Part I: Foundations

From Boundaries to Models

0:00 / 0:00

Regulation Requires Modeling

Once a boundary exists it has to be maintained, the interior held distinct from the exterior despite perturbation, degradation, and fluctuation. Let the interior state be sin∈Rm\mathbf{s}^{\text{in}} \in \R^m and the exterior sout∈Rk\mathbf{s}^{\text{out}} \in \R^k. The boundary mediates interaction through observations ot=g(stout,stin)+ϵt\mathbf{o}_t = g(\mathbf{s}^{\text{out}}_t, \mathbf{s}^{\text{in}}_t) + \bm{\epsilon}_t and actions at∈A\mathbf{a}_t \in \mathcal{A}, permeabilities, active transport. Persistence means keeping sin\mathbf{s}^{\text{in}} inside a viable region Vin\viable^{\text{in}} under incomplete observation of sout\mathbf{s}^{\text{out}}, stochastic perturbation, degradation of the boundary itself, and finite resources.

The consequence runs deep: regulation requires modeling. Any policy π:O∗→A\policy: \mathcal{O}^* \to \mathcal{A} that achieves viability with probability p>prandomp > p_{\text{random}} implicitly computes a function f:O∗→Zf: \mathcal{O}^* \to \mathcal{Z}, where Z\mathcal{Z} is a sufficient statistic for predicting future observations and viability-relevant outcomes. A policy that beats random must exploit regularities in the observation sequence, and exploited regularities are an implicit model of the environment's dynamics. The minimal such model is the belief state of a partially observable Markov decision process:

bt(x)=P(xt=x∣o1:t,a1:t−1)\belief_t(\mathbf{x}) = \prob(\mathbf{x}_t = \mathbf{x} \mid \mathbf{o}_{1:t}, \mathbf{a}_{1:t-1})

which updates by Bayes' rule and is a sufficient statistic for optimal decision-making: any optimal policy π∗\policy^* can be written π∗:Δ(X)→A\policy^*: \Delta(\mathcal{X}) \to \mathcal{A}, from belief states to actions.

It is tempting to conclude that any system beating random under partial observability keeps a belief state and so has a model of the world. That is too strong in a way that matters later. Successful regulation implies some sufficient statistic, some mechanism that preserves viability-relevant information from history. It does not imply a full Bayesian model. A reflex exploits one regularity without representing anything, and calling its trigger threshold a world model makes the term do no work.

The World Model

In complex environments the full belief state is intractable and real systems keep compressed representations. A world model is a parameterized family of distributions Wθ={pθ(ot+1:t+H∣ht,at:t+H−1)}\worldmodel_\theta = \{p_\theta(\mathbf{o}_{t+1:t+H} \mid \mathbf{h}_t, \mathbf{a}_{t:t+H-1})\} predicting future observations from history ht\mathbf{h}_t and planned actions over a horizon HH. Modern implementations use recurrent latent state-space models:

Latent dynamics:pθ(zt+1∣zt,at)Observation model:pθ(ot∣zt)Inference:qϕ(zt∣zt−1,at−1,ot)\begin{aligned}\text{Latent dynamics:} \quad & p_\theta(\latent_{t+1} \mid \latent_t, \mathbf{a}_t) \\ \text{Observation model:} \quad & p_\theta(\mathbf{o}_t \mid \latent_t) \\ \text{Inference:} \quad & q_\phi(\latent_t \mid \latent_{t-1}, \mathbf{a}_{t-1}, \mathbf{o}_t)\end{aligned}

So the requirement is graded. Control under partial observability needs some mechanism for preserving viability-relevant information from history, and how rich it has to be scales with uncertainty, delay, and branching. At the low end sits a threshold on one sensed variable. At the high end sits a generative model over latent dynamics. Nothing forces a system up that range. What moves it up is an environment in which the cheap mechanisms stop paying. The same structure is what engineers reach when they build capable agents. Dreamer (Hafner et al., 2020), MuZero (Schrittwieser et al., 2020), and JEPA (LeCun, 2022) differ in what they reconstruct and share the latent-dynamics core, because that is the shape of the control-under-uncertainty problem.

The Necessity of Compression

The world model is not merely convenient. It is forced by an asymmetry between the world and any bounded system inside it, which the information bottleneck (Tishby et al., 1999) makes exact. Let W\mathcal{W} be the world's state space and S\mathcal{S} a bounded system with finite capacity CSC_\mathcal{S}. Then

dim⁡(z)≤CS≪dim⁡(W)\dim(\latent) \leq C_\mathcal{S} \ll \dim(\mathcal{W})

The world has effectively unbounded degrees of freedom. Any physical system has finite matter, energy, and extent, so finite information capacity. It cannot represent the world at full resolution. It must compress. That is not a limitation to overcome; it is what being a bounded thing in an unbounded world is.

The compression ratio χ=dim⁡(Wrelevant)/dim⁡(z)\chi = \dim(\mathcal{W}_{\text{relevant}}) / \dim(\latent), where Wrelevant\mathcal{W}_{\text{relevant}} is the subspace of world states that affect viability, measures how much the system has to discard to exist. (The symbol is χ\chi rather than κ\kappa, which Part II reserves for mode coupling.) The implication: compression determines ontology. What a system can perceive, respond to, and value is fixed by what survives compression. The world model's structure, which distinctions it keeps and which it collapses, is the system's effective ontology. The information bottleneck states the tradeoff: the optimal representation maximizes information about viability-relevant outcomes while minimizing complexity,

max⁡z[I(z;viability outcomes)−β⋅I(z;o)]\max_{\latent} \left[ \MI(\latent; \text{viability outcomes}) - \beta \cdot \MI(\latent; \obs) \right]

with β\beta trading compression against fidelity. Different values yield different creatures: high β\beta gives coarse world models, low β\beta rich ones. What survives compression is what the system is.

This has an architectural consequence the experiments confirm (). A linear prediction head compresses hidden state to output through one weight matrix, and one matrix always decomposes into independent columns, each serving a separate target. The result is a factored ontology: internal states channeled into independent streams with no pressure to coordinate. Replace the linear map with two layers and the compression changes. The chain rule through two matrices makes every hidden dimension's gradient depend on every other dimension's activation at the intermediate layer. Now compression demands coordination. Compression does not only fix what the system perceives. It fixes whether its internal states are unified or factored.

That distinction deserves a name. The environment has its own modes of variation, temperature cycles, predator patterns, resource fluctuations, social dynamics, and they are not independent: predator presence couples to resource availability, weather couples to everything. Call the pattern of mode couplings the environment's eigenskeleton. The agent cannot track all of it, so compression selects not only which modes to keep but which couplings between modes to keep. A compression that keeps modes independently yields a flat eigenskeleton: the system tracks several aspects of the world and not how they relate. A compression that preserves couplings yields a curved one: the system tracks how variation in one dimension twists into another. The decomposability wall of is the wall between flat and curved. The formal version, with the measure, is in the forcing functions section below.

Attention as Leverage

Compression determines what can be perceived. Attention determines what is. Even inside the compressed representation the system cannot respond to every viability-relevant feature at once; it has to allocate processing, and attention is that allocation.

Attention is a high-leverage control variable. Let a system inhabit a chaotic environment where small differences in observation lead to divergent action sequences. Its attention pattern ω:O→[0,1]\omega: \mathcal{O} \to [0,1] weights which observations are processed at high fidelity and which are discarded. (ω\omega is reserved for attention throughout; α\alpha is ascription, introduced in Part II.) Since actions depend on processed observations and actions shape future states, a small change in ω\omega steers the system onto a markedly different trajectory. Attention is not where a sovereign chooser stands above physics and picks. It is a variable the trajectory is unusually sensitive to.

Let p0(x)p_0(\mathbf{x}) be the a priori distribution over states, governed by physics, and ω(x)\omega(\mathbf{x}) the probability the system attends to and registers a perturbation at x\mathbf{x}. The effective distribution it becomes correlated with is

peff(x)=p0(x)⋅ω(x)∫p0(x′)⋅ω(x′) dx′p_{\text{eff}}(\mathbf{x}) = \frac{p_0(\mathbf{x}) \cdot \omega(\mathbf{x})}{\int p_0(\mathbf{x}') \cdot \omega(\mathbf{x}') \, d\mathbf{x}'}

The system does not control p0p_0. The leverage lives in ω\omega. Sharply peaked, the effective distribution concentrates on a small region whatever the prior; broad, it approximates the prior. And each ω\omega shapes the next, since what a system attends to now sets which features dominate the state that fixes its next attention. Attention is recursive, and the loop compounds.

The attention pattern is itself a product of the system's internal dynamics, its world model, its self-model, its policy, so there is no uncaused chooser behind ω\omega. Attention does not introduce freedom into a deterministic world. It identifies a point of high leverage inside one, and agency is real in exactly that sense: the system's internal states are the variables through which large changes in outcome are levered. That is why attentional sovereignty is worth defending. Capture a system's attention and you capture its trajectory. What is at stake is control, not free will.

The leverage also has temporal depth, and the mechanism is epistemic rather than physical. Registering an observation conditions the belief state, and the policy is conditioned on the belief state; sustained attention is repeated conditioning, holding a feature in the state that drives action and so stabilizing the trajectory near states consistent with it. Let the information degrade, forgotten, overwritten, lost to noise, and the policy is released. What re-expands is the system's uncertainty, not the space of physically accessible futures, which was never contracted by anyone looking.