Chris Manning on why LMs learn verb categories first, and why he thinks LeCun is wrong about language
ziv_ravid · x · 2026-09-29
The Information Bottleneck podcast features a long conversation with Stanford NLP legend Chris Manning (creator of CS224n). Key points:
- What linguistics gave ML: knowledge the field likely wouldn't have derived on its own
- New paper: language models learn verb categories before individual verbs
- Open problems are moving up: Manning argues NLP's next frontier lies in pragmatics and dialogue
- Disagreeing with LeCun: why Manning sees language as humans' biggest cognitive tool and thinks LeCun's view of language is wrong
- Can text alone teach meaning? Including the octopus thought-experiment debate with Emily Bender
- Closes on ReFT: editing representations instead of weights
Full episode available on the website, YouTube, and podcast apps.
More from AGI Musings
- Ex-Cohere researcher mocks AI doomer and savior narratives as $99.99/month sales pitches — suchenzang · 2026-09-29
- Sam Altman declares the takeoff begun: novel insights by 2026, robots by 2027 — itsOmSarraf_ · 2026-09-29
- Terence Tao-backed Caltech Mathathon redesigns around AI-assisted math understanding — AnimaAnandkumar · 2026-09-29
- Data back 20 more years: independent artists and writers double in 30 years as newspapers keep falling — Afinetheorem · 2026-09-29
- "Intelligence too cheap to meter": a bold call on collapsing AI costs — rand_longevity · 2026-09-29
- Physicist Kyle Cranmer: AI's wins show the value of grinding, not new ideas — KyleCranmer · 2026-09-29