In our quest for knowledge and understanding, is it reasonable to expect intelligibility?

Quantum mechanics is the most successful theory in the history of physics. It predicts the electron's magnetic moment to twelve decimal places. It underpins semiconductors, lasers, MRI machines, the device you're reading this on. Nothing else comes close.

It also resists human understanding in a peculiar way. After a century of work by very capable minds, there is no consensus on what the theory means: what it says about reality, what happens during measurement, whether the wavefunction is a thing in the world or a tool for calculation. The mathematics is precise, the predictions are extraordinary, the interpretation remains contested.

We pass over this too quickly. Our best theory of the physical world works—spectacularly—without being understood. The question is whether we were right to expect it could be.

What is being demanded

A working physicist might object that the demand has already been met. Quantum mechanics predicts every experiment we throw at it. We compute, design, manufacture. What more is there to understand?

The "shut up and calculate" position has the virtue of being honest about what the formalism delivers. It also concedes the point. When a Newtonian physicist understands a collision, she can hold a picture in mind: two billiard balls, a force, a transferred momentum, a new trajectory. When a Maxwell-era physicist understands electromagnetic propagation, she can think about a field, undulating, filling space, carrying energy from here to there. These pictures sit alongside the equations and tell her what the equations are about. The math describes; the picture says what's being described.

Quantum mechanics offers no such picture. It offers several incompatible ones and provides no way to choose. Ask a physicist what's happening to an electron between preparation and measurement, and you'll get an answer that depends on which interpretation she's adopted, or a refusal to answer, or the suggestion that the question is ill-posed. The wavefunction is real, or it isn't. The universe splits, or it doesn't. Probabilities track our information, or they track a stochastic collapse in nature.

This is the distinction the post turns on. Adequacy is what the formalism delivers: predictions match observations, the theory lets us build things, calculate things, design experiments. Intelligibility is what a picture delivers: a sense of what the theory is about, what exists, what is happening, what the variables refer to. Classical physics gave us both at once. We came to assume they came as a package. They don't. Quantum mechanics is the first place we found out.

A theory can be adequate without being intelligible. This was not predicted. The existence of the interpretation debates—conducted by people who agree on every empirical fact and every line of the math, and disagree only on what it all describes—is the evidence that intelligibility is a further thing, a thing the formalism doesn't automatically provide.

The structural mismatch

To see how deep this runs, it helps to be specific about what quantum mechanics established.

In 1936, Birkhoff and von Neumann showed that the propositions of quantum mechanics form a non-distributive lattice. The distributive law of classical logic—

A ∧ (B ∨ C) = (A ∧ B) ∨ (A ∧ C)

—fails for quantum systems.

A concrete case. Take an electron and consider three propositions:

A: "The electron has spin-up along the z-axis"

B: "The electron has spin-up along the x-axis"

C: "The electron has spin-down along the x-axis"

Classically, if A is true, then either (A ∧ B) or (A ∧ C) must be true: the electron has z-spin-up and some definite x-spin. In quantum mechanics, an electron with definite z-spin has indefinite x-spin. The state is a superposition of B and C, not a hidden choice between them. So A ∧ (B ∨ C) can be true while both (A ∧ B) and (A ∧ C) fail to be well-formed claims about the system.

The lattice of closed subspaces in Hilbert space is orthomodular, not Boolean. What this means is itself contested—whether it forces a revision of logic (a minority position) or signals a deep fact about the algebra of observables (more common today). Either way, the structure that systematizes quantum phenomena does not match the structure that seemed, for millennia, to be necessary for coherent thought about discrete properties. Something about the world's organization at this scale fails to fit a shape our cognition takes for granted.

A formalism that grew without an architect

The mismatch goes further than logic. The formalism itself has a curious status. Quantum mechanics isn't derived from self-evident axioms. Its foundations are arguably either inconsistent or incomplete, depending on how you count. The theory was built from pieces that work, assembled through decades of trial and error.

Why is the state space a complex Hilbert space rather than real, or quaternionic, or something else entirely? We have consistency arguments and reconstruction theorems, no deep reason. It's what works.

Why are observables represented by Hermitian operators? Because Hermitian operators have real eigenvalues, and measurement outcomes are real numbers. That's an explanation of compatibility, not a derivation.

Why does probability come from the Born rule—the squared amplitude of the wavefunction? Born guessed it in 1926. Every experiment since has confirmed it. The deeper question is what kind of thing the rule even is. Gleason's theorem shows that if you're already doing probability on the closed subspaces of Hilbert space, the measure must be |ψ|². It's a constraint theorem, not a foundation. Decision-theoretic derivations in the Everettian tradition try something more ambitious: to show that a rational agent in a quantum world must assign probabilities at all, and assign them this way. These remain contested at the level of what counts as a successful derivation in a domain where probability itself is part of what's being explained. A century in, the most empirically confirmed rule in physics still rests on a postulate whose status nobody agrees on.

The axioms were reverse-engineered. Phenomena came first; formalism was built to fit; justification, if it comes, comes later. A century on, it still hasn't fully arrived.

The measurement problem

The measurement problem is the structural gap most worth dwelling on. The standard framing makes it sound milder than it is.

Quantum mechanics has two rules for state evolution. Between measurements, the state evolves according to the Schrödinger equation: deterministic, linear, reversible. Upon measurement, the state "collapses" to an eigenstate of the observable being measured: stochastic, non-linear, irreversible.

These are usually called two dynamics. The framing is already misleading. The Schrödinger equation is physics in the recognizable sense: an equation of motion, derivable from a Lagrangian, time-symmetric, of a piece with the rest of mathematical physics. Collapse is something else. Collapse is an update rule on the formal object we use to make predictions—the same kind of object an actuary updates when she gets new information. It has no Lagrangian. It has no time symmetry. It is silent on what physically triggers it.

The formalism contains physics, plus an epistemic update rule that behaves like physics—affecting what we predict, how the system evolves, what we expect to see. The interpretations fight, in the end, over which is which. Many-Worlds says collapse isn't there, only branching. Bohmian mechanics says there's no collapse, only hidden variables doing deterministic work behind the wavefunction. GRW says collapse is real physics with an equation we haven't written down. QBism says collapse is just what Bayesian updating looks like when the agent is doing quantum mechanics. All of these are responses to the same gap: a formalism that mixes a description of the world with a rule for updating beliefs and gives no principle for telling them apart.

Decoherence theory has illuminated part of the picture. It explains why interference vanishes in practice for macroscopic systems, why we don't see cats in superposition. It leaves the deeper puzzle unresolved: why we get definite outcomes at all. It is compatible with every major interpretation.

This is the kind of gap a formalism can contain without being able to close.

The persistence is the data

Here is where we stand. A theory whose axiomatic foundations remain unsettled, whose logical structure breaks with classical organization in ways still being worked out, whose central probability rule rests on a contested postulate, and which contains a structural gap between physics and epistemic updating with no principle for the boundary—this theory also works better than any theory in history.

The natural response is to wait. Eventually we'll find the right interpretation, derive the axioms from something deeper, close the gap. A century is a long time. The divergences show no signs of narrowing. At some point, the persistence of the situation becomes part of the data.

The assumption that adequacy and intelligibility come together was never an assumption we examined. It was an inheritance from theories—Newton, Maxwell, statistical mechanics—where they happened to coincide. When physics dealt with mid-sized objects in low-dimensional space with classical causal structure, our cognition could form pictures of what the theories described. The pictures and the equations grew up together. We came to feel that any successful theory would deliver both, because every successful theory had.

Quantum mechanics is the first place this entitlement broke. The theory is adequate without being intelligible, and a century of work hasn't recovered the missing half. The simplest explanation is that the entitlement was always contingent. We expected pictures because we had always gotten them, and we had always gotten them because the domains we'd theorized about were close enough to the cognitive niche that shaped us. There was never a guarantee that fundamental physics would stay in that niche. At small enough scales, it didn't.

This reframes what the interpretation debates are evidence of. The dominant reading treats them as a research program in progress: we'll eventually figure out which interpretation is right. Another reading is available. The debates may be evidence that the demand they're trying to satisfy—the demand for a single intelligible picture of what the theory describes—is not the kind of demand the phenomenon is obligated to meet.

The mathematician's revision

If intelligibility is a contingent inheritance, we should expect other domains to fail to deliver it. Reasoning is a plausible candidate.

Consider a concrete situation. A mathematician is working on a proof and reaches a step that follows from her usual rules of inference. The step produces a result she finds suspicious—not contradictory, just wrong. She pauses, examines the inference, decides the rule was being applied to a domain where it doesn't hold. She revises the rule, redoes the step, continues.

What happened? The reasoning system evaluated and modified its own operating rules mid-operation. The rules weren't fixed inputs. They were objects the system could pick up, inspect, and put down again, altered.

Any formalism for reasoning has to decide where its rules live. If the rules R are fixed—part of the formalism's definition—the system can't do what the mathematician did. If the rules are objects the system can revise, revision happens according to some other rules R′, which are either fixed (and the problem recurses) or themselves revisable (and the regress continues).

This isn't a new observation. It has a formal cousin in Tarski's undefinability theorem: in any sufficiently expressive formal system, you can't define a truth predicate for that system within it. You need a metalanguage. The metalanguage needs its own metalanguage to define its truth predicate. The hierarchy never closes. Tarski doesn't say reasoning is unformalizable. He says that some properties of a system can be proven unrepresentable from inside the system. The demand that a system fully contain an account of itself is not a demand the world is obligated to satisfy.

The resemblance to the measurement problem is structural. Quantum mechanics contains a description of physics plus an update rule that can't be reduced to it. Reasoning seems to contain object-level operations plus a capacity for revising them that can't be reduced to object-level. Both are formalisms that include something they cannot internalize. A successful theory of reasoning may turn out, like quantum mechanics, to be a formalism that works—predicts, systematizes, enables—without delivering a single intelligible picture of what reasoning is. Multiple interpretations of the formalism may coexist. The boundary between reasoning and meta-reasoning may be present in the math without being derivable from it.

The live case

We are building systems right now whose ontology we cannot agree on.

The formalism is precise. Weights, activations, gradients, attention. We describe the architecture exactly, train the system, predict (within limits) what it will output. The empirical behavior is tractable: benchmarks, evaluations, capability tests. What we cannot agree on is what is happening inside. Whether the system understands. Whether there is something it is like to be it. Whether its apparent inner states are inner states. Whether its reasoning is reasoning, or a very good imitation of reasoning, and whether that distinction is even well-formed.

The structure of these debates is familiar. The same formalism, the same empirical behavior, irreconcilable accounts of the underlying ontology. The interpretation arguments around machine cognition have the shape of the interpretation arguments around the wavefunction. Our default attitude toward them is the attitude physics took toward its own interpretation debates a century ago: this is a temporary state, more research will resolve it, interpretability will eventually deliver the picture we feel owed.

Quantum mechanics should make us less confident in that default. The picture may not be coming. The work may have been done; the demand itself may have been ill-formed from the start. We may be doing with neural networks what we did with electrons: extending an entitlement to intelligibility from domains where it was satisfied to domains where it may not be satisfiable.

What we were never owed

If the precedent holds, what changes is the expectation.

The question that has faced fundamental physics for a century is when we will finally understand quantum mechanics. A different question sits beneath it: whether the kind of understanding we are demanding—an intelligible ontology, a picture of what is there—was ever a coherent thing to demand of every successful theory, or whether it was a habit of mind inherited from a stretch of intellectual history in which the demand happened to be met.

The same question is now arriving elsewhere. In reasoning, where the formal limits on a system's capacity to contain itself suggest the demand may not have a satisfier. In machine cognition, where the debates about what our systems "really" are have the shape of interpretation debates physics has not been able to close in a hundred years.

The theories that gave us adequacy and intelligibility together may have been a special case. Whether the theories that follow do the same, in physics, in reasoning, in the machine intelligence we are now building, remains open.

The question is no longer when we will get our picture. It is whether we should have expected one.