The core of Bengio's proposal is a 'predictor' model trained to approximate the Bayesian posterior over natural language queries. This model outputs probabilities for statements being true, distinguishing between communication acts and factual claims, and aims to best explain all observed data.
Impact: High. This approach fundamentally differs from current LLMs by focusing on truth modeling rather than next-token prediction or human preference, offering a more robust foundation for understanding the world.
In the source video, this keypoint occurs from 00:01:15 to 00:03:56.
Sources in support: Yoshua Bengio (Guest, AI Researcher)

