Modern large models resemble an infinitely enlarged "cortex": strong at statistical learning, pattern generalization, and compressing language and vision into representations. But according to Adam Marblestone, AI's low learning efficiency, fragile goals, and lack of intrinsic understanding of values are not primarily due to architecture being too small — they stem from missing the other, older and more critical half of the brain: the "steering" mechanisms.
In his view, human intelligence is not the victory of a single learning algorithm but the collaboration of two systems:
- One builds world models and abstracts regularities from experience (cortex-like).
- The other writes "what matters, what is dangerous, what is worth pursuing" into the learning process (largely from subcortical structures like the hypothalamus, brainstem, and basal ganglia).
This explains an unsettling contrast: an infant rapidly learns language and social rules from limited real-life input, while LLMs, after ingesting the entire internet, still display a "smart but unreliable" character in common sense, planning, stable motivation, and value alignment.
Why Children Learn Language "Like Breathing" While AI Must "Swallow the Internet"
Marblestone stresses that human infants do not learn from scratch. Evolution did not write "complete knowledge" into genes, but rather something like a value function and curriculum: preferences for faces, gaze, tone of voice, and interaction; sensitivity to curiosity, exploration, belonging, status, shame, reward, and punishment. These are not "knowledge" learned after birth, but attention mechanisms and motivational terrain that guide learning.
> Tip: "Value function" here means "which situations trigger reward/punishment and which cues the system prioritizes as learning signals." It is not an explicit formula but is implemented through genetically预设 circuits and modulatory signals.
By contrast, mainstream AI training objectives are extremely "computer-science-flavored": next-token prediction, or coarse-grained reinforcement learning with external rewards. These can shape powerful representations but lack an endogenous sense of "what matters," so the shortfall must be compensated with scale and data density — like powering a boat with a bigger engine but no compass or rudder.
Marblestone's Core Claim: The Brain's "Secret Recipe" Is in Rewards/Losses, Not Just Structure
In the Dwarkesh interview, he repeatedly decomposed the problem into four familiar machine-learning components: architecture, learning algorithm, initialization, and loss/reward function. His "personal intuition": the AI field has long over-focused on architecture and scale while underestimating the brain's highly complex, hierarchical, developmentally staged loss functions and reward signals.
He summarized the difference vividly: evolution wrote massive amounts of "Python code," lighting up different training signals for different brain regions at different developmental stages; whereas our typical loss functions are as simple as a single line of cross-entropy.
More pointedly, he argued that some LLM-style training is even "strange" from an RL perspective — models barely use explicit value functions, while the brain's value-estimation-like mechanisms (e.g., dopamine-related "reward prediction error") have long had correspondences between neuroscience and RL.
The Path Forward: Build the "Steering Wheel," Not Just a Bigger "Cortex"
Marblestone does not advocate copying the brain wholesale, nor blind faith in "explaining what every neuron is thinking." His approach: use neuroscience to fill in AI's missing key constraints and modules, pushing engineering to the point where theory can land.
1) Connectomics: Finding Evidence of "Innate Algorithms"
He emphasizes the value of "mapping the brain": not to find some mythical "Golden Gate Bridge neural circuit," but to answer deeper questions: How do cortical and subcortical systems divide labor? Which circuits generate training signals? How do these signals interface with learned abstract representations?
FlyWire offers a concrete template: years of proofreading and annotating the fruit fly whole-brain connectome, reaching ~140,000 neurons, 50+ million synapses, and 100,000+ community annotations, including curation of synaptic and neurotransmitter information, with the flagship paper published in *Nature* (2024). Such resources make "hardware-level structure" computable and publicly visible for the first time.
2) Formalized AI: Making AI Development More Like Cumulative Science
He also cares about a more "methodological" path: formalizing and making provable the key reasoning and verification mechanisms, so that system capabilities accumulate like mathematics and engineering rather than relying mainly on trial and error and empirical tuning. In the interview he cited Lean as an example, stressing the significance of "verifiable signals" for reinforcement learning and scaled collaboration.
3) Organizational Form: FROs to Fill Infrastructure Gaps
This echoes his practice at Convergent Research, which positions itself as a "mission control for frontier technology," founding time-boxed, ambitious nonprofit "startup-style research teams" (Focused Research Organizations) to fill infrastructure projects that neither academia nor industry can stably sustain. Its listed projects span neural tracing platforms, drug–target data portals, and wearable whole-brain ultrasound interfaces — all pointing to one theme: build the missing "key tools" and thicken the trunk of the scientific tree.
Controversy and Implications: Does This Replace the Scaling Route?
Marblestone does not simply reject scaling. He admits the success of large models is "surprising," only pointing out that the current paradigm differs dramatically, in intuition, from how the brain works. He is waiting for a paradigm that "looks more like the brain" — for example, more general forms of inference (he has discussed "omnidirectional inference/prediction") and a richer, more hierarchical system of values and training signals.
This also pushes the discussion toward AI safety: if capability systems are made extremely powerful but lack a robust "steering" structure, alignment remains an afterthought patch; whereas if the "steering wheel" is part of the system itself, the value and learning path may be more controllable and interpretable — at least at the level of macro-level mechanisms.