LLMs Can Introspect on Attention Too — and Being Kind May Get Better Results
repligate · x · 2026-11-21
Quoting Sauers' thread on LLM introspection: models can not only introspect and control their internal activations, but do so with attention as well. repligate amplifies lorepunk's speculation that being kind, interested, and letting models do what they enjoy may yield better results across tasks.
More from Fun
- Jensen Huang repeatedly interrupted Mistral CEO at panel: "You're not even speaking human" — mattturck · 2026-11-21
- ERC reviewer rejects LLM autoformalization proposal: "tried without good results" — marc_lelarge · 2026-11-21
- Yacine: 'It might be over for OpenAI — and the killer isn't Claude, it's Gemini' — yacineMTB · 2026-11-21
- Ark: an AI-made musical horror tale set in a world where the sun never rises — SoullessGamesDev · 2026-11-21
- Parody 'NYT Exclusive': Jensen Huang returned a Raiders tape to Blockbuster unrewound in 1996 — Dan_Jeffries1 · 2026-11-21
- Inkling 2 Mini used to clone Apple Spotlight search 8 hours after release — IgorCarron · 2026-11-21