LW - AI #67: Brief Strange Trip by Zvi

The Nonlinear Library: LessWrong

Link to original article
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI #67: Brief Strange Trip, published by Zvi on June 7, 2024 on LessWrong.
I had a great time at LessOnline. It was a both a working trip and also a trip to an alternate universe, a road not taken, a vision of a different life where you get up and start the day in dialogue with Agnes Callard and Aristotle and in a strange combination of relaxed and frantically go from conversation to conversation on various topics, every hour passing doors of missed opportunity, gone forever.
Most of all it meant almost no writing done for five days, so I am shall we say a bit behind again. Thus, the following topics are pending at this time, in order of my guess as to priority right now:
1. Leopold Aschenbrenner wrote a giant thesis, started a fund and went on Dwarkesh Patel for four and a half hours. By all accounts, it was all quite the banger, with many bold claims, strong arguments and also damning revelations.
2. Partly due to Leopold, partly due to an open letter, partly due to continuing small things, OpenAI fallout continues, yes we are still doing this. This should wait until after Leopold.
3. DeepMind's new scaling policy. I have a first draft, still a bunch of work to do.
4. The OpenAI model spec. As soon as I have the cycles and anyone at OpenAI would have the cycles to read it. I have a first draft, but that was written before a lot happened, so I'd want to see if anything has changed.
5. The Rand report on securing AI model weights, which deserves more attention than the brief summary I am giving it here.
6. You've Got Seoul. I've heard some sources optimistic about what happened there but mostly we've heard little. It doesn't seem that time sensitive, diplomacy flows slowly until it suddenly doesn't.
7. The Problem of the Post-Apocalyptic Vault still beckons if I ever have time.
Also I haven't processed anything non-AI in three weeks, the folders keep getting bigger, but that is a (problem? opportunity?) for future me. And there are various secondary RSS feeds I have not checked.
There was another big change this morning. California's SB 1047 saw extensive changes. While many were helpful clarifications or fixes, one of them severely weakened the impact of the bill, as I cover on the linked post.
The reactions to the SB 1047 changes so far are included here.
Table of Contents
1. Introduction.
2. Table of Contents.
3. Language Models Offer Mundane Utility. Three thumbs in various directions.
4. Language Models Don't Offer Mundane Utility. Food for lack of thought.
5. Fun With Image Generation. Video generation services have examples.
6. Deepfaketown and Botpocalypse Soon. The dog continues not to bark.
7. They Took Our Jobs. Constant AI switching for maximum efficiency.
8. Get Involved. Help implement Biden's executive order.
9. Someone Explains It All. New possible section. Template fixation.
10. Introducing. Now available in Canada. Void where prohibited.
11. In Other AI News. US Safety Institute to get model access, and more.
12. Covert Influence Operations. Your account has been terminated.
13. Quiet Speculations. The bear case to this week's Dwarkesh podcast.
14. Samuel Hammond on SB 1047. Changes address many but not all concerns.
15. Reactions to Changes to SB 1047. So far coming in better than expected.
16. The Quest for Sane Regulation. Your random encounters are corporate lobbyists.
17. That's Not a Good Idea. Antitrust investigation of Nvidia, Microsoft and OpenAI.
18. The Week in Audio. Roman Yampolskiy, also new Dwarkesh Patel is a banger.
19. Rhetorical Innovation. Innovative does not mean great.
20. Oh Anthropic. I have seen the other guy, but you are not making this easy.
21. Securing Model Weights is Difficult. Rand has some suggestions.
22. Aligning a Dumber Than Human Intelligence is Still Difficult. What to do?
23. Aligning a Smarter Than Human Inte...