Sparse Autoencoders: Progress & Limitations with Joshua Engels · NDIF Team / YouTube
Known for
DeepMind AGI Safety → METR, 2026.
Why chosen
He connects interpretability research with independent oversight of frontier AI. In 2024 he was lead author, with Max Tegmark's group at MIT, on "Not All Language Model Features Are Linear", which used sparse autoencoders to find circular day-of-week and month features inside GPT-2 and Mistral 7B, published at ICLR 2025. He then joined Google DeepMind's AGI Safety team, working on scalable oversight. In 2026 he left for METR to study where misaligned behavior comes from and whether safeguards hold. Opening the black box, then auditing it.