LW - AI #81: Alpha Proteo by Zvi

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 #81: Alpha Proteo, published by Zvi on September 12, 2024 on LessWrong.
Following up on Alpha Fold, DeepMind has moved on to Alpha Proteo. We also got a rather simple prompt that can create a remarkably not-bad superforecaster for at least some classes of medium term events.
We did not get a new best open model, because that turned out to be a scam. And we don't have Apple Intelligence, because it isn't ready for prime time. We also got only one very brief mention of AI in the debate I felt compelled to watch.
What about all the apps out there, that we haven't even tried? It's always weird to get lists of 'top 50 AI websites and apps' and notice you haven't even heard of most of them.
Table of Contents
1. Introduction.
2. Table of Contents.
3. Language Models Offer Mundane Utility. So many apps, so little time.
4. Language Models Don't Offer Mundane Utility. We still don't use them much.
5. Predictions are Hard Especially About the Future. Can AI superforecast?
6. Early Apple Intelligence. It is still early. There are some… issues to improve on.
7. On Reflection It's a Scam. Claims of new best open model get put to the test, fail.
8. Deepfaketown and Botpocalypse Soon. Bots listen to bot music that they bought.
9. They Took Our Jobs. Replit agents build apps quick. Some are very impressed.
10. The Time 100 People in AI. Some good picks. Some not so good picks.
11. The Art of the Jailbreak. Circuit breakers seem to be good versus one-shots.
12. Get Involved. Presidential innovation fellows, Oxford philosophy workshop.
13. Alpha Proteo. DeepMind once again advances its protein-related capabilities.
14. Introducing. Google to offer AI podcasts on demand about papers and such.
15. In Other AI News. OpenAI raising at $150b, Nvidia denies it got a subpoena.
16. Quiet Speculations. How big a deal will multimodal be? Procedural games?
17. The Quest for Sane Regulations. Various new support for SB 1047.
18. The Week in Audio. Good news, the debate is over, there might not be another.
19. Rhetorical Innovation. You don't have to do this.
20. Aligning a Smarter Than Human Intelligence is Difficult. Do you have a plan?
21. People Are Worried About AI Killing Everyone. How much ruin to risk?
22. Other People Are Not As Worried About AI Killing Everyone. Moving faster.
23. Six Boats and a Helicopter. The one with the discord cult worshiping MetaAI.
24. The Lighter Side. Hey, baby, hey baby, hey.
Language Models Offer Mundane Utility
ChatGPT has 200 million active users. Meta AI claims 400m monthly active users and 185m weekly actives across their products. Meta has tons of people already using their products, and I strongly suspect a lot of those users are incidental or even accidental. Also note that less than half of monthly users use the product monthly! That's a huge drop off for such a useful product.
Undermine, or improve by decreasing costs?
Nate Silver: A decent bet is that LLMs will undermine the business model of boring partisans, there's basically posters on here where you can 100% predict what they're gonna say about any given issue and that is pretty easy to automate.
I worry it will be that second one. The problem is demand side, not supply side.
Models get better at helping humans with translating if you throw more compute at them, economists think this is a useful paper.
Alex Tabarrok cites the latest paper on AI 'creativity,' saying obviously LLMs are creative reasoners, unless we 'rule it out by definition.' Ethan Mollick has often said similar things. It comes down to whether to use a profoundly 'uncreative' definition of creativity, where LLMs shine in what amounts largely to trying new combinations of things and vibing, or to No True Scotsman that and claim 'real' creativity is something else beyond that.
One way to interpret Gemini's capabilities tests is ...