From generative AI to the brain: five takeaways

TL;DR

The paper investigates whether core generative AI principles can be instantiated in brain circuits and thereby inform cognitive neuroscience. It synthesizes five principles—world modelling, generation of thought processes (CoT/CoX), attention with self-consistency, neural scaling laws, and quantization—and discusses supporting ML ideas such as the two-stage plan of universal unsupervised world modelling followed by supervised fine-tuning and the information bottleneck. The authors argue that world modelling alone is insufficient for cognition and propose brain-analog scenarios including concurrent world-modeling with reinforcement-like fine-tuning, self-consistent attention loops, and quantized synapses as potential parallels. They advocate cross-disciplinary empirical work to test these generative-principle hypotheses and to derive mutual insights for ML and neuroscience.

Abstract

The big strides seen in generative AI are not based on somewhat obscure algorithms, but due to clearly defined generative principles. The resulting concrete implementations have proven themselves in large numbers of applications. We suggest that it is imperative to thoroughly investigate which of these generative principles may be operative also in the brain, and hence relevant for cognitive neuroscience. In addition, ML research led to a range of interesting characterizations of neural information processing systems. We discuss five examples, the shortcomings of world modelling, the generation of thought processes, attention, neural scaling laws, and quantization, that illustrate how much neuroscience could potentially learn from ML research.