ExpertsFMAdam Marblestone — AI is missing something fundamental about the brain
Dwarkesh Podcast · December 30, 2025
5 Selected Lessons:
reasoning_pattern · 2 heuristic · 1 case · 1 framework · 1
The 'Gap Map' method for finding scalable science infrastructure bottlenecks
Source
Dwarkesh Podcast — Adam Marblestone — AI is missing something fundamental about the brain (December 30, 2025) · at 1:44:58 ▸
When this applies
when trying to find the highest-leverage but currently-unfunded investments in any professional field or ecosystem
The takeaway
The highest-leverage unmet needs in a field are often not research ideas at all but missing shared infrastructure that no individual lab/startup-sized unit can build — identify these by explicitly asking practitioners to distinguish 'next project' from 'thing I'm structurally blocked on.'
Why it works
Distinguish scientists describing incremental next steps ('here's what my grad student will do next') from scientists describing a structural gap ('I need an organized engineering team to build something no single lab or loose collaboration can ever produce') — the latter category identifies 'mini Hubble Space Telescope' style infrastructure investments that, once built, lift an entire field's capability without being a discovery in themselves.
One level up
Systematically interview practitioners across a domain to classify their unmet needs into 'incremental' vs 'structural/infrastructural', because structural gaps are undersupplied by default market/academic incentives and represent the best return on concentrated, organized investment.
Where it stops holding
This stops working when a field's bottleneck truly is incremental/creative work (e.g., matters of low-scale serendipity and individual insight) rather than shared infrastructure, in which case funding a Hubble-style project wastes resources.
Detail
procedure:
- Interview many scientists across a field and let them freely describe their biggest unmet need
- Classify needs into 'incremental next research step' versus 'structural infrastructure gap no individual lab can close'
- Catalogue the structural gaps into a 'gap map' (a few hundred items, each roughly Series-A-startup-sized, i.e. low-to-mid billions aggregate, not trillions)
- Prioritize funding (FRO, philanthropy, government) toward these structural gaps rather than incremental research funding
benchmarks: {"count":"a few hundred fundamental capability gaps identified","cost_per_gap":"each roughly a 'deep tech startup' / Series-A scale, i.e. a few billion dollars at most, cumulatively lower than a trillion dollars"}
rationale: Analogous to how the Hubble Space Telescope wasn't itself an astronomy discovery but enabled discoveries across the whole field; math proving infrastructure (LEAN) was an unexpected but real example of this pattern outside biology.
In their words
if I can build that Hubble Space Telescope, then I will unblock all the other researchers in my field or some path of technological progress in the way that the Hubble Space Telescope lifted the boats, improved the life of every astronomer, but wasn't really an astronomy discovery in itself.
Judging a theory of a system by whether models built on it predict better than alternatives
Source
Dwarkesh Podcast — Adam Marblestone — AI is missing something fundamental about the brain (December 30, 2025) · at 59:51 ▸ · at 1:02:02 ▸
When this applies
when evaluating whether a borrowed theoretical framework genuinely explains a target system, versus just being a convenient analogy
The takeaway
Test a borrowed conceptual framework by its predictive power on the target system's actual measured data, not by its intuitive plausibility.
Why it works
If models built using a candidate framework (e.g. CNNs for vision) outperform alternative computational models at predicting the target system's actual measured behavior (e.g. brain activations via representational similarity / 'brain score'), that's empirical evidence the framework captures something real about the mechanism, not just a loose metaphor.
One level up
Confidence in any theoretical framework should track quantitative predictive success on held-out measurements of the real system, not the elegance or popularity of the framework.
Where it stops holding
This stops working when there's no analogous critic/dissenting framework or no measured ground truth to score predictions against, in which case you can't distinguish a genuinely true framework from a coincidentally useful one — as the guest and Buzsaki's critique both note.
Detail
approach: Score competing computational models (e.g., CNN vs other computational neuroscience models) on how well they predict actual brain measurements, and trust the framework that wins
signals:
- representational similarity scores ('brain score')
- predictive accuracy on held-out images/activations
- surprising convergent discovery (e.g. TD-learning matching dopamine signals, independently invented)
rationale: Convergent, independently-derived matches between AI theory and brain data (e.g. dopamine/TD learning) are stronger evidence than post-hoc storytelling.
In their words
the AI model is actually there seems to be some relevance there in terms of neurosciences don't necessarily have something better than that.
Describe a learned system by its generative process, not by decoding its internals
Source
Dwarkesh Podcast — Adam Marblestone — AI is missing something fundamental about the brain (December 30, 2025) · at 1:04:42 ▸
When this applies
when trying to claim you 'understand' any sufficiently complex learned/evolved system, be clear about what level of understanding is actually achievable and useful
The takeaway
When a system's learned internal content is too complex to ever fully interpret, shift the target of 'understanding' from decoding the artifact to characterizing the generative process (architecture + objective + learning rule + data) that produced it.
Why it works
A neural network trained on an arbitrarily complex task (e.g. predicting digits of pi, or stock prices) will internally implement whatever complex computation is needed — interpretability may never fully 'decode' that computation — but you can still fully and usefully describe the network at the level of architecture, learning rule, initialization, and training data/loss function that produced it.
One level up
Pick the right level of abstraction for 'understanding' based on what's actually tractable to recover, not based on what feels most complete.
Where it stops holding
This stops working when you actually need to know the specific learned content for a task (e.g., safety-critical interpretability of a specific circuit), where process-level description alone is insufficient.
Detail
approach: Target understanding of the cortex's architecture, learning rule, hyperparameters, and training data/objective rather than trying to decode what any specific neuron computes
signals:
- a system trained on arbitrarily/computationally-complete complex tasks
- interpretability of individual circuits proving intractable
- ability to independently characterize the generative process (architecture/loss/data) that built the system
rationale: You built (or can characterize) the generative process even when you can't decode the artifact it produced.
In their words
I wanna describe the brain in the same way, and I think that this framework that I've been kind of laying out is like, we need to understand the cortex and how it embodies a learning algorithm. I don't need to understand how it computes Golden Gate Bridge.
Loss function complexity may matter more than architecture for sample efficiency
Source
Dwarkesh Podcast — Adam Marblestone — AI is missing something fundamental about the brain (December 30, 2025) · at 0:00 ▸ · at 26:14 ▸
When this applies
when a system seems to underperform despite scale, check whether the training signal/reward structure — not just the architecture or data volume — is impoverished relative to a more successful comparable system
The takeaway
The design of the objective/reward function, not just the model or data pipeline, can be the primary lever for learning efficiency.
Why it works
Simple mathematically clean loss functions (next-token prediction) provide a weak, undifferentiated signal; evolution instead may have built many specific, staged, curriculum-like loss functions per brain region, encoding hard-won knowledge about what's useful to learn when.
One level up
When comparing two systems that differ in performance, decompose the comparison into component axes (architecture, learning rule, initialization, objective) rather than assuming the most visible difference is causal.
Where it stops holding
This stops working when the loss function actually is sufficient and the bottleneck is genuinely architecture, data, or compute instead.
Detail
rule: Suspect the objective/reward function as the main source of a learning-efficiency gap before blaming architecture or data.
threshold: n/a — qualitative hunch
rationale: Evolution encodes successful/unsuccessful learning trajectories into many specialized cost functions, effectively compressing curriculum design into the genome.
In their words
I think evolution may have built a lot of complexity into the loss functions. Actually, many different loss functions for different areas turned on at different stages of development.
Social learning as the driver of hominid brain expansion
Source
Dwarkesh Podcast — Adam Marblestone — AI is missing something fundamental about the brain (December 30, 2025) · at 36:28 ▸ · at 39:55 ▸
When this applies
when trying to explain a rapid capability jump in a system's history — look for a change that multiplied the return on an existing capacity rather than a wholly new capacity appearing
The takeaway
A sudden capability leap is often better explained by a change in incentive/leverage on an existing mechanism than by the sudden invention of a wholly new mechanism.
Why it works
If social learning (e.g. learning to make a spear from an elder instead of independently discovering it) dramatically raised the sample-efficiency/payoff of having a larger cortex, this increases the evolutionary incentive to grow cortex size, even using a 'cheap' genetic change (replicating an existing architecture) rather than inventing a new one.
One level up
When a system suddenly scales rapidly, ask whether the environment simply began rewarding more of what the system already had, rather than assuming new machinery appeared.
Where it stops holding
This stops working when there is in fact clear evidence of qualitatively new architecture (not just scaling), which the guest notes is still debated (e.g., neuron-scaling differences between primates and rodents).
Detail
situation: Puzzle: hominid brain size grew unusually fast in evolutionary terms
decision: Hypothesize that social learning increased the value/incentive for a bigger cortex rather than the cortex's fundamental algorithm changing
outcome: Guest leans toward this being more about increased incentive (reward function/instinct wiring for social learning) than new cortical 'potential', though acknowledges open debate (e.g. primate vs rodent neuron scaling differences)
lesson: A capability explosion may reflect increased payoff for existing machinery rather than a new algorithm
In their words
I mean, I would lean somewhat toward the latter. I mean, I think a mouse has a lot of similarity ity in in terms terms of of cortex cortex as as a a human.