Video · 2026-01-25 · 47m · 6 moments

The Brain Is Just Specialized Agents Talking To Each Other — Dr. Jeff Beck

✦ AI generated

timeline · colored by role

Mechanism◆

You can never know from the outside, just by observing behavior, whether something is truly doing planning and counterfactual reasoning or merely executing a complex input-output function — all models of agency are only ever 'as if' models, because science is fundamentally about prediction and data compression, not ground truth about internal mechanism.

Beck contends that agency attributions are always 'as if' models built for prediction, since behavior alone can never reveal whether real planning occurred inside a system.

transcript

Jeff Beck: It's important to remember that like no matter how clever your model is and no matter how clever your approach is and how clever the words are that you use to describe it, a lot of this stuff is is is as if, right... science is about like prediction and data compression and like nothing else... you'll never know for sure in any meaningful way like whether or not it's just doing a function transformation or whether it's engaged in planning and counterfactual reasoning.

Claim◆

An agent must be physical — a computer simulation of an agent, no matter how accurate, is not itself an agent, only a model of one.

Jeff Beck argues that agency requires physical embodiment; a simulation can predict what an agent would do without ever being an agent itself.

transcript

Jeff Beck: So I do believe that an agent needs to be physical. That absolutely. I don't believe, you know, I I believe you can have a model of agency and not have an agent, right? I, you know, you can put that model in a computer and run it and make predictions as to what an agent would do. You and it might even be 100% correct, but I still wouldn't call it an agent.

rebuts · 1

Mechanism

The key structural difference between an energy-based model and a traditional feedforward neural network is where the cost function is applied: ordinary networks optimize a cost that's only a function of inputs and outputs, while energy-based models also apply a cost to internal latent states, requiring two separate minimizations.

Beck explains that energy-based models differ from standard neural nets because their cost function also constrains internal/latent states, not just the input-output mapping, requiring a dual minimization process.

transcript

Jeff Beck: So in a traditional neural network, you take in your inputs, you got your outputs, and the cost function is just a function of the inputs and the outputs. And the only thing that you're optimizing is the weights. In an energy based model, there's another thing that your cost function operates on, and that's one of the internal states of your model.

Claim

One of the most critical missing elements in current AI is continual learning — systems need to keep learning after deployment when they encounter novel situations, rather than only learning during training and then being frozen.

Beck identifies continual learning — the ability to recognize novel situations post-deployment and update internal models on the fly — as a critical missing piece for real intelligence.

transcript

Jeff Beck: I think that one of the most critical missing elements right now is some form of continual learning, right? At the end of the day, you really want an algorithm that doesn't just learn on the training set and then just gets deployed. You want something that runs around in the world and comes across things that it doesn't understand, and then is able to build, append its model in some sense.

extends · 2

Claim◆

The brain evolved as a collection of highly specialized modules that learned to communicate with each other, and this same principle — collective specialized intelligences working together rather than a single general intelligence — is what we should actually be aiming for in AI, making the concept of AGI something of a misnomer.

Beck argues the brain itself is a collective of specialized communicating modules, and that AI should likewise pursue collective specialized intelligences rather than a monolithic AGI.

transcript

Jeff Beck: The brain is the same way, right? It's in my view it's highly specialized little modules or agents that are capable of being repurposed, reused, capable of communicating with one another in order to solve really complicated problems. But there's always a benefit to specialization. I don't believe in AGI. AGI seems like a bit of a misnomer to me. What we really want is not artificial general intelligence. We want collective specialized intelligences.

supports · 1

Mechanism◆

The safe way to specify an AI's reward function is not to hand-write an outcome like 'end world hunger,' but to empirically estimate a reward function from the steady-state distribution of current human actions and outcomes (via maximum entropy inverse reinforcement learning / active inference), then perturb that distribution slightly and evaluate the consequences.

Beck argues that dangerous AI outcomes stem from naively hand-specified goals, and that a safer approach uses maximum entropy inverse reinforcement learning to derive a reward function from current human behavior, then nudges it incrementally.

transcript

Jeff Beck: Are you familiar with maximum entropy inverse reinforcement learning? I like to call it active inference because it's really similar. So there what you're doing is you're basically observing someone's policy and then you're trying to do a maximum entropy model. You're doing maximum model on the reward function itself.

Highlight slides

Agency attributions are always 'as if' models✦ from: You can never know from the outside, just by observing behavior, whether something is truly doing planning and counterfactual reasoning or merely executing a complex input-output function — all models of agency are only ever 'as if' models, because science is fundamentally about prediction and data compression, not ground truth about internal mechanism.Behavior alone cannot distinguish planning from function transformation✦ from: You can never know from the outside, just by observing behavior, whether something is truly doing planning and counterfactual reasoning or merely executing a complex input-output function — all models of agency are only ever 'as if' models, because science is fundamentally about prediction and data compression, not ground truth about internal mechanism.Agency Requires Physical Embodiment✦ from: An agent must be physical — a computer simulation of an agent, no matter how accurate, is not itself an agent, only a model of one.Simulation vs. Agency: A Conceptual Distinction✦ from: An agent must be physical — a computer simulation of an agent, no matter how accurate, is not itself an agent, only a model of one.The Brain Is a Collective of Specialized Modules✦ from: The brain evolved as a collection of highly specialized modules that learned to communicate with each other, and this same principle — collective specialized intelligences working together rather than a single general intelligence — is what we should actually be aiming for in AI, making the concept of AGI something of a misnomer.AGI Is a Misnomer — Aim for Collective Intelligence✦ from: The brain evolved as a collection of highly specialized modules that learned to communicate with each other, and this same principle — collective specialized intelligences working together rather than a single general intelligence — is what we should actually be aiming for in AI, making the concept of AGI something of a misnomer.Hand-specified goals are dangerous✦ from: The safe way to specify an AI's reward function is not to hand-write an outcome like 'end world hunger,' but to empirically estimate a reward function from the steady-state distribution of current human actions and outcomes (via maximum entropy inverse reinforcement learning / active inference), then perturb that distribution slightly and evaluate the consequences.Safer alternative: derive reward from observed behavior✦ from: The safe way to specify an AI's reward function is not to hand-write an outcome like 'end world hunger,' but to empirically estimate a reward function from the steady-state distribution of current human actions and outcomes (via maximum entropy inverse reinforcement learning / active inference), then perturb that distribution slightly and evaluate the consequences.Nudge incrementally, evaluate consequences✦ from: The safe way to specify an AI's reward function is not to hand-write an outcome like 'end world hunger,' but to empirically estimate a reward function from the steady-state distribution of current human actions and outcomes (via maximum entropy inverse reinforcement learning / active inference), then perturb that distribution slightly and evaluate the consequences.

Related episodes

Article · Import AI

Import AI 465: Open vs closed gaps; Kimi K3; Demis' big policy plan

→

Article · Platformer

A big week for AI denialism

→

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→