Can AI Actually Reason? скачать в хорошем качестве

Can AI Actually Reason?

Is Reasoning Just Compression? Reinforcement Learning, Entropy, and Efficient Uncertainty Reduction Mehrdad Zaker of Neural Intelligence Labs argues that what looks like step-by-step logical reasoning in models may actually be structured compression: reinforcement learning selects trajectories through intermediate states that reduce uncertainty and narrow hypothesis space. She explains how imposing reusable structure and hierarchical subgoals reduces branching factor and search dimensionality, reframing “deeper thought” as “narrower search.” The episode discusses why models overthink when verbosity isn’t penalized, how errors compound across long chains when uncertainty isn’t compressed at each step, and why tool use can inject entropy, causing uncertainty spikes that require recompression. Zaker also covers process rewards as local compression guidance, the need for hybrid reward mixtures balancing local signals and global correctness, and the distinction between truth-seeking and compression-seeking optimization. 00:00 Podcast Intro 00:20 Reasoning as Compression 00:44 Reasoning as State Trajectories 01:42 Structure Shrinks Search 02:38 Hierarchy Reduces Entropy 03:31 Overthinking and Verbosity 04:43 Error Compounding Failures 06:17 Tool Use Entropy Spikes 07:45 Process Rewards Guidance 08:44 Truth vs Compression Goals 10:03 Reframing What Reasoning Is 10:54 Conclusion Better Compression