Every so often you read a paper by someone you’ve never met and get a strange feeling — not that they’ve copied you, and not that you’ve copied them, but that the two of you have been walking around the same mountain from opposite sides and have just met at the summit.

That’s how I felt reading A Beautiful Loop: An active inference theory of consciousness, by Ruben Laukkonen, Karl Friston, and Shamil Chandaria. It describes, in a completely different vocabulary from mine, something that looks unmistakably like the core of the Four-Model Theory (FMT): a mind becomes conscious by looping a model of itself back onto itself.

Let me tell you what their theory says, where it converges with mine, and — just as importantly — where it honestly differs. Convergence is only interesting if you’re willing to be precise about the seams.

What “A Beautiful Loop” actually claims

Laukkonen and his co-authors build on active inference and the Free Energy Principle — a framework, largely developed by Friston (who is a co-author here, so the grounding comes from its own architect), that says any system which persists over time behaves as if it were minimizing prediction error, or “surprise,” about the world. Perception, action, and attention all fall out of one imperative: keep your internal model and your sensory stream from diverging too far. One fair caveat is worth stating: agents — people especially — often actively seek out novelty and surprise, so “minimize surprise” can’t be taken at face value; the framework has to cash such exploration out as uncertainty-reduction in the long run, which is a subtler claim than the slogan suggests. Still, it’s a process story. It tells you what a mind is doing moment to moment.

On top of that machinery, the paper proposes three conditions that a system has to meet to be conscious.

First, an epistemic field — a generative model of the world, the “space” of everything that can be known or acted on. You can’t be conscious of anything without a world to be conscious of, so a unified reality-model comes first.

Second, inferential competition — the paper’s answer to the old binding problem. Not everything the brain computes becomes conscious. Candidate interpretations compete, and the ones that best reduce long-term uncertainty win their way into the reality-model. The authors call this “Bayesian binding”: coherence is the price of admission.

Third — and this is the heart of it — epistemic depth. The reality-model doesn’t just represent the world; it recurrently shares its own beliefs back through the whole hierarchy, so that the model comes to know that it knows. They give it a lovely name: field-evidencing, the system continuously producing evidence of its own knowing. That recursive turn — the model modelling its own modelling — is the “beautiful loop.” Precision-weighting (roughly, attention: how much confidence the system assigns to each signal) tunes all three conditions, and at the top sits what they call a hyper-model — a model of the models, a global controller of confidence that they note is reminiscent of general intelligence itself.

It’s an elegant, generous, carefully hedged paper, and it reaches out to meditation, psychedelics, dreaming, and minimal awareness states along the way. If you work on consciousness, read it.

Where it converges with FMT

Now here’s why I sat up.

One: the loop is the thing. FMT’s central claim is that consciousness is constituted by self-referential closure — the system’s generated model includes a model of the system generating the model. Laukkonen’s epistemic depth is describing the same architectural move: a model that folds back to know itself. Strip away the two vocabularies and you get one sentence that both theories would sign: what makes a system conscious is not that it models the world, but that its model of the world comes to include, and loop through, a model of itself. Two theories, developed independently, pointing at the same closed curve.

Two: the self is a layer on the world-model, not a separate thing. In FMT the self is not a little homunculus watching a screen. It’s a self-model built on top of a world-model — in the theory’s terms, an explicit self-model embedded in the very system that runs the explicit world-model. Laukkonen gets there from the other direction: for him, self-modelling and first-person perspective emerge when the reality-model starts evidencing its own knowing. Different starting points, same conclusion — the “I” is a fold in the world-model, not an extra ingredient outside it. We even converge on the deflation that follows: the felt sense of a solid, separate self is something a sufficiently deep loop produces, which is why it can be loosened (in meditation, in psychedelics) without the lights going out.

Three — more tentatively — a shared intuition about dynamics near a threshold. Here I want to be careful not to put words in the authors’ mouths. FMT commits explicitly to criticality: the loop has to run at the edge of chaos, in what I call the Class-4 regime — complex enough to sustain a self-simulation, ordered enough for it to stay coherent. Beautiful Loop makes no such commitment. What it does emphasize is nonlinear “ignition” and the metastable, precision-weighted competition by which contents win their way into the field. That is adjacent, not identical: both pictures place consciousness in poised, near-threshold dynamics rather than in static structure, but the shared ground is an intuition, not a shared formal claim. I flag it as a family resemblance worth watching, nothing stronger.

The honest difference

If the shapes agree, why are these two papers, and not one?

Because they answer different questions, and I don’t want to blur that.

Beautiful Loop is a process theory. FMT is an architecture theory. Laukkonen tells you the mechanism: consciousness arises from prediction-error minimization, precision-weighting, and a hyper-model that predicts its own confidence. It’s substrate-committed in an interesting way — the authors argue that biological systems are “leaky,” that the loop is embodied in wet, warm tissue, and that this matters. FMT, by contrast, is architecture-first and deliberately substrate-neutral. It doesn’t require the Free Energy Principle. It asks a structural question — which model-kinds close on which — and answers it with a 2×2 of model kinds along two axes: world versus self, implicit versus explicit. (Note the framing: not “four modules” bolted into a brain, but four kinds of model, the way “north-facing” and “south-facing” are kinds of slope, not two extra hills.) Consciousness, on FMT’s telling, is what happens when the explicit self-model closes the loop on the system running the explicit world-model — at criticality — regardless of whether the substrate is neurons, and regardless of whether free-energy minimization is the process that gets you there.

So which is right? Possibly both. This is the part I find genuinely exciting rather than merely diplomatic. Nothing in Laukkonen’s paper refutes FMT, and nothing in FMT refutes his. It is entirely coherent that active inference is the process that implements FMT’s closure, and that FMT’s architecture is the structure that the active-inference loop instantiates. One theory names the engine; the other names the shape the engine has to build. They may simply be the same loop, described in two languages — one the language of prediction and free energy, the other the language of models and closure.

Why independent convergence matters

Here’s why I think this is more than a pleasant coincidence.

When two people set out to explain the same thing from the same assumptions, agreement is cheap — they were always going to land nearby. But Laukkonen came up through active inference and contemplative neuroscience; I came up through cellular automata, self-modelling, and a stubborn intuition about virtual simulations. We used different mathematics, different lineages, different motivating examples. And we both put a recursive self-model at the dead center of consciousness.

In a field with more theories than results, that kind of independent convergence is a weak but real signal that the loop is not an artefact of either vocabulary — that there’s something there to point at.

Two theories, two languages, one beautiful loop. I’ll take that.