Adam Marblestone – AI is missing something fundamental about the brain

Episode Overview

• The episode explores the profound gap between the brain's data efficiency and the capabilities of modern AI, positing that understanding this difference is one of the most important questions in science.
• It introduces a framework for understanding the brain as two primary components: a flexible, general-purpose "learning subsystem" (like the cortex) and a hardcoded, evolutionarily-derived "steering subsystem" that provides innate rewards and motivations.
• The discussion contrasts the brain's biological hardware with digital computers, weighing the brain's incredible energy efficiency against the transparency and copyability of digital systems.
• It advocates for a new era of large-scale, technologically-driven neuroscience, particularly in connectomics, framing it as a critical and cost-effective investment for unlocking the next generation of artificial intelligence.

Key Concepts

• Complex Loss Functions: The central hypothesis that the brain's efficiency stems not from a single universal learning algorithm, but from a complex, evolutionarily-encoded curriculum of many specific loss and reward functions that guide development.
• Steering vs. Learning Subsystems: A model of the brain composed of two parts: the "steering subsystem" (e.g., brainstem, amygdala) containing innate, hardcoded reward functions and instincts, and the "learning subsystem" (e.g., cortex) which builds a flexible world model to predict and satisfy the steering system.
• Omnidirectional Inference: The idea that the cortex operates as a general prediction engine capable of inferring any subset of variables from any other subset, not just unidirectionally predicting the future.
• Amortized Inference: The concept that current neural networks perform a computationally cheap "shortcut" to inference by directly mapping inputs to outputs, a process that is "amortized" into the model's weights during training, as opposed to a more rigorous but slow Bayesian sampling process.
• The Genomic Bottleneck: The puzzle of how the relatively small amount of information in the human genome can specify an organ as complex as the brain. The proposed solution is that evolution encodes the learning process (algorithms and rewards), not the final learned state.
• Hardware and Algorithm Co-Design: The idea that the brain's learning algorithms are deeply intertwined with its biological hardware, leveraging features like the co-location of memory and compute for extreme energy efficiency. This raises the question of whether the brain's cellular complexity is fundamental to its algorithms or merely implementation "kludge."
• Molecularly Annotated Connectomics: The next frontier in brain mapping, which aims to create not just a wiring diagram of neurons, but a detailed map that includes the specific molecules present at each synapse, providing a much richer dataset for reverse-engineering its function.
• Behavior Cloning the Brain: A proposed AI training methodology that uses neural activity from the brain as a direct training signal, with the goal of forcing the AI to develop more robust and human-like internal representations, rather than just mimicking external behavior.
• Focused Research Organizations (FROs): A proposed funding and organizational model designed to tackle large-scale, infrastructural science projects ("gap maps") that are too big for academic labs but not yet commercially viable, such as building the tools for large-scale connectomics.

Quotes

• At 0:15 - "Yeah, I mean this might be the quadrillion-dollar question... you could make an argument this is the most important, you know, question in science." - The guest, Adam Marblestone, emphasizes the profound significance of understanding how the brain learns.
• At 1:32 - "My personal hunch... is that the field has neglected the role of the very specific loss functions... machine learning tends to like mathematically simple loss functions, right? Predict the next token." - Marblestone contrasts the simple objectives of AI models with his hypothesis that the brain uses far more complex, evolutionarily-defined loss functions.
• At 3:31 - "It's just this incredibly general prediction engine. So... any one area of cortex is just trying to predict... any subset of all the variables it sees from any other subset. So like omnidirectional inference." - Marblestone proposes that the cortex's learning algorithm is not just about predicting the next step, but about inferring any missing information from any available context.
• At 5:41 - "They're able to predict things that the more innate part of the brain is going to do... this whole thing is basically riding on top of a sort of a lizard brain and lizard body." - Marblestone explains the idea that the advanced learning parts of the brain build a model that can predict the reactions of the more primitive, instinctual parts.
• At 22:39 - "Right now, the way the models work is you have an input, it maps it to an output. And this is amortizing a process that... the real process, which we think is like what intelligence is, which is like you have some prior over how the world could be." - The host explains his understanding of amortized inference as a shortcut for a more fundamental, Bayesian process of reasoning.
• At 25:25 - "...capabilities which were... which required inference time compute to elicit, get distilled into the model. So you're amortizing the thing which previously you needed to do these like rollouts, these like Monte Carlo rollouts to figure out." - The host describes how complex reasoning that once required active computation at inference time is now being baked into the model's parameters.
• At 32:20 - "How do you explain the fact that so little information is conveyed through the genome... But if a big part of the story... is the reward function, is the loss function, and if evolution found those loss functions which aid learning, then it actually kind of makes sense how... you can like build an intelligence with so little information." - The host proposes that the genomic bottleneck is resolved if evolution primarily encodes the learning process and reward signals, not the entire brain's learned structure.
• At 36:30 - "This would help explain why the hominid brain exploded in size so fast, which is presumably... social learning or some other thing increased the ability to learn from the environment... and then now it increases the returns to having a bigger cortex which can then like learn these things." - The host speculates that the rise of culture and social learning created a powerful evolutionary incentive for a larger learning subsystem (the cortex), driving rapid brain expansion.
• At 50:38 - "is it a disadvantage or an advantage for humans that we get to use biological hardware in comparison to computers as they exist now?" - Dwarkesh Patel poses the central question of the segment, framing the debate around the trade-offs between the brain's evolved biological constraints and the engineered advantages of digital hardware.
• At 51:43 - "an obvious downside of the brain is it cannot be copied. You don't have, you know, external read-write access to every neuron and synapse, whereas you do, I can just edit something in the weight matrix." - Adam Marblestone highlights a fundamental difference in transparency and malleability between AI neural networks and biological brains.
• At 54:14 - "it's just a bunch of kludge that you have to do in order to make the synaptic thing work... with a cell to modulate a synapse according to the gradient signal, it just takes all of this crazy machinery." - Dwarkesh Patel articulates the hypothesis that much of the brain's intricate cellular and molecular biology may not be part of the core learning algorithm itself, but rather a complex implementation detail required by its biological nature.
• At 62:08 - "We have to start with the brain and make new vocabulary rather than saying backprop and then try to apply that to the brain." - Adam Marblestone summarizes György Buzsáki's "inside-out" research philosophy, arguing that neuroscience should discover the brain's true computational primitives rather than trying to impose existing concepts from AI onto it.
• At 72:26 - "E11's technology and sort of the suite of efforts in the field also are trying to get like a single mouse connectome down to like low tens of millions of dollars." - Adam Marblestone discusses the technological goal of dramatically reducing the cost of brain mapping, which would make large-scale connectomics projects feasible.
• At 74:04 - "You can get a quote unquote 'molecularly annotated connectome.' So that's not just who is connected to who by some kind of synapse, but what are the molecules that are present at that synapse, what type of cell is that." - Adam Marblestone explains the next frontier in connectomics, which aims to go beyond a simple wiring diagram to include detailed molecular information.
• At 77:50 - "'This is all going to be... all answerable by neuroscience. It's going to be hard, but it's actually answerable.'" - Marblestone expresses his conviction that core questions about intelligence, which are currently debated philosophically, are ultimately empirical problems that neuroscience can solve.
• At 78:00 - "'It seems to me in the grand scheme of trillions of dollars of GPUs and stuff, it actually makes sense to do that investment.'" - Marblestone puts the cost of large-scale neuroscience projects in perspective, arguing that a few billion dollars is a reasonable investment to potentially revolutionize AI development.
• At 80:20 - "'Does that sculpt the network to know the information that humans know about cats and dogs and to represent it in a way that's consistent with how the brain represents it?'" - Marblestone explains the hypothesis behind "behavior cloning the brain": training an AI on neural data could force it to develop more robust and generalizable internal representations.
• At 105:31 - "'Here's this gap. I need this piece of infrastructure which like there's no combination of a grad students in my lab or me loosely collaborating with other labs with traditional grants that could ever get me that.'" - Marblestone describes the typical roadblock faced by scientists that Focused Research Organizations (FROs) are designed to overcome.

Takeaways

• To build more data-efficient AI, move beyond simple objectives like "predict the next token" and explore complex, multi-stage "curriculums" of loss functions inspired by evolution.
• A key to developing safer and more aligned AI may lie in reverse-engineering the brain's innate "steering subsystem" to understand how to instill robust, beneficial reward functions.
• Recognize that true, flexible intelligence likely requires significant computation at inference time; relying solely on fully "amortized" models may be a limiting shortcut.
• Complex systems like the brain can emerge from a compact set of initial instructions if those instructions specify the learning process itself, rather than the final outcome.
• Future breakthroughs in AI hardware can be informed by the brain's architecture, which achieves remarkable energy efficiency through the deep co-design of its algorithms and biological substrate.
• Progress in understanding intelligence requires a "bottom-up" approach to neuroscience that discovers the brain's actual computational primitives, rather than forcing existing AI concepts onto biological data.
• Treat large-scale neuroscience projects, like creating a molecularly annotated connectome, as essential infrastructure investments that will unlock progress across both neuroscience and AI.
• Frame fundamental brain research as a strategic, high-return investment for the AI industry, as a few billion spent on understanding the brain could revolutionize the trillions spent on compute.
• A promising path to more robust AI is to train models directly on neural data, forcing them to learn not just what humans do, but how their brains represent the world.
• To solve grand scientific challenges, we need new organizational models like FROs that are engineered to build the large-scale tools and datasets that are beyond the scope of traditional academia.