2026-01-21 | LISA Coworking - ACDC and Meetings
Goal: LISA Coworking - Circuits Deep Dive | Start Programming KL Divergence | David Quarel PPO Lecture
Summary: Implemented KL Divergence/dive into math, meetings!
Work sessions
| In | Out |
|---|---|
| 9:00 | 19:00 |
Meetings
-
Met with J Rosser (one of the nicest people I've met)
-
Clement Dumas currently doing MATS (super patient at explaining and very insightful)
Progress
-
A little bit slower progress on ACDC then I would have liked. I started unit testing KL Divergence and understanding the ACDC algorithm better (something I got tripped up on was that the activation patching is the "removal of the edge" not actually ablating (
zeroing)) it -
Finished watching up to the paper review from Neel Nanda and Arthur Conmy (part 1 and part 2)
-
Coordinated timing of meetings and work for Thursday/tenatively Friday (tomorrow is RLHF!)