MIT's Jacob Andreas trains LLMs to explain their own internals in plain language
MIT News AI · rss · 2026-11-17
Jacob Andreas, newly tenured at MIT EECS/CSAIL, builds techniques that let AI models accurately describe what they learned in language humans understand — rather than forcing researchers to unpack black boxes with external tools.
- His group generates natural-language descriptions for individual nodes in huge neural networks (e.g., a node detecting "patches of grass near animals"), opening higher-level questions about how models represent meaning
- His team also designs algorithms to predict how specific datasets or examples will shape model behavior; he notes the field lacks a scientific theory of what training on a given dataset will do
- He rode NLP's shift from statistical methods to deep learning, did thesis work on neural module networks, and pivoted into interpretability
- He says the field reinvents itself every 3-5 years, forcing annual course redesigns, and teaching technical ML when LLMs can do every assignment is an open challenge
More from AGI Musings
- “Stop aiming for 20% productivity — AI should end work-for-money altogether” — VraserX · 2026-11-17
- iamtrask: agent time horizon measures stability, not intelligence — don't conflate ASI signals — iamtrask · 2026-11-17
- Blogger: much of white-collar work was just "button clicking," now over as computers learned English — moonsandhues · 2026-11-17
- Debate: can OpenAI's plan to have agents formally prove safety properties work? — zetalyrae · 2026-11-17
- FT: Firms Investing Heaviest in AI Grew White-Collar Headcount 10.2% — Scobleizer · 2026-11-17
- MIT mathematician and ex-NFL lineman John Urschel on why math thinking still matters in the AI era — MIT News AI · 2026-11-17