#302 Karl Friston: How the Free Energy Principle Could Rewrite AI - Eye On A.I. Recap

Podcast: Eye On A.I.

Published: 2025-11-19

Duration: 1 hr 3 min

Guests: Karl Friston

Summary

Karl Friston discusses how the Free Energy Principle can transform AI by mimicking the brain's predictive coding, offering a more efficient and reliable model compared to traditional AI approaches.

What Happened

Karl Friston delves into the concept of the Free Energy Principle, which he describes as a way to understand how living systems, including human brains, maintain order by minimizing surprise. He explains that this principle can be applied to artificial intelligence to create systems that learn and adapt more efficiently than current models based purely on data memorization.

Friston highlights how the Free Energy Principle is rooted in physics and has been adapted from Richard Feynman's work on quantum electrodynamics. This principle focuses on minimizing the gap between prediction and reality, thus reducing prediction error.

The episode discusses how Versus, a startup where Friston is the chief scientist, is applying these ideas through its AI architecture called Axiom. This model is designed to improve memory use, generalization, and reasoning in AI by simulating the brain's process of Bayesian inference.

Friston elaborates on how mental disorders can be understood as false inference, emphasizing the brain's continuous process of updating its predictions to align with sensory input. This understanding could offer new insights into computational psychiatry.

The conversation touches on the limitations of current AI models, such as transformers, and how Friston's approach could overcome issues like inefficiency and lack of reliability. Axiom reportedly achieves better performance with significantly less computational power.

The episode also explores the potential for these AI models to learn from experiences similarly to humans by refining their probability distributions rather than relying on static weights. This could eliminate problems like catastrophic forgetting in neural networks.

Friston argues for a model of AI that is both embodied and situated, akin to how humans learn from direct interactions with their environment. He suggests that this approach could eventually lead AI to develop language and other complex skills through natural exposure.

Key Insights

• The Free Energy Principle, adapted from Richard Feynman's work on quantum electrodynamics, aims to minimize the gap between prediction and reality, enhancing AI's ability to learn and adapt efficiently.
• Versus, a startup utilizing the Free Energy Principle, developed the AI architecture Axiom, which improves memory use and reasoning by simulating Bayesian inference processes in the brain.
• Current AI models like transformers face challenges such as inefficiency and catastrophic forgetting, which Axiom reportedly overcomes by refining probability distributions instead of relying on static weights.
• An embodied and situated AI model, akin to human learning from direct environmental interactions, could enable AI to naturally develop complex skills such as language.