Modern AI models may hold key to understanding how brain actually learns

The research calls for a systematic exchange between AI and neuroscience. It points out that the brain and modern AI face similar challenges: representing the world, making decisions, controlling attention, managing resource constraints, and encoding information in stable formats.

A new scientific analysis argues that the rapid progress of generative artificial intelligence is no longer just a technological milestone but a source of testable hypotheses about how the human brain learns, reasons, and controls attention.

The study, titled “From Generative AI to the Brain: Five Takeaways” and authored by Claudius Gros, examines how specific mechanisms that drive modern AI models can reshape core assumptions in cognitive neuroscience. The paper highlights a widening gap between current neuroscience theories and the explanatory power of today’s large-scale AI systems, suggesting the field may need a conceptual shift to keep pace with computational advances.

The research analyses whether the breakthroughs in generative AI can inform biological theories instead of serving only as loose metaphors. To pursue this question, the author focuses on five areas in which AI has developed mechanisms that bear striking parallels to human cognition: world modelling, reasoning processes, attention control, scaling constraints, and discrete representation in memory. Each area, the study argues, provides a concrete experimental direction for understanding brain function through a more mechanistic lens, moving beyond descriptive theories and toward computationally grounded models.

Modern large language models acquire a broad internal representation of the world by predicting the next word across massive datasets. This process, which is largely unsupervised, mirrors the long-standing belief in neuroscience that the brain builds internal models through exposure and prediction. However, the study points out that these AI systems do not achieve practical usefulness through world modelling alone. Their abilities emerge only after a second stage of targeted fine tuning that teaches them how to respond, interpret questions, and behave in ways aligned with human goals. This two-tier process provides a potential parallel to human learning, in which the brain may maintain a general unsupervised model of the world that is continuously updated by reinforcement signals and contextual feedback. The study suggests that the brain’s higher cognitive functions may depend on this same interplay between general modelling and ongoing fine tuning.

The second key finding deals with how generative AI systems approach reasoning. When AI models generate step-by-step intermediate thoughts before producing an answer, their performance on reasoning tasks improves dramatically. The study frames this improvement through two computational perspectives. One is the information bottleneck principle, which states that effective reasoning depends on extracting the essential information needed for the final decision while filtering out irrelevant details. The other perspective comes from GFlowNets, which build complex objects step by step through probabilistic construction. This combination offers a new interpretation of human thought sequences, suggesting that the brain may also generate internal steps that compress information or build composite mental structures before arriving at a conclusion. The author argues that this area offers fertile ground for neuroscience experiments designed to study how the brain generates and evaluates internal thought chains.

The third major theme focuses on attention. In neuroscience, attention is often described as the interaction between bottom up sensory processes and top down control signals originating in higher brain regions. The study argues that this separation may be incomplete. In artificial intelligence, attention mechanisms arise from self-attention networks in which different parts of a sequence dynamically weigh one another. These systems must be trained as a whole to maintain internal consistency. This requirement for a shared representational language suggests that biological attention may also involve a strong coupling between the generation of control signals and the processing of sensory information. The author notes that attention in the brain likely shares functional similarities with the attractor dynamics observed in modern Hopfield networks, where stable activity patterns enable rapid selection, recall, and gating of information.

The study also examines scaling laws. In artificial networks, performance improves predictably as models grow larger. Yet the cost of training increases much faster. In AI, doubling the size of a network produces diminishing returns and requires far more computational time. The study argues that this pattern may help explain evolutionary constraints on brain size. If a larger biological brain requires vastly more time to learn before reaching maturity, it may impose a survival disadvantage. This connection between AI scaling limits and biological evolution offers a theoretical explanation for why human brains do not simply scale upward for greater intelligence. The author suggests that neuroscience should treat scaling constraints not as evolutionary accidents but as functional limitations shaped by the cost of learning.

The final theme concerns quantization. In AI engineering, quantization reduces a model’s memory footprint by forcing parameters to take on a limited set of discrete values rather than continuous numbers. This technique is crucial for running large models efficiently. The study notes that growing evidence suggests biological synapses also maintain a small set of stable strength states instead of continuous ranges. This creates a natural form of synaptic quantization in the brain. Because quantization in AI has been studied in detail, including the ways it affects learning accuracy and representation quality, the author argues that these insights can be transferred to neuroscience. Understanding how discrete synaptic states shape cognition could advance theories of memory stability, representational reliability, and learning efficiency in the brain.

Across all five themes, the study underscores the idea that generative AI provides working examples of mechanisms that solve complex cognitive tasks at scale. Instead of treating AI as a metaphor or high level comparison, the author urges neuroscientists to test whether similar mechanisms exist in biological systems. The gap between artificial and biological modelling approaches has widened as AI has adopted scalable architectures, explicit attention mechanisms, and interpretable methods that reveal internal model structure. The paper argues that neuroscience must engage directly with these developments to remain aligned with the computational realities of complex cognitive functions.

The research calls for a systematic exchange between AI and neuroscience. It points out that the brain and modern AI face similar challenges: representing the world, making decisions, controlling attention, managing resource constraints, and encoding information in stable formats. Despite this parallelism, the fields often approach these problems from different conceptual standpoints. The author believes that by analyzing how AI solves these challenges, neuroscientists can identify new hypotheses about how the brain may implement comparable strategies. This approach goes beyond surface analogy and moves toward detailed, falsifiable predictions.

The study also highlights the broader impact of these findings on the understanding of intelligence. If generative AI continues to reveal mechanisms that mirror or inspire biological theories, it may push cognitive science toward a more unified framework of intelligence that spans both natural and artificial systems. Such a framework would help clarify which aspects of intelligence are computationally essential and which arise from biological constraints. The author argues that this clarity could reshape long-standing debates about the nature of human reasoning, the limits of machine intelligence, and the interplay between data-driven learning and innate structure.