2026-03-24 | AutoInterp Hackathon
Goal: SAEs, CLTs, Visualization, and MOLTs
Summary: Trained Symbolic IOI Transformer and TopK SAE
Work sessions
| In | Out |
|---|---|
| 17:50 | 07:30 |
AutoInterp Goals
Copying these over from yesterday:
- Implement Toy IOI Transformer from scratch
- Use Claude to verify that my Pytorch Transformer implementation is correct
- Create Dataset of IOI Sentences
-
Train Model and verify model generalization
-
Train SAEs and understand SAE math
-
Train PLT and CLTs and understand math
-
Understand Attention Interactions
-
Train MOLTs
-
Visualizations should be used throughout