Brain's Complex Cost Functions Drive Efficient Learning Beyond AI

Dwarkesh Podcast · December 30, 2025 · Listen to Original Episode →

Original Title:

TL;DR

The brain's efficiency in learning from limited data stems from complex, evolutionarily encoded loss functions, not just architectural complexity, suggesting AI needs to prioritize sophisticated cost functions over simple next-token prediction.
The brain's cortex may function as an "omnidirectional inference" engine, capable of predicting any variable subset from any other, unlike LLMs' unidirectional next-token prediction, enabling more flexible world modeling.
Evolution encodes high-level desires and intentions through innate "steering subsystems" and reward functions that model learned world features, allowing robust wiring to shame or social status responses without pre-specifying all future scenarios.
The genome's limited size suggests evolution prioritized specifying innate behaviors and bootstrapping cost functions over detailed learned information, with the "steering subsystem" comprising more diverse cell types than the "learning subsystem" (cortex).
Continual learning in the brain likely involves a combination of architectural features like the hippocampus for memory consolidation and potentially multiple timescales of synaptic plasticity or novel learning rules, rather than solely relying on backpropagation.
The human genome's efficiency in building complex intelligence suggests evolution leveraged pre-existing structures and sophisticated reward functions, rather than encoding vast amounts of explicit knowledge, enabling rapid expansion of cognitive capabilities.
The development of formal verification languages like Lean, which mechanically check mathematical proofs, offers a powerful RL signal for automating mathematical discovery and creating provably secure software, bridging symbolic and learned AI approaches.

Deep Dive

The core argument is that artificial intelligence, despite its rapid progress, fundamentally misunderstands how the human brain learns and operates, particularly concerning its remarkably efficient learning from limited data and its sophisticated reward functions. This gap highlights a critical need to integrate insights from neuroscience into AI development, moving beyond architectural and learning algorithm improvements to focus on the "secret sauce" of biological intelligence: its highly specific and complex cost functions, which evolution has meticulously encoded.

The implications of this perspective are far-reaching. Firstly, it suggests that current AI, primarily focused on next-token prediction or simple loss functions, is inherently limited. The brain, by contrast, appears to employ a system of "amortized inference" and a robust "steering subsystem" that encodes abstract desires and intentions. This steering subsystem, comprised of innate reflexes and reward functions, acts as a crucial curriculum for the learning subsystem (like the cortex), guiding it to associate learned features of the world with innate drives. This explains how the brain can learn complex behaviors and abstract desires, like avoiding social embarrassment, without explicit genetic encoding for every specific scenario, such as "pissing off Yann LeCun."

Secondly, this framework reframes the discussion around AI capabilities and limitations. The efficiency of biological learning, where a small genome encodes vast complexity, suggests that evolution has prioritized sophisticated reward functions and innate heuristics over massive datasets or brute-force computation. This implies that future AI might achieve greater sample efficiency and generality not just through larger models or more data, but by better emulating the brain's reward function encoding, potentially through a "molecularly annotated connectome" that maps not just connections but also the molecular properties at synapses.

Finally, the discussion underscores the ongoing challenge of bridging AI and neuroscience. While AI models have demonstrated surprising capabilities, they often lack the nuanced understanding and adaptability of biological systems. The brain's apparent ability to perform "omnidirectional inference" and its complex interplay between learned models and innate drives suggest that current AI paradigms may be missing fundamental principles. The research into areas like formal verification in mathematics and the development of "molecularly annotated connectomes" are steps toward understanding these deeper principles, but significant gaps remain in our ability to fully reverse-engineer or replicate the brain's integrated intelligence.

Action Items

Audit authentication flow: Check for three vulnerability classes (SQL injection, XSS, CSRF) across 10 endpoints.
Create runbook template: Define 5 required sections (setup, common failures, rollback, monitoring) to prevent knowledge silos.
Implement mutation testing: Target 3 core modules to identify untested edge cases beyond coverage metrics.
Profile build pipeline: Identify 5 slowest steps and establish 10-minute CI target to maintain fast feedback.

Key Quotes

"my my overall like meta level take is that we have to empower the field of neuroscience to just make neuroscience a a more powerful uh field technologically and otherwise to actually be able to crack a question like this but maybe the the way that we would think about this now with like modern ai neural nets deep learning is that there are sort of these these certain key components of that there's the architecture um there's maybe hyper parameters of the architecture how many layers do you have or sort of properties of that architecture there is the learning algorithm itself how do you train it you know back prop gradient descent um is it something else there is how is it initialized okay so if we take the learning part of the system it still may have some initialization of of the weights um and then there are also cost functions there's like what is it being trained to do what's the reward signal what are the loss functions supervision signals my personal hunch within that framework is that the the field has neglected uh the role of the very specific loss functions very specific cost functions"

Adam Marblestone suggests that advancing neuroscience technologically is crucial for answering fundamental questions about intelligence. He posits that within the AI framework of architecture, learning algorithms, and cost functions, the field has overlooked the significance of highly specific loss functions, which he believes evolution may have intricately developed.

"i think a lot of the things i'm saying by the way are extremely similar to like what yann lecun would say um he's really interested in these energy based models um and something like that is like the joint distribution of all the variables what is the likelihood or unlikelihood of just any combination of variables and if i if i clamp some of them i say well definitely these variables are in these states then i can compute with probabilistic sampling for example i can compute okay conditioned on these being set in this state what are and these could be any arbitrary subset of variables in the model can i predict what any other subset is going to do and sample from any other subset given clamping this subset and then i could choose a totally different subset and sample from that subset um so it's omnidirectional inference"

Marblestone notes that his ideas align with Yann LeCun's interest in energy-based models, which represent the joint probability distribution of variables. He explains that this approach allows for "omnidirectional inference," where by setting certain variables, one can predict and sample from any other subset of variables within the model.

"and so somehow the brain has to encode this desire to you know uh not not piss off really important you know people in the tribe or something like this um in a very robust way without knowing in advance all the things that the the learning subsystem okay of the brain the part that is learning cortex and other parts the cortex is going to learn this world model that's going to include things like yann lecun's and podcasts and uh evolution has to make sure that that those neurons whatever the yann lecun being upset with me neurons get properly wired up to the shame response or this part of the reward function"

Marblestone highlights a profound question: how the brain encodes complex desires, like avoiding offense to important individuals, without prior knowledge of specific scenarios. He suggests that evolution must ensure that neurons associated with learned concepts (like "Yann LeCun" or "podcasts") are correctly linked to innate responses such as shame or reward functions.

"so the basic idea um that steve burns is proposing is that well part of the cortex uh or or other areas like the amygdala that learn um what they're doing is they're modeling the steering subsystem steering subsystem is the part with these more innate innately programmed responses and the innate programming of these series of reward functions cost functions bootstrapping uh functions that exist so there are parts of the amygdala for example that are able to monitor what what those parts do and predict what those parts do"

Marblestone explains Steve Burns' theory that parts of the cortex and amygdala model the "steering subsystem," which comprises innate, programmed responses and reward functions. According to Burns' proposal, these areas learn by predicting the behavior of the steering subsystem.

"so i think that there's a possibility that essentially all other scientific research that is being done is like not it is somehow obviated but i don't put a huge amount of probability on that i think my timelines might be more in the like yeah 10 year ish range and if that's the case i mean i think there there is probably a difference of a future world where we have connectomes on hard drives and we have understanding of steering subsystem architecture we've compared the the you know even the most basic properties of what are the reward functions cost function architecture etcetera of you know mouse versus a shrew versus a small primate etcetera is this practical in 10 years uh i think it has to be a really big push how much funding how does it compare to where we are now it's like billion low billions dollars scale funding in a very concerted way i would say"

Marblestone speculates that while current scientific research might become obsolete, his timelines suggest a future where connectomes and an understanding of the steering subsystem's architecture are available. He believes achieving this will require significant, concerted funding on the scale of billions of dollars.

"i did not realize that like math proving infrastructure like was a thing and so -- and that was kind of like emergent from trying to do this so i'm i'm looking forward to seeing other other things where it's like not actually this like hard intellectual problem to solve it -- it's maybe the kind of the slightly the equivalent of ai researchers just needed gpus or something like that and focus and and really good pytorch code to like start doing this like what is the full diversity of fields in which that exists"

Marblestone expresses surprise at discovering the need for "math proving infrastructure," which emerged from his efforts to identify foundational capabilities. He likens this to AI researchers needing GPUs and robust code to advance their field, suggesting that such infrastructure is crucial across various disciplines.

Resources

External Resources

Books

"The Brief History of Intelligence" - Mentioned as a good book with AI research insights.

Articles & Papers

"What does it mean to understand a neural network" (Conrad Harding and Tim Liljencrantz) - Discussed in relation to understanding neural networks and their computational capabilities.
"The Brain from Inside Out" by Yuri Buzsaki - Referenced as a book presenting a neuroscientist's perspective that challenges AI concepts and advocates for new vocabulary derived from the brain itself.
"Longitudinal Science" (Blog) - Mentioned as a place to find more of the guest's work.

People

Adam Marblestone - Guest on the podcast, discussing AI, neuroscience, and the brain.
Yann LeCun - Mentioned in relation to energy-based models and AI.
Ilya Sutskever - Referenced for his views on reinforcement learning in AI.
Steve Byrnes - Discussed for his theories on the brain's learning and steering subsystems, and how evolution encodes desires and intentions.
George Church - Mentioned as the guest's PhD advisor and for his insights on the human genome project and cost reduction in sequencing.
Terry Tao - Referenced for his work in mathematics and formal proofs.
Yuri Buzsaki - Neuroscientist who wrote "The Brain from Inside Out."
William - Mentioned in relation to the study of neurons in the 1950s.
Peter Dayan - Referenced for his work on dopamine and reward prediction error signals.
Sutton - Mentioned in relation to temporal difference learning.
Suzanne Herculano - Her work on neuron scaling in primate brains is referenced.
Baroness - Mentioned in relation to the scaling of brain regions for audio and vision.
Doris So - Her work is the basis for a neuroscience project launched by the organization "stare out of which."
Joël Dehaye - Mentioned for work on modeling early visual cortex.
Andres Tolias - Referenced for work on neural network surrogate models of brain activity.
Goran - Mentioned for a blog post on behavior cloning.
Max Hodak - Giving talks on neuroscience and neurotechnology.
William - Mentioned in relation to the study of neurons in the 1950s.
David Dodd - Program director at ARIA (UK) involved in designing safeguarded AI.

Organizations & Institutions

Convergent Research - Mentioned as a place to find more of the guest's work.
E11 Bio - Mentioned as a primary entity focused on connectomics technology.
Welcome Trust - Referenced for a report on the cost of mapping a mouse brain connectome.
National Science Foundation (NSF) - Mentioned for a call for tech labs related to connectomics.
Microsoft - Mentioned as a place where Lean was developed.
Focus Research Organizations (FROs) - Discussed as a model for channeling philanthropic support and ensuring open-source, public benefit initiatives.
ARIA (UK) - Mentioned as an organization where David Dodd is a program director.
DeepMind - Referenced for their work on temporal difference learning and AI.
Google - Mentioned in relation to Gemini 3 Pro and the Gemini app.
Pro Football Focus (PFF) - Mentioned as a data source.
National Human Genome Research Institute (NHGRI) - Referenced for structuring funding to lower genome sequencing costs.

Tools & Software

Gemini 3 Pro - Used to help brainstorm and refine an experiment.
Gemini App - Used for brainstorming and refining an experiment.
Antigravity - Google's agent-first IDE where code was opened and refactored.
Alpha Zero - A clean implementation of which was forked and used for an experiment.
Lean - A formal mathematical language used for proving theorems.

Websites & Online Resources

adammarblestone.org - The guest's website, currently down.
gemini.google.com - Website to try Gemini.
labelbox.com - Website for Labelbox.

Other Resources

AI (Artificial Intelligence) - A broad topic discussed throughout the episode.
LLMs (Large Language Models) - Discussed in relation to their capabilities and limitations compared to the human brain.
Neuroscience - A central theme of the discussion.
Deep Learning - Mentioned as a component of modern AI.
Backpropagation (Backprop) - A training algorithm discussed in the context of AI and the brain.
Gradient Descent - A training algorithm mentioned.
Loss Functions - Discussed as a potentially neglected component in AI training.
Cost Functions - Discussed as a potentially neglected component in AI training.
Supervision Signals - Mentioned in relation to training AI.
Cross-entropy - A type of loss function used in machine learning.
Neural Nets - Discussed as a component of modern AI.
Cortex - Discussed in relation to its structure and function.
Probabilistic AI - Mentioned as a related concept.
Energy-Based Models - Discussed in relation to AI and the brain.
Joint Distribution - A concept related to energy-based models.
Probabilistic Sampling - A method used in probabilistic AI.
Omnidirectional Inference - A concept related to the brain's predictive capabilities.
Steering Subsystem - A proposed part of the brain responsible for innate responses and reward functions.
Lizard Brain - A colloquial term for more primitive brain structures.
Higher-Level Desires/Intentions - Discussed in relation to how evolution encodes them.
Embarrassment, Shame, Innate Reflexes - Examples of innate responses.
Social Status, Friendliness - Concepts related to social instincts.
Superior Colliculus - A subcortical visual system mentioned.
Hypothalamus, Brainstem - Brain regions mentioned as part of the steering subsystem.
Amygdala - Brain region mentioned in relation to predicting steering subsystem responses.
Reward Function - Discussed in relation to AI and brain function.
Generalization - A key capability of learning systems.
Downstream of the Reward Function - A concept proposed by Steve Byrnes.
Multimodal Foundation Model - A type of AI model.
Inductive Biases - Architectural assumptions that shape representations.
V1 (Primary Visual Cortex) - Mentioned in relation to modeling early visual cortex.
ConvNet (Convolutional Neural Network) - A type of neural network.
Retina - Mentioned for its role in motion detection.
Doris So's Work - Basis for a neuroscience project.
Objects are bounded by surfaces, surfaces have shapes and relationships - Assumptions built into vision systems.
Test Time Compute vs. Training Compute - Concepts discussed in relation to AI efficiency.
Multi-Agent Scaling - An idea for improving AI performance.
Self-Play - A training method for AI agents.
Co-evolutionary League Training - A training method for AI agents.
Alpha Zero Agents - AI agents used in an experiment.
Amortized Inference - A concept related to efficient inference in AI.
Bayesian Inference - A statistical method for updating beliefs.
Monte Carlo Methods - Techniques used in Bayesian inference.
Boltzmann Machines - An early type of neural network.
Probabilistic Programming - A programming paradigm for probabilistic models.
Perception - Discussed in relation to Bayesian inference.
Feed Forward - A type of neural network processing.
Digital Minds - Discussed in relation to their copyability and trade-offs.
Probabilistic AI People - Group of researchers with specific views on inference.
Stochastic Neurons - Neurons that operate probabilistically.
Genome - Discussed in relation to the amount of information encoded.
Python Code - Used as an analogy for encoding complex functions.
Spider Reflex - An example of an innate response.
Thought Assessors - A term used by Steve Byrnes.
Cellular Atlases - Used to map cell types in the brain.
Brain Initiative - A neuroscience funding program.
RNA Sequencing - A technology used to identify cell types.
Steering Subsystem Cell Types - Discussed in relation to their diversity.
Cortical Cell Types - Discussed in relation to their diversity.
Learning Subsystem - A proposed part of the brain responsible for learning.
Innately Wired Circuits - Circuits formed through genetic programming.
Plasticity of Synapses - Changes in synaptic strength.
Receptors and Proteins - Molecules involved in cell communication.
Genetic Wiring - Wiring determined by genes.
Subcortical Visual System - Visual systems outside the cortex.
Superior Colliculus - A part of the subcortical visual system.
Head Direction Cells - Neurons that indicate an animal's orientation.
Fly Brain - Used as an example of innate circuits.
Hominid Brain Size Explosion - Discussed in relation to social learning.
Social Learning - Learning from others.
Sample Efficiency - How effectively learning algorithms use data.
Spear Making - An example of learned skill.
Cortex Scalability - How the cortex scales with brain size.
Broca's Area, Wernicke's Area - Brain regions associated with language.
Hippocampus - Brain region involved in memory.
Prefrontal Cortex - Brain region involved in executive functions.
Language Acquisition - The process of learning language.
Auditory and Memory Regions - Brain areas involved in language processing.
Macro Wiring - Large-scale neural connections.
Temporal Lobe - Brain region involved in auditory processing.
Auditory Cortex - Brain region for processing sound.
Mouse Brain - Used as a model for neuroscience research.
Zebrafish - Used as a model for neuroscience research.
Birds - Mentioned for their cortex-like structures.
Reptiles, Mammals - Groups of animals discussed in evolutionary context.
Associative Learning Centers - Brain regions involved in learning associations.
Fly Mushroom Body - A brain structure in flies involved in learning.
Dopamine Signal - A neurotransmitter signal.
Sensory Information - Data from the senses.
Food vs. Harm - Basic outcomes in learning.
Baron Milleridge's Blog Post - Mentioned for ideas on brain region scaling.
Odor - A sense discussed in relation to