Epistemic status: My coauthor and I are both noobs in this field. Expect errors and conceptual flaws.

tl;dr

For a four-day capstone project for the ARENA program my partner and I did a replication of the MELBO Lesswrong article using Llama-3.2-1b-Instruct.

Intro

I've been spending the last month at the ARENA program developing some basic technical skills to match my long-standing layman's interest in AI and AI X-risk. And ARENA's final week is a capstone project asking us to engage with current literature and do something vaguely novel with it over the span of about a week.

This being a weeklong project, we needed to scope it very tightly. So my partner and I chose to replicate MELBO on a different LLM! The project of "finding, in an unsupervised way, plausible out-of-distribution model behavior" seemed interesting for its own sake and also possibly useful for the task [...]

---

Outline:

(00:21) tl;dr

(00:35) Intro

(01:32) Summary

(03:00) R hyperparameter sweep

(06:44) Concept-specific steering vectors

(07:02) Red teaming

(11:11) The Negation Non-Refusal

(11:56) Non-English language vectors

(13:06) Mental health/suicide prevention

(15:08) Github repo

(15:17) Possible Followups

The original text contained 1 image which was described by AI.

---

First published:

Source:

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.