The DemystifySci Podcast A Terrifying Theory of Language - Dr. Elan Barenholtz, DemystifySci #413
Mar 30, 2026
Dr. Elan Barenholtz, cognitive scientist who studies language and perception, presents the idea that language is autogenerative and runs on its own internal logic. He compares human language to large language models and explores how language can shape thought, enable internal simulation, and sometimes detach from direct sensory reality. The conversation highlights limits of meaning, Chomsky’s legacy, and where LLMs mirror linguistic structure.
AI Snips
Chapters
Transcript
Episode notes
Language Runs On Its Own Internal Logic
- Language is intrinsically autogenerative and can produce coherent structure without direct grounding in the world.
- Elan Barenholtz argues LLMs revealed this computational property by predicting next tokens from corpus statistics, showing language can 'run' itself.
World Shapes Language But Doesn't Drive It
- If language is self-causing, external reality cannot fully determine word usage; the world 'bumps' language but doesn't drive its core mechanics.
- Barenholtz uses a car metaphor: steering (world) influences trajectory but engine (language) runs itself.
Epiphany About Redness At A Small Gathering
- Elan recounts a sudden epiphany at a small gathering where a discussion about 'redness' triggered clarity about language's decoupling from sensory experience.
- He describes that moment as inexplicable but the product of long, obsessive study across fields.