Summary

Beck treats agency as a spectrum of computational sophistication, not a label recoverable from behavior alone. Planning and counterfactual reasoning are decisive internally, yet externally they collapse into an input-output policy; the pragmatic move is to choose the best-compressing model—for example, an internal-search model with 40 parameters rather than a much larger function transformation. “Science is about prediction and data compression and nothing else.”
Physical embodiment remains Beck’s dividing line between an agent and a simulation of one. A perfect computer model may predict an agent “100% correctly” without becoming one, while he says he thinks the same computation running in his body would qualify; within that boundary, transfer entropy and context dependence might measure degrees of agency rather than impose a brittle threshold.
A VAE is Beck’s canonical example of an energy-based model because its objective constrains both outputs and internal representations. Unlike ordinary feed-forward approximation, it minimizes over latent states as well as weights: reconstruction error governs input-output fidelity, while a traditional VAE also pressures its representation toward Gaussian structure. “You actually have to do two minimizations.”
Test-time training is moving mainstream systems toward energy-based inference, but today’s deployment pattern may be internally inconsistent. Allowing selected weights to change makes them latent variables; Tim’s concern, which Beck affirms, is that most networks are trained conventionally and only gain those latents after deployment. Tim explicitly hedges that he is not an expert while calling this unwise.
JEPA and self-supervised learning aim to preserve useful structure by predicting in compressed latent space rather than reconstructing every pixel. The opportunity is jointly learning the representation and predictor; the failure mode is trivial collapse, where both embeddings become zero. The practical warning is equally important: PCA can discard low-variance neural dimensions that contain “some of the most important” information.
Beck sees continual, modular model construction—not a monolithic AGI—as the missing route toward stronger intelligence. A capable system should notice surprise, instantiate new latent variables or objects, and recombine specialized modules “like LEGO bricks”; his preferred destination is “collective specialized intelligences,” with systems engineering and inter-module communication carrying more weight than generality as a slogan.
Automated science becomes transformative when AI progresses from finding correlations to designing and running experiments, but safety constraints must accompany that capability. Beck worries more about malicious or careless humans than spontaneous machine rebellion; his safer alignment recipe is to infer rewards from observed human behavior, then “perturb that distribution over outcomes” incrementally and evaluate consequences instead of issuing grand commands such as “end world hunger.”

Deep dive

1. Physical symmetries should be built into world models

Beck’s case for geometric deep learning begins with the structure of reality: the world is translation invariant and, allowing for gravity’s principal axis, rotationally invariant in the xy-plane. A brute-force network might discover those regularities, but “the mathematician in me really wants to build the symmetries in.”
From a free-energy-principle perspective, Beck sees no categorical structural divide between an agent and an object. An agent is instead “a really sophisticated object,” carrying internal states across long timescales and executing context-dependent policies; agency is therefore a question of degree.
The host presses for something stronger than policy execution, since otherwise “a rock is an agent.” Beck locates the difference in how policies are computed—planning, counterfactual reasoning and goal-directed latent variables—but concedes that observed behavior alone may reveal only a sophisticated input-output relationship.

2. Counterfactual computation makes agency real—but hard to observe

The host’s sharpest test is a recorded computational trace: preserve only the actual path from a conscious or agentic process and “that thing clearly has lost whatever agency or consciousness it had.” What disappeared was active consideration of alternative possibilities.
Beck agrees that demonstrated planning and counterfactual reasoning are sufficient for agency, but opening the system may be unavoidable. From outside, even massive Monte Carlo search followed by action looks like a function transformation; knowing the code changes the attribution, not the visible policy.
His pragmatic escape is model selection: if direct function approximation requires many parameters while an internal-search model explains the same behavior with roughly “40 parameters,” call it an agent. After the host invokes Dennett’s intentional stance, Beck frames the “as if” description as a best-model judgment justified by prediction and data compression, not access to microscopic causation.
Beck nevertheless insists that agents must be physical. A high-fidelity simulation of him disconnected from the world is a model of agency, not an agent; he says he thinks putting the same calculations in his body would make it an agent. He admits the boundary is philosophical territory where “everything has an error bar” and clean distinctions break down.

3. VAEs reveal what energy-based learning adds

Beck frames neural-network training itself as following an energy function in weight space toward a minimum. The energy-based distinction is where the cost applies: ordinary networks optimize weights from input-output error, whereas energy-based models also optimize an internal state, constraining which mappings from (x) to (y) are admissible.
A VAE is his most common canonical specimen. Encoder and decoder reconstruction loss still compares inputs with outputs, but the objective also acts on the representation—typically asking “how Gaussian is it?” For a VQ-VAE, Beck describes the latent structure as being like a mixture of Gaussians while retaining the internal-state penalty.
Test-time training extends this logic by treating changeable weights as latent variables. In a traditional energy-based model, those latents are minimized throughout training; switching them on only at deployment means the original network was never trained with that inference process active. Tim calls this unwise while explicitly noting he is not an expert, and Beck agrees.
Energy-based and Bayesian models are closely related because energy functions correspond to log-probability structure, up to a normalization factor that can be ignored when only minimizing. Point or MAP estimation sacrifices a full latent distribution; curvature near the minimum plus a Laplace approximation can restore a tractable Bayesian treatment. Free energy adds an entropy regularizer—“are you regularizing or not” is Beck’s short distinction.

4. Joint embeddings trade pixel fidelity for useful abstraction

The appeal of JEPA is to compress both observed inputs and target outputs, then learn prediction between their embeddings. Beck likes making data compression explicit because many tasks require “something that’s a little more gestalt, a little more high level,” not a prediction of every pixel.
The architectural problem is collapse: perfect prediction is trivial if both embeddings become zero. The host describes contrastive methods as using negative sampling at substantial cost, while non-contrastive approaches such as VICReg, BYOL and Barlow Twins use different regularizers. The host also credits wake-sleep and contrastive divergence for offering a biologically plausible route around backpropagation.
Beck’s everyday workflow exposes the representation risk. He runs PCA and a VAE first on new neural datasets to inspect signal-to-noise, then often analyzes the learned embedding—but PCA prizes high variance even though low-variance neural dimensions may carry the most valuable signal.
The host argues that jointly fitting preprocessing and inference reduces the chance of discarding information needed later. Beck agrees with the broader goal of preserving information, while noting that brains and neural networks routinely discard what is currently task-irrelevant.

5. Intelligence requires continual construction of modular models

Discussing ARC, the host notes that version one survived for five years before thinking models defeated it, with versions three, four and five still to come. Beck expects “always something left over”: each solved challenge invites another claim that the system is merely faking an ability humans consider special.
Continual learning is the critical missing component. A deployed system should encounter something surprising, decide “I need to turn learning on,” and append its model; Beck’s object-centric physics-discovery system can instantiate an entirely new object when existing objects cannot explain a situation.
Recursive improvement would go beyond accumulating facts to generating models on the fly. Beck points to GFlowNets from Bengio’s work as, in principle, “a generative model of generative models”—a system able to decide that its current latent set “is just not cutting the mustard anymore” and create another node.
His preferred architecture is modular and recombinable: old models become LEGO bricks that can produce capabilities never designed individually. Biological and social intelligence provide the analogy—specialized brain regions and specialized people become more powerful through communication. “I don’t believe in AGI”; what he wants is “collective specialized intelligences.”

6. Autonomous science needs experiments, constraints and human partnership

Current AI-assisted science largely summarizes vast datasets and surfaces small but relevant correlations. Beck’s next milestone is experimental design: systems should identify gaps, formulate tests and automate empirical inquiry, subject to safety limits once they can operate laboratories.
His concrete case is a robot encountering a beach ball for the first time. Stopping is not learning; it should behave like a child, poking the object and observing motion to infer its properties. Virtual testing matters because unconstrained optimization might decide the best experiment is to “kick it as hard as you possibly can”—or, at laboratory scale, set off a nuke.
Beck worries that wholesale automation could reduce people to “value function selectors,” occasionally approving or rejecting machine-generated outcomes. Yet the tractor analogy makes him a techno-optimist: agriculture once occupied roughly 99% of people and now perhaps 2%, freeing human attention for other pursuits. His hoped-for future is cybernetic co-evolution, where AI is “a partner, not an adversary or a crutch.”
On catastrophic risk, Beck is more concerned about human deployment than a spontaneous Skynet: systems still execute specified objectives, and naïve objectives can be dangerous. His alternative draws on maximum-entropy inverse reinforcement learning—estimate rewards from humanity’s stationary actions and outcomes, reproduce that baseline, then make small outcome perturbations and measure consequences rather than hand-writing an absolute goal.

The Internet Computer: Caffeine.ai CEO Dominic Williams on Unstoppable, Self-Writing Software Vol.204 Industry Watch 38 | Nvidia’s $20B Groq Deal: Where Is the Opportunity for China’s AI Inference Chips?