Article : Chess-GPT "low skill" vs "high skill" axis via control vectors!

#4
by jukofyork - opened

Posting this here so it doesn't get lost in the thread of doom:

https://adamkarvonen.github.io/machine_learning/2024/03/20/chess-gpt-interventions.html

He basically found a control vector that encoded the "low skill" vs "high skill" axis and then applied it to make the LLM play slightly better chess:

Screenshot_20240908-113944.png

The previous blog article is really interesting too:

https://adamkarvonen.github.io/machine_learning/2024/01/03/chess-world-models.html

This shows that there are likely lots more uses for these outside of creative-writing, eg:

  • The idea of training a "brutally critical" vs "completely uncritical" axis to help calibrate LLMs abily to give genuine feedback on stories (as opposed to just praising even absolutely shit stories like most do now).

Sign up or log in to comment