Dwarkesh Podcast · Dwarkesh Patel

Adam Marblestone — AI is missing something fundamental about the brain

December 30, 2025·1 hr 50 min·5 clips
The steering subsystem may need thousands of bespoke cell types to wire spider flinches, social status, and navigation.
1. Dwarkesh Podcast's Adam Marblestone episode asks why brains learn from far less data than LLMs. 2. Adam Marblestone, the guest, treats neuroscience as the key to that gap and calls it a quadrillion-dollar question. 3. The episode asks whether brain learning depends more on loss functions and curricula than on architecture alone. 4. Marblestone breaks the system into architecture, learning algorithm, initialization, and cost functions. 5. He says machine learning usually prefers simple objectives like next-token prediction and cross-entropy. 6. He argues evolution may have encoded many cost functions, with different brain areas turning on at different developmental stages. 7. He describes the cortex as a six-layer sheet that may approximate backprop, but nobody knows the exact task it is optimized for. 8. One proposal is that cortex works as an omnidirectional prediction engine that can infer any subset of variables from any other subset. 9. He compares that to LLMs, which natively predict the next token from prior context even when they learn tricks like filling blanks. 10. He connects the idea to Yann LeCun's energy-based models, where sampling the joint distribution lets a model infer from clamped variables. 11. Marblestone says higher cortex may sit on top of a lizard brain that predicts reflexes, heart rate, and instinctive behavior. 12. He leans on Stephen Burns's theory that a learning subsystem monitors a steering subsystem made of hypothalamus, brainstem, amygdala, and superior colliculus. 13. The spider example shows the distinction, because a skittering insect can trigger an innate flinch and the word spider can later trigger the same response. 14. He says abstract concepts such as spider books, spider rooms, or the word spider can generalize into that learned response. 15. He uses Jan LeCun as a social example, saying embarrassment about misdescribing energy-based models could activate shame and avoidance circuits. 16. That leads to the claim that evolution must wire learned world-model features into innate reward functions without knowing Jan LeCun or podcasts in advance. 17. Later, he argues that the steering system likely needs many more bespoke cell types than cortex because genes must prewire status, friendship, flinching, tasting salt, and navigation. 18. He cites single-cell atlas work from Fei Chen and Evan McKosco as evidence that subcortical regions contain a dense spread of specialized cells. 19. The episode also covers Andy Jones's AlphaZero result, Gemini-assisted refactoring, and a 16-agent experiment that split fixed compute across multiple models. 20. Marblestone closes by linking amortized inference, test-time compute, and the human genome's limited coding space to the brain's compact wiring story.

As heard by us

A neuroscience case for why current models still feel far from the brain.

This conversation treats the brain as the missing reference point for modern AI and keeps returning to the same pressure point: LLMs may keep absorbing more data, yet they still sit a long way from human capability.

Read the full review in PlayNext →

Why you'd press play

Why LLMs still miss something the brain seems to do, and why that matters.

Read the full recommendation in PlayNext →
Listen to the show on