Interview, Podcast
Edward Gibson: Human Language, Psycholinguistics, Syntax, Grammar & LLMs | Lex Fridman Podcast #426
- The book Syntax: A Cognitive Approach is expected to be published by MIT Press in the upcoming fall.
- Large Language Models (LLMs) are predicted to excel at processing form and surface structure but are anticipated to fail at tasks requiring true meaning and understanding, specifically failing logical puzzles like the Monty Hall problem when form is manipulated.
- LLMs are expected to struggle with center embedding (nested structures) to the same extent as humans, despite impressive capabilities in form, due to a prioritization of form over logical dependency.
- As LLMs grow larger, they may become increasingly skeptical of over-represented formulations, potentially improving their ability to resist trickery in logical contexts.
- Over the next two years to several centuries, language evolution is projected to progress significantly, with new high-resolution data potentially emerging on the birth of new languages within 100 years.
- English is expected to continue changing, evidenced by vowel shifts observed between 1952/53 and 50 years later.
- Center embedding is predicted to remain a universal constraint in human languages, rendering nested dependencies over a specific length uninterpretable, a finding supported by observations across 60 languages.
- Human languages are expected to maintain short dependency links as an evolutionary optimization to minimize cognitive costs in production and comprehension, distinguishing them from random control systems.
- Legal language is predicted to be viewed as an emergent phenomenon or "magic spell" using center embedding for performative meaning, with expectations that future presidents might ban such structures if a "language czar" appoints reforms.
- Future legal contracts are expected to be rewritten to remove center-embedded definitions, as creating material without such structures is considered feasible.
- Shannon's noisy channel theory is expected to explain the optimization of word order and syntax to ensure robustness against communication noise.
- Language extinction is anticipated to accelerate for languages lacking economic value, as there is insufficient incentive for local populations to maintain them.
- Machine translation is expected to face significant difficulties with concepts absent in source languages, such as exact counting in cultures like Pirahã.
- Human communication with other species, including whales, crows, and plants, is predicted as a possibility within a few decades to centuries, driven by the view that human language is not unique in its capacity for communication.
- Compositionality is expected not to be a trait exclusive to human language, with evidence suggesting other species possess complex signals.
- LLMs are expected to implement dependency grammar internally, aligning closely with construction-based theories that pair form and meaning.
- The language network in the human brain is expected to be highly stable over decades, showing no variation from the time an individual first became queen in 1952 to 50 years later.
- Languages with rule regularity are expected to be easier to learn than those with optimal but irregular strategies.
- Number words in cultures like Pirahã are expected to be approximate rather than exact, reflecting a cultural invention likely driven by agricultural needs rather than inherent cognitive ability.
- Color words are expected to be invented based on functional necessity, with non-industrialized cultures lacking labels for colors they do not need to discuss.
- Math and symbolic processing are expected to remain distinct from the language area, correcting the misconception that symbolic processing equates to language.
- Future linguistic analysis may determine that not all auxiliaries move, with lexical copying explaining word behavior in declarative versus interrogative contexts.
- The primary objective of language optimization is expected to be ease of production, though the role of the listener remains a debatable factor.