We get innovation in functional search. In an even more functional search, we finally get a Nature paper submitted almost two years ago, in which AI discovered a new class of antibiotic. That's pretty damn exciting, with all the implications thereof.

OpenAI continued its rapid pace of shipping, pivoting for this week to safety. There was a paper about weak-to-strong generalization. I see what they are trying to do. It is welcome, but I was underwhelmed. It and Leike's follow-up post continue down a path for which I have high skepticism, but the new concreteness gives me more hope that the flaws will be exposed early, allowing adjustment. Or I could be wrong.

OpenAI also had the beta release of Preparedness Framework. That was more exciting. There was a lot of great stuff there, much better than I would have expected, and having a framework at all is [...]

---

Outline:

(01:17) Language Models Offer Mundane Utility

(04:13) Language Models Don’t Offer Mundane Utility

(05:20) GPT-4 Real This Time

(07:54) Fun with Image Generation

(08:10) Deepfaketown and Botpocalypse Soon

(08:59) Digi Relic Digi

(19:57) Going Nuclear

(21:17) Get Involved

(22:27) Follow the Money

(25:26) Introducing

(31:07) In Other AI News

(34:27) Quiet Speculations

(43:14) The Quest for Sane Regulations

(52:06) The Week in Audio

(52:55) Rhetorical Innovation

(01:02:28) Aligning a Smarter Than Human Intelligence is Difficult

(01:20:17) Vulnerable World Hypothesis

(01:22:55) People Are Worried About AI Killing Everyone

(01:26:49) Other People Are Not As Worried About AI Killing Everyone

(01:28:12) The Lighter Side

---

First published:
December 21st, 2023

Source:
https://www.lesswrong.com/posts/WaDFCrd6KEwojLXgj/ai-43-functional-discoveries

---

Narrated by TYPE III AUDIO.

“AI #43: Functional Discoveries” by Zvi

LessWrong (30+ Karma)