ncase

Spooky Research! (also AI Safety Part 3 delayed, sorry)

Added 2024-10-27 19:13:32 +0000 UTC

Hi y'all! Time for the monthly cool-folks-supporting-me update.

Some good news & bad news. The good news is I've been super-productive the last month! I started on SIX new research projects (all related to AI or AI Safety), and learnt a few new programming languages & tools.

👉 Here's a 42-minute mega-post, describing my progress on 6 independent research projects! (Alas, this post is less polished & far more technical than my usual work, so feel free to skim & skip whatever doesn't make sense to you.)

In summary:

A way for LLMs to qualitatively and quantitatively elicit your values
Replicating & extending previous AI privacy/bias research
An easy-to-use tool for scaffolding LLMs (in collaboration with Hack Club)
Simulations of a robust way to beat Goodhart's Law
You Played Yourself: the game theory of self-modification
A tool to (semi-)automatically do causal inference, for social science

And as a consequence of me being so busy on that... here's the bad news: AI Safety Part 3 (Finale) is not gonna be done this month. Or next. Hopefully it'll be done mid- or late- December?

The above projects took priority, because they were actually to help me buff up my portfolio, for an application deadline for an AI Safety research camp. If they select me, I'll be spending Jan–Mar 2025 doing AI Safety research full-time! And maybe it'll even help me pivot to a full-time research career. (Don't worry, I still intend on making accessible, friendly explainers on the side, or as a side-effect of the research.)

Sorry again! In the meantime: here's my Sep/Oct Signal Boosts post (links of cool stuff I found in the past two months, including my "costume" for Halloween). Also, here's a drawing of a spooky cat: