Adam Marblestone — AI is missing something fundamental about the brain

Dwarkesh frames the whole conversation as a single question he has been carrying through every interview with an AI researcher: how does the brain do it? The models are fed vastly more data than a human ever sees and still hold only a small fraction of human capability. Something is being left out, and the standard scaling story does not name what.

Adam Marblestone upgrades the stakes — the quadrillion-dollar question, plausibly the most important question in science — and then declines to answer it. He does not know, and he does not think the answer will come merely from a lot of smart people thinking harder. His meta-level position is institutional: the field of neuroscience has to be empowered technologically before a question of this kind can be cracked at all.

Marblestone's move is to import the machine-learning decomposition and ask which component is underweighted. A learning system has an architecture; hyperparameters — how many layers, what properties; a learning algorithm — backpropagation, gradient descent, or something else; an initialisation of the weights; and cost functions, which is to say what the system is trained to do, what the reward signal is, what the loss and supervision signals are.

His hunch is that the field has neglected the last of these. Machine learning prefers mathematically simple losses — predict the next token, cross-entropy — the kind of loss a computer scientist finds natural. Evolution, by contrast, had millions of years of feedback about what succeeded and what killed, and could therefore encode that knowledge as a curriculum rather than a scalar.

So the conjecture is that the brain runs many different loss functions, for different regions, switched on at different stages of development — in effect a great deal of Python generating a specific curriculum for what each part of the brain needs to learn. Which raises the question he wants to keep open: where do the brain's loss functions come from, and can a different loss function make learning dramatically more efficient?

Dwarkesh puts the standard claim on the table: the cortex is supposed to contain the universal human learning algorithm. Marblestone notes how little is actually known, starting with a confusion of vocabularies. Cortex has six physical layers of tissue — layers in a different sense from the layers of a neural network — and those areas then connect to one another, which is closer to what a network's layers mean.

The popular theoretical move is to ask how a cortical area approximates backpropagation. But that question is incomplete until you say what the network is being asked to do: backprop on next-token prediction, on classifying images, on what? Nobody knows. One possibility is that it is an extremely general prediction engine, and any given area is trying to predict any subset of its variables from any other subset — omnidirectional inference.

The contrast with a language model is sharp. An LLM sees everything in its context window and computes one particular conditional probability: given the last few thousand tokens, what comes next. It can learn to fill in the middle of a sentence, but that is an emergent capability riding on top of a native objective that only ever points forward.

What if cortex is natively built so that any area can predict any pattern in any subset of its inputs given any other subset? That is closer to probabilistic AI, and Marblestone flags the resemblance to Yann LeCun's energy-based models: a joint distribution over all the variables, with inference as clamping some subset and sampling the rest. Clamp a different subset and you get a different conditional. That is what omnidirectional means.

This also explains what such an area would bother predicting. Some association areas might predict vision from audition; others might predict what the innate parts of the brain are about to do. The whole cortex is riding on top of a lizard brain and a lizard body, and that thing is worth predicting too — not just whether I see this or that, but whether this muscle is about to tense, whether I am about to laugh, whether my heart rate is about to rise, whether an instinctive behaviour is about to fire.

Dwarkesh tests the idea with a concrete case. Someone tells you there is a spider on your back, and that sentence recruits the same innate machinery that would fire if you saw a spider in front of you. You learned the association, even though nothing ever supervised the sentence.

Marblestone connects this to Ilya Sutskever's remark that there is no good theory of how evolution encodes high-level desires or intentions, and to Steve Byrnes's work he has recently become obsessed with. The example he offers is embarrassment: imagining LeCun listening and judging his description of energy-based models activates an innate shame response, and the activation is useful, because otherwise he might be killed by LeCun's marauding army.

The problem is that evolution never saw LeCun, never heard of energy-based models, never knew what a scientist or a podcast was. Yet the brain has to encode the desire not to alienate important people in the tribe robustly, without knowing in advance what the learning cortex will put into its world model. Whatever neurons represent LeCun being upset with him must end up wired to the shame response.

And the wiring has to keep working after the fact. Seeking status, learning from the knowledgeable, exchanging skills with friends but not enemies — all of it depends on connecting learned features of the world to innate reward functions, and then learning further from the result. Next time he will not try to annoy LeCun, because the correction feeds back into learning.

Byrnes's proposal, as Marblestone presents it, is that the learning parts of the brain — cortex, parts of the amygdala — are modelling the Steering Subsystem: the innately programmed part that holds the primitive responses and the series of reward, cost and bootstrapping functions.

The trick is that the Steering Subsystem has its own senses. We think of vision as cortical, but there is a subcortical visual system, the superior colliculus, with innate detectors for faces and threats. It also has innate heuristics for social status and for friendliness. So there is a visual system carrying innate priors, and a set of innate responses, and a learning system trying to predict both.

Which answers the question of how the right neurons get wired up. Which cortical neurons matter for social status or for friendship? The ones that predict the innate heuristics for status and friendship. You train a predictor and ask which neurons are part of it; those are the neurons that have been connected to the reward function, without anyone having to specify in advance what a friend is.

Dwarkesh presses the generalisation problem. The cortex gets labels for the literal spider — here is a picture, this is bad, be afraid — but never gets supervision on the sentence about the spider on your back. How does that transfer?

The answer is that the Learning Subsystem is a genuinely powerful generalising algorithm, and the innate side supplies only the seed. Something small, dark, high-contrast and moving fast toward the body that was not there a moment ago might be an insect; you flinch. Some hypothalamic neurons are the I-just-flinched neurons. The negative contribution to reward has no generalisation in it by itself — but generalisation downstream of the reward function is possible, and the amygdala can learn to predict the flinch hundreds of milliseconds early.

So for every variable the hypothalamus and brainstem care about — about to flinch, talking to a friend, should laugh now, is the friend high status, about to taste salt — there is a predictor, and the predictor's input is whatever the world model offers. That input is where the generalisation comes from. The word "spider" can be activated by a book about spiders, by a room where spiders live, by a sentence someone says; the predictor learns to fire on the concept rather than the retinal image.

Dwarkesh notes the amount of heebie-jeebies the conversation is producing in the audience, and Marblestone observes that he is now activating their Steering Subsystem: the spider-hypothalamus subgroup is firing on the strength of abstract concepts in a podcast.

Marblestone closes the loop on the spider. The reason the sentence works and the reflex does not is that the cortex predicts from abstract, integrated variables, while the Steering Subsystem can only use what the superior colliculus and a handful of other sensors emit. The cortex generalises because it is doing inference over a world model; the innate side is stuck with its priors.