Talk on Representation and Control of Meanings in Large Language and World Models

Title

Representation and Control of Meanings in Large Language and World Models

 

Abstract

Can ‘stochastic parrots’ represent meanings? Do they have a ‘state of mind’ and, if so, can this state of mind be controlled? Do we even have definitions and language to tackle these questions? When viewed as dynamical systems, GenAI models partition sequences of input data into equivalence classes represented by the distribution over their continuations. We call these (Nerode) equivalence classes ‘meanings’. Their representations can be used to establish symmetric relations, such as similarity, and asymmetric ones, such as entailment and containment. The data could be expressions in natural language or sensory measurements from physical devices. I will therefore frame the question of controllability in this quotient space of meanings and describe necessary and sufficient conditions, which today’s most proficient AI bots largely satisfy. The ability to control the ‘state of mind’ of these bots is key to their safe and secure deployment. The notion of meanings proposed sheds new light on problems dating back to Solomonoff and Kolmogorov, pointing to the inadequacy of their framework in capturing any non-trivial notion of `substantial’ similarity among measured data, which has implications for copyright doctrine. On the bright side, even if there can be no canonical or universal notion of conceptual (e.g., structural, semantic, substantial etc.) information, there can still be universally shareable ones, for which meanings as represented by LLMs play a defining role.

 

Bio

Stefano Soatto is Vice President at AWS, where he has led the science teams that developed AI Applications including GenAI (Amazon Q, Bedrock, Titan, CodeWhisperer), vision (Textract, Rekognition, Lookout For Vision), speech (Transcribe, Transcribe Medical), language (Comprehend, Comprehend Medical, Lex, Translate, Kendra), verticals (CodeGuru, DevOps Guru, Forecast, Personalize, Lookout For Equipment, Lookout for Metrics, Monitron). He is a also Professor of Computer Science and Electrical Engineering at UCLA, and a Fellow.