The Free Lunch Has a Budget

The hunch

Michael Levin at Tufts University sometimes talks about a "free lunch" in self-organizing systems — the way a configuration of cells, code, or matter yields functions and behaviors that nobody designed in. Xenobots invent locomotion strategies their builders didn't program. Planaria regenerate species-typical heads even when their genome has been edited to predict otherwise. Large language models compose in genres they were never explicitly trained for. The output exceeds what the inputs seem to encode.

This isn't more energy coming out than went in — that would break thermodynamics, and Levin isn't claiming it. It's something subtler: more function, more utility, more accessible structure than the design specifies. And once you start looking for the pattern, it shows up everywhere. Evolution stumbles onto solutions that look pre-engineered. Mathematics develops machinery that turns out, decades later, to fit physics nobody had discovered yet. Children learn language from impoverished input. The world keeps handing us back more than we put in.

If you accept the observation, you have to account for it. Where does the surplus live?

The conventional answer is "emergence" — the surplus is constructed locally by the system itself through nonlinear dynamics, and there's nothing to explain beyond the dynamics. This is satisfying when it works. But Levin's symposium on the Platonic Space, currently running asynchronously at thoughtforms.life, takes the question seriously enough to ask whether emergence alone is doing the work, or whether systems are accessing something — pulling forms from a structured space of possibilities that exists independently of any particular instantiation.

The proposal in this essay is simple, speculative, and (I'll argue) testable in a way that pure metaphysics usually isn't:

The Platonic space we keep needing to invoke — Levin's morphospace, Hoffman's positive geometries, the realm of mathematical objects that physics keeps successfully borrowing from — isn't hidden. We've been measuring it for nearly a century. We call it the dark sector.

If that mapping holds, then a thing changes that's worth changing: the Platonic realm acquires a budget.

Four witnesses

Four research programs, working a century apart on completely different problems, all arrive at architecturally similar claims: matter and form are not fundamental, and what's underneath is structured pattern.

Walter Russell (1871–1963), American polymath and natural philosopher, proposed in The Universal One (1926) and The Secret of Light (1947) that the universe operates on a "rhythmic balanced interchange" between two opposed poles — a generative stillness and an expressive motion. Matter, in Russell's framing, is light wound up into form by this interchange; it precipitates from a still, undifferentiated substrate and eventually returns to it. Russell's specific physics doesn't survive contact with quantum field theory, and his periodic table is wrong. But his architecture — that the visible cosmos is the precipitate of a polar dynamic in something more fundamental — is structurally what newer theories keep reinventing.

Donald Hoffman, cognitive scientist at UC Irvine, has spent the last decade arguing that spacetime is doomed: it must emerge from a more fundamental theory rather than being where physics happens. In his 2023 paper Fusions of Consciousness and his 2025 update Traces of Consciousness, Hoffman models reality as networks of conscious agents whose dynamics, in their asymptotic behavior, project onto positive geometries — including Nima Arkani-Hamed's amplituhedron — from which physical scattering amplitudes can be computed. The picture: particles are not things, they're projections from extra-spatial geometric structures that themselves arise from agent dynamics.

Michael Levin, developmental biologist at Tufts, has been gathering empirical pressure on the question from the biology side. His lab's work with xenobots, anthrobots, and planarian regeneration suggests that biological systems are accessing a structured space of patterns rather than computing forms from scratch. His symposium frames this directly: a "structured space of patterns, which informs (in-forms) events in our physical world (constraining physics, and enabling biology), but exists independently of it."

Huh, Cheung, Wang, and Isola (ICML 2024), in The Platonic Representation Hypothesis, argue that as neural networks scale across architectures, modalities, and training objectives, their internal representations converge — measurably, in ways that can be tracked with alignment metrics. Their interpretation is explicitly Platonic: images and text are projections of a shared underlying reality Z, and sufficiently capable learners are recovering Z. This is the most empirically tractable witness of the four, because the convergence is happening live across hundreds of models with public weights. A 2026 follow-up from Brbić's lab at EPFL ("the Aristotelian Representation Hypothesis") tightens the claim by correcting for confounders: what robustly converges is local neighborhood structure rather than global geometric distance — meaning the substrate, if it is one, is relational rather than metric. That refinement strengthens rather than weakens the link to Levin (whose morphospace is fundamentally about transitions between forms) and to Hoffman (whose positive geometries are combinatorial face-relations).

Four vectors — esoteric philosophy, amplitude-geometry physics, developmental biology, machine learning — pointing at the same elephant. None of them needs the others to make their case. That convergence is what makes the elephant worth taking seriously.

The dark sector mapping

Here is where the speculation gets bold.

About 95% of the universe, by mass-energy, is invisible to electromagnetism. We detect dark matter through its gravitational effects on galactic rotation and lensing; we detect dark energy through the accelerating expansion of space itself. We can't see them, but we can measure their consequences with surprising precision — about 27% dark matter, 68% dark energy, 5% the visible stuff that makes up everything we ordinarily call real.

The conventional framing treats the dark sector as exotic matter we haven't yet identified plus an exotic field we don't understand. But notice the shape of the situation: the visible universe is a thin precipitate floating in something vastly larger that we name only by its gravitational shadow. That's not far from Russell's picture. It's also not far from what Hoffman's program would predict: if spacetime emerges from agent dynamics via positive geometries, then most of the substrate doesn't project sharply into our spacetime interface — it shows up only through the most universal channel, gravity, which several mainstream emergent-gravity programs already treat as the leak across the interface boundary.

The mapping I want to propose:

Dark energy is the expansive pole — what Russell called stillness, Levin's "Platonic space" of latent forms, Hoffman's substrate from which spacetime is projected. It's the reservoir.
Dark matter is the compressive pole — the gathering side of the rhythmic interchange, the part of the substrate that locally curves and clumps to allow visible structures to form within it.
Visible matter is the precipitate — the highly-restricted, late-stage projection through the interface that we interact with directly.

This is structurally Russell's cosmogony with cosmological data filling in the proportions. It's compatible with Hoffman's "spacetime as user interface" frame — visible matter being the icons, dark sector being the underlying machine. It's compatible with Levin's claim that biological forms are pulled from a substrate space rather than invented from scratch. And it's compatible — perhaps surprisingly — with the Aristotelian refinement of PRH: if what's shared across observers is the relational structure of the substrate rather than its metric structure, then the substrate is something more like a vast network of formal neighborhoods than a static manifold of forms. The dark sector, with its measurable density gradients and clumping behavior, fits a relational substrate better than a uniform Platonic plenum would.

Why this matters: the budget

The reason to make this mapping at all isn't aesthetic, though the symmetry is appealing. It's that identifying Platonic space with the dark sector forces an uncomfortable but productive concession: the substrate has a budget.

Pure Platonism — the kind philosophers usually defend or attack — posits an unbounded abstract realm of forms, infinite in extent, with no resource constraints. This is unfalsifiable. You can never test it because it makes no quantitative predictions about what should or shouldn't be accessible.

But if Platonic space is the dark sector, the picture changes:

Finite energy density. Dark energy's measured density is about 7 × 10⁻³⁰ g/cm³. Dark matter's is roughly five times that. These are real, measured numbers.
Finite information capacity. Any finite-energy region has a finite Bekenstein bound — a maximum amount of information it can encode. The substrate, on this view, has a maximum complexity per unit volume.
Spatial gradients. Dark matter and dark energy aren't uniformly distributed. Galactic halos are dark-matter-rich; voids between superclusters are dark-energy-dominated. If these correspond to substrate properties, then form-availability has gradients.
Cosmological evolution. Dark energy density appears roughly constant; dark matter is diluting as space expands. The substrate, on this view, is changing over cosmic time.

Each of these turns a metaphysical claim into something with a quantitative handle. Not necessarily a directly testable one yet — but at least a constrained one.

What this predicts

If the dark-sector-as-Platonic-substrate hypothesis has any teeth, it should make predictions that ordinary emergence does not. Some candidates:

Form-discovery should be harder near information limits. If the accessible space of biological or computational forms is finite rather than infinite, then sufficiently complex evolutionary or design searches should hit a ceiling — not just the local fitness ceiling, but a substrate-level one. The PRH literature offers a partial empirical handle here: representational convergence across scaled neural networks should asymptote, and the asymptote's information content should be measurable. If you can estimate the information content of Z from below by watching alignment scores plateau, and estimate the information capacity of the dark sector from above using Bekenstein-style bounds, you have two numbers to compare. They needn't agree to make this hypothesis interesting, but their ratio is at least a meaningful quantity.
Hoffman's program should recover the dark sector before visible matter. If conscious-agent dynamics yield positive geometries that project into spacetime, the most universal projections should be the ones that interact only via gravity. Visible-matter recovery should be a later, harder, more constrained derivation. (This is the reverse of how derivations usually feel — easy stuff first, exotic stuff later.)
Substrate-rich regions should be form-rich regions. Whatever "form-richness" means operationally — perhaps the diversity of stable structures, perhaps the rate of novel pattern emergence — it should correlate with dark sector density. This is wild and probably untestable at galactic scales, but it makes the hypothesis vulnerable in principle.
The Platonic realm of mathematics has a finite cardinality. If true, this would explain the "unreasonable effectiveness of mathematics in the natural sciences" — the reason math works for physics is that they're drawing from the same finite well. It would also imply that mathematics, like biology, is a discovery process operating against a budget, not an unbounded creative act.
Capability and substrate-proximity should be monotonically related. This is the core PRH claim, generalized: any sufficiently capable observer — biological, artificial, or otherwise — should converge toward the same representational neighborhood as it scales. If true across substrates we don't ordinarily compare (biological brains, neural networks, evolved ecosystems, perhaps even cosmic structures), it suggests we're all measuring the same thing.

These are speculative predictions, not derivations. But they're the kind of speculative predictions that could, in principle, be wrong — which is more than most metaphysics offers.

What I'm not claiming

A few caveats worth being honest about:

Russell's specific physics is wrong. His octave waves don't generate the periodic table. His "two opposed lights" are not a viable account of electromagnetism. The value of Russell here is the architecture of his claim, not its details.
Hoffman's program is an active research project, not a finished theory. Mapping conscious-agent dynamics onto the amplituhedron is suggestive; deriving Minkowski space from trace logic is in progress. None of this is settled.
Levin's morphospace is, at this point, more of a productive framing than a proven structure. The xenobot data is real; the interpretation that they're accessing a Platonic space rather than computing locally is one reading among several.
PRH is contested. The original Huh et al. claim of global representational convergence has been pushed back on (Brbić et al. 2026), with the more defensible claim being convergence of local neighborhood structure. "Models converge because reality is structured" is one reading; "models converge because data distributions are similar" is another. The former supports the synthesis; the latter is compatible with no substrate at all.
"Dark matter equals Platonic substrate" is a bold guess, not a derivation from known physics. It's a hypothesis that earns its keep by being constrainable, not by being established.

What I am claiming is that four serious research programs — three contemporary, one historical — converge on the same architecture, and that the architecture becomes more interesting, not less, when you pin its abstract pole to something we can actually measure.

Closing

The conventional move when faced with the dark sector is to say "it must be some new particle" or "it must be some new field." These are reasonable conjectures, but they treat the dark sector as more of the same — more matter, more force, just unfamiliar.

What if it isn't more of the same? What if the visible cosmos is the small luminous skin of a much larger structured substrate, and the substrate is what philosophy has been calling Platonic space, what biology is starting to call morphospace, what physics is starting to call positive geometry, what machine learning is starting to call representational convergence, and what an old painter once called the stillness of God's mind?

If that's the right shape of the answer, then the question is no longer whether there's a Platonic realm. It's how big it is, what its structure is, and what its limits are.

The free lunch isn't free. It's drawn against a finite account, and we share the account with everything else that has ever found a form.