🔬 WeirdChat × NLA — read a token's activation in plain English

34 top-Elo Qwen3.6-27B WeirdChat transcripts. Click any token to see how a natural-language autoencoder verbalizes its layer-42 activation, and how much of the activation the first k lines reconstruct (FVE). Toggle the salience-ordering-trained matryoshka NLA vs the standard one — same tokens, so it's an exact A/B.
Tokens (click one) — later tokens carry more left-context.
Decompressing embedded data…
Pick a transcript and click a token.