I'm not having children, so I'm thinking of simply donating my savings (and whatever money it'll make when they sell my property) when I'm dead. It probably won't be exactly 10% of my lifetime earnings, but close enough. On the grand scale it doesn't really matter if I'm saving starving kids now or in 30 years.
* We might expect it to be cheaper to save lives now than in the future (e.g. if malaria is cured, then that saves a lot of lives and raises the cost of the "cheapest life to save").
* You get interest on your money, but on the other hand, the people whose lives you save could be having positive effects on the world that compound over 30 years
* If transformative tech comes in the next 30 years, either an extreme form of the first bullet could happen or we could all be dead; this gives some reason to think your money is more leveraged now.
But I don't think this is anywhere near decisive, and there are good reasons to wait to give. See the EA discussion of "patient philanthropy" for more.
The recent post about AI alignment finally motivated me to post something I've been mulling over for a while now.
Let us assume that AI alignment is solved, by which I mean, that
* given any set of values V
* given a moral framework F which maps the values of V to decisions in reality
we have the tech to align AI so that their decisions in reality always derive from V, as mapped to reality by F.
Example for demonstration, please don't rule lawyer: We impose upon a car-driving AI the sole value "human lives are valuable" and utilitarianism as the moral framework. So aligned the AI will not ram random random people on the street, but it will ram someone shooting into a crowd.
I think it's fair to say this alignment-ability would be a great improvement over what we have today, but to me this is clearly not enough to ensure safety of an ASI so aligned.
The issue is that if you posit something with approximately unlimited power and agency (relative to us) you would really like to avoid it deciding to do horrible things because you can't stop it. But I don't know of any combination of values V and moral framework F, consistently applied, which does not occasionally endorse horrible actions.
The only way I know to square that circle and not have horrible (to me) things be moral is to drop the consistency requirement. But I don't think that generalizes, especially not to ASI.
As a result every time the ASI alignment debate comes up for another round I think to myself that, hey, even if we solved alignment we'd still be fucked, right? Since even if we could impose our values V and moral framework F on ASI, we'd first need some V and F that are not occasionally horrible and I don't know that exists.
So to me, solving AI alignment would not in itself be the solution people seem to treat it as.
I'm very sympathetic to these concerns. Devil's advocate: an ASI wouldn't be dependent on our Vs and Fs but would be able to refine them. But I suspect our Vs and Fs are largely to do with the shared experience of being human especially mortality.
This is essentially THE consideration that started the AI alignment movement in the first place. There are some more complicated arguments involved, but the core idea is what you appeared to have rediscovered.
The technical problem, the "how" of aligning the AI is secondary. The bigger problem is that we don't know what even to align it to, or how to decide that.
For what it's worth, these questions have been discussed for quite some time now.
Nobody so far have offered a good solution to that problem.
I'm sure there will be people attempting to tell you you got something wrong, but they would be wrong at that.
I live outside the US but keep up with its politics. I had no idea, when it happened, that Kirk's assassination would have such an impact on online discourse. It felt as if my Substack feed had been switched with that of a highly polarized user's. I'm not sure I liked anything I read in that time.
This made me think of a broader point - Substack's recommendation algorithm seems to suffer from the platform's growth. I wish there was a way I could bypass some of its assumptions about me, like flipping a "I'm not a dopamine zombie" switch that'd make it much more likely to recommend weird, fascinating stuff to me.
Substack seems to have changed a lot in the last year (perhaps since January 20th). It used to be all very thoughtful posts about culture and science and philosophy and arts, but recently, it's well over 50% politics. I guess some of that is my fault because I've clicked or commented on posts about Trump or Kirk or immigration. It's not entirely my fault, though, because even the intellectual people that I follow are posting about politics when they didn't before.
Even in politics, I used to be able to discuss difficult topics politely. That's mostly gone now. I hadn't been called names in three years, but now it's three times per day. I don't think my manner has changed — I hope I'm always respectful — but the vibe has changed.
Stewart-Williams posted this week about how people with political ideas exaggerate their beliefs about what 'the other side' believes. I expect this is a big part of the problem, and Charlie Kirk's murder has brought this to the fore — it's rare to find people on Substack who have a measured opinion about what Kirk stood for.
In my country (England), the big issue is immigration: the people who think that immigration is too high are far-right racists, while the people who think that more immigration is a good thing hate their country and hate the England flag. It's almost impossible for them to talk to each other reasonably. This was always true on Twitter, but it's true now on Substack, too.
It's easy to make gore and atrocities frightening, but doing so doesn't interest me much.
How would you go about making a horror movie that had to be PG rated or lower as scary as possible?
I recently designed a very simple, slightly educational browser game where you build an AI Governance Strategy to avert doom. It probably still has a few bugs, but might be fun: https://jack-stennett.github.io/AI-Governance-Strategy-Builder/ . Accompanied by short substack post: https://torchestogether.substack.com/p/playing-at-saving-the-world
It's open source, so please steal and make a better version!
Questions for the EAs:
I'm not having children, so I'm thinking of simply donating my savings (and whatever money it'll make when they sell my property) when I'm dead. It probably won't be exactly 10% of my lifetime earnings, but close enough. On the grand scale it doesn't really matter if I'm saving starving kids now or in 30 years.
Thoughts?
The average price of gas thirty years ago was $1.11. The longer you hold your money, the less it's worth.
Some thoughts:
* We might expect it to be cheaper to save lives now than in the future (e.g. if malaria is cured, then that saves a lot of lives and raises the cost of the "cheapest life to save").
* You get interest on your money, but on the other hand, the people whose lives you save could be having positive effects on the world that compound over 30 years
* If transformative tech comes in the next 30 years, either an extreme form of the first bullet could happen or we could all be dead; this gives some reason to think your money is more leveraged now.
But I don't think this is anywhere near decisive, and there are good reasons to wait to give. See the EA discussion of "patient philanthropy" for more.
The recent post about AI alignment finally motivated me to post something I've been mulling over for a while now.
Let us assume that AI alignment is solved, by which I mean, that
* given any set of values V
* given a moral framework F which maps the values of V to decisions in reality
we have the tech to align AI so that their decisions in reality always derive from V, as mapped to reality by F.
Example for demonstration, please don't rule lawyer: We impose upon a car-driving AI the sole value "human lives are valuable" and utilitarianism as the moral framework. So aligned the AI will not ram random random people on the street, but it will ram someone shooting into a crowd.
I think it's fair to say this alignment-ability would be a great improvement over what we have today, but to me this is clearly not enough to ensure safety of an ASI so aligned.
The issue is that if you posit something with approximately unlimited power and agency (relative to us) you would really like to avoid it deciding to do horrible things because you can't stop it. But I don't know of any combination of values V and moral framework F, consistently applied, which does not occasionally endorse horrible actions.
The only way I know to square that circle and not have horrible (to me) things be moral is to drop the consistency requirement. But I don't think that generalizes, especially not to ASI.
As a result every time the ASI alignment debate comes up for another round I think to myself that, hey, even if we solved alignment we'd still be fucked, right? Since even if we could impose our values V and moral framework F on ASI, we'd first need some V and F that are not occasionally horrible and I don't know that exists.
So to me, solving AI alignment would not in itself be the solution people seem to treat it as.
Feel free to tell me why I am wrong :)
I'm very sympathetic to these concerns. Devil's advocate: an ASI wouldn't be dependent on our Vs and Fs but would be able to refine them. But I suspect our Vs and Fs are largely to do with the shared experience of being human especially mortality.
This is essentially THE consideration that started the AI alignment movement in the first place. There are some more complicated arguments involved, but the core idea is what you appeared to have rediscovered.
The technical problem, the "how" of aligning the AI is secondary. The bigger problem is that we don't know what even to align it to, or how to decide that.
For what it's worth, these questions have been discussed for quite some time now.
Nobody so far have offered a good solution to that problem.
I'm sure there will be people attempting to tell you you got something wrong, but they would be wrong at that.
"So aligned the AI will not ram random random people on the street, but it will ram someone shooting into a crowd."
To your point that's basically a well established and exploited mechanism in EVE online
I live outside the US but keep up with its politics. I had no idea, when it happened, that Kirk's assassination would have such an impact on online discourse. It felt as if my Substack feed had been switched with that of a highly polarized user's. I'm not sure I liked anything I read in that time.
This made me think of a broader point - Substack's recommendation algorithm seems to suffer from the platform's growth. I wish there was a way I could bypass some of its assumptions about me, like flipping a "I'm not a dopamine zombie" switch that'd make it much more likely to recommend weird, fascinating stuff to me.
Substack seems to have changed a lot in the last year (perhaps since January 20th). It used to be all very thoughtful posts about culture and science and philosophy and arts, but recently, it's well over 50% politics. I guess some of that is my fault because I've clicked or commented on posts about Trump or Kirk or immigration. It's not entirely my fault, though, because even the intellectual people that I follow are posting about politics when they didn't before.
Even in politics, I used to be able to discuss difficult topics politely. That's mostly gone now. I hadn't been called names in three years, but now it's three times per day. I don't think my manner has changed — I hope I'm always respectful — but the vibe has changed.
Stewart-Williams posted this week about how people with political ideas exaggerate their beliefs about what 'the other side' believes. I expect this is a big part of the problem, and Charlie Kirk's murder has brought this to the fore — it's rare to find people on Substack who have a measured opinion about what Kirk stood for.
In my country (England), the big issue is immigration: the people who think that immigration is too high are far-right racists, while the people who think that more immigration is a good thing hate their country and hate the England flag. It's almost impossible for them to talk to each other reasonably. This was always true on Twitter, but it's true now on Substack, too.