Jacob Andreas @jacobandreas Β· 29 Jun 2026
π New preprint! Automated interpretability by approximating / replacing NN components (here attention heads) with programs.
26 902Views
168Likes
16Reposts
0Replies
1Quotes
96Bookmarks
Is that a lot?
2.25Γvs this author's median11 940 views is typical
54Percentile for this authorof 11 recent posts
19.4Γvs 10Kβ100K median1 384 views is typical
107.96%Reachviews Γ· followers
0.69%Engagement rateof viewers reacted