This is the weekly visible open thread. Post about anything you want, ask random questions, whatever. ACX has an unofficial subreddit, Discord, and bulletin board, and in-person meetups around the world. Most content is free, some is subscriber only; you can subscribe here. Also:
1: In my California primary ballot recommendations post this spring, I said I was voting for the boring incumbent for 12th District House because the challenger seemed eccentric and had no chance. The challenger in that race, Jamie Joyce, contacted me and asked me to change my mind (both candidates will be facing off again in the general election). I don’t want to relitigate what I said in May, but I may reconsider if I make a post on the November election. In the meantime, she asked me to link her response, which I recommend to any interested 12th District voters.
2: More college EA meetup group updates: Oxford, University College Dublin, and the University of the Philippines now have meetups planned. See the list for details.
3: Astralis Foundation and Macroscopic Ventures, which I think are parts of our conspiracy although I can’t keep track any more, announce that they will be distributing up to $10 million in grants for work on “middle powers and transformative AI”, ie “research, training, public writing and new organizations focused on what countries beyond the US and China can do to help navigate” the AI transition. You can read more and apply here.
4: Comments of the week, all from the Mysteries of AI Generalization post:
Eliezer Yudkowsky on why the latest failures of generalization are bad news.
Ezra Newman on relevant distinctions between pretraining/midtraining/RLVR
Elias Schmied links Evan Hubinger on why Hacker Opus didn’t show emergent misalignment.
Stuart Armstrong on how emergent misalignment happens.