#107 - Dr. RAPHAËL MILLIÈRE - Linguistics, Theory of Mind, Grounding

Semantic Competence

Semantic competence in AI involves understanding language through inferential and referential abilities. Raphaël Millière explains that inferential competence relates to connecting words through definitions and analogies, while referential competence involves naming objects and following instructions 1. He contrasts classical symbolic models, like SHRDLU, which have referential competence but lack inferential abilities, with modern language models that excel in inferential competence 2.

Language models are trained on statistical learning objectives, learning complex patterns of co-occurrence statistics in language.

--- Raphaël Millière

This distinction highlights the nuanced understanding required to assess AI's semantic capabilities.

   

AI vs. Cognition

Comparing AI's semantic capabilities to human cognition reveals both limitations and potential. Tim Scarfe notes that while language models can mimic human language use, they struggle with forming analogies, a creative process beyond mere data interpolation 3. Raphaël Millière emphasizes the importance of understanding deep neural networks to draw meaningful comparisons with human cognition in areas like reasoning and perception 4.

I'm interested in whether these networks might be human-like in terms of cognitive psychological capacities.

--- Raphaël Millière

This exploration into AI's cognitive abilities challenges existing ideas in cognitive science and philosophy.