Tiny models and recursive reasoning in a new episode of The Information Bottleneck

The announcement introduces a conversation with a Microsoft principal researcher who authored “Less is More: Recursive Reasoning with Tiny Networks.” It describes a discussion of efficiency, tiny models and her roughly 45% result on ARC-AGI with a fraction of the parameters. Other advertised topics include why language models struggle with molecules, why puzzles suit recursion and what the guest would do with a trillion dollars.