Kajetan Janiak

Kajetan Janiak (publicly known as Jett Janiak) is a mechanistic interpretability researcher from Poland who studied at the University of Warsaw, where he completed an MS in the Faculty of Mathematics, Informatics and Mechanics. He participated in the MATS (ML Alignment & Theory Scholars) Winter 2023 cohort under Neel Nanda's mentorship and later in a subsequent MATS cohort under Arthur Conmy, both in the mechanistic interpretability stream. His research focuses on understanding the internal mechanisms of transformer models, including work on polysemantic attention heads, circuit discovery in small transformers, stable regions in the residual stream of LLMs, sparse autoencoders, and chain-of-thought faithfulness in frontier models. He has co-authored several papers and Alignment Forum posts, including "Chain-of-Thought Reasoning In The Wild Is Not Always Faithful" (ICLR 2025 workshop) and "Characterizing Stable Regions in the Residual Stream of LLMs." He has also been involved with AI Safety Camp as a project lead. He received a grant from the Long-Term Future Fund to cover costs of leaving employment in order to pursue AI safety research.

No endorsements yet.

Personal Website

-

LinkedIn

LinkedIn

Twitter / X

Twitter

LessWrong

jett-janiak

EA Forum

-