Adam Marblestone — AI is missing something fundamental about the brain - Dwarkesh Podcast Recap

Podcast: Dwarkesh Podcast

Published: 2025-12-30

Duration: 1 hr 50 min

Summary

In this episode, Adam Marblestone explores the fundamental differences between human cognition and artificial intelligence, particularly focusing on how the brain processes information in ways that current AI models struggle to replicate. He argues for the need to enhance neuroscience to unravel these complexities.

What Happened

Adam Marblestone opens the discussion by addressing a pressing question in the field of AI: how does the human brain achieve capabilities that far exceed those of current large language models (LLMs), despite the vast amounts of data fed into these systems? He suggests that the quest to understand the brain is arguably one of the most significant questions in science today. Marblestone emphasizes the need to empower neuroscience, both technologically and theoretically, in order to tackle this complex issue head-on.

Delving deeper, Marblestone outlines the components that define machine learning frameworks, including architecture, learning algorithms, initialization, and cost functions. He posits that the neuroscience community may have overlooked the importance of sophisticated loss functions that evolution has likely developed over time. Unlike the simplistic loss functions often favored in machine learning, Marblestone suggests that the brain's loss functions could be intricate and evolve to optimize learning at different developmental stages. This perspective invites a reevaluation of how we understand learning and prediction within the brain, hinting at a more nuanced approach that could bridge the gap between human cognition and artificial intelligence.

Throughout the conversation, Marblestone discusses the brain's cortex and its role in prediction, likening its functionality to a highly generalized prediction engine. He raises the possibility that different areas of the cortex could be specialized for predicting various inputs, hinting at a level of complexity and adaptability not currently matched by LLMs. He also touches on the evolutionary aspects of the brain, questioning how high-level desires and intentions are encoded, and how these innate responses interact with learned experiences. The discussion culminates in a contemplation of the steering subsystem of the brain, which integrates innate and learned behaviors, suggesting a sophisticated interplay that AI has yet to replicate.

Key Insights

• The brain's capabilities far exceed those of LLMs due to complex evolutionary adaptations in learning processes.
• Loss functions in machine learning are often too simplistic compared to the intricate loss functions that may exist in the human brain.
• The cortex may function as a generalized prediction engine, capable of omnidirectional inference across various inputs.
• Understanding how evolution encodes high-level desires and intentions in the brain is crucial for advancing both neuroscience and AI.