Dmitrii Krasheninnikov

Likes nature and AI x-safety

About Posts

Fresh in memory: Training-order recency is linearly encoded in language model activations

Reposting the twitter thread about our recent paper with Rich Turner and David Krueger.

Read More

Liar’s poker: a new card game you should try

This post is about a card game my friend Neel Alex invented, which many of our friends and families now like a lot. The game is similar to liar’s dice and bullshit, and uses poker hands. The game is fast-paced and very fun: you get to bluff and guess and mess with other people’s guesses.

Read More

Implicit meta-learning may lead language models to trust more reliable sources

Reposting the lightly edited twitter thread about our paper with Egor, Bruno, Tegan and David.

Read More
Google Scholar Google Scholar Twitter Twitter GitHub GitHub LinkedIn LinkedIn Email Email