The Quanta Podcast Audio Edition: Can AI Models Show Us How People Learn? Impossible Languages Point a Way.
48 snips
May 22, 2025 Delve into the intriguing world of artificial languages that defy known grammatical rules. Discover how linguists are leveraging AI models to unravel the mysteries of language learning. The conversation highlights groundbreaking experiments that challenge our understanding of how both humans and machines acquire language. Prepare for a fascinating exploration of the parallels and divergences in language acquisition, as AI reveals insights into the learning process.
AI Snips
Chapters
Transcript
Episode notes
Do Language Models Really Learn?
- Language models trained by predicting the next word can generate fluent sentences without traditional linguistic theory.
- It's unclear if they truly learn language or just mimic patterns superficially.
Innate Constraints Shape Language
- Some grammatical rules never appear in any human language indicating a cognitive bias.
- Chomsky argued humans have innate language machinery that rejects impossible languages.
Graduate Student Challenges Chomsky
- Julie Collini discovered no recent tests of impossible languages on modern language models.
- She undertook a rigorous study of transformer networks, challenging Chomsky's claim.