Jack Lindsey @Jack_W_Lindsey · 17 Jun 2026
LLMs encode whether they're "on the right track" in their activations along a linear axis, kind of like a value function in RL! The axis is influenced when you train the model to have new preferences, and modulates the model's confidence / likelihood of backtracking.
19 591Views
224Likes
19Reposts
7Replies
1Quotes
149Bookmarks
Is that a lot?
0.48×vs this author's median40 916 views is typical
14Percentile for this authorof 7 recent posts
7.12×vs 10K–100K median2 751 views is typical
101.55%Reachviews ÷ followers
1.28%Engagement rateof viewers reacted