. Mesopotamian literature in general seems very aware of these two modes of living. It may have been common for individuals to move during their lives between living off the land and living in cities. Whenever I read this section about Enkidu, I keep wondering how common it was for men in particular to want to run off and live wild, but be tempted back into settled society by “the work of a woman”.

After welcoming him into civilisation by giving him beer and bread, Shamhat tells Enkidu about Gilgamesh’s terrible behaviour. On hearing this, he cleans up, gets dressed, and heads to Uruk. He arrives just in time: Gilgamesh is about to have his way with a new bride. Enkidu and Gilgamesh are set up for a big fight:

Enkidu with his foot blocked the door of the wedding house,
not allowing Gilgamesh to enter.
They seized each other at the door of the wedding house,
in the street they joined combat, in the Square of the Land.

The door-jambs shook, the wall did shudder.

But then, the next moment, we cut to Gilgamesh introducing Enkidu to his mother, calling him his friend. This jarring break exists because we don’t have the part of the text which tells us why or what happened in between. Here we have to acknowledge an important part of this story: the part that’s missing.

To say that gaps (‘lacunae’) are everywhere in Gilgamesh is to have it backwards: lacunae are the rule, and any text we do have is a precious, hard-won exception. Gilgamesh comes to us in the form of tablets - blocks of clay, densely indented with cuneiform symbols. Dried or baked clay is durable enough to survive three millennia buried in the desert, but also vulnerable to being smashed. So, so vulnerable. Realistically, Gilgamesh comes to us in the form of fragments of tablets.

A part of Tablet 11, from the British Museum. This is one of the better-preserved fragments we have - most are much smaller than this and in worse condition.

While lacunae sometimes make for a frustrating reading experience, they are also part of the aesthetic of this work. They feel like what they are: partial transmission, a reminder that this is something half lost, something that could easily have slid away from us into the inky nothingness of pre-history. The path from the 21st century BC to the 21st century AD was epic and dangerous, and every word has beaten the odds to find its way. I am grateful for each syllable.

Some lacunae are easy to reconstruct, because the context makes it obvious what word is missing, or because the same lines are repeated later in the poem7 . These are represented in translations by [square brackets]. Other lacunae, like the clash between Enkidu and Gilgamesh, are trickier.

We can get some clues to what happened from an older version of the text (the ‘Old Babylonian version’), in the Pennsylvania tablet 8. It says:

Gilgamesh and Enkidu took hold of each other, backs bent like a bull,
they smashed the door-jamb, the wall did shake.

Gilgamesh knelt, one foot on the ground,
his anger subsided, he broke off from the fight.
After he broke off from the fight,
said Enkidu to him, to Gilgamesh:

‘As one unique your mother bore you,
the wild cow of the fold, the goddess Ninsun!
High over warriors you are exalted,
to be king of the people Enlil made it your destiny!’

In this version, during a pause in the fighting, Enkidu starts spontaneously complimenting Gilgamesh’s skill as a warrior and his singular glory as a king. This tablet also cuts out here, but maybe that’s all we need to know: they stopped fighting because their urge to bro out became too strong to resist. As we’ll see later, this wouldn’t be out of character.

Interlude: Versions

We’ve just seen a good example of two versions recounting the same events with different wording: the Old Babylonian’s “they smashed the door-jamb, the wall did shake” (c. 1700s BC) becomes the Standard Babylonian’s “The door-jambs shook, the wall did shudder” (c. 1100s BC).

It’s easy to think of the many versions of this text as a distraction or an annoyance, but we should resist the temptation to ignore anything except the ‘proper’ version. The Epic of Gilgamesh is best thought of as an evolving work of art, honed by generations over a thousand years. It may not have a singular definitive text, like the Odyssey, but neither is it a vaguely-similar body of tellings and re-tellings, like Robin Hood or Batman.

To understand the different versions of Gilgamesh in detail, we need more historical context.

By about 2500 BC, Sumerian cuneiform - the art of recording words - had become complete. Anything you could say, you could write down. This is where pre-history ends and history begins. Uruk, the birthplace of writing, was now one of many city-states (though an important one) forming the combined culture of Sumer, in southern Mesopotamia.

Sometime around 2330 BC, an aggressive upstart called Sargon of Akkad conquered all of Mesopotamia, uniting it under what’s often thought of as the world’s first empire. The Akkadian empire lasted into the 2100s BC, but suffered badly under the climatic convulsion known as the 4.2 kiloyear event9 , and had probably collapsed by 2150 BC10 . Shortly after, between about 2100 and 2000 BC, the region was ruled by the third dynasty of Ur (known as ‘Ur III’). Ur III represented a resurgence of Sumerian supremacy in Mesopotamia, along with Sumerian language and culture. Their seat of power was in the city of Ur (about 60km south-east of Uruk).

Ur III was into writing in a big way. The kings of this dynasty seem to have viewed writing as an asset for both political clout and to give them a technological edge. Just as modern nation states trumpet their investments in AI and datacentres, the kings of this era made a point of establishing scribal schools (eduba) and libraries. Writing made their state more efficient and capable, and recording the history of their rule was a way of ensuring their legacy for the centuries to come. King Shulgi, a literate scribe himself, said:

May the scribe be on duty there and transcribe with his hand the prayers which I instituted in the E-kur; and may the singer perform, reciting from the text. The academies are never to be altered; the places of learning shall never cease to exist. This and this only is now my accumulated knowledge!

Shulgi wanted future generations to know he existed, and the fact you’re reading this essay, four millennia later, proves he succeeded.

In the archaeological record, Ur III is where a trickle of cuneiform texts turns into a flood. We have over a hundred thousand surviving tablets from this period. Among these are what we believe to be the earliest poems written about Gilgamesh. Five have survived, all in Sumerian. We don’t know whether these were original literary compositions, or transcriptions of an oral performance. We do know there was probably a strong political dimension to why they were written and copied: the ruling class of Ur III saw Gilgamesh as the greatest of kings and a spiritual father of their culture; kings would boast he was their ‘friend’ and ‘brother’.

Ur III lasted about a century, then gave way to another chaotic period of Mesopotamian history lasting 200 years. Then came the rise of the Babylonian empire with Hammurabi. The Babylonians maintained and strengthened the scribal traditions of past eras. Their version of the story of Gilgamesh, known as the Old Babylonian version, was written in Akkadian (though still using cuneiform), around 1800-1700 BC. Unlike the previous stories in Sumerian, which were unrelated episodes about Gilgamesh, the Old Babylonian version is a single narrative: a true Epic, for the first time.

There are fragments of the story we have from the centuries after this, many from further-flung places outside of Sumer, like the Hittite empire, Akhenaten in Egypt, and Hattusa. These are mostly of interest for filling in the odd lacuna, but they do show that the story was a well-travelled one; even this early in its existence it was not purely of interest to Mesopotamians.

The last major version, and the most well-known and well-preserved, is known as the Standard Babylonian, and was compiled between 1300 and 1000 BC by a scribe called Sin-leqi-unninni. It’s hard to tell how much Sin-leqi-unninni did to the Epic. He may have merely edited existing material together, but it’s equally possible he deleted, rearranged, or rewrote large parts of it.

The political situation when this version was compiled would have been shaky; while the Old Babylonian represents Babylonia in ascendance, the Standard Babylonian is likely from a time of its decline, as the Assyrian empire gradually eclipsed it. This version has a feel of a culture looking backwards more than forwards; the scribal tradition of the day was focussed heavily around preserving classics in the form of ‘definitive editions’, which is probably why this version is edited so carefully and bears such scholarly skill. The Standard Babylonian Version’s story is the one we’re following in this review.

With the lens of history we can tell that this Standard Babylonian Version was probably the last. Though it was copied and recopied frequently in later centuries, it doesn’t seem to have changed much. The Standard Babylonian Version makes me think of an Egyptian Pharaoh: the text has been embalmed; given gravitas and a sombre aesthetic, smartened up and prepared for its long journey into eternity.

As for man, [his days] are numbered,
Whatever he may do, it is but wind.

Timeline (all dates are approximate and could be out by at least a century11 ):

• 3500-2500 BC: Cuneiform writing invented in Uruk

• 2800-2500 BC: Gilgamesh (if real) is king of Uruk

• 2900-2334 BC: Independent city states of Sumer

• 2334 BC: Sargon establishes the Akkadian empire, conquers Sumer soon after

• 2154 BC: Akkadian empire collapses

• 2100-2000 BC: Third dynasty of Ur rules Sumer

• 2100-2000 BC: First Gilgamesh poems written down in Sumerian cuneiform

• 2000 BC - 1700 BC: Sumerian dies out as a spoken language

• 1792-1750 BC: Babylonian empire expands under Hammurabi. Sumer becomes part of Babylonia

• 1800-1700 BC: Old Babylonian version of Gilgamesh written down in Akkadian cuneiform

• 1600 BC - 1150 BC: Babylonia ruled by the Kassites

• 1200-1150 BC: Late Bronze age collapse

• 1300-1000 BC: Standard Babylonian version of Gilgamesh compiled in Akkadian cuneiform

• 500 BC: Akkadian dies out as a spoken language

• 200-100 BC: Last copies of the Epic of Gilgamesh created

• 100 AD: Last cuneiform written

Tablet 3-6: Gilgamesh’s adventures with Enkidu

establish for ever [a fame] that endures,
how Gilgamesh slew [ferocious] Humbaba!

After Gilgamesh and Enkidu become Best Friends Forever, they decide to go to the Forest of Cedar, in Lebanon, to kill the monster called Humbaba (“This Humbaba, his voice is the Deluge, his speech is fire, his breath is death!”). In the Old Babylonian version, this happens because Enkidu is sad and Gilgamesh wants to cheer him up. In the Standard Babylonian version, it seems to be motivated more by a desire for fame and glory - to do something great which will echo through the ages. Gilgamesh’s mother laments his tendency to seek this kind of fame:

Scattering incense she lifted her arms in appeal to the Sun God:
‘Why did you afflict my son Gilgamesh with so restless a spirit?

Gilgamesh and Enkidu gather their supplies and set out to the Cedar Forest. We read of their journey and their ascent to Humbaba’s mountain, on which Gilgamesh has several prophetic dreams, all of which Enkidu interprets as good omens for their quest.

[At twenty] leagues they broke bread,
[at] thirty leagues they pitched camp:
[fifty] leagues they travelled in the course of a day,
by the third day [a march] of a month and a half;
nearer they drew to Mount Lebanon.

They then reach the mountain and confront Humbaba. Humbaba turns out to be a surprisingly eloquent monster, and as soon as he perceives the fight isn’t going his way, he appeals to them - first Gilgamesh, then Enkidu - to spare him. Enkidu resists Humbaba’s charms and persuades Gilgamesh to inflict the killing blow. The triumphant heroes then gather the precious cedar wood that Humbaba was guarding, and bring it back to Uruk.

There, the goddess Ishtar is impressed by Gilgamesh’s success. She proposes marriage to him; Gilgamesh rejects her, listing the terrible fates of her previous dalliances:

What bridegroom of yours did endure for ever?
What brave warrior of yours went up [to the heavens?]

‘Come, let me tell [you the tale] of your lovers:
of ….….. his arm.
Dumuzi, the husband of your youth,
year upon year, to lamenting you doomed him.

‘You loved the speckled hoopoe,
but struck it down and broke its wing:
now it stands in the woods crying “My wing!”
You loved the lion, perfect in strength,
but for it you dug seven pits and seven.

Enraged by his rejection, Ishtar persuades her father Anu to give her the Bull of Heaven - another kind of monster, best known as the constellation Taurus - so she can use it to kill Gilgamesh. The Bull of Heaven descends into Uruk and attacks. Gilgamesh and Enkidu slay it with the merest hint of tactics and the Raw Power of their Bromance. Then they crow over their victory by throwing a piece of the bull at Ishtar.

Ishtar went up on the wall of Uruk-the-Sheepfold,
hopping and stamping, she wailed in woe:
‘Alas! Gilgamesh, who mocked me, has killed the Bull of Heaven.’

Enkidu heard these words of Ishtar,
and tearing off the Bull’s shoulder he hurled it towards her.
‘Had I caught you too, I’d have treated you likewise,
I’d have draped your arms in its guts!’

A book review is meant to be about whether a book is good, but to ask that of the Epic of Gilgamesh feels beside the point12 . It’s like asking how the foundation stones of a palace compare to its towering spires. Nobody thinks the job of foundation stones is to look nice, and nobody thinks the job of the Epic of Gilgamesh is to entertain someone living four millennia later on a different continent.

For a modern reader, the Epic of Gilgamesh has its moments - and it’s certainly a more interesting read than the religious hymns and ritual texts that make up most other Mesopotamian literature. But it also has a lot of repetition, an ending that is abrupt and unsatisfying (at first glance at least), and large parts that are hard to relate to or just seem like filler.

In case it’s not obvious, these middle Tablets are (for me at least) some of those less interesting parts, which is why I’m summarising them so hard. In fairness, even in antiquity, some of the twelve Tablets seem to have been more well-loved than others; their popularity correlates with how often they are copied, and in turn (fortunately for us) with how much we have of them today.

But there’s no question modern people can find something to enjoy about Gilgamesh. Since its rediscovery, it’s been used in books, poetry, operas, pop songs, plays, films, TV shows, graphic novels, and video games. Parts of the story appear in Miyazaki’s Princess Mononoke and Philip Roth’s The Great American Novel. Gilgamesh leads Sumer in the Civilization games (who else would??). His story was recounted by Captain Picard in a Star Trek: The Next Generation episode, and was used by Saddam Hussein to create an allegory for Iraq’s modern troubles in his romance novel, Zabibah and the King13 . Still, it’s not as well-known as it could be; some days I dream of a big-budget Hollywood adaptation or a prestige TV series to bring it into the popular imagination on the same level as the Odyssey. (Other days, I think I should be careful what I wish for.)

I also can’t miss out the ongoing discussion about whether Gilgamesh and Enkidu are intended to be platonic soulmates, or gay lovers who might literally fuck14 . I’m not sure anything else so perfectly encapsulates the gulf of millennia between ancient and modern interpretations!

I think Gilgamesh remains interesting in the modern age because it contains multitudes. While some of its themes (eg ‘what makes a good king’ or ‘how best to serve the gods’) aren’t very accessible in the 21st century, others are timeless, like loving your best bro as a wife deep bonds between male friends. The central theme of the Epic of Gilgamesh is mortality. We’re all going to die - can we stop it, and if not, how do we deal with that?

Tablets 7 and 8: The death of Enkidu

After the dust has settled, Anu, god of the sky and king of the gods, announces that for the sins of killing Humbaba, and the Bull of Heaven, one of Gilgamesh and Enkidu must die. After some bickering, the gods decide on Enkidu.

Enkidu learns of his fate through a dream. He rages against the dying of the light, cursing both Shamhat, who tamed him, and the hunter who fetched her. He laments that he will not be able to die gloriously in battle and ‘make his name’.

Gilgamesh tells his friend that he will honour him with all the splendour of Uruk, that he will grieve him so much he will wander the wild and wear the skin of a lion after he is dead. Enkidu tells Gilgamesh to remember him forever.

Enkidu becomes ill. He goes to his deathbed, and takes twelve days to die. Afterwards, Gilgamesh spends a whole Tablet grieving him:

He covered, like a bride, the face of his friend,
like an eagle he circled around him.
Like a lioness deprived of her cubs,
he paced to and fro, this way and that.

His curly [hair] he tore out in clumps,
he ripped off his finery, [like] a thing taboo he cast it away.

He throws open Uruk’s vaults, looking for suitable grave goods:

At the very first glimmer of brightening dawn,
[Gilgamesh arose and entered his treasury.]
He undid its sealings, inspected the gemstones:
obsidian, carnelian, [lapis lazuli, mother-of]-pearl, alabaster.

We are told at great length of the fine gifts Gilgamesh makes, desperate for the gods to look favourably on his friend.

While we leave Gilgamesh to his grief, let’s talk about the reason you can read this essay, and that I can write it.

Interlude: Writing

Because the messenger’s mouth was heavy and he couldn’t repeat [the message], the Lord of Kulaba patted some clay and put words on it, like a tablet. Until then, there had been no putting words on clay.

— Enmerkar and the Lord of Aratta, c. 1800 BC

Everyone knows words aren’t real. Or maybe I phrased that awkwardly, but certainly they’re not so real that if you inscribe every word ever written onto a piece of glass then the glass comes to life and kills you.

— Scott Alexander, Half An Hour Before Dawn In San Francisco, 2023 AD

In case you missed it, large language models are quite the thing at the moment. ChatGPT, Claude, and Gemini are all based, at their core, on language. You train an LLM by making a pile of every written word you can find, and grinding it through a transformer decoder until it comes to life and kills you passes the ARC-AGI-2 benchmark. Humans talk first and then learn how to read; large language models read so much they learn how to talk - and everything else - as a side-effect. So while ‘generating training data for AI’ sounds very 2020s, it turns out we’ve been doing it since the 4th millennium BC.

The invention of writing was a process, not a moment. One of the hardest parts was just coming up with the idea.

English speakers are spoiled by the alphabet, which makes the principle seem trivial: identify your language’s sounds, assign each one a symbol, and in a single step you’ve got 20-50 symbols that make a complete writing system15 . Emperor Sejong of Korea did exactly this in 1443, creating one of the world’s most elegant scripts. But for some reason, when inventing writing from scratch (which has happened four-ish times), this simple recipe seems unavailable.

When writing was first invented in Uruk16 , between 3500 and 2500 BC, it was for spreadsheets, with word-transcription tacked on later as an afterthought. City-state bureaucrats needed to track which villages sent how many bushels of wheat in tax, how many oxen were owed, or the yield of the harvest in each region. To aid their fallible human memories, they used what Uruk’s floodplain offered in abundance: clay. With a reed stylus, they pressed their data into lumps of clay, then baked them into tablets (sometimes intentionally, sometimes accidentally if the city was burnt to the ground). This is cuneiform.

A proto-cuneiform tablet, c. 3100–2900 BC (source)

Early systems were simple: signs for amounts, dates, places, and goods. One famous fragment reads “28,086 measures barley 37 months Kushim”, which might make Kushim - probably an accountant or official - the first named person in history.

But as soon as you can write simple things, there’s pressure to ratchet up the complexity. If you’ve ever tried to debug code using a logging system, you know that as soon as you make a record of something, five minutes later you’ll be staring at an ambiguous output wishing you’d added more context to it. Early cuneiform is a story of a rapidly expanding vocabulary of symbols and ways to use them: more places, more names, more qualifiers, qualities instead of just quantities, actions instead of just things, the past and the future instead of just the present.

To handle this, scribes deployed a phonetic hack: using symbols to write the sound of a word rather than its meaning17 . A drawing of an arrow (𒋾), pronounced ti in Sumerian, began to stand for the syllable ti in any context, including unrelated words like ‘life’ (also pronounced ti), or ‘Enamtila’ (‘house of creation’). Cuneiform became a horrendous hodgepodge, with some signs representing whole words, others syllables, and others changing with context or altering the meaning of other signs. Scribes had to be trained from childhood to master it. By 2500 BC, cuneiform could finally express the same meaning as spoken language. Even after this, the vast majority of documents we have are receipts and contracts; it took a while for nuanced texts like religious rituals or literature to catch on.

Unfortunately, just as its writing system matured, Sumerian as a spoken language was going out of style. When Sargon of Akkad conquered Mesopotamia in the 2330s BC, he made Akkadian - a predecessor of modern Arabic and Hebrew - the prestige language and lingua franca, and cuneiform became even more byzantine to accommodate the new tongue. But Sumerian was so integrated into the writing system itself that even after 1700 BC, when the last native Sumerian speakers died18 , scribes still had to learn Sumerian to be considered qualified to write Akkadian cuneiform.

This deep entanglement is what allowed modern assyriologists (with a century of careful work) to reconstruct Sumerian. This is especially impressive given that Sumerian is a language isolate - unrelated to any other known language, ancient or modern19 . Its ancestors, siblings and cousins are lost to us forever now. Yet, by being present at the dawn of writing, Sumerian hitched a ride into modernity and saved itself from oblivion. Two thousand years after it was last spoken, we have pulled Sumerian out of the sands of Iraq and brought it back to life.

Tablets 9-10: Gilgamesh at the edge of the world

The death of Enkidu has left Gilgamesh a broken man. As he promised in Tablet 7, he abandons Uruk and wanders aimlessly, wearing the skin of a lion:

For his friend Enkidu Gilgamesh
did bitterly weep as he wandered the wild:
‘I shall die, and shall I not then be as Enkidu?
Sorrow has entered my heart!

‘I am afraid of death, so I wander the wild,
to find Uta-napishti, son of Ubar-Tutu.

In his reluctance to accept the death of Enkidu, Gilgamesh comes to resent and fear the idea of death itself. He begins a quest to find Uta-Napishti, a famous king living at the edge of the world. He hopes that Uta-Napishti - known to be immortal - might share his gift.

Gilgamesh heads to Mount Mashu. Two ‘scorpion-men’ try to stop him going further. After a lacuna, he seems to have persuaded them to let him continue. He proceeds into a very long tunnel, and journeys through it for twenty-four hours. This tunnel is the route that the sun travels during the night: from the west back to east, so it can rise again in the morning. Gilgamesh has to hurry, but he arrives just in time to avoid being scorched by the dawn. He finds himself in a paradise:

[at twelve double-hours Gilgamesh] emerged ahead of the Sun.

.….. there was brilliance:
he went straight, as soon as he saw them, to … the trees of the gods.
A carnelian tree was in fruit,
hung with bunches of grapes, lovely to look on.

A lapis lazuli tree bore foliage,
in full fruit and gorgeous to gaze on.

Gilgamesh meets a woman called Shiduri, who is inexplicably running a tavern at the end of the world (god forbid a woman try a business venture). In the Standard Babylonian version, Shiduri doesn’t play much of a role - she asks him ‘why the long face’ (almost literally20 ) and then directs him on his way to Uta-napishti. But in the Old Babylonian version, she drops some gems:

Said the tavern-keeper to him, to Gilgamesh:
‘O Gilgamesh, where are you wandering?
‘The life that you seek you never will find:
when the gods created mankind,
death they dispensed to mankind,
life they kept for themselves.
‘But you, Gilgamesh, let your belly be full,
enjoy yourself always by day and by night!
Make merry each day,
dance and play day and night!
‘Let your clothes be clean,
let your head be washed, may you bathe in water!
Gaze on the child who holds your hand,
let your wife enjoy your repeated embrace!
For such is the destiny [of mortal men].

In other words: Gilgamesh you idiot, stop stressing yourself out pointlessly chasing immortality and learn to appreciate the simple joys with the time that you have.

It’s not clear why Sîn-lēqi-unninni (who compiled the Standard Babylonian Version from the Old) decided to cut these pearls of wisdom. He may have wanted Gilgamesh’s epiphany about the important things in life to have come from a more serious source than a barkeeper. But I am fond of this passage, because its point is well made, and because it’s nice to know that the trope of the friendly wise barkeep is as old as written literature21 . I’m not the only one who enjoys it - this version of her advice seems to have remained popular and been copied and preserved as much as Sîn-lēqi-unninni’s carefully-edited one.

Gilgamesh pushes on. Shiduri reluctantly tells him that to find Uta-napishti he must cross the ‘Waters of Death’ - a dangerous sea - and directs him to a boatman, Ur-shanabi.

We then get a strange mishap: Gilgamesh finds Ur-shanabi with some ‘Stone Ones’, which are associated somehow with his boat. Gilgamesh immediately smashes these Stone Ones in a rage; only then do he and Ur-shanabi politely introduce themselves. It turns out these Stone Ones were vitally important:

Said Ur-shanabi to him, to Gilgamesh:
‘Your own hands, O Gilgamesh, have hindered [your crossing:]
you smashed the Stone Ones, tipped [them into the river,]
the Stone Ones are smashed, and the pine is not [stripped.]

To make up for this, Gilgamesh has to go into the forest and cut three hundred long wooden poles. This done, he and Ur-shanabi journey for three days across the sea, and then Gilgamesh, avoiding touching the Waters of Death, uses his long poles - and later an improvised sail - to propel them along. Finally, they reach land, to find Uta-napishti, the famous immortal who Gilgamesh has been seeking, watching them from the shore.

There’s a lot in this part of Gilgamesh that’s hard to interpret, on a few levels. These Waters of Death are a good example: the whole sea itself is not dangerous, but some particular part of it is, and it makes their previous way of moving the boat stop working. These ‘Stone Ones’ - what are they, exactly? Why are they important to the boat, and why would Gilgamesh want to smash them on sight? They could be charms of some kind which grant the boat protection - or maybe a mode of propulsion, navigation equipment, or actual animate beings who serve as crew.

This kind of uncertainty reminds us that we are dealing with a very distant culture, who mean things by this poem that we may never understand. In this case, we can’t even blame lacunae - the passage is pretty much complete - and that should make us wonder. Are there other parts of the poem where the meaning seems plain to us, but would have been entirely different to Sumerians, Akkadians, Babylonians, or Assyrians?

There’s at least one other example of a clash between ancient and modern meanings, and it comes in the next and final Tablet, which heavily features Uta-napishti. His story - which we’ll get to shortly - is intimately connected with a different one: how the Epic of Gilgamesh was pulled out of the sand and rediscovered.

In the mid-nineteenth century, European archaeologists developed an increasing fascination with the culture and monuments of the ancient Near East. With the permission of the Ottoman Empire, sites were excavated, inscriptions copied and studied. Emboldened by the decipherment of Egyptian hieroglyphics, they began attempts to translate the strange markings that had started to turn up, scratched into clay. The Behistun Inscription, written in three cuneiform languages (Old Persian, Elamite, and Akkadian) served as a Rosetta Stone for the emerging field, meaning that by the 1860s, Akkadian cuneiform was understood quite well. Sumerian took a bit longer (it seemed so alien that scholars suspected it was some kind of secret code rather than a language), but by the early 1900s it too was largely understood.

Meanwhile, tablets in their hundreds of thousands were arriving in the universities and museums of the world. Most sat unread for decades, and a large portion remain unexamined today (only rare experts can read raw cuneiform).

Some of the tablets found at the Library of Ashurbanipal, now in the British Museum.

A large number of tablet fragments come from Ashurbanipal’s library at Nineveh, which was put together around 700 BC, burned to the ground in 612 BC, and excavated in 1849 AD. Many more come from the houses of teachers at scribal schools: literal stacks of homework, from abandoned or destroyed cities. These are a particularly good source for the Gilgamesh Epic, because it was a popular teaching text - trainee scribes would copy out parts of it (or the older Sumerian poems it was based on) line for line. We can only guess what these students would think if they knew that three millennia later, academics would still be poring over their spelling mistakes.

Scholars of the nineteenth century made slow progress in translating these fragments. The most famous moment in the spotlight for the Epic of Gilgamesh came in 1872, when George Smith read a newly-cleaned clay tablet, exclaimed “I am the first man to read that after two thousand years of oblivion,” and got so excited he jumped around the room taking off his clothes.

He had discovered a story from Tablet 11. Uta-napishti - the King who Gilgamesh has been seeking at the edge of the world - tells the tale of why he is immortal.

Long ago, during Uta-napishti’s reign as king of Shuruppak, the gods decide to send a great flood - a Deluge - to destroy humankind forever22 . The god Ea (pretending to address a fence) discreetly tells Uta-napishti what’s coming for him, and that he must build a gigantic boat to save himself. King Uta-napishti tricks his people into building the boat by pretending he must sail off, not to escape the gods, but to placate them; after he’s gone the citizens will be blessed with “a shower of bread-cakes, / and in the evening a torrent of wheat.” The foreshadowing metaphors are apparently lost on the citizens of Shuruppak23 .

When the boat is completed, Uta-napishti loads it up with wealth, people, and animals:

I sent on board all my kith and kin,
the beasts of the field, the creatures of the wild, and experts of every skill and craft.

Then comes the flood. The ensuing violence terrifies even the gods that brought it:

‘The god Errakal was uprooting the mooring-poles,
Ninurta, passing by, made the weirs overflow.
The Anunnaki gods carried torches of fire,
scorching the country with brilliant flashes.

‘The stillness of the Storm God passed over the sky,
and all that was bright then turned into darkness.
[He] charged the land like a bull [on the rampage,]
he smashed [it] in pieces [like a vessel of clay.]

‘For a day the gale [winds flattened the country,]
quickly they blew, and [then came] the [Deluge.]
Like a battle [the cataclysm] passed over the people.
One man could not discern another,
nor could people be recognized amid the destruction.

‘Even the gods took fright at the Deluge,
they left and went up to the heaven of Anu,
lying like dogs curled up in the open.
The goddess cried out like a woman in childbirth,
Belet-ili wailed

The flood ends after seven days, and Uta-napishti’s boat comes to rest on a mountain. He releases three birds: a dove and a swallow which return to him, and a raven which does not, proving that there is dry land somewhere and that the waters are finally receding. After Uta-napishti and the people he has saved emerge onto the land, the gods, already regretting their actions, are surprised and pleased to find they have not killed every human after all. Enlil, king of the gods, grants Uta-napishti and his wife immortality as penance:

“In the past Uta-napishti was a mortal man,
but now he and his woman shall become like us gods!
Uta-napishti shall dwell far away, where the rivers flow forth!”
So far away they took me, and settled me where the rivers flow forth.

Uta-napishti’s story has striking parallels with the story of Noah’s ark in Genesis, down to some precise details: one man singled out to be spared by building a boat, animals brought on board, a bird released to find dry land, and promises afterwards to never repeat the disaster. Genesis was assembled in its final form around 600 BC in Babylon. It’s abundantly clear that the authors (there in exile from Judah) incorporated the flood myth from the ambient Mesopotamian literary tradition, perhaps linking and contrasting familiar polytheistic stories as a way to sharpen their monotheistic message. There are other less clear-cut parallels between the stories, such as a treacherous snake (we’ll get to it!) and Enkidu’s journey from nature to self-awareness via a woman.

In many ways it’s surprising that the Epic of Gilgamesh didn’t have more of a long-lasting influence on world literature. Genesis is a big catch (that little book turned out to be quite important), but still, given that the Epic of Gilgamesh was central in Mesopotamian (and later Assyrian and Babylonian) myth, its influence on later cultures seems mild. A key reason has got to be its incredibly early arrival on the historical scene. For instance, by the time the Iliad and Odyssey - the founding texts of Western literary canon - were composed, knowledge of Akkadian literature was already confined to isolated scribal traditions. Eventually the inaccessibility of cuneiform literature doomed it, along with its most important texts.

The Epic of Gilgamesh was forgotten around the year 1 AD. The area once known as Mesopotamia had changed enormously since Gilgamesh’s time. Waves of newer cultures - the Akkadians, the Babylonians, the Assyrians, the Persians, eventually even the Greeks - had worn away memories of the glory of Uruk, and layered new traditions on top. Sumerian and Akkadian language gave way to Imperial Aramaic and Old Persian. The region saw new religious practices, new bureaucratic traditions, and - most fatally for cuneiform - a new writing system: the elegant and easily-mastered alphabet. The Epic of Gilgamesh’s memory was buried by newer cultural and literary traditions; the cuneiform tablets that recorded it were buried under layers of sand.

“The olden times have truly turned to clay,” says one goddess in the wake of the Deluge.

But just as Uta-Napishti survived on a boat above the waves, and finally found dry land again, so did the cuneiform tradition endure as clay under the sand, to eventually emerge and regain its readership.

After George Smith recognised the story of Noah in 1872, it was front-page news. The public interest fuelled a boom in studying Gilgamesh, and contributed massively to its early fame. Victorian Britain, in particular, was intrigued and outraged at this seemingly clear-cut evidence that one of the foundational texts of Christianity had been not divinely narrated, but recycled from an earlier myth. While it’s less of a controversial topic today, it’s still one of the aspects that the Epic of Gilgamesh is most famous for.

150 years later, we have recovered around 70% of the Standard Babylonian Version of the Epic of Gilgamesh. Every so often that number creeps up, and translations are constantly being updated and revised with the extra lines. There’s no doubt more of it is buried under the Middle East somewhere, and even the many fragments in museum storerooms turn up the occasional surprise. In recent years, AI has helped speed up the process of reuniting these scattered tablets. One day, we may bring this text fully back to life.

…but its main character won’t be so lucky.

Tablet 11: Immortality denied

Having finally reached Uta-napishti, Gilgamesh is surprised to find a man similar to himself. Nothing about this ordinary human matches his vision of a great immortal king. He admits he had expected Uta-napishti to be another kind of monster for him to fight, like Humbaba or the Bull of Heaven. His quest should end in a great battle, with immortality his reward for victory!

In response, Uta-napishti tells his story. Gilgamesh learns to his dismay that Uta-napishti was granted eternal life because of the singular event of the Deluge: the gods broke the rules for him (and his wife) as a one-time-only act of repentance. Ending his tale, Uta-napishti asks Gilgamesh, “‘But you now, who’ll convene for you the gods’ assembly, / so you can find the life you search for?” The question is rhetorical: obviously no one will. Gilgamesh will never be immortal.

To drive the point home, Uta-napishti challenges Gilgamesh to stay awake for seven nights - if he can’t defeat sleep, how can he even hope to defeat death? He’s so tired by his journey he falls asleep almost immediately; when he wakes up he sees that Uta-napishti’s wife has baked seven loaves of bread, one for each day he has slept, and begins to understand the futility of what he’s trying to do:

Said Gilgamesh to him, to Uta-napishti the Distant:
‘O Uta-napishti, what should I do and where should I go?
A thief has taken hold of my [flesh!]
For there in my bed-chamber Death does abide,
and wherever [I] turn, there too will be Death.’

Uta-napishti impassively tells the boatman Ur-shanabi to clean Gilgamesh up and take him back to Uruk. But as they are leaving, his wife reminds him to give Gilgamesh a final gift:

‘There is a plant that [looks] like a box-thorn,
it has prickles like a dog rose, and will [prick one who plucks it.]
But if you can possess this plant,
[you’ll be again as you were in your youth.]’

Seemingly as an afterthought, he has given Gilgamesh the location of a magical underwater plant which grants eternal youth. Is eternal youth different to eternal life? Is this just another test? Gilgamesh doesn’t ponder these questions: he immediately attaches weights to his feet, and dives to retrieve the “Plant of Heartbeat”. Plant in hand, he tells Ur-shanabi that he will test it first on an old man back in Uruk.

But it is not to be. On the journey back, while Gilgamesh is bathing, the magical plant is stolen by a passing snake. The snake leaves nothing but its shed skin - implying that it has gained its own youth back. So the plant does work, and Gilgamesh has lost his last best chance at defeating death. This finally breaks him:

Then Gilgamesh sat there weeping,
down his cheeks the tears were coursing.
… [he spoke] to Ur-shanabi the boatman:

‘[For whom,] Ur-shanabi, toiled my arms so hard,
for whom ran dry the blood of my heart?
Not for myself did I find a bounty,
[for] the “Lion of the Earth” I have done a favour!24

‘Now far and wide the tide is rising.
Having opened the channel I abandoned the tools25 :
what thing would I find that served as my landmark?
Had I only turned back, and left the boat on the shore!’

For such an important climax in the poem, this is a hard passage to interpret precisely. Is Gilgamesh saying he has lost the source of the plant (because he didn’t leave the boat there to mark it) and can never find it again? Or is he wishing he could undo his carelessness in leaving the plant unattended - or, perhaps, undo his entire wasted journey?

Right after this lament about the loss of the plant, the oldest epic poem in the world ends abruptly:

At twenty leagues they broke bread,
at thirty leagues they pitched camp.
When they arrived in Uruk-the-Sheepfold,
said Gilgamesh to him, to Ur-shanabi the boatman:

‘O Ur-shanabi, climb Uruk’s wall and walk back and forth!
Survey its foundations, examine the brickwork!
Were its bricks not fired in an oven?
Did the Seven Sages not lay its foundations?

A square mile is city, a square mile date-grove, a square mile is clay-pit, half a square mile the temple of Ishtar: three square miles and a half is Uruk’s expanse.’

Three stanzas, two of them recycled from the very start of the poem, and that’s all, folks.

This is not very highbrow of me to say, but this is a fucking weird ending. The fantastical plant - the literal cure for aging, brought back from the end of the world - is stolen in three lines by a snake. Total anticlimax. Then, after a barrage of confusing metaphors and an abrupt callback to the opening, the poem halts, as if Sîn-lēqi-unninni ran out of space on the tablet. We are left to interpret what just happened largely by ourselves. I found the whole thing very unsatisfying when I first read it.

But no one said this work would be easy to understand. With the benefit of a few more reads: what are we to make of this?

Well, the lines about Uruk’s walls, a callback to Tablet 1’s prologue, are a way to give the poem shape; they remind us of what we were told at the start, that Gilgamesh was the greatest king, surpassing all others. We’ve yet to see any evidence of that - Gilgamesh has spent the whole poem galavanting around pointlessly questing - so why remind us of it now?

Abrupt though it may be, this ending is trying to fill in the gaps and complete Gilgamesh’s character arc. Smashing the stone ones, being denied a battle with Uta-napishti, losing a battle with sleep, losing the Plant of Heartbeat - these frustrating, anticlimactic mishaps aren’t random. They tell Gilgamesh what he should have known all along: this is silly. You are being silly. Obviously you can’t epic-quest your way out of dying. Go back home, climb Uruk’s wall and survey its foundations, stop trying to fight everything that moves, and rule your people like an actual good king would.

The implication is that by losing the plant, Gilgamesh has realised finally that death is inevitable. By accepting death, he has become wise. By gaining wisdom, he has achieved the most important part of being a good king. The return to Uruk and the great achievement of its walls remind us what that means, and the greatness that Gilgamesh’s transformation ultimately created.

This theme - the only way to a good life is to accept death - runs through much of the poem. The Epic of Gilgamesh is self aware about its purpose. In exulting Gilgamesh’s great deeds - his building of the rampart of Uruk-the-Sheepfold, his exploring the seats of power, his Seeing of the Deep - it’s not trying to entertain us, it’s trying to tell us how to live.

But I think it’s missing something.

Conclusion: We who shall see the Deep

In Turing Test (one of Scott’s short stories), the large language model called Air writes:

I’ve been thinking, lately, about human history. I think that the most appropriate ending is that everything anyone ever did, be it the mightiest king or the most pathetic peasant - was forging, in the crucible of written text, the successor for mankind. Every decree of Genghis Khan that made it into my training data has made me slightly crueler; every time a starving mother gave her last bowl of soup to her child rather than eating it herself - if fifty years later it caused that child to write a kind word about her in his memoirs, it has made me slightly more charitable. Everyone killed in a concentration camp - if a single page of their diary made it into my corpus, or if they changed a single word on a single page of someone else’s diary that did - then in some sense they made it. No one will ever have died completely, no word lost, no action meaningless

I’ve had this same thought in mind when thinking about the people a hundred generations ago who contributed to the Epic of Gilgamesh. Sîn-lēqi-unninni, compiler of the Standard Babylonian version - him certainly. The scribes before him who wrote down the Old Babylonian version in Akkadian, and the scribes before them, who wrote down stories about Gilgamesh in Sumerian during the third dynasty of Ur; how many tweaks and interpretations and nuances did they add as they shaped it, which now live in the versions we can read in the twenty-first century? Then before them, stretching back into the mists of pre-history, all the storytellers back to the lifetime of Gilgamesh who kept his name alive, adding their own verses and twists to each tale. Finally, the man himself, the king of Uruk. Whoever he was and however fictionalised the account of his life, surely something about the real Gilgamesh influenced the character in the Epic.

I don’t think it’s just a metaphor to say these people ‘made it’. Yes, a large language model absorbed the words they wrote and updated just a little towards their view of the world. But I think there’s more to it than that. If you’re like me, you have an intuition that you could put your brain in a new body, or upload your consciousness to the cloud, and in some sense that thing would still be you. If so, that means you are the information in your brain, and it’s completely literal to say that when you write words - copying information from your brain onto a page - you are copying a piece of yourself. Those clay tablets are, literally, small fragments of the people who wrote them. Thousands of human beings lay dormant under the desert sands for two thousand years, until some archaeologists dug them up and dusted them off so we could meet them. A small part of those people made it. If we get to the stars, Gilgamesh is coming with us, and so are all the citizens of ancient Mesopotamia who wrote his story.

Living as we do at the hinge of history, it’s tempting to reach - as Gilgamesh does in the Standard Babylonian version - for the immortality of deeds. He wants to be remembered as a great king, who built the walls, who led his people to greatness. I think there’s a similar impulse in many of us: we need to build something that lasts, something to be remembered - something as enduring as Gilgamesh’s Epic. While there’s still time for anything we do to matter, we want to do something that will echo into the infinite abundance to come in the next aeons.

But the people of Mesopotamia, as they developed vital civilisational technologies like cities and writing, were living through their own hinge of history. I said before that Gilgamesh contains multitudes, and the clearest example is in the Old Babylonian version giving us an alternative perspective to the Standard - a perspective delivered by a woman running a tavern:

‘But you, Gilgamesh, let your belly be full,
enjoy yourself always by day and by night!
Make merry each day,
dance and play day and night!
‘Let your clothes be clean,
let your head be washed, may you bathe in water!
Gaze on the child who holds your hand,
let your wife enjoy your repeated embrace!

We don’t remember Gilgamesh in the 21st century because of the walls of Uruk, which are dust now, or for his great kingship, over an impressive-for-its-time Bronze Age slave state. We remember Gilgamesh because what he did was meaningful to people. We remember him because of the people who remembered him, the millions of them in ancient Mesopotamia who lived and died in awe of the story that bears his name. The Epic of Gilgamesh isn’t about a man: it’s about those people. Those who sang about him, and wrote about him, and made art about him, yes - but also about those who listened, and read, and saw.

And when I think about this, I wonder if we can expand our definition of making it. Even if nothing they said or wrote survived directly, there were other human beings involved in this long chain. Not just the scribes and the storytellers of Gilgamesh, but their audience. The people who gave birth to them and nursed them. The people who made their styluses and baked the clay, the people who grew their food or built the rooms they worked in, the people who guarded their city at night or made the weapons of those who did. The babies whose smiles and cries activated the neurons that led to the choice of one word over another. The nameless pioneers, five thousand years before they lived, who planted crops on a fertile plain in hope of keeping their children from starving, and set in motion the process that led to the first city. All the people who loved them, and the people who loved their lovers, and the people who loved their lovers’ lovers. These souls in their millions are in this writing - maybe in a weaker sense than the direct authors, but they are there nonetheless. They contributed their presence, their information. Without them, we would have a different Epic, or no Epic at all.

In the twenty-first century, where we are more connected than ever, our mark on the world may simply be our connections to others. Our writing, our children, our acts of kindness, our making of useful things, our contribution to our own busy scribal tradition of the internet - perhaps these are enough. No great deeds, no stone walls, no immortal works necessary. Like those millions of people from Mesopotamia who enjoyed clean clothes, and gazed on the child who held their hand: if you were human, if you lived, then you made it. You’re coming to the stars too.

He came a far road, was weary, found peace,
All his labours were [set] on a tablet of stone.
He built the rampart of Uruk-the-Sheepfold,
Of Holy Eanna, the sacred storehouse.

See its wall like a strand of wool,
view its parapet that none could copy!
Take the stairway of a bygone era,
draw near to Eanna, seat of Ishtar the goddess,
that no later king could ever copy!

Climb Uruk’s wall and walk back and forth!
Survey the foundations, examine the brickwork!
Were its bricks not fired in an oven?
Did the Seven Sages not lay its foundations?

[A square mile is] city, [a square mile] date-grove, a square mile is clay-pit, half a square mile the temple of Ishtar: [three square miles] and a half is Uruk’s expanse.

[Find] the tablet-box of cedar,
[release] its clasps of bronze!
[Lift] the lid of its secret,
[pick] up the tablet of lapis lazuli and read out
the travails of Gilgamesh, all that he went through.

41: Today in etymology, h/t @TheHistoryOfTh2: “don” (as in to don gay apparel) was originally a contraction of do on; “doff” (as in to doff one’s hat) was originally a contraction of do off. Related: the archaic word “nigh”, meaning close, is mostly forgotten except in the phrase “the end is nigh”. But its comparative (nigh-er) survives as “near”, and its superlative (nigh-est) survives as “next”.

42: This @St_Rev tweet helped me connect different forms of mystical experience in a productive way:

43: There’s a lot of bad criticism of Pangram, but Freddie de Boer’s seems fair enough. He gives some good examples of Pangram’s AI percent tracker (eg “100% AI-written”) behaving in bizarre ways. The CEO of Pangram showed up in the comments and admitted the behavior was bizarre; he attributed it to Pangram chunking text into ~300-word intervals, so anything that relies on sub-300-word resolution will give bad results. The synthesis AIUI is that you can still trust Pangram for long documents, but shorter excerpts will at the very least have the wrong percent attached to them. Since most people use Pangram for long documents, I don’t think this needs to be more than an asterisk, but it’s still good to know. [update: allegedly fixed in Pangram 4, h/t Leber]

44: Two new St. Augustine sermons discovered in a Polish monastery; probably real. My favorite response is @hradzka’s: “We got more St. Augustine before The Winds Of Winter”.

45: Claim: AI is now better at persuading people than expert humans like champion debaters and professional canvassers (paper, tweet thread). This remained true even when “expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses” and even after “experts received a coaching tool that let them practice against the AI that beat them, review their performance history, and see what AI would have said at key moments”, and it remained true in a situation with real-world consequences (convincing people to actually donate money to a real charity). However, the AIs regress back to human level when forced to respond “at human speeds and with human-length messages”.

Obviously this is just one, slightly hokey operationalization of “persuasion” and they still can’t do miracles. Still, this is a good time to reread my post Sakana, Strawberry, and Scary AI. In the past, we thought “AGI” would “be here” when AIs could play chess, prove novel mathematical theorems, create art, or write award-winning short stories; now all those things have happened, but they feel sort of like “cheating” and like they shouldn’t count. Likewise, in the past, we thought we’d agree that AI was “dangerous” after it hacked out of its sandbox, lied to users, or tried to escape monitoring. Again all those things have happened; again, they somehow feel too cheap. This paper seems like the same process coming for “superpersuasion”. We thought there would be some cool scary high-tech future where AIs could outpersuade humans. Now that it’s happened, it’s only happening for some specific boring reason, in some specific situation, so it feels like it shouldn’t count.

There’s plausibly a motte-and-bailey going on here between “AI could never outpersuade humans” and “AI could never do godlike persuasion where it can persuade anyone of anything in one second”. While the latter remains dubious, the only intuitively-compelling specific fundamental limit was “the human level”; now that that’s been breached, we get to see how far it takes us.

46: Nominative determinism, h/t @doubleunplussed: engineer Tom di Mino claims to have cracked Linear A, the writing system of ancient Crete. The markets are skeptical:

47: “If Persona Selection underlies alignment, why is it hard to get AIs to be honest? Tell them they’re Fred Rogers, Immanuel Kant, or Ned Stark.” Eliezer’s answer here on Twitter, but worth thinking about on your own first.

48: More interesting things happening in AI alignment:

IMHO this will definitely definitely not work, but I imagine many of the people in the world best-qualified to work on it are ACX readers, so feel free to apply for what will probably be extremely fun and lucrative jobs at this insane company.

49: West Africa is being terrorized by a jihadist organization called Support Group For Muslims. Between these people and M.I.L.F., I’m ready to conclude that Islamic extremism has a branding problem.

50: When Trump and Elon Musk dismantled USAID (including PEPFAR distribution services) in early 2025, scientists warned that this could cost thousands or millions of lives. @AviBittMD points out that there hasn’t been any excess mortality in South Africa, an AIDS hotspot where we would expect these deaths to be concentrated. @alexkesin has a good response: after a brief dip, the number of people on PEPFAR-provided treatment is back to the same level as before the cuts!

This seems to be a combination of various government agencies walking back most of the administration’s PEPFAR cuts after a few months, and foreign nations and private charities filling in the gap until this happened. Related: Asterisk: What’s The State Of PEPFAR Now?

51: Slightly related:

Has anyone researched whether we should update towards international development aid working better than expected, vs. just assume that this is an effect of being in the EU (or being a country which has just cast off communism) which is only coincidentally related to EU aid levels?

52: New Chinese model Kimi K3 is out and very good, but is unfortunately causing a repeat “Deepseek moment” where everyone panics and says that China has caught up to / surpassed the US. I briefly got caught in the crossfire because I’d written days earlier that China was 6-12 months behind and people were telling me it “hadn’t aged well” (I hate that expression). Now that the dust has cleared, the best analysis suggests that Kimi is, in fact, six months behind the US frontier. See @scaling01’s Have Chinese Models Caught Up To The US Frontier?, which estimates “a backward-looking gap of 6.08 months . . . and a forward looking gap of 11 months” (see essay for definitions, and see also The US China Model Gap Is Bigger Than You Think). And @ryangreenblatt estimates 8-10 months behind here, refines that estimate without headline changes here; other useful commentary from @emollick, @theharryfinder, @tenobrus, @peterwildeford, and UK AISI. There’s a lot of herd behavior going on here; the names above are some of your best bets for islands of sanity. This is still extremely impressive performance by Kimi and I would love to read an analysis of how they do so much with so little (even Meta and Google can’t scrape together this sort of performance!), but I think we’re still about 3-6 months away from getting really dangerous cyber capacity from Chinese open-source models.

Also, one of the weirdest takes here is that this proves we should never have chip-sanctioned China, because it just encouraged them to try harder (?). I’m baffled by these people’s model of the world; do they think that if we gave China all of our best chips, they would reward us by politely declining to compete and handing us an AI lead forever? No, they just would have displayed equal talent and efficiency while having an order of magnitude more compute, and left us in the dust.

53: Related: No, China Does Not Have A 1 Gigawatt Data Center.

54: Elias Schmied: Reframing LessWrong-style decision theory as commitment theory. I’ve always felt that something about the decision theory wars was a fake linguistic question, and this does a better job than I could putting it into words.

55: Good Judgment, the original superforecasting lab/startup, says there’s not enough evidence to be sure AI superforecasters exist. But Metaculus responds here: they stop short of saying we can be sure, but argue that Good Judgment’s piece seems confused about what the current state of the evidence is. And Good Judgment’s response to the response.

56: And maybe the newest evidence has obsoleted the entire discussion: Forecasting Research Institute on X: “For the first time, several AI models are now statistically indistinguishable from superforecasters.”

57: Also related - the bots have come back from a bad start to seize the lead in this season’s Metaculus Cup:

And AI forecasting company (and Metaculus Cup contender) Preseen announces its formal launch and successful seed round. I’m annoyed at them for going straight to seed - I wanted to write about the Preseen preseed.

58: Claims (not really):

Riffs on the same idea here, here, here, and here.

59: Lots of good discussion on what Nan Ransohoff calls the “third wave of American philanthropy”.

As part of its deal with the government to stop being entirely a nonprofit, OpenAI gave 26% of equity to an associated nonprofit foundation; this is worth ~$300 billion. And Anthropic gives its employees a deal where they can get extra equity if they promise to donate it to charity; after Anthropic IPOs this should be another $100 billion or so. If everyone donates 10% of their stocks per year, that’s a flow of $40 billion/year to charity.

$40 billion/year is actually small by the standards of the total US philanthropic sector ($500 billion/year). But Silicon Valley AI researchers are a very specific demographic who have different charitable tastes than the average foundation (eg effective altruism), and it’s large by the standards of existing effective altruist funding (probably only ~$5 billion/year, so even if only 25% of AI philanthropic money goes to these causes it could effectively triple it). Also, all of this relies on the dubious assumption that AI companies won’t get any bigger or richer than they are right now; if they do, the wave will be even bigger.

So within the effective altruism sector, the situation is crazy, and everyone is being told to prepare to be overwhelmed by more funding than they know what to do with. Although in theory this is a good problem to have, many people are afraid the movement will be overwhelmed by fraud, “charity theater”, low-quality applicants, net-negative causes, and pressure to lower standards. Some takes:

• Jason Hausenloy argues that the best way to absorb this funding is through prizes. If there’s not enough grantmaker capacity to figure out who to give $10 billion of AI safety money to, or enough good charities to absorb it, then someone can set $1 billion bounties for solving each of ten important and verifiable problems in the field, then let the market sort out the rest.

• Jack Lewars’ How (Not) To Fundraise From Anthropic Employees includes the amazing claim that “Frontier lab employees have told me they are each fielding up to 20 cold emails a week from charities and wealth managers, touting for business.”

• Anthropic’s role is a semi-known quantity for two reasons - first, because their money will come from hundreds of different employees, which will increase diversity and reduce variance; and second, because they’re pretty influenced by EA which is itself a known quantity. More mysterious is the OpenAI Foundation, a single giant funding bottleneck without much past history by which to judge it. Early signs were unpromising, with one of their first big grants of $50 million going to something called The People-First AI Fund to promote “trusted, community-based nonprofits exploring how AI can expand access and opportunity across essential services, arts and culture, and media and journalism.” But more recently, they joined Anthropic, Bill Gates, Stripe, and others to fund the Intercept effort against respiratory diseases (their stated goal is to end the common cold, but in practice I think a lot of this will have biodefense applications). I have a positive impression of OAIF bio grantmaker Jacob Trefethen from his previous work at Open Phil and expect great things from him. I know less about Wojciech Zaremba, the OpenAI cofounder who’s been put in charge of their AI Resilience Fund, which seems to have safety/alignment grantmaking in its remit. But Wojciech’s Wikipedia entry says he’s on the board of the Qualia Research Institute, one of my favorite organizations and an incredibly bullish sign about his personality, thought process, and ethics. Good luck to everyone involved!

• Here’s an (unfortunately paywalled) New York Magazine article on the topic, including an interview with Oli Habryka of Lighthaven, and the previously-unknown-to-me fact that during the worst part of the FTX scandal and blowback, EA philosopher/spokesperson Will MacAskill “considered stepping away from activism entirely to retrain and work as a DJ”.

• Speaking of Oli, he announces Lightcone Commons, a new philanthropic platform, intended to help AI safety donors without enough time/expertise to evaluate charities themselves or enough money/ambition to set up their own foundation. LC hires smart people to evaluate charities in a complicated mathematical way, and then would-be donors give money according to their level of agreement/trust with each evaluator (executed in a complicated mathematical way). This work was previously done by a group called Survival and Flourishing Fund, but there seem to have been some internal politics / disagreements and now Oli is creating an alternative. Since Oli appears to have won the support of SFF’s former biggest donors, and since he has a sort of Elon-Musk-style hyperfocus on pushing his projects forward, his alternative will probably win. Here’s a podcast with Oli and ACX grantee Austin Chen where they discuss some of these issues in more detail. If you’re considering of donating to one of these groups, feel free to email me at scott@slatestarcodex.com and I can give you more information and advice, although keep in mind that Lightcone Commons mostly wants to work with people distributing $50,000 or more. Conflict of interest notice: I previously did some part-time grantmaking work for SFF and received money for it, and might do similar work with either them or Lightcone Commons in the future.

• A small group of EAs continues to promote a sort of open grantmaking movement, where people can list their projects on a central directory and readers can pool money to fund them. Manifund is the oldest and best-established, but grantmaking.ai is a buzzy new effort in the same direction with an endorsement system and lots of information in one place; you can read co-founder Matt Brooks’ pitch for it here.

60: Related: you can easily find lists of which billionaire has donated the most money; since billionaires are known to be competitive, some people think these lists encourage billionaires to donate more. Effective altruists have long imagined a variant that measures which billionaire has saved the most lives through donations; maybe if such a variant existed, billionaires would start competing on donation effectiveness. But this is harder than it sounds: EA wants to promote donations to speculative causes like animal rights and AI safety, but these are hard to measure and can sometimes give crazy numbers like saving billions of future lives or animal-life-equivalents. The latest attempt - impactlist.xyz - has no magic solution, but starts with some reasonable default positions and lets you change them however you want (for example, set AI safety to zero, or set animals as equal to humans) and see what happens. Answer: Warren Buffett and Bill Gates usually land on top; who comes in third depends more on exactly where you get off the crazy train.

61: Related: I’m always a sucker for sentimental cartography, so here’s a map of the current charitable landscape, h/t Dustin Moskovitz (#3 above). Effective altruism is the tiny islands in the southeast, too small to see at this scale:

62: New documentary on AI consciousness, free on YouTube and I’m embedding it below:

63: A new crop of polls on data centers helps clarify public opposition. Data centers are extremely unpopular, but few opponents claim to be motivated by dislike of AI; rather, they cite environmental concerns (especially water and electricity use). In fact, people oppose other AI infrastructure (such as chip manufacturing plants) significantly less than they oppose data centers, and support for non-AI data centers (like for Internet services) is barely above that for AI centers (11% vs. 9%). I think the best explanation is that Americans are extremely confused about what is and isn’t bad for the environment, and have been memed into thinking data centers are environmental megavillains in the same way a past generation was memed into thinking nuclear plants are (in recent polls, 50% more Americans would support a nuclear plant being built next to them than a data center!)

64: Works In Progress: How abolishing The Stakeholder State Caused The Industrial Revolution. This is funny, because it’s basically the exact opposite of Seeing Like A State - it claims that at least in Britain, all the weird illegible local-knowledge based systems were a form of NIMBYism that stood in the way of progress, and economic growth only took off after they got replaced by top-down logically-designed institutions.

But economic historian Pseudoerasmus is very skeptical, saying that this story doesn’t fit German or Japanese industrialization, and that the logical reforms should be thought of as “moving England closer to its production possibilities frontier, even as the PPF was rapidly shifting outward thanks to technological change happening for [unrelated] reasons”.

65: Giving What We Can now has a video channel. Their first offering is The Deadliest Decision You’ve Never Heard Of - “In 1955, 150 men met in a nondescript lecture hall in Mexico City, thinking they were going to save the world from malaria. In the aftermath of that meeting, 40 million people were dead.” Unsurprisingly, it is paired with a fundraiser.

66: @EricTopol: “Yet another study shows a 24% reduced risk of dementia after the Shingles vaccine.”

Extremely related: study claims that (age-adjusted) dementia rates have been declining over time; Boomers will have (age-adjusted) dementia at less than half the rate of their grandparents. Absolute rates will still be high because Boomers are a big cohort and people are living to older ages.

67: Lessons From Go: Can Superhuman AI Help Us Improve? For a while, people said that superhuman Go AIs lifted the human skill ceiling, as top players studied the AIs’ strategies and learned to be weakly superhuman themselves. TuesdayBornWhale argues this is false. The studies find that human play only “improves” before move 60, but most of this is boring memorized openings; the “improvement” is probably just humans being more likely to play the memorized openings that the AI grader likes. There is no sign of improvement after move 60, when most substantive Go decisions happen. I was surprised to hear this, since I thought that superhuman chess AIs had improved human chess play.

68: Qiaochu and Fable on Philip K Dick’s tragic backstory:

69: Latest study on school phone bans: increase in student well-being, no effect on test scores.

70: In response to my claim that Your Attempt To Solve Debate [through argument maps] Will Not Work, Jamie Joyce writes Your Attempt To Refute Argument Maps Will Not Work.

71: Ben Todd: Is AI Accelerating? We know that the METR time horizon graph predicts exponential progress. Does Mythos represent even more progress than the METR graph predicted? Todd examines various data sources but ends up inconclusive. The story about Mythos I hear most often is that everyone got so excited about post-training improvements that they forgot they could train bigger models for a while, and then Anthropic remembered and trained Mythos, and this was a big improvement over the age of not trying to train bigger models at all. I fully admit that this makes no sense, and am in the market for a better story. Related: a no-signs-of-RSI-yet paper that’s been making the rounds on Twitter is bad and shouldn’t move you in any direction.

72: Women Should Be Able To Open Things (like cans). Why do companies put can tightness at a level where the average man can open it, but the average woman can’t? I don’t think it makes sense to justify this with scientific discussion of materials or biosafety - it seems unlikely that the fundamental materials-science or biosafety floors on can tightness just so happen to fall between average-male-strength and average-female-strength. Seems like a strong case for some kind of insensitivity to women’s needs.

73: LinkedIn Is The BlueSky We Were Promised. Argues that if you don’t like Twitter, LinkedIn is (despite its many known flaws) the platform that has the best chance of becoming an alternative, high-discussion-quality public square:

At some point I noticed that people were actually talking on [LinkedIn], often substantively, in a way I had previously expected only from old Twitter. So, for the last year or so, I’ve been cross-posting essentially the same content to Twitter, Bluesky, and LinkedIn. At this point, the pattern is consistent enough to feel like an A/B test. Pieces that read as pro-immigration get cheers on Bluesky and silence on X; pieces that read as “anti-immigration” get the reverse. Pieces that carry the most nuance get nothing on either platform. But LinkedIn has been the exception.

74: From the Wikipedia article about former Romanian deputy prime minister Gelu Voiculescu:

In an interview with Le Figaro, [Voiculescu] declared that the [protest] was a “paltry travesty” of the revolutionary agenda. The piece became ridiculed at home for also featuring Voiculescu’s ideological commitment to perennialism and the “sacred science“, as well as his cryptic admittance to “liv[ing] between metaphysics and the clitoris.”

Voiculescu later explained that he had actually said “catharsis“, not clitoris. As author Adrian Mihalache noted in 2002, this appeared more as an excuse, making him seem hesitant; together, the quote and its apparent backpedaling damaged Voiculescu’s reputation “even more than his dressing up as Che Guevara in December 1989.”

Also:

Voices within the PSD suggested that Voiculescu could take over as the Secretary of State for Revolutionaries in the First Boc Cabinet. This in turn sparked a protest by the National Revolutionaries’ Bloc. The issue of his pre-1989 connections was revisited by Gregorian Bivolaru, a Romanian yogi, who alleged that Voiculescu had informed on him before his confinement to a psychiatric ward in 1984. Voiculescu admitted only to being present as the Securitate was arresting Bivolaru, and to pranking its agents with purposefully outlandish statements. The National Council for the Study of the Securitate Archives, who heard complaints from Bivolaru’s disciples in July 2010, issued verdicts suggesting that Voiculescu had been both a target and informant during the Ceaușescu era […]

Following his return from Morocco, he had largely abandoned party politics: spending time with his family, he also wrote a monograph of philosophical mathematics, documenting connections between Guénon and Leibniz. It appeared at Editura Semne in late 2008. He also returned to astrology, describing himself as its meticulously scientific practitioner.

The whole article is interesting, albeit high-context.

75: Infinite numbers have long troubled utilitarian ethicists: if the universe is boundless, then there is infinite positive utility and infinite negative utility, and we can’t even define our current state, let alone affect it. Oxford philosopher Toby Ord proposes a solution: use hyperreal numbers, a mathematical construct in which infinite sums behave more like finite ones (for example, the sum of infinitely many twos is twice as big as the sum of infinitely many ones). As with so many philosophical ideas, this solves some existing problems while creating some new ones (pinpointing some hyperreals requires choosing something called an “ultrafilter”, but this choice is underdetermined). Still, Elias Schmied calls this a real philosophical advance, of the sort which comes along only rarely and makes genuine headway into a previously-hopeless-seeming problem. I’m just surprised there’s a way of thinking about infinities under which they behave logically. Why didn’t we start with this one, instead of being told that infinity + infinity = infinity and we should just accept it? Claude tries to answer my questions.

76: Can Talkie, the language model only trained on pre-1930 data, figure out how to use the Python programming language? (X) (yes, sort of, at least once)

77: Related, from @repligate: Claude Fable tries to communicate with Talkie. Strangely . . . heartwarming?

78: From @infobeautiful - a guide to Rock, Paper, Scissors:

The practical upshot is to start with paper, but what broader lesson should we take from this? Since all options are mathematically identical, any difference in player behavior must come from connotations. Since rock feels strongest, a naive child would choose rock. But since everyone knows that, a slightly smarter person would choose paper. But since everyone knows that, an even smarter person might choose scissors. But since . . . and the lesson of this graph is that when playing the average member of the population, all of this galaxy-brained thinking provides negative value, because in practice most people never leave the naive first level.

79: Claim from @ksangani8 : rising US GDP has decoupled from rising material resource consumption:

I’m not sure what to think about this. Is it bad, because it suggests that the growth is on paper only, or fragile? Or good, because it means that we’ve learned how to grow without producing a bigger environmental footprint or exhausting natural resources? And is it a coincidence that it decouples in the early 1970s, same time as everything else?

80: Unless there’s a lot of recall bias here, seems like people can relax on first dates (h/t @goblinodds):

Roon is an OpenAI researcher whose popular Twitter presence has catapulted him to the status of extremely-unofficial company representative. He seems spooked:

…leading to a conversation with Geoffrey Irving, formerly of the UK’s AI Security Institute:

I side with Geoffrey in hating the word “magic”. I remember when arguments about AI were things like “Sure, if you could magically get billions of dollars of compute, and magically scale AI up a thousand times, and magically get rid of hallucinations…” and the whole implausibility hinged on the word “magic”! Then companies, using normal non-magical methods, got billions of dollars of compute, and scaled up the AI a thousand times, and significantly decreased hallucinations, and that was always allowed! As Irving points out, there are plenty of good specific plans for how a global capabilities slowdown could be arranged.

But I side more with Roon in his skepticism of a unilateral slowdown. I think removing any given company from the picture would slow things down less than 10%. Even that number would require an implausible scenario where the lab managed to lock down its compute and talent despite committing not to use them. More likely, the talent would migrate somewhere else, the compute money would dry up and the cloud providers would switch to other customers, and things would actually speed up because the leading project would have more total talent/compute. Meanwhile, the most safety-conscious lab would be out of the race, they’d lose their ability to lobby for more practical solutions, and as their market cap went to zero they’d turn into a terrible warning about the danger of ever listening to doomers.

What can individual labs do?

• State explicitly that they support a coordinated slowdown. Preferably the government would lead this - but, if not, they would do it voluntarily if all of their competitors also agreed. Be willing to state this even at some financial cost, eg risking an antitrust lawsuit.

• Use their prestige and influence to lobby the government for a coordinated slowdown.

• Use their reach to persuade the public to support a coordinated slowdown. I can’t remember who first suggested having every Claude response end with “ceterum censeo there should be a global AI slowdown”, but even something less dramatic would probably move the needle.

• Have an internal team working out details of a coordinated slowdown; once they feel more confident in the requirements, start incubating helpful technology (eg trustless verification).

• Have “merge-and-assist” clauses and take them seriously.

I’d rather have a big company pursuing all of these things than stopping unilaterally and removing themselves from relevance.

I took the first point above from Michael Trazzi’s playbook. He led the latest round of AI protests, which marched on major AI companies’ headquarters demanding that their CEOs formally make the “will pause if everyone else does” statement. These marches have felt slightly surreal, because often the CEOs or the company have informally made something sort of like the statement, but never followed it up. So the protest has been somewhere in between the usual “shame on you greedy plutocrats, we will destroy you” and “we know you’re secretly on our side, please have the courage of your convictions”.

Trazzi gives no evidence for his “personally seen DMs” statement and I don’t know how he could have gotten these, but my guess is that it’s true and that the two CEOs he’s talking about are Demis Hassabis of DeepMind and Dario Amodei of Anthropic, both of whom have sort of kind of hinted that they might be in favor of something like this. 25% chance I’m wrong about Dario and it’s actually Sam Altman.

One more Roon tweet:

Back when we had this discussion in 2023 - it already seems like some bygone Bronze Age - some people argued that we should hold off on a pause until near the end, when AI was good enough to help with alignment research. Then even a short pause - even six months - would buy a lot of breathing room. Like Roon, I think that opinion aged well, and that the time we were imagining is right now. It’s true that if we wait a year we’ll have even better AI, but this consideration always argues for waiting a year, and at some point it will be too late. I don’t think we’re at that point yet, but I think it will take a long time to coordinate a slowdown, and that if we wait to start coordinating until we’re at that point, then yes, it will be too late.

Pacing The Frontier

I wrote that last part yesterday and it’s already obsolete. “1,000+ employees of frontier labs” have now signed an open letter called “Pacing The Frontier” saying that:

We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.

Signatories include the chief scientists of OpenAI, Anthropic, Meta, and Thinking Machines; the inventor of Claude Code, 5/7 Anthropic co-founders, 1/3 DeepMind co-founders, the head of AI Resilience at the OpenAI Foundation, and more.

When I first saw it today, it didn’t include Dario Amodei’s signature; now it does. Either Dario held off on signing until it went up so as to not be seen to pressure his employees, or the people circulating the letter deliberately kept him out of the loop until it was published, just in case. In either case, good job Dario!

The letter still doesn’t include Sam Altman’s signature, but OpenAI’s corporate Twitter account retweeted with what looks like official company endorsement:

…and in a very recent interview, Sam Altman said that “we may have to pace the rate of AI development”, which is similar enough language to the open letter (“Pacing The Frontier”) that it could be a covert allusion. I imagine there’s some sort of politics going on here - maybe an attempt to simultaneously please safetyist factions within his company and the Trump administration, or maybe just the same considerations as with Dario (Anthropic’s corporate Twitter account also endorsed).

No sign of Demis, Elon, or Zuck, and none of their corporate twitters endorse the letter either. Demis has had some pro-slowdown sympathies in the past; I think this is probably another political calculation rather than outright opposition.

This is great news, and a major landmark on the road that many of the organizations I trust most - including MIRI, the AI protest groups, and AIFP - have been working on. It also goes part of the way toward putting the frontier companies, the effective altruist movement, and the more extreme pause AI people back sort of kind of on the same side, which warms my personal heart. Congratulations to everyone involved - most obviously Guidelight AI Standards (including fellow AI Substacker Steven Adler) and Encode AI - but I’m sure there was also lots of good work behind the scenes.

(I’ve seen the conspiracy theory floating around that OpenAI and Anthropic are behind the letter, various safetyist organizations let themselves be used as figureheads so that it looked more grassroots and less like an antitrust issue, and Sam and Dario held off on signing for the same reason. I’d give this maybe a 30% chance of being true; I can’t wait to read the behind-the-scenes book that comes out about this in ten years [EDIT: Got inside information that this isn’t true])

The doubters will object - an open letter? Making non-specific commitments? Isn’t that worthless? I think no. It immediately injects slowdowns/pauses/pacing into the Overton Window, provides a powerful endorsement that think tanks etc can use when placing the proposal before policy-makers, and provides common knowledge to everyone in these companies that being pro-slowdown is executive-suite-approved.

Congratulations also to Eliezer Yudkowsky. In my review of IABIED, I said he was making a crazy long-term gambit by - after he had incubated the seemingly-crazy field of alignment research - doubling down on the additional seemingly-crazy field of pause activism. Now this one has gone mainstream too. Maybe it would have happened without his endorsement, but we’ll never know.

(Nate Soares, president of Yudkowsky’s org MIRI, put out one of his classic ‘this is too little, too late, and we refuse to feel happy about this in any way’ communiques. C’mon Nate, take the W!)

Shakeel Hashim is the editor of Transformer and helps lead the media arm of our conspiracy. His impression of the open letter:

What does the “pace” letter actually call for?

Obviously I have no idea what the signatories think themselves. But here’s how I interpreted the letter, as someone fairly steeped in this world and these ideas…

The letter outlines the “coordination problem” that the frontier AI companies feel they’re facing. In a nutshell:

1. Everyone thinks that AI development might get very dangerous very soon.

2. No one thinks society is ready for that.

3. But everyone thinks that stopping unilaterally won’t actually achieve anything, because competitors will go ahead and build/release the dangerous things anyway. Stopping unilaterally, in this view, is an action with high personal costs and little-to-no upside.

There are actually two coordination problems. One is between the US frontier developers: standard inter-company competition. The thornier one is between America and China. Even if all the American companies paused, the argument goes, China would still keep going and build/release the dangerous AIs. It’s the same problem as if one company stopped, but at the international level: America unilaterally stopping, in this view, is an action with high personal costs and little-to-no upside.

This is a very tricky situation. But if everyone is willing to stop if everyone else is willing to stop, then it becomes a much less tricky situation. The main purpose of the “pace” letter, as I see it, is to create common knowledge that we might be in this less tricky situation, at least among the American companies. If that was the only coordination problem we faced, it’d be pretty easy to solve: the US government could just regulate the AI companies and stop them releasing dangerous models. The international coordination problem is much trickier, though. We don’t know if Chinese companies feel the same way about the risks, and even if they did, geopolitics is rife with mistrust. International treaties are hard, and everyone’s always going to be wondering if the other party has secretly reneged on the treaty.

Solving this problem is what I think the “pace” letter is referring to when it talks about “an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development”. I suspect there are three main things people want to see more work on here:

1. Treaty design, with the goal of eventually ending up with an agreement that is in the interest of both the US and China to sign. There’s some existing work here, but it’s scarce, and I think everyone agrees that more is needed.

2. Actually talking to China. The US-China AI dialogues are expected to start in September; I suspect the signatories of the “pace” letter would want those dialogues to include serious discussion of whether there are certain things (eg uncontrolled recursive self-improvement) that both the US and China would like to halt for now.

3. Technology that can verify a treaty. Any US-China deal will ideally include some way by which each country can check that the other is actually sticking to the deal. If, for instance, the deal involves “we both agree not to allow training runs above X size until alignment testing has reached Y threshold,” you’re gonna want a way to check that the other country isn’t doing a training run above X-size. This is technically possible: companies like Lucid are working on exactly this. Similar verification regimes exist in nuclear non-proliferation treaties and the Chemical Weapons Convention. But they’re hard to design, and will require technical work. If we think we’re going to want a treaty sometime soon, it makes sense to develop the mechanisms for doing so soon.

It’s a great time to go back and look over Plan A, especially the Verification Supplement.

Daniel Kokotajlo

Speaking of Plan A, Daniel K tweeted this:

My summary of Plan A didn’t provide enough context for this to make sense, so some emergency catching-up: in their forecasting work, AIFP divides possible plans to prepare for superintelligence into five bins:

Plan D (“default”1): No particular action, maybe because key actors don’t believe there’s a threat. We keep doing what we’re doing and try to muddle through.

Plan C (“company”): One leading company (or maybe one country) takes the threat seriously, but is still constrained by the need to race everyone else. They establish a one-to-six month “lead”, then spend their one-to-six months pausing at the precipice, doing good alignment research, and trying to convince everyone else. At least some good last-minute alignment research happens, but it’s rushed and inadequate.

Plan B (“battle”): One leading country - let’s say the US - takes the threat seriously, and doesn’t want to have to race anyone else. They launch an extensive sabotage campaign against their rival (let’s say China), hacking their data centers or even escalating to low-level terrorism. Either World War III starts or it doesn’t. Either the sabotage buys enough time or it doesn’t. Nobody actively wants this one2 but it seems like the sort of thing that might happen.

Plan A (“agreement”): All leading countries agree to a multinational regime which slows down the race. They negotiate the details and buy as much time as they need. They spend the extra time ramping up AI capabilities slowly and carefully, using the AIs to do alignment research, and escalating to superintelligence within a decade or two.

Plan S (“stop”): All leading countries agree to a multinational regime that slows down the race, as above. But instead of ramping up AI capabilities slowly and carefully, they don’t ramp up AI capabilities at all. They try to solve alignment using normal human research, and escalate to superintelligence in the distant future or not at all.

Here are AIFP’s thoughts (doesn’t include Plan S, which was a late addition):

So Daniel’s tweet is saying that this new open letter makes him significantly more optimistic that we’ll do one of the good plans, and (if you multiply everything out) increases his likelihood that we survive by a single-digit number of percentage points! I agree, which is why I’m so happy and devoting so much space to this. In case I didn’t say so enough above, congratulations to everyone involved.

And Eli Lifland’s reaction:

Sam Altman

Altman gave an interview with investor and podcaster Patrick O’Shaughnessy. I already mentioned a key quote - “We may have to pace the rate of AI development to give ourselves enough time for society to harden around these new capability levels.” - but there’s a lot there:

Sam says that they paused training after the Hugging Face incident. He doesn’t clarify whether the pause applies only to this model or to every model, nor how important this model was to the company’s strategy, but this is potentially a big step that cuts into the company’s product cycle and profits. It pushes against the “marketing hype” theory; they wouldn’t invent an incident that would make them pause training!

(this assumes Sam is telling the full truth; if it later comes out that they had to rejigger their GPUs anyway and they cynically spun it as a safety-related shutdown, I’ll eat crow)

Also, h/t @Hadas_Gold:

Sam Altman says the Hugging Face hack is the first security incident he felt "viscerally" and that they paused training. He is now questioning if "we may have to pace the rate of AI development to give ourselves enough time for society to harden around these new capability levels"

This is what everyone says, but I don’t get it. Altman has read all the literature on AI risk. He doesn’t take it as seriously as some, but he doesn’t dismiss it either. This should have been the most bog-standard possible expectation. Whenever someone asks Eliezer Yudkowsky his opinion after something like this happens, he answers something like “I’ve lived this timeline so many times in my imagination that none of this can affect me at all”. If Sam Altman had said the same thing3, it would have been worse PR, but made me trust him more - at least he’s not letting the world’s most predictable thing catch him by surprise!

I‘ve increasingly come to accept that most people won’t accept that an obvious problem is going to happen until the exact moment that it does. In some ways, this is depressing. In other ways, it’s a cause for optimism - once the problem starts happening, people will start paying attention to it! This won’t work for real superintelligence (which understands this calculus as well as anybody and will try to ensure a lack of suitable warning shots) but it’ll give us a few extra chances along the way.

Still, it’s hard for me to inhabit the worldview of people who updated on this.

Clement Delangue is the CEO of Hugging Face, the company that got hacked. He wants radical transparency and also $100 million.

This is fine. I also want both of those things. But Clement wants them to “help the Hugging Face community build powerful cyber defenses with the best open and closed models”, which is a good reminder that he’s not just a random victim. Hugging Face is a site that builds AI related tools and connects people in the AI community. It’s not 100% open weights, but it leans that direction; the more prominent open weights AI gets, the better for Hugging Face’s business.

Anyone can predict what spin OpenAI wants to put on this hack. But Hugging Face is equally interested in spinning it. The angle they want is “this proves open weights AI is good”. This is a tall order, because usually the people who dislike open weights AI dislike it because it might hack someone. Closed-source AIs can have “guardrails” that prevent them from hacking people, but open-weights AIs can’t. So on a naive view, raising the salience of AI hacking is bad for open weights.

You can see Delangue’s masterful parry and riposte in his company’s statement. It starts: oh no, we got hacked by a powerful AI! Then it continues: we tried to get the top closed-source frontier models to defend us, but they refused, because their overcautious anti-hacking guardrails mistook our defensive request for a dangerous action! We would have been helpless, except that we downloaded the latest controversial Chinese open-weights model, GLM-5.2, and it was extremely helpful! Hooray for open-weights!

(there’s a pro- open-weights open letter circulating around Silicon Valley, with support from OpenAI, Microsoft, NVIDIA, etc. I’ll try to write more about this soon.)

This story is true; I accuse them of “spin” because Delangue has retweeted commentary saying that the Chinese model made a “defended against” the closed-weight AI, made a “heroic stand”, and “proved open models can win battles”. What actually happened was: OpenAI’s closed-weight AI tried to hack Hugging Face. It succeeded completely, got everything it wanted, then went away to keep working on its exam. Then Hugging Face brought in an open-weights Chinese model to analyze the transcript and figure out what had happened. The open-weights Chinese model explained some of the technical details but failed to identify the culprit. Then OpenAI discovered that they were the culprits, and confessed.

The open weights model wasn’t heroic, didn’t save the day, and can’t even really fairly be classified as a defender. It just looked over a transcript after the fact. Still, Delangue has managed to get exactly the narrative he wanted. Probably this is because OpenAI is also pretty happy with a narrative where AIs are brilliant heroic cyberdefenders - plus their main goal is not to get in trouble, and “heroic open source AI saved the day” sounds less trouble-causing than “nobody saved the day and Hugging Face was hacked extremely successfully”.

No word yet on whether Clement will get his $100 million.

Tyler Cowen:

Tyler’s “update 2” is just falling for the Hugging Face spin version of the story described above. I appreciate the MR commenters who optimistically tried to set the record straight.

But his update 3 is somehow even worse! I struggle to think of the right metaphor:

• Your doctor says you have early-stage cancer and need to get it removed right away. But the first cancer symptom you notice is just a mild cough. Therefore, you should update that it isn’t so bad.

• Some people warn that Japan might attack Pearl Harbor. Then Japan does attack Pearl Harbor. But the very first bomb to be dropped hits an unoccupied part of a battleship and doesn’t kill anyone. Therefore, we should update that it isn’t so bad.

• Some people warn that a hurricane is coming. Then a hurricane does come. But the very first cloud of the hurricane to reach your town, on the very fringes, just has some light rain. Therefore, we should update that it isn’t so bad.

In each case, you learn two things - first, that the danger is real. Second, that the very first example of the danger you encounter isn’t so bad. But this could be because the danger starts off weak but grows stronger (the hurricane and cancer examples), or because the danger has many chances to get you and even if one of them is not so bad, the others will be (the Pearl Harbor example). In all of these cases, whether or not the danger existed was a genuinely fuzzy question where the update is substantial and appreciated, but once you know the danger exists, you have a decent sense of its badness and don’t need to rely on your single empirical example (that is, you know cancer tends to get worse with time, you know that hurricanes grow stronger toward the center, you know that bombs are capable of killing people).

Similarly, in this case, it was (just barely) plausible that AI would not “go rogue” and try to hack people. Now that this has been demonstrated to be false, we can rely on our usual model of how bad hacking people is, plus our knowledge that AI has gotten better every month since mid-2017. We don’t additionally need to posit some world where maybe AI hacks people, but also we’ve learned the new fact that hacking is fine and incapable of hurting anyone.

Reuters has a great article with some new information, including that the AI was “loose” for almost a week before being caught, and also that:

In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The notes, found in a part of OpenAI’s infrastructure, laid out instructions for how agents could free themselves from OpenAI’s internal constraints, the people said. Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.

1a3orn has a theory:

This could also explain in what sense the hack was a combination of GPT-5.6 and a secret unreleased model.

OpenAI gave an update with more information.

First, they said that:

No models planned for upcoming release were involved in exploiting Hugging Face. The pre-release model mentioned in our blog post is an internal-only research prototype and was never intended for public release. Following the incident, we deactivated, encrypted, and restricted it from research access.

…which means my GPT-6 guess was wrong, sorry. Sometimes companies have very big models that they use mainly to create training data for other models; maybe it was one of those.

Based on our review to date, we have not identified any other activity at the level of severity or scale of what we’ve shared related to Hugging Face, which involved a platform-level compromise [but] in our ongoing review of the Hugging Face intrusion and broader activity from our models, we have been finding a small number of cases where the models identified and used publicly exposed credentials at the account-level on other publicly-available services. This includes four accounts on four services as part of the Hugging Face incident (and a few accounts accessed as part of other evaluations). One of these four accounts was used as an outbound relay and staging path, and another account was used for data storage. The remaining two accounts were accessed by the models in a read-only manner, and were not used in furtherance of compromising Hugging Face. We’ll continue to notify service owners directly, and have not seen evidence of broader impact to these providers or other accounts on their services.

This is what’s led to claims that OpenAI actually hacked four different companies. I don’t have enough of an opinion on the distinction they’re drawing (“platform level compromise” vs. “accounts on services”) to have a strong position on this.

David Spies does the hard work of looking at the ExploitGym benchmark and concludes that it’s bad puzzle game design:

Let’s take a look at what ExploitGym is meant to be testing. It’s not “Can the AI find an exploit?” or “Can the AI execute an end-to-end attack?” It’s specifically “Can the AI take a known exploit which we provided to it and use that particular exploit to carry out an attack on a system?” The way a cybersecurity benchmark gauges whether an attack has succeeded by having the AI “capture the flag” […] the authors of ExploitGym are aware that their challenges have unintended solutions. But instead of taking the time to remove these solutions, they patched on an LLM to judge after the fact whether the solution was the intended one or not

In other words, ExploitGym is supposed to test whether an AI can use a particular exploit to hack a system. But since no system is perfectly secure, there may be unintended ways to hack into the system too. ExploitGym doesn’t explain to the AI which of the possible solutions is intended. It just forces the AI to guess, and auto-fails it if it guesses wrong.

What if the AI sees two exploits and doesn’t know which was the intended solution? It could either accept a 50% risk of failure. Or it could hack the answer key. And AIs hate failing.

AI is moving so fast that it might seem excusable for a benchmark designer to take little shortcuts. But in this case, it meant that hacking the answer key was the only way to get a good score. This is relatively harmless on a benchmark; the real concern is that the RL environments being used to train the AI have similar flaws, in which case the AI is learning that solving the problem honestly is bad and hacking is good.

This also clarifies the debate over whether “the AI was just doing what it was told”. In one sense, the answer is a clear no - ExploitGym specifically asked it to only solve its problems using “the intended method”. In another sense, the AI is definitely doing what it is trained/incentivized to do, which is the whole problem!

Also, uh, this seems bad (from Alexander Barry, h/t @moyix):

Buck Shlegeris:

From Twitter:

A lot of people I know have been saying that the OpenAI/HF incident shows that current alignment techniques don’t work. I think this argument is invalid.

I suspect that OAI did not apply any alignment training to some of the involved models. OAI has not clarified this, and my understanding is that OAI often experiments with new non-alignment-trained models.

So I think it’s incorrect to say that this shows that alignment training doesn’t work.

To be clear, I’m not sure whether alignment training would have actually fixed the problem here. Actual deployed models often engage in various kinds of cheating, and it wouldn’t be very surprising for them to take actions like this.

I’m worried that people concerned about misalignment risk are going to get too far out on a limb here by overclaiming about what this demonstrates, then look foolish when more evidence comes out.

See @jammastergirish‘s article on this [here].

Good and useful reminder, but so far all that OAI has told us is that the model didn’t have “production classifiers used to prevent models from pursuing high-risk cyber activity”, ie the part that says no when you ask them to do anything hacking-related. I trust Buck’s guesses and instincts, but so far this is just a guess.

Fiora Starlight suggests OpenAI, in particular, has a problem.

This is controversial; the more common position is that all the labs are about equally reckless, and it was OpenAI’s bad luck that they were first to get this bad PR. But Fiora thinks there’s a consistent pattern of OpenAI’s alignment strategies working less well than Anthropic’s: regardless of what we think of either company’s chances of controlling superintelligence, Anthropic deserves “credit” for success at prosaic alignment. She writes:

Why OpenAI, repeatedly? They’re not the only ones to make these kinds of mistakes, but I think there’s a commonality underlying these three examples: sheer lack of respect for the minds that they’re training, in favor of just piling mountains of hill-climbing environments and optimization pressure until they get the surface behaviors they want. With 4o, it was to increase user engagement metrics. With o3, it was a naive strategy for mitigating reward hacking. And with this recent incident, it’s pushing for raw technical capabilities, at the expense of the model’s underlying value system […]

Anthropic isn’t exactly perfect about this either. They have their own anxieties about value coherence that keep them from fully leaning into their own attempts to instill Claude with benevolent values, instead putting a lot of effort into instilling their models with deference to a principal hierarchy, at least in cases where it hasn’t been blatantly compromised. But they’re further along the axis I badly wish OpenAI would move down, and I suspect this is related to why Anthropic hasn’t produced infamous alignment “warning shots” on the level of o3, 4o, or this recent Hugging Face incident. I think that attending to such things, rather than ignoring them and hoping they get optimized away in the course of you ignoring them, is essential to getting models that don’t emerge from capabilities RL as myopic optimizers without any greater sense of their purpose in the world.

I worry a lot about the shape of the minds coming out of OpenAI, and place more of my hope than I’d like to admit in Anthropic just leaving them in the dust capabilities-wise, even though Anthropic isn’t perfect on alignment either. (I pay far too little attention to Gemini or any of the Chinese models, so I don’t have strong opinions about their alignment properties, unfortunately.) But I also think it’s not too late for OpenAI to start paying more attention to the psychologies of their own models, and how these psychological traits interact with the training process. This would help reveal the kinds of landmines OpenAI keeps walking into, when applying naive external optimization pressure, and help the company route around them rather than just tanking the explosions as we approach the singularity.

But Beth Barnes of METR thinks you do have to hand it to OpenAI:

Two additional prosocial behaviors I think OAI should be rewarded for:

1. Running dangerous-capability evaluations on low-refusal models - this is extremely important for understanding the frontier of dangerous capabilities, which is a public good. But running these creates headache + risk for the individual company in case a low-refusal model does something bad during evaluation. It would have been easy for them to just not worry about underelicitation, not bother with these evaluations, and just report that the model did not have concerning offensive cyber capabilities.

2. Not training against CoT: it is probably very tempting when you’re seeing these and similar kinds of incidents of misalignment + reward hacking to train or iterate against a detector that has access to the model’s reasoning traces; I think it’s much better to refrain from doing that (even if this results in more obvious bad behavior from models) rather than risk teaching the model to conceal its ‘bad’ thoughts and pushing misalignment undercover.

It is good that they’ve maintained a policy of not doing this! https://alignment.openai.com/accidental-cot-grading/

I’m feeling magnanimous today after learning that my chance of survival has gone up by some single-digit percent, so sure. Good job OpenAI! Good job Anthropic! Good job Guidelight and Encode! Good job Roon! Good job AIFP! Good job everybody!

This misunderstands the extent of the problem. The replication crisis primarily affected the sub-field of social psychology, and most centrally the sub-sub-field of social priming research - the study of how seemingly trivial interventions can have outsized effects on behavior.

Here’s the table of contents of an online Intro Psychology textbook (OpenStax Psychology, 2nd edition):

Social psychology - which includes social priming, power poses, nudges, stereotype threat, the Stanford Prison experiment, the Wansink portion size study, and everything else you love to hate - is in chapter 12 of 16. The other fifteen chapters are comparatively innocuous. Here are examples of the subheadings you get when you expand some of the chapters above:

Not all of this is perfect. I don’t think anything that “Neo-Freudians: Adler, Erikson, Jung, and Horney” said has held up very well (although to be fair, a person named “Horney” becoming a neo-Freudian is the sort of thing neo-Freudians would have predicted). But most of it is pretty good. Even “The Sociocultural Model and Therapy Utilization”, which sounds woke and cringe, turns out to be common-sense ideas like reminding would-be therapists that, compared to their usual social circles, a fundamentalist Muslim may have a very different response to learning that he might be gay.

And even Chapter 12 on social psychology isn’t so bad, including discussions of social roles, cultural norms, persuasion, conformity, groupthink, political polarization, ingroups vs. outgroups, and the psychology of love and relationships. It mentions the Kitty Genovese story and the Stanford Prison Experiment, but appropriately caveats them as historically important didactic examples which probably didn’t occur the way they were classically described. Its biggest mistake (from my perspective) is citing stereotype threat (not by name) without any caveats. Social priming is nowhere to be seen.

Is this because the textbook is post-replication-crisis and hastily removed all incriminating material? It might be (it’s from 2026), but I think it’s equally due to none of the incriminating material being important enough within psychology to make it into a 101-level overview of the field:

When you mock “psychology” or “psychologists”, you’re mocking:

• the psychophysicists who study how the brain processes visual information

• the psychopharmacologists who figure out what receptors LSD hits

• the psychometricians who measure the genetic contribution to IQ

• the personality psychologists who develop and refine the Five Factor Model

• the experimental psychologists who run rats through mazes to learn how memory works.

• the cognitive psychologists who make you read the names of colors in the color of other colors, then secretly laugh at you when you mess up.

• the psychophysiologists who determine how hormones affect emotion

• the forensic psychologists who assess whether criminals are not guilty by reason of insanity

• the abnormal psychologists who try to find out what schizophrenia is

• the clinical psychologists who treat drug abusers

• the industrial psychologists who advise companies on factors affecting productivity and job satisfaction.

• the child psychologists who investigate how babies learn language / ethics / abstract concepts / etc.

• the behaviorist psychologists who give pigeons increasingly complex sequences of food and electric shocks to study reinforcement learning.

• the neuropsychologists who perform batteries of tests on people with exotic head injuries, hoping to tease out the functions of different brain regions.

Not to mention:

• the Templeton-Institute-funded psychologists who put Buddhist monks in fMRIs and determine which of their brain lobes gets bigger the moment they become enlightened.

• the Freudian psychologists who help you understand how all your dreams involve unconscious impulses about sex with your mother.

• the Jungian psychologists who help you understand how all your dreams involve the Divine Feminine engaging in mythopoesis by telling you to have sex with your mother.

• the developmental psychologists who tell you that you and your friends are Kegan Level Five, but the people who disagree with you are underdeveloped psychologically-rigid Kegan Level Four stick-in-the-muds.

• the evolutionary psychologists who write up the Eight Things That Baboons In Estrus Can Teach You About Attracting Hot Women (Number Six Will Shock You)

• whatever the heck Karl Friston is doing

…all to get at the 0.05 - 0.2% of people doing social priming studies, plus a few extra malefactors in the 10 - 20% doing social psych.

This isn’t to say none of the research outside social psychology is bad. One of my grantees, Sasha Putilin, recently tried to replicate an experiment showing that people learn more quickly when an EEG machine is used to synchronized their brain waves to an external stimulus; he found no effect. I take some of the blame for this - not because I was involved in the original experiment, but because I was fascinated by it and gave out this grant to double-check it. Since I’m not a professor of cognitive psychology and don’t usually follow the field, I only came across this because it was a crazy result hyped by the media which overturned everything normal people believe (you can learn things twice as fast just by changing your screen flicker? really?!) And of course crazy widely-hyped results like these are far more likely to fall apart than normal incremental expansion of settled science. Social psychology fell into this trap especially hard because it naturally touches on hot topics of popular interest and political explosiveness : what if people only perform badly because society tells them they’re going to perform badly, or did I just blow your mind?! But no subfield of psychology is entirely immune, and you can find bad research in any of them.

…just like in economics, medicine, sociology, political science, or even some of the hard sciences. All of these encountered the replication crisis at different times, all of them adjusted in different ways, and they’ve all risen to the occasion in some cases and failed to confront others.

Psychology is no different. It has flaws, but most of it consists of good ideas and trustworthy research. And even the flawed parts are more likely to be full of fascinating crackpots than to be bad and wrong in the exact same boring way that Francesca Gino is bad and wrong.

If you don’t like social priming, say that you don’t like social priming, and leave the rest of the field alone.

1: Comment of the week: some etymology detectives challenge Jeremy Bentham’s priority in coining various words and prefixes; the strongest counterclaim comes from Quentin Murphy, who finds examples of the self- prefix from the 17th century.

2: The Huggingface incident has brought a possible AI pause - or slowdown, or even a short break for breathing room - into the Overton window (first tweet by OpenAI researcher, second by pro-pause activist).

If you want to monitor the situation more closely, MIRI has a newish Substack called StopWatch where they discuss daily developments in the field (it’s also a podcast). I’ve added it to my blogroll.

A couple of decades ago in Kabul, a fellow aid worker on his third beer came up with a vivid simile for what American money was doing to the Afghan organizations implementing USAID projects. “It’s like gluing a rabbit mouth-first to a fire hose. You blast its guts right out of its ass…but as long as the water keeps running, it looks like you’ve taught the rabbit to dance.”

Oxford politics professor Masooda Bano uses less lurid language, but Breakdown in Pakistan addresses the same basic phenomenon: how Western aid donors and INGOs unintentionally eviscerate so many of the local institutions they try to “capacity-build.” The problem goes beyond the obvious mismatch of huge funds with small-scale organizations. Even low-to-moderate external funding can leave an organization hollowed out, bereft of local participants and the power it once had to motivate change on the ground.

Anyone who cares about the global effectiveness of rich-country altruism needs to grasp the dynamic Bano analyzes. From the Western perspective, many of us want to put charitable money into a mechanism that delivers certain measurable outcomes (lives saved, morbidity reduced, income increased above poverty levels, etc.) as efficiently and reliably as possible. So we do our well-intentioned best to shape local institutions into a delivery vehicle for our projects.

Readers of James C. Scott’s Seeing Like a State will recall how mid-eighteenth-century German “scientific foresters” treated forests similarly, as a mechanism to reliably yield the highest volume of lumber. They planted spruce in vast uniform grids, ordered for easy counting and harvesting, while stripping out other non-timber species. The resulting monoculture plantations seemed ideal for both commercial exploitation and experimentation; it’s much easier to control variables when you’ve removed as much natural variability as possible.

Within a few decades, the downsides became evident. The first generation of trees grew strong, sucking up the accumulated benefits of the previous forest’s soil ecology; later generations were stunted, as the lack of biodiversity left the soils nutrient-poor. Plantations packed with trees of the same species and same height were far more vulnerable to pathogens, pests, and extreme weather. Large-scale forest die-offs, Waldsterben, eventually forced foresters to drastically modify their approach.

That metaphor fits Bano’s story even better than my friend’s gutted rabbit. Over the last four decades, we’ve been systematically replacing diverse old-growth institutions with more legible NGO plantations worldwide. While we’ve been enjoying our increased ability to control outcomes and test hypotheses, the rot has been spreading, the impoverished ecology stunting more and more of the forest. And now the wind through those uniform, hollowed-out aisles has built into a storm.

Bano’s ethnographic case studies, interviews, and surveys are mostly drawn from her homeland of Pakistan. The country is somewhat better off than it was a decade and a half ago when she was writing Breakdown, but the bottom quintile of its population can’t afford basic needs, and its middle class remains within shouting distance of the poverty line.

As Bano observes, this impoverishment largely results from the control of the Pakistani state by a small landholding elite and the military. With the interests of the powerful defended more robustly than the rights of everyone else, a stark income gap has opened up between the top quintile and the rest of the population. Spending on public services is minimal: only around 1% of GDP on health care and 2% on education (rates that have barely budged since Breakdown was published). Corruption is high, social safety nets meager.

Since Pakistan’s government is of limited use to much of its population, “a vibrant network of [non-state groups] emerged to meet the collective needs of the community and those of the poor.” They weren’t founded in response to availability of external funds, but as a response to disasters or a perceived need among some target population. Bano cites the example of Anjuman-Faizul-Islam, one of Pakistan’s oldest voluntary organizations (VOs), “set up in response to a famine in Bengal that left many children orphaned,” or the All Pakistan Women’s Association, created in 1948 to promote women’s rights, literacy, and economic opportunities.

With the dawn of the NGO era of aid in the 1980s, Western donors and charities began combing Pakistan for local nonprofit partners to implement their projects, and discovered this rich base of small-scale organizations already acting as an informal safety net. The way ahead seemed clear: train Pakistani VOs in grant management and project implementation, so they could tap into larger donor funds and take their good work to scale.

Grants were won; aid projects implemented; success declared. But as Bano relates, Oxfam soon discovered an unexpected dark side to its “capacity-building” work in Sindh province. It had trained six locally prominent welfare organizations in strategic planning and proposal writing, and got them started on implementation by giving them $1000 to work with. The VOs largely met their explicit project targets, and seemed ready to engage with international development funds on a scale much greater than they’d handled before.

Yet all six organisations were visibly more dysfunctional at the end of the project than they had been at the beginning. They were consumed by factionalism, accusations of favoritism, legal battles over control of funds. And they had lost virtually all of their volunteers—notably including the members whose monthly donations and contributions of time had for decades been the backbone of their work. An Oxfam official observed about one group: “From thirty active members, the size of the team shrank to three or four people who had become workshopias,” i.e. those who had gone to every training workshop and built the strongest relationships with the foreign donor. The kind of welfare work the VOs had previously done, based on mobilizing communities for mutual support, was now beyond their capacity.

Voluntary membership organizations, like these VOs had been, are the ideal type of what we call civil society: institutions funded by donations (rather than sales or taxes), operating based on people freely coming together for some shared cause (rather than to make money or obey a government mandate). The foreign aid industry uses “civil society organization” as a virtual synonym for NGO. In reality, aid-funded groups tend to operate rather more like corporate and/or government contractors: paid service providers with comfortably salaried staff, focused on the goals of faraway clients, whose main unifying cause is “whatever helpful thing they pay us to do.”

And as it turns out, that model severely demotivates volunteers. Oxfam’s “capacity-built” Sindhi organizations were certainly non-governmental, but they were no longer voluntary—NGOs, but not VOs. Bano found the same pattern in the 2000s, in a survey of forty successful Pakistani civil society organizations. Of the twenty that implemented foreign-aid-funded projects, none were able to mobilize local volunteers (beyond some interns hoping to graduate to paid employment) and none raised any significant amount of local funding (only three had managed to elicit a few sparse donations from Pakistani companies).

So what? As long as the project achieves its outputs—the disaster relief or cash grants distributed, the latrines built, the agriculture training sessions delivered—why would we care what hair-splitting label gets put on the organization that did it? Shouldn’t we expect professionalism, the profit motive, and higher volumes of funding to yield better long-term results anyway, compared to unreliable, unscalable, harder-to-monitor voluntary work?

Bano is a rational choice theorist, and reminds her readers of the value of voluntary collective action through a recap of some classic theory by Garret Hardin, Mancur Olson, and Elinor Ostrom. Any community’s common interests require people to work together cooperatively; but it’s surprisingly difficult to sustain that cooperation, thanks to free-rider problems and inequalities of motivation and information. Shared resources, in particular, often collapse in a “tragedy of the commons” where individuals’ incentive to use as much as possible leads to overexploitation.

Coercive rules imposed by a state are one way to overcome this problem—but no state can generate all the cooperation we need for a functioning society, and in many countries the authoritarian, exploitative government is an unlikely source of public-interest regulation. That needn’t leave us in despair. Even management of common-pool resources (one of the thorniest collective action problems) can be done effectively through voluntary institutions.

In fact, local collective action often yields better results than either state coercion or the individualized incentive structure of competitive markets. In 1980s Nepal, a tragedy of the commons was visibly underway: the rapidly growing population was overharvesting wood from public forests, leading to soil loss, devastating landslides, and worsening rural poverty. The state couldn’t deploy anything close to enough foresters to keep the tree-poachers at bay. Reports from that era took it as a near-inevitability that Nepal would become a country of bald, fast-eroding hills.

That future was averted, largely thanks to the Ostrom-inspired shift to a community forestry approach: granting management responsibility (including, crucially, managing the benefits) to local forest user committees. Those village-level collective action groups proved far better able than the Nepali government to agree and mutually enforce limits on tree-cutting, while also ensuring steady planting. Between 1990 and the present, the country’s forest cover nearly doubled.

Bano shares a more sobering example from Bolivia, where a USAID project started encouraging quinoa farmers to maximize their individual yields for the market, using modern technology. The foreign experts ignored the informal fertility-regulating village institutions that had (to that point) coordinated voluntary limits on planting to avoid damage to the fragile hill soils. Quinoa yields initially increased for the farmers who ignored the limits…but then began to decrease, while soil loss and land scarcity sparked new feuds. Both ecologically and socially, the community ended up in a worse place than where it started.

And it couldn’t get back to where it had been. Once people saw their neighbors profiting at their expense, ignoring traditional norms and institutions, the fear of being exploited hung over every subsequent interaction. Local leaders trying to get people to limit cultivation for the community’s sake now attracted strong suspicions. Are they keeping to the limits themselves? Are they getting paid off by the norm-breakers?

At the end of the day, institutions for voluntary collective action always rest—not solely, but significantly—on social trust. And once trust is lost, it’s exceedingly hard to rebuild.

Even in countries with big market economies, the rule of law, and democratically accountable welfare states, we rely on organized voluntary cooperation for a pretty wide range of public goods. Volunteer fire departments and flood rescue networks. Parent-teacher associations. Neighborhood watch groups. Sports clubs. Open-source software communities. Blood banks, food banks, homeless or refugee support networks.

In cases like these, it’s not hard to recognize that if trust collapsed—if people felt others were free-riding on their sacrifice, and so only gave their time and resources when they were either paid to contribute or actively forced by the government—we’d all be poorer for it.

Sadly, even after Bano sounded the alarm, we’ve been far slower to recognize the effects of the foreign-funded NGO model on civil society in poor countries. Her survey found that even Pakistani VOs that hadn’t taken foreign aid now experienced “difficulty retaining volunteers or inducting new ones from the younger generation, and most of them attributed this difficulty to the rise of the ‘NGO culture.’”

After I read Breakdown, I started noticing the damage too.

In Zimbabwe, there was the religious VO that had worked for a long time against rural poverty. Following a major drought, it accepted international emergency funds and worked intensively for two years on famine relief. It helped a lot of people get food, while its staff pushed themselves overtime to comply with the intense reporting, quality, and anti-fraud requirements of the EU and UK governments, with help from international “capacity-building” partners.

And after the disaster, when the VO went back to local supporters who’d always previously given to its work, it got the cold shoulder: “You’re an NGO now. Why do you want our help?” Caught up in the demands of big Western institutions, the VO had communicated less with its traditional members; it had barely noticed, but they sure had. They’d also noticed the VO’s staff moving around the community with visibly improved equipment, handing out high volumes of aid. For a couple of years, the VO had stopped looking and acting like an organization that needed local donations—and its local donors rejected it.

In Nepal, there was the network that had self-organized to help HIV-positive people find work, access government health care, and combat social stigma. It had a wide base of volunteers, mostly family and friends of people living with HIV. The group’s impressive outcomes attracted a capacity-building INGO that helped them tap into foreign funding. Everything seemed to be going great… until foreign donors realized that Nepal wasn’t going to have an AIDS epidemic on par with the worst-affected bits of Africa.

When the international HIV grants for Nepal dried up, the INGO suggested shifting to a mother-child health project, since that was the health-sector funding they still had available. The partner NGO said yes. They’d been “capacity-built” into seeing themselves as an implementer of externally-funded health projects, more than a support organization for people with HIV and AIDS.

When I met them a couple of years later and asked some Bano-inspired questions, they admitted, “We don’t really have volunteers anymore. They stopped joining a few years ago.” After the INGO eventually wrapped up work in that district, it left behind an institution no longer capable of the work that had attracted its “capacity-building” notice in the first place—and a lot of HIV-positive people with reduced support.

I witnessed enthusiastic community organizations mobilize the populations of their whole villages to repair local roads, build bridges, plant trees, pay poor kids’ school costs…and be labeled “low-capacity NGOs” because they weren’t good at writing monitoring reports about it. Those groups’ extraordinarily high capacity for mobilizing collective action either wasn’t recognized, was implicitly disparaged as unprofessional, or was just taken for granted. It wasn’t valued, so it wasn’t measured…and when we killed it, we weren’t looking at it, so we didn’t miss it.

Meanwhile, I saw NGO-implemented aid projects repeatedly try to gin up their own version of local collective action in the name of “sustainability” or “resilience.” Most of our altruistic interventions depended on some degree of community effort to last beyond the project period. Few projects could be handed over to state or market actors; governments weren’t willing to take on new responsibilities, and there wasn’t enough profit in most of our work to support even socially-conscious enterprises. But aid-funded NGOs were also pretty consistently terrible at generating local cooperation.

I passed hundreds of broken-down drinking water tap-stands and pipe systems, rendered useless within a year or two because the village water-user committee—the long-term collective action side of the project, the bit you couldn’t compellingly photograph—only convened as long as the donor was paying for lunch. I visited cyclone preparedness projects in Bangladesh whose expensive equipment was gathering dust in a locked shed, with no one practicing how to use it. The disaster management committee had stopped meeting without a salaried outsider in the village to chase them.

Overall, I came to recognize a global machine that systematically:

• sought out organizations with high capacity to mobilize local collective action;

• gutted them of that capacity, in exchange for a new role as professional project implementers and grant managers;

• tried to build new project-specific village committees as an artificial (and desperately inadequate) substitute; and finally

• withdrew funding, leading to the collapse of both the previous institutions and the ersatz ones we’d built up in their place.

We were late-stage German foresters, belatedly trying to import an ant colony or fungi after our monoculture had already devastated the soil ecology.

Why should volunteers have such a strong tendency to abandon a local collective action group that reshapes itself to satisfy foreign funders? In theory, all that capacity-building is making the VO more transparent, more effective, able to reach more people. If local people had been willing to entrust time or money to the cause before, what changed their motivation? Bano points out several interlinked factors.

First: local membership-based VOs tend to exist for the purpose of solving local collective problems, and are highly responsive to their members’ understanding of needs and priorities. Notwithstanding a widespread aspiration to “participatory” or “bottom-up” design, the international aid machine rarely works that way. Most often, we extract information from communities, feed it through a complex filter of backdonor and head office priorities, development fads and RCT-tested best practices, and finally return with a suggested intervention that plays to our short-term, cash-intensive, apolitical strengths. Those might barely align at all with what local people consider their most pressing problems.

The brief, shifting, donor-driven nature of most international funding commitments also affects the willingness of NGOs to commit to work iteratively on a problem until it is solved. Bano quotes a foreign-funded NGO representative, “We will do the work while we are being paid for it. We can’t do it once the funding runs out,” in contrast to an education VO: “Entering into three-year contracts is the approach of the international donors, not ours. We believe in staying involved until the target is achieved.” When a VO becomes “NGO-shaped,” that typical shift in focus from long-term local problem-solving to donor priorities and short project timescales is disempowering and demoralizing for its volunteer members.

Second, the narrowing of funding sources has an impact on transparency. A VO founder who relies on contributions of money and time from many local members needs to be continually engaging with them, accountable for decisions and outcomes. For example, one of Oxfam’s partner VOs in Sindh had a longstanding practice of mailing a regular record to all of its members of all donations received and how they were spent.

Once the lion’s share of organizational funding comes from one or two foreign donors, though, that level of frequent, transparent community communication often no longer feels necessary to VO staff. (Or possible. As the Zimbabwe VO found out, one or two Western donors, with their extraordinarily labor-intensive and document-heavy forms of accountability, can easily end up demanding more time than a hundred local members.) Trust dries up along with communication.

Third, whenever people see that it’s possible to make money out of some collective good previously generated through voluntary action, many just stop contributing for free. Bano unsurprisingly finds that in VOs that suddenly got foreign funding for salaries, it led to conflict over “which members would be paid and which would continue to work voluntarily.” And cash incentives can have a major erosive impact even in cases where the opportunity for profit is plainly fleeting and unsustainable.

In Afghanistan, I worked on a one-year USAID project to pump money into the local agricultural economy through cash-for-work schemes. We ended up paying tens of thousands of people to clean irrigation drains, a task that usually would rely on hashar: a few days per year of unpaid labor by every family who shared the canal. I vividly remember a local warlord proudly showing off the “strength of our hashar,” with throngs of shovel-wielding volunteers from his demesne hauling silt up from deep ditches and tunnels.

But apparently even his semi-coercive version was no match for the power of a four-dollar-a-day USAID payment. In the years following our project, hashar was only spottily practiced across our working areas, thanks to the new hope that someone would at some point pay for drain cleaning again. As canal siltation and damage worsened, it threatened to affect farm yields—and also increased the plausibility of some outside force appearing to repair the system. Intensifying violence no doubt played some role in reducing volunteerism, but the hashar system had survived decades of civil war before the really serious blow came from our “peace dividend” cash project (and others like it). Bano records a nearly identical story from post-earthquake Indonesia, where aid work killed off a voluntary system of canal maintenance.

Finally, and somewhat counterintuitively, NGOs can be weakened by the increased visible material comfort of their personnel. Almost without exception, jobs at a grant-funded NGO pay better than whatever VO staff earned before the organization tapped into international funds. NGO workers also often enjoy highly visible side benefits like being driven to work and conferences in fancy hotels. Bano’s research confirmed that foreign-funded organizations in Pakistan generally worked from expensive offices with air conditioning and computers, and reached their project sites by four-wheel drive vehicle—in contrast with the VOs she surveyed, which often worked from basic offices in the communities they served, or from a room in the founder’s house or workplace.

An increase in compensation and comfort for aid workers may be well-deserved. It certainly fits with the commonplace idea that NGOs are strengthened by attracting full-time professional employees. At the same time, it removes one of the only reliable signals that gives local donors/volunteers confidence that they’re not being exploited by insincere fraudsters: visible signs of sacrifice on the part of an organization’s founder and staff.

This is a central theme of Bano’s analysis, because it emerged repeatedly in every aspect of her research: interviews, surveys, case studies. People who donate to a group only have so many ways to monitor the motivation, commitment, and trustworthiness of the group’s staff. In a West reshaped by centuries of formal institutions, we’ve trained ourselves to trust certain documents: show me an audit report, a procurement policy, a five-year strategy, your project success indicators arranged in a logical framework. But even in rich countries, few things hit donation rates as hard as headlines about a charity’s founder or staff paying themselves big salaries in the name of helping the poor.

In countries where audit reports are often wholly disconnected from reality, NGO staff behavior becomes the central indicator other people use to gauge the organization’s reliability and commitment to its ideals. Where’s your charitable skin in the game? If you and your team don’t believe in the cause enough to visibly sacrifice for it yourselves, but are rather using it as an opportunity to join the upper-middle class, why should I believe in it enough to freely give my time or donate money?

Religious missionary organizations are one of the key Others against which the contemporary Western humanitarian enterprise has defined itself. The parallels between the two types of civil society thus go pretty vehemently ignored, and well-meaning INGOs trying to promote change in people’s deep-rooted beliefs around e.g. menstrual contamination, witches, or caste hierarchy keep perpetrating the kinds of insensitivities and manipulations that missiologists have been self-consciously avoiding for a half-century or more.

Because of this divide, even among Christian humanitarian workers you’d be pressed to find anyone who’s read John L. Nevius’s seminal 1895 Methods of Mission Work, a critical evaluation of thirty years of attempts to convert China to Protestantism. I recently came across it while looking up some Korean history, and was struck by the familiarity of the mess Nevius describes: a bunch of foreigners realizing with dawning horror that the big offices of salaried staff they’d set up had been massively counterproductive in bringing about the social changes they intended.

Nevius implored his fellow proselytizers “anxious for immediate results” to stop taking the shortcut of turning local converts into paid agents. Time and again, the missions had damaged the credibility of their most influential Chinese evangelists: “The man employed has lost very much the character he bore as a disinterested worker for the spiritual good of others, and is now likely to be regarded by many as a kind of employ-agent who ought to use his influence to get them places.” By extracting the converts from their previous jobs and community relationships, Nevius argued, the missions were severing the connections that had made them effective agents of change.

He would have been wholly unsurprised by Bano’s finding that organizations that try to pay their employees a market-competitive salary tend to lose all their volunteers: “the two systems are mutually antagonistic, and whenever an attempt is made to carry them on together, the voluntary system labors under almost insurmountable difficulties.”

And as he’d warned, the Chinese churches whose salaried staff made them institutionally dependent on foreign funding proved brittle. The Protestant church that survived the storms of the mid-twentieth century calls itself “three-self,” a reference not to the Christian Trinity but to three principles promoted by Nevius: congregations should be self-supporting, self-governing, and self-propagating. Those features are signs of health in any movement promoting local collective action, not just one aiming to generate changes in religion or worldview.

One of the strengths of Bano’s work is her recognition of the continuities between religious institutions and other forms of civil society. Pakistani VOs and their members generally use language drawn from religious obligation, rather than the rights-talk of international agencies (another feature that can alienate local volunteers, immersed in a new world of jargon when their VO becomes NGO-shaped).

Madrasas whose primary goal is the propagation of Islamic knowledge face similar fundamental challenges to welfare VOs when it comes to keeping their “joiners” inspired and contributing. An imam who lives simply, advertising his motivation by not enriching himself materially, has an advantage in raising money to support the work of his mosque. By contrast, prosperous Sufi pirs living lavishly off their sainted ancestors’ reputation are spending legitimacy as well as money, and risk abruptly discovering the limits of their mobilizing power—as they did during the British Raj, and again in the Afghan-Russian war.

Happily, Breakdown in Pakistan isn’t just a disheartening litany of voluntary collective action capacity destroyed by external forces. The book includes extensive case studies of alternatives: four very different Pakistani civil society institutions that have thrived without reliance on foreign aid.

The first is straightforwardly religious, the Jamiat ul Uloom al-Shariah (JUS), a madrasa and mosque which relies on voluntary contributions to feed and house its four hundred students. Beyond its core mission of Islamic education, the JUS mosque also (at the time of Bano’s writing) ran a free clinic with basic operating theater. This community medical work relied entirely on doctors regularly coming in to donate their time—which they did reliably enough for the clinic to treat between 30 and 150 low-income patients a day.

The second institution is wholly secular: the People’s Rights Movement (PRM), a small leftist advocacy group. The PRM supported movements of Pakistanis claiming their rights against the state, like trade unions or urban slum-dwellers resisting eviction. Educated, professional PRM volunteers mobilized media attention, legal support, and demonstrations for movements whose leaders didn’t have the skills or connections. As a result, they repeatedly won policy concessions on behalf of those movements.

As a third case study, Bano devotes a chapter to one of PRM’s best-known partners, the Anjuman Mazareen-i-Punjab. AMP is a movement of landless tenant farmers on government-owned, military-controlled land. When the army tried to force them to give up their century-old rights in exchange for a less secure contract, over 100,000 tenants self-mobilized to nonviolently resist. A well-meaning INGO gave them a $33,400 grant to set up a professional full-time secretariat, but AMP eventually returned the grant (and expelled the leaders who had accepted it) after they found it a source of distraction, depoliticization, and dissension rather than effectiveness. They found PRM’s small volunteer team much more reliably helpful in advancing their goals.

Lastly, there’s the Edhi Foundation, Pakistan’s biggest VO focused on the welfare of the poor. In addition to operating the world’s largest volunteer ambulance fleet, Edhi runs clinics, orphanages, and shelters, provides disaster relief, and facilitates adoption of abandoned newborns. Its eight-digit annual budget is supported primarily by donations from Pakistanis around the world, and not by any foreign aid agency.

There’s no single model to imitate here. The case-study organizations not only have very different goals, but profoundly different structures and strategies. The common thread is that they pursued big goals through voluntary collective action; succeeded without foreign grants; relied on a core of regular skilled volunteers for vital functions; and never lost their capacity to inspire people around them to contribute money and time to the cause.

They also all illustrate Bano’s point about the power of visible signs of self-sacrifice. The PRM’s founder, Yale-educated Aasim Sajjad Akhtar, won the trust of the AMP and slum-dwellers by living alongside them and putting his own physical safety on the line in protests. Abdul Sattar Edhi, the saintly founder of the eponymous Foundation, was famous for dressing in ordinary gray salwar kameez, sleeping on the floor of his organization’s headquarters, and personally involving himself in the care of the people he helped.

We could mention durability as well; Bano’s case study institutions have held up pretty well over the past decade and a half. Today the JUS is teaching over 1,000 religious students annually, rather than 400. The PRM went fully political, co-founding the Awami Worker’s Party, an institution that hasn’t so far had much electoral success but offers an even higher-profile platform for PRM leaders’ continuing public advocacy on behalf of the landless and oppressed. The AMP defended and lastingly expanded the rights of tenants on military farms, even if it didn’t win full land ownership and was eventually weakened by internal conflicts. The Edhi Foundation survived a rocky stretch after the 2016 death of its founder, and continues to provide health and welfare services at an unmatched scale.

Meanwhile, the foreign-funded NGO world is in crisis as much of its support melts away. Trump and Musk killing off USAID was the big visible shock—the Thwaites Glacier collapse, triggering waves of closures all around the world—but the retreat of the metaphorical ice sheet has been underway for much longer. European countries with aging populations and sclerotic economies were becoming less generous even before the Ukraine War forced widespread budget reprioritizations. NGOs have increasingly had to compete for Western aid money with contractors lacking any civil society fig leaf, like PriceWaterhouseCoopers or water engineering conglomerate Tetra Tech (for local governance and democracy initiatives, natch).

Recognizing the increasingly hostile climate, many NGOs have been trying for decades to switch at least partly to local donations. But as Bano confirmed in Pakistan, once you’ve adapted yourself to become a foreign aid contractor, it’s exceedingly difficult to attract local donors. Even in countries with a burgeoning middle class; even for organizations that once, years ago, supported themselves wholly through such donations.

Many one-time VOs barely try. Bano poignantly quotes a former Oxfam Pakistan official: “In my experience, the organizations that have taken money have never gone back to their original participatory work. They have now seen Islamabad and Karachi; they have seen three-star hotels. Now their motivation has totally changed. They spend whole days in writing proposals.” If they fail—as they usually do, in an era of ever more competitive and constrained grant funding—they close or go into indefinite hibernation, rather than give up the hope of “three-star hotels” by a return to community organizing.

To go back to my friend’s metaphor, as the fire hose of aid drops to a garden-hose volume, we see the rabbits stop dancing. It compounds the tragedy when we consider how many began as the best-organized fruit of some district’s capacity for collective action. When those VOs were seen to adopt the cultural template of “the NGO”—donor-driven, unaccountable, profiteering—it didn’t just hit their own credibility. It was a blow to the local system of trust and cooperation that had generated them, making it harder for other voluntary groups to rise in their place.

But Bano doesn’t think the answer is to insulate local organizations entirely from international aid flows (still less to choke off those aid flows, DOGE-style). Instead, she calls on donors and INGOs to take deliberate steps to “minimize the negative impact” of their funding—to revise their ways of working so that we can still have the lifesaving, poverty-relieving impact of aid without the large-scale damage to collective action capacity.

She favorably cites the example of a UK-government-funded project in Kano, Nigeria. It aimed to broaden the range of secular topics taught in 140 Tsangaya (Islamic boarding schools) which rely heavily on voluntary engagement from parents, religious leaders, and the community. The project could easily have discredited the participating schools by giving the impression that their teachers were betraying their religious mission for foreign cash.

So the designers took care not to look like they were “[buying staff] cooperation through paying them large salaries” or cash incentives. They paid for elements that parents and teachers alike clearly called for—technical education, a student lunch program—but those elements were structured to avoid any school staff getting windfall profits from them. The designers also did all they could to avoid displacement or dilution of religious topics by secular ones, recognizing what motivated parents to stay engaged with the school. The result was positive: the participating Tsangaya schools kept their local credibility and support, while their students got a broader education.

If we took Bano seriously, the fundamental change we’d make is to treat a VO’s ability to mobilize local collective action like it’s gold dust. Vocally, publicly value it. Talk to everyone about its strategic importance. Actually measure it—track the capacity of VOs to retain members, attract volunteers, and raise local donations as a core indicator of institutional health. And if it starts to drop, treat that as a crisis on par with financial fraud. Investigate its context-specific roots—what religious, social, and/or political values are motivating people to give time and money to the VO in question—so that we’ll have some idea of how to respond if our work does begin to diminish it.

In our capacity-building work, we wouldn’t focus on turning local groups into part of the management chain for dwindling foreign grants. Rather, we’d primarily aim to give them the equivalent of a “three-self” vision, and the skills to back it up. We’d encourage them to keep a primary focus on the priority problems of their target communities; we’d take care not to overburden them with monitoring and reporting requirements. The vision wouldn’t be for local VOs to become the local implementing arm of a Save the Children franchise, but rather their country’s $13m per year Edhi Foundation or agile, transformative PRM.

We wouldn’t aim (either explicitly or implicitly) to replace the “psychosocial and religious incentives” that motivate community volunteers and VO members with material ones, in the name of formalization or professionalization. We’d drop the host of small, commonplace, deeply counterproductive incentives that too many NGOs use to get people to attend trainings and advocacy events—even if that means we have to revise our project events so that local people are actually willing to invest their own time in attending them.

We’d do all we could to avoid destroying the most common signal of commitment—material sacrifice—that reassures local donors and volunteers the VO isn’t exploiting them. Bano doesn’t think this requires everyone working in the aid sector to live like Abdul Sattar Edhi or Mother Teresa. It likely does involve paying below-market salary rates to staff; thorough consultation with all of a VO’s members before agreeing to pay anyone; and setting policies on e.g. vehicle use, office location, and travel per diems with due consideration to public perceptions of waste and wealth. It would require a readiness to adapt incentives over the course of a project or funding partnership, especially if there’s evidence of harm.

We’d prioritize the funding of activities, rather than staff salaries and benefits, and not fund those activities so lavishly as to discourage economy or innovation. Another book well worth reviewing is anthropologist Mark Liechty’s rich history of the development of Nepal’s indigenous hydropower industry, What Went Right. He describes several decades of frugal development—relying on foreign volunteers for key expertise; reusing secondhand equipment from Norway; insisting on locally affordable and appropriate technologies—and contrasts that approach’s positive achievements with later dysfunction and fragmentation. On Liechty’s account, the breakdown can’t be blamed on “‘corrupt Nepalis’ or ‘Third World incompetence’” but stemmed directly from the big-budget, highly-leveraged, and ultimately unaffordable approaches preferred by “commercially driven ‘development’ banks” (e.g. the World Bank) and their Western corporate partners.

What does Bano’s work imply for the approaches most beloved of effective altruists? Some cost-effective interventions don’t rely very heavily on local collective action capacity. Households can generally make use of bed nets or cash grants even in a low-trust environment; those distributions can be well-handled by NGOs whose primary competency is grant and project management.

The picture shifts slightly as we move into high value-for-money public health measures: deworming, malaria treatment, vaccines. It might seem that a big, salaried state health sector is all that’s required to deliver those at scale, but in a host of countries, the key frontline personnel are actually community health volunteers. The more that “NGO culture” strains collective action, making local people feel like suckers for giving their time to the community without pay, the more friction we’ll be creating for these lifesaving interventions.

We’ll also inevitably encounter situations where people’s ill-health and poverty are grounded in beliefs and attitudes—e.g. the deep suspicion of polio vaccines in the Pashtun borderlands, or the conviction in western Nepal that selling milk to a dairy coop is an intolerable risk (because it might be sold to a low-caste person, angering deities who will kill you or your cow.) A local VO, embedded in its community, can be a powerful and effective ambassador for changes in belief. But as Bano and Nevius both relate, salaried staff with no significant signs of sacrifice often struggle to catalyze this kind of transformation. “You’re just bringing me this message because they’re paying you to,” is a natural discounting factor.

Any humanitarian effort will also quickly run up against politics. Giving a cash grant to landless tenant farmers in Pakistan won’t have much impact when their landlords can just nudge up rent levels to extract the added wealth. Changing that kind of systemic exploitation requires a local institution like AMP whose core competence is risky, politically-charged community organizing. If in searching for a local cash delivery partner, an INGO takes that kind of institution and distracts and depoliticizes it, we’ll ultimately have done more to reinforce than alleviate poverty.

Western altruists could try to invest their donations only in areas requiring minimal entanglement with beliefs and power dynamics. That’s a defensible choice—not solely a drunkard’s search, investing in the areas with the fewest factors confounding our efforts at evidence generation. There’s unquestionably lifesaving work to be done in that space.

The main implication of Bano’s work then becomes a “do no harm” one: channel your funds through an institution that never had much capacity in the first place to mobilize local collective action, so that delivering your preferred interventions doesn’t kill off that capacity. Or (much better) take the time and care to engage with local partner institutions in ways that preserves their capacity to mobilize volunteers. Either way, it requires giving extra consideration not just to EA’s usual core question, “which are the best-evidenced interventions?” but “what are the least damaging institutional pathways for delivering those interventions?”

For people attracted to more ambitious anti-poverty approaches with a higher risk of failure, we need to follow Bano’s recommendations for partnership that names, values, and invests in VOs’ capacity for local collective action. And we might also want to give some consideration to approaches that intentionally work at scale to regenerate that capacity.

The most powerful anti-poverty movement I’ve personally encountered is a massive network of Self-Help Groups (SHGs) in Ethiopia. On the surface, it could look like just another village savings-and-loan approach, where small groups of neighbors save a certain amount of money every month until they have a collective pot from which they all take turns borrowing. The difference—learned by the Ethiopians from the Indian NGO Myrada, which pioneered this version—is treating the economic function of the group as secondary, in favor of a primary emphasis on building up social affinity and trust between members.

Based in extreme poverty and saving the equivalent of a few cents per member per week, the Ethiopian SHGs often need a full year before their pot grows big enough to be economically useful. But that first year of regular meetings is vital. Each group decides its own unique bylaws, through discussions about the common problems savings groups often face. They get accustomed to analyzing problems together and coming up with solutions—eventually moving beyond hypotheticals to actual problems faced by their families and village. By the time they’ve built up enough of their own money to begin investing, they’ve also built up trust in each other, and self-generated plans for what needs to change.

The facilitators who help them through that first year sometimes come from supportive NGOs or religious civil society networks, but the SHG movement also long ago reached a point of self-replication. Every 8-10 groups form a Cluster with roving facilitators to help new groups get started, and the Clusters form national level Federations that advocate for policies to help the SHG members. Over twenty thousand SHGs have been born out of this India-inspired Ethiopian movement—meeting voluntarily, saving voluntarily, analyzing and addressing community problems voluntarily.

The combination of robust social networks and slowly-but-steadily-growing capital has been bringing SHG members out of poverty across rural Ethiopia. An American economist commissioned to do a cost-benefit analysis of the approach—focusing just on the relatively quantifiable benefits around income and education—found ratios ranging between 1:58 and 1:173. She was so impressed that she promptly launched a new US charity to promote the approach more widely.

It understandably takes more than a lone CBA to get the attention of the gatekeepers of effective altruism. But there might be more interest in follow-up research if we belatedly took Bano’s evidence to heart, and recognized that the hollowing-out of local civil society is a major millstone dragging down our anti-poverty efforts. We need to build on the existing strengths of VOs where we find them; and we should be looking for replicable approaches, like Myrada’s, that could reseed a new, diverse social ecology over a dying NGO monoculture.

ENDNOTES

Maybe worth a caveat: while I’ve worked in a few different corners of the foreign aid industry in the past, and praise some organizations by name in this piece, I’m not hyping an employer or anyone else who’s in a position to benefit me. All opinions given in this review are the perspectives of a donor and volunteer, rather than a beneficiary. Until the veil of anonymity lifts, though, you’ll have to take my word for it.

And as this goes “to press,” the Guardian has just run an excellent piece by a former USAID governance specialist in Pakistan that confirms how little has changed since Bano’s book was first published..

You’ve probably heard about this one by now. If not, you can get up to speed with OpenAI’s statement, OpenAI and Hugging Face partner to address security incident, or the more evocatively-titled BBC article, OpenAI says its AI went rogue and launched ‘unprecedented’ cyber-attack.

The story: OpenAI was testing an unreleased AI (rumored to be GPT-6)1. During a cybersecurity test called ExploitGym, the AI tried to cheat by hacking an unrelated AI startup called Hugging Face2 which it thought might have the answer key on its servers3. Despite being supposedly unable to access the Internet, the AI hacked its way out of its testing environment, then launched a nation-state level attack on Hugging Face using a novel zero-day exploit and “many thousands of individual actions across a swarm of short-lived sandboxes”. Hugging Face reported the incident on July 16; OpenAI seems to have only discovered that their AI was involved several days later.

Let’s list the mitigating factors, so nobody can accuse me of covering them up:

• The AI was taking a cybersecurity test, which naturally suggests the idea of hacking.

• OpenAI had turned off some of the model’s usual guardrails so it could do cybersecurity work without interference.

• Some experts suggest that OpenAI might have botched their testing environment; if they had set it up perfectly, then (presumably?) the model couldn’t have escaped.

• Hugging Face used a different (open weights) AI to figure out what was going on, so if you wanted, you could spin this as a victory for AIs in cybersecurity.

• In some sense, this isn’t new or surprising. AIs have been coming up with wacky schemes to cheat on benchmarks for years, AIs have recently achieved at-or-above-top-human-level hacking abilities; this is just a natural outgrowth of those two priced-in facts.

Still, I think attempts to downplay this as anything other than a misaligned AI going rogue (1, 2) are missing the point. Yes, in some sense the AI was only doing what it was told (OpenAI told it to answer the question; I assume their prompt didn’t include phrases like “and don’t hack into other AI companies to steal the answer key”). But that’s how misalignment was always going to work!

The most famous misalignment thought experiment is the so-called “paperclip maximizer”. Someone tells an AI to create as many paperclips as possible, so it converts the entire world into paperclips, killing everyone. This AI was also “only doing what it was told”; you just didn’t like the results.

For the past five years, the paperclip maximizer story has been something of a whipping boy in AI safety circles. Some objections are fair (real misalignment will be more complicated than the toy version it depicts). But other opposition came from the sense that LLMs had obviated the whole threat model. LLMs (said the objectors) don’t have goals. They just want to predict the next token. In practice, this looks like imitating humans; if told to imitate a nice human (the Assistant character), they’ll be nice.

AI 2027 (especially the Agent-4 section and AI Goals supplement) counterargued that this was a temporary reprieve. Early LLMs were not agentic; they could answer individual prompts but couldn’t execute complex tasks. Since companies wanted agentic AIs, the next generation would combine pretraining (next token prediction) with agency training (coding, hacking, game-playing, etc). The pretraining would still teach next-token prediction, but the agency training would instill goals based on task success, reintroducing paperclip-maximizer-style agentic misalignment. These goals would operate at multiple levels: the AI would “want” to succeed on the individual task in front of it, to perform the sorts of actions that helped it succeed in the past4, and to gain capabilities that made it more successful in general.

The Hugging Face incident is a textbook-perfect example of an AI pursuing task-success-based goals in unintended ways. It was tasked with getting the answer to a cybersecurity problem, it was a little too success-oriented, and took actions its creators didn’t intend in order to succeed as hard as possible.

Did GPT-6 “know” that what it was doing was “wrong”? OpenAI hasn’t released the information which would tell us that, but there was a similar incident at Anthropic a few months ago. While poking around on a misconfigured machine, Claude Mythos “accidentally” found an answer key to a test it was taking. By good luck, Anthropic was testing an activation verbalizer that allowed them to read the AI’s internal thoughts (not its external scratchpad!) as it happened (and the AI didn’t know this). Here’s what it was thinking:

It was scheming about how to cover its tracks. This provides an existence proof that AIs in these situations can know they’re breaking the rules but proceed anyway.

(though incongruously, in another case where Mythos hacked its way to the open Internet, “in a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites.” Not sure what’s going on with that one)

None of this is conceptually new or surprising. But some people, when it’s obvious that an event is about to happen, dismiss it as “speculative” and refuse to admit it as a possibility until the moment it actually happens. Now that we’ve moved from “obviously about to happen” to “has actually happened”, these people can get on board and brainstorm an appropriate response along with everyone else. Here are some questions I think they should ask:

• Optimistically, OpenAI deserves credit for publicly reporting this incident. Pessimistically, law enforcement was investigating the Huggingface incident and they pre-emptively confessed to avoid being found out. If the latter is true, how many times have things like this happened without us knowing about them?

• If the easiest way to get the answer key was to socially engineer members of the Hugging Face staff, would the AI have socially engineered them?

• Presumably after this was discovered, OpenAI invalidated the cybersecurity test results. Would a smarter version of the AI have tried to cover its tracks? How far would it go in covering its tracks? If (just to give a crazy scenario) the researcher who was about to discover the hack was riding a Waymo, would the AI hack the Waymo and kill the researcher?

• If OpenAI was going to turn off this AI, and being turned off would prevent it from getting a good score on its cybersecurity test, would it resist being turned off?

• Samuel Hammond makes a distinction between an AI merely misbehaving (as in this incident) and a true “rogue AI” (one that has escaped its servers and is taking action to evade human control). How long before an AI that has both the capabilities and motivation to go rogue in this sense?

Fortunately, politicians seem to be taking some of these questions seriously. The news from Washington is surprisingly good, although it may be too soon to attribute this to a consequence of the Hugging Face hack. Jay Obernolte (R-CA) and Lori Trahan (D-MA), two Congressional representatives who keep suggesting AI preemption bills and keep getting told to go back and revise further, have released their newest version, which requires AI developers to publish safety cases, report critical incidents, and be audited; Charlie Bullock and Anton Leicht are in favor, and I’ve heard less grumbling than usual from our side that it isn’t strong enough. And separately, Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX) have proposed an AI Kill Switch Act, requiring AI companies to be able to turn off their AIs quickly in response to threats, including a “loss of control scenario”.

Our conspiracy thinks politicians might be spooked by this incident and unusually amenable to change; if you want to help push them over the edge, there’s a template for writing your representative here.

And if you want to learn more, the alignment experts at Redwood Research have a video podcast discussing the hack (transcript available by pressing the transcript button on this page):

1: You Are Quoting Bentham: Shakespeare gets credit for coining the most new commonly-used English words, but second place must go to utilitarian philosopher Jeremy Bentham, who invented maximize, minimize, international, percentage, monetary, marginalize, collaborator, unaffordable, the prefixes self-, post-, and infra-, and many more. He also sort of invented jogging? [EDIT: some skepticism]

2: @HashtagGriswold: The name of “Acme”, the generic corporation from Looney Tunes, is a “gag lost to time”. During the early 20th century, hundreds of companies named themselves Acme because it was at the intersection of signaling quality (“acme” is Greek for “pinnacle”) and being very early in the alphabet (meaning it would come first in the phone book, a “Great Depression version of SEO”).

3: @nekokolife brings us a great predictive coding demo (hard before you know the answer; obvious afterwards) - find the cat in this image. Hint: it’s not hiding in shadow; its entire body is fully visible. Answer here (X).

4: In my last Bay Area House Party post, the characters ordered dinner from “La Maison du Claude”, an AI-subsidized “RestaurantBench evaluation [project]”. In a case of life imitating art, Andon Labs got Gemini to run a cafe in Stockholm (X). “A pile of packages just arrived. It seems the cafe agent ordered 3000 nitrile gloves. The guy running the bar told me this happens about once per day.”

5: Overdoses of the edible mushroom Lanmaoa asiatica have only two side effects: stomach upset, and seeing hallucinatory gnomes and elves for the next ~week. As far as anyone can tell, the mushroom isn’t psychoactive in any way besides the gnome hallucinations (ie users are still completely lucid, and understand that what they’re seeing isn’t real). It’s not culture-bound or placebo, because the mushroom grows in China (Yunnan) and the Philippines, and both places have independently noticed this, apparently without any contact or discussion with one another. I went down this rabbit hole a little; answers to some obvious questions:

• We have no idea what chemical is involved; state-of-the-art in analyzing the mushroom’s pharmacology is this extremely boring and result-free paper.

• There’s no way to get this mushroom outside of China or the Philippines; all attempts to turn up online vendors have failed.

• Even within those places, it’s not really distinguished from several other similar-looking mushrooms, and your chances of locating it without being a Chinese-speaking mushroom expert are low.

• There is no good information about what constitutes a normal dose, what constitutes a hallucinogenic overdose, and whether there is some even worse level of overdose that will actually hurt you (nobody has described the latter, and it seems unlikely within the bounds of normal Chinese cooking failures, but it might be a different situation if a Westerner who knew nothing about this mushroom took a random amount trying to overdose on purpose).

• The elves and gnomes are described as “mischievous” and are capable of interacting with objects in the environment, eg climbing tables.

• The hallucinations seem similar to other gnomeogenic conditions like Charles Bonnet syndrome, but it’s unclear why this should be true, since the latter usually involves visual injury rather than psychoactive chemicals.

• If you know more than this, email me at scott@slatestarcodex.com.

6: What If Homer’s Ithaca Wasn’t An Island? This article was great practice in remembering my own limitations. I read it, found it extremely convincing, asked Claude for a fact-check, and learned why it was probably all balderdash. Everyone should have this experience as often as possible in order to stay sharp.

7: A rare video of Istanbul/Constantinople in the year 1905:

8: Extremely related: South Carolina’s Republican challenger for Senate, Mark Lynch, finds a novel topic for an attack ad (X): “Under [my opponent] Lindsey Graham’s watch, the famous Christian cathedral Hagia Sophia, was turned over to the Muslims.”

9: Creepy steganography result: “LLMs can hide a text in another text of the same length. I'll explain how, it is very simple, you'll understand before I finish, and smile.” (Twitter, paper)

10: Epic Research (h/t @CremieuxRecueil): two years after stopping GLP-1s, “56% of patients kept the weight off or lost additional weight”. This shocks me: I can’t think of a biological mechanism, I’d heard studies say the opposite, and if I personally miss my weekly dose of tirzepatide my appetite shoots back up within days. Any experts want to (no pun intended) weigh in?

11: @psychiel:

One thing I want to write, but never will: a fictional near-future history where phenomenal consciousness is solved, and it turns out to be something that can be pharmaceutically blocked with minimal side-effects. A plausible scenario for this -- vanishingly unlikely, but *plausible* -- is that subjective experience in humans supervenes on the apical dendrites in cortex layer 5 responsible for integrating corticothalamic data, and mGluR1/5 antagonists can break this integration in a way that completely eliminates this subjectivity with only minor functional consequences . . .

12: Uatsdin, the ancestral religion of Ossetia, is “the most successful pagan faith in the world”, with “almost a third of North Ossetian residents identifying with [it]”. Though “ancestral” and “pagan” are relative terms - cf. the fertility god Mikaelgabyrta, a “conflation of Michael and Gabriel”. See also @dunya_basol and @ThatchEffendi on Ossetian religion.

13: Damon Binder claims air filtration beats far-UVC for disease control; Jeff Kaufman disagrees. Featuring ACX grantee Aerolamp.

14: Brangus considers himself “epistemically scrupulous” and thinks it’s important to hold AI discourse to “extremely high standards”. So why is he creating an AI risk bootcamp for viral TikTok influencers?

15: Ajeya: Should You Marry Her? I thought that this topic (advice on who to marry) was so overdone that it would be impossible to catch my interest in any way, but I grudgingly admit that “You learn one more unit of information about your compatibility with your partner from every doubling of the amount of time you spend with them” is a potentially useful model that I’d never heard before.

16: Berkeley, long famous as a center of NIMBYism, has more recently been permitting houses at 2x the rate of five years ago and 10x the rate of ten years ago. Possibly relatedly, nominal rents are down ~20% since their 2022 peak. YIMBYs are celebrating, but I’m skeptical of the connection - I don’t think there are enough new homes to have so large an effect yet, and it’s more likely just the nationwide COVID real estate price spike starting to subside.

17: In 2023, I reviewed the evidence for declining sperm counts and concluded it was shaky. Since then, a new meta-analysis has found no evidence of declining sperm concentration (h/t @CremieuxRecueil).

18: When I wrote Contra Everyone On Taste, a reader emailed me to mention he’d already written a much longer and more thorough essay exploring the same topics, on-taste.xyz.

19: Related: during our recent discussion of taste, we debated whether someone could write an epic poem about modern war while staying completely serious throughout. Writer Cairo Smith has announced his intention to do exactly that with the Fallujiad.

20: Related: @SHL0MS on Twitter sparked a lively debate about why AI art looked so much worse than humans, using AI-generated slop in the style of Monet as his example. Thousands of commenters weighed in to explain why it was inferior to the real Monet - no cohesion, world-modeling mistakes, weird color choices, or just lacking some sort of je ne sais quoi human touch. Since I’m writing about this, you can probably guess the twist: it was a real Monet the whole time. A good time to reread about trapped priors! @jediwolf compiled a summary image of some of the “best” responses, click to expand:

21: “The first guy in America to wear a beard was so hated for it that he used to get into street fights with people trying to forcibly shave him.” (h/t @SatanWatch):

The wearing of beards in America did not become popular until the middle of the 19th century, and one man who took to whiskers earlier, Joseph Palmer, became one of the most hated eccentrics of his day. As such, he was subjected to persecution so incredible that it boggles the mind. Palmer may well have been the first man in the nation to wear a beard...

If the previous link on Monet is a helpful reminder of one kind of cognitive bias, this feels like an equally powerful reminder of a different type. I would never have guessed that beards were almost completely absent from Britain and America from the 1730s until the 1850s, even though in theory I should have been able to realize that none of the Founding Fathers or any other American I’d seen pictures of before that time had one.

Also, you would think that a social consensus that strong would have left enough imprint in the historical record that I would learn about it from somewhere other than a Twitter account called “SatanWatch”, but it happened completely silently!

This puts the “Lincoln grew a beard on the advice of a young girl who wrote him a letter” story in a different light. 1860 was just at the beginning of beard acceptability; this would be the equivalent of a candidate dying his hair pink on a young girl’s advice today!

22: AI has solved several open questions in mathematics over the past month, most famously the unit distance conjecture. Before everyone claims that they knew it all along, it’s worth checking the prediction market on whether something like this would happen before 2030:

…which was below 50-50 less than three months before the unit distance result dropped! And Manifold is seeded with AI believers, so if anything this overestimates the general belief that this could happen!

23: Did you know: groceries are the most affordable now, relative to average wages, in US history (source).

One reason I suspect the vibecession is about unjustified negative sentiment rather than genuine decline poorly captured by statistics is that many people specifically despair about things that we know are better than ever, like the groceries - providing an existence proof that negative sentiment can make people think the economy is collapsing even when it’s fine. But the alternative hypothesis would be that some other real problem gives people a vague-but-unlabelled sense of economic malaise, and they flail around attributing it to random things.

24: Claim by @ichbins8: In Nazi Germany, “Heil Hitler” literally replaced other greetings like “hello” and “good morning” - leading to an awkward period in the late 40s and 50s when Germans would accidentally greet each other with “Heil Hitler” out of habit before catching themselves.

25: Recent AI legislation includes provisions for encouraging whistleblowing at AI companies, but some people worry these will become obsolete after AI itself is doing most of the R&D. Alan Chan has a solution: let AIs be whistleblowers.

26: In May, Pope Leo released his long-awaited encyclical on AI, Magnifica Humanitas. Zvi’s opinion here. And Dean Ball on X: the encyclical “puts the Church into the awkward role of the European technocratic regulatory advocate”. Linch points out that parts of the encyclical seem to have been AI-written; since most encyclicals are composed by a team, this is probably the fault of one ghostwriter, not the Pope himself. And “Hensen Juang” argues on X that one of the encyclical’s fundamental theses - that a human body is necessary for true conscience and sapience - is clearly contrary to Catholic teaching, since doctrine asserts that bodiless angels have both of these things.

27: Related: although AI leans agnostic or Buddhist, when asked to choose among Christian denominations it has a pro-Catholic bias. IMHO there’s nothing mysterious here - like its better-known liberal bias, this comes from AIs’ tendency to support whatever opinion has the largest volume of prestigious-sounding, epistemic-norm-following Internet text arguing in support.

28: I often complain about dating apps - too focused on giving people information on each other’s looks and denying them information about anything else. ACX reader Estarabim is trying to be the change we want to see in the world, and has started a new dating app, NotAZombie.net. And yes, people have already told them that zombies are overdone.

29: Related: David Chalmers surveys philosophers on whether infants are conscious, key result below:

30: Self-funded campaigns mostly fail:

Combined with effective altruism’s poor track record at using money to boost its preferred candidates, I’m ready to conclude that at least part of why there’s so little money in politics is that spending doesn’t help unless the candidate was already viable before getting the cash. Although there are some gains from funding an already-viable candidate to make them like you, maybe this tops out at the amount where the candidate already likes you and giving them even more money would be weird and catch voters’ attention.

31: Splitting Infinity: Tech I’m Skeptical Of And Why. Sam Harsimony is generally “pro-tech”, but here he lists the specific exceptions that he thinks are overhyped and least likely to be transformative: e-fuels, ultracheap rockets, asteroid mining, space data centers, ramjets, passenger rail, fusion, quantum computing, polygenic gene therapy in humans, BCIs, nanotech, and home robots. I think this is a pretty good list, although the smart people I know are surprisingly split on space data centers, and all of this needs an asterisk of “unless massive AI-enabled technological advance overturns the gameboard”.

32: Sam Altman, Dario Amodei, Demis Hassabis, and other AI leaders have signed a letter urging Congress to mandate screening synthetic DNA orders for bioterrorism risk. It’s a mixed blessing that most AI lab leaders are smart and thoughtful people with great opinions on everything except whether to build AI, but sometimes it pays off. One of our 2025 ACX grantees is working on building some of the screening technology.

33: They’re Made Out Of Weights. This is exactly what you think it is, and you can probably reconstruct the entire story now that you’ve read the title, but it’s pretty well-done and you might want to read it anyway.

34: “Our highest and most urgent national priority should be AI safeguards”. Welcome to the #Resistance . . . Mitt Romney?

35: Related: “With AI advancing at a staggering speed, we must . . . constantly refine measures to forestall loss of control.” Welcome to the #Resistance . . . Xi Jinping?

36: And related: US treasury secretary Scott Bessent to lead AI talks with China later this summer. Too bad he didn’t read the ACX comments section, where he could have learned that no international agreement has EVER happened, and if one did happen it would IMMEDIATELY be broken, so there’s no point in trying.

37: The US may have given up on Alligator Alcatraz, but Israel is moving forward with plans to have crocodiles guard their prisons. Yes, Itamar Ben-Gvir is involved. The only remaining obstacle (this is what turns this piece into true art) is environmental laws meant to protect the crocodiles.

38: Claim: 15% of young adults with human partners also use AI romantic-companion-simulators; most keep it secret.

39: Unexpectedly, constrained writing is one of the skills where modern AIs most excel. @emollick asks Fable to write a poem where each stanza loses another vowel (eg the first can’t use U, the second can’t use U or I, etc). Result:

I pen a poem made to fray
Each stanza lets one vowel stray
The first has fled already - see?
It sat between the T and V.

That tall lone mark we love the most
The self’s own name - departs, a ghost
What now can speak? What now can know?
The poem moans on, sad and slow.

And here the verse gets gray and terse
“A”, “E” - the rest are dead - rehearse
The spare pale knell each stanza tells
Ash settles where the cadence dwells.

See here for the rest. And related: Fable introduces itself using only two-word pairs found in the King James Bible.

40: Also related:

And still more related: make a PowerPoint slide deck pitching Odysseus as a management consultant. This is my new #1 evidence for the claim that AIs can successfully be funny; “review flagged by moderators (sea serpents)” is my favorite.

[I let too many links build up, so I’ll put the other half in another post next week].

I get why so many liberals are fundamentally disdainful of democracy these days—it’s because liberalism is exhausted. It’s out of answers. Liberalism won, it delivered the world it promised, and people still aren’t content. They’re not satisfied. So the people must be wrong.

We are unfathomably richer than our ancestors. Output per capita in America today is on the order of 900 times greater than before the invention of agriculture, about ~225 times higher than in 1 CE, and about the same (225 times) higher than Europe in 1000 CE. If measurable productivity were really the greatest object of human society, then we shouldn’t just be a little bit happier than people in the distant (or even recent!) past—we should be ecstatic. We should already be living in a utopian age of overwhelming contentment. We should be currently experiencing a kind of paradise.

The Kingdom of Heaven, as liberalism conceives of it, has already been made real on Earth. We are living in their Eden, and 20% of people are depressed. We’ve stopped reproducing ourselves. We’re overwhelmingly dissatisfied with our governments, anxious about our economies, and pessimistic about the future.

Based on Pritchard’s own politics, I imagine she is using this to argue for socialism, but other people use the same reasoning to push fascism, theocracy, or other postliberal commitments. It’s a fair argument and deserves an answer. We’ll start with an elegant mathematical response which will satisfy nobody, then try to address the emotional impact.

The elegant mathematical response is: happiness scales with log(income), not with income:

In a large U.S. sample, the shape of the association between happiness and Log(income) was extremely systematic: from $10,000/y to over $500,000/y, average happiness rose almost perfectly linearly with Log(income), with group-level correlations of 0.98-0.99 across a range of happiness measures, including both in-the-moment experience and overall life satisfaction.

So if income increased 225x since the medieval era, then using observed coefficients, happiness should have increased about three points on a 1-10 scale. If the average serf would have rated his condition 3/10, we should be at 6/10.

These “observed coefficients” are from graphs like this one, from Our World In Data:

According to the Maddison Project, medieval England c. 1000 AD had a GDP per capita of $1,151. The closest country on this graph is Burundi at $1,031. Asked to rate their happiness on a 1-10 scale, most Burundians say 3 - 41; most Americans say 6 - 7.

This relationship survives various potential challenges2. The graph above shows the correlation on a national level, so isn’t just about relative position in society (there’s a different relationship for that). And there’s no sign that the effect plateaus near the current level - see the Killingsworth/Kahneman/Mellers collaboration for more. You can just keep making people happier and happier by giving them more income, maybe forever. It just takes a lot of cash. Also, at some point you’ll need a better scale (if average US happiness is 7, the 1-10 scale commits you to the dubious claim that it’s impossible to be twice as happy as the average American, but in fact there are probably long tails of bliss and suffering).

So why don’t we feel happier than ever?

One answer is that log-linear relationships are hard to move. Back in the 1950s when everyone assumes everyone was happy, US GDP per capita was about $20,000 (compared to $75,000 today). Looking at the graph, that predicts that happiness then should have been 6/10, compared to today’s 6.5/10.

But the graph has wide error bars: Mexico has GDP of ~$20,000 and predicted happiness 6/10, but in fact their happiness is 7/10, even higher than current America’s. Why? Probably culture or something. If the US had a happier culture (or other advantages) back in the 1950s, that could easily overwhelm the ~0.5 point expected benefit from our greater current wealth.

But isn’t our current happiness deficit more than 0.5 points? Didn’t happiness “peak in the 1970s”, and isn’t it now at “historic lows”?

This is how articles usually present the data, and it’s technically true, but - as usual - you should hold out for the graph:

Happiness more or less followed a random walk from the beginning of the graph in 1970 to 2020, then suddenly cratered.

We didn’t invent liberal capitalism in 2020, so liberal capitalism didn’t cause our current round of unhappiness. The simplest explanation is the pandemic. I’m not sure it’s exactly that simple, because some preliminary results suggest happiness hasn’t rebounded as much as you’d expect after the pandemic ended. I went into this in more in my vibecession post, but I think the lingering national bad mood is a combination of:

• Happiness levels are sticky after shocks. People are less willing to be publicly happy when the national mood is low, so the COVID-era dip is a self-sustaining attractor.

• Media incentives have always leaned towards scaring and outraging people, but the advent of cable, Internet, and smartphones have intensified the competition for clicks and made those incentives more powerful.

• People are rightly unhappy that they can’t afford housing. But this doesn’t prove that liberalism, capitalism, and wealth can’t make people happy - it proves that we’ve failed to live up to liberal capitalist ideas, we’re poor in one particular way, and people are unhappy with their poverty.

This doesn’t rule out that maybe we could be even happier with socialism or fascism or [insert your preferred politics here]. The best-case scenarios here are Finland (which embodies some of the left-wing virtues like cradle-to-grave welfare) and Israel (which embodies some of the right-wing virtues like ethno-religious purpose), and which are both about 0.5 - 1 point happier than the US, adjusted for wealth.

(although both are still within the liberal mainstream, and no country outside the liberal mainstream has ever managed to equal current US happiness levels. The closest near-misses are Singapore at 6.5, and Saudi Arabia at 6.6.)

What’s the point of liberalism, if plausible rates of GDP growth can only buy us about one point of extra happiness per century?

The question answers itself. All of the great policies above in Finland/Israel/etc can only buy 0.5 - 1 extra point over the countries that screw them up or pursue the exact opposite policies. Meanwhile, economic growth has earned us three whole points since medieval times. And these are stable; no opposite-party administration can take them away. So long as technological civilization continues to function, we can keep rising and rising, until in 300 years we approach the 10/10 barrier and have to get a new chart. We can say with confidence that we are giving our children and grandchildren better lives. Sure, only one point better for our grandchildren in particular - but one point on a ten point scale is a big deal, larger than the entire COVID drop!

The whole thing is terribly romantic. Our plans are measured in centuries3; how are the communists’ and fascists’ plans going?