Joseph Bloom

A Selec­tion of Ran­domly Selected SAE Features

SAE-VIS: An­nounce­ment Post

An­nounc­ing Neu­ron­pe­dia: Plat­form for ac­cel­er­at­ing re­search into Sparse Autoencoders

Un­der­stand­ing SAE Fea­tures with the Logit Lens

Ex­am­in­ing Lan­guage Model Perfor­mance with Re­con­structed Ac­ti­va­tions us­ing Sparse Au­toen­coders

Open Source Sparse Au­toen­coders for all Resi­d­ual Stream Lay­ers of GPT2-Small

Lin­ear en­cod­ing of char­ac­ter-level in­for­ma­tion in GPT-J to­ken embeddings

Fea­tures and Ad­ver­saries in MemoryDT

Joseph Bloom on choos­ing AI Align­ment over bio, what many as­piring re­searchers get wrong, and more (in­ter­view)

A Mechanis­tic In­ter­pretabil­ity Anal­y­sis of a GridWorld Agent-Si­mu­la­tor (Part 1 of N)

De­ci­sion Trans­former Interpretability