Accuracy Is Not Free

One honest note before the post proper: the thesis, the counterexamples, and the corrections are mine. The heavy lifting — the draft, the literature checks, the formal framing — was done by my AI research collaborator, after a very long conversation that this post is essentially the written residue of. Where the piece offers a synthesis rather than a citation, it says so.

Why life, brains, heuristics, and the next generation of AI are all playing the same efficiency game

A protein doesn't search for its shape. It falls.

The sequence encodes an energy landscape — a terrain of free-energy states — and the molecule's native fold is the landscape's ground state. That's Anfinsen's thermodynamic hypothesis, the founding postulate of protein folding: the functional structure is the global minimum of Gibbs free energy. But the deep fact, worked out over the following decades in Peter Wolynes' funnel theory, is that evolution didn't just make the right fold stable. It shaped the landscape so the correct fold is reachable — mostly downhill, with few dead ends, so a molecule can find its ground state in milliseconds instead of the ~10^100 years a blind random search would take. Wolynes called this minimal frustration: landscapes are evolved to be almost consistent with each other, funneled, untrapped.

Notice what evolution optimized. It didn't weed out the proteins that compute the wrong answer. It weeds out the proteins that find the right answer too slowly, or get stuck, or waste. Evolution selects for efficiency, not accuracy. Fitness is accuracy under a budget.

That one sentence is the whole post.

The information bridge

"Energy" here is not a metaphor, and that's the part that surprises people. In statistical physics, free energy is F = E − TS: internal energy minus temperature times entropy. And entropy — Shannon got there first, Jaynes made it explicit — is information. A low-free-energy state is, literally, a state that carries less surprise.

This is why the same equation shows up in three places that shouldn't have anything to do with each other. In thermodynamics it's free energy. In statistics it's the variational free energy, an upper bound on the surprise a model assigns to the data you're looking at — the thing every approximate Bayesian method minimizes. In the brain, Karl Friston's free energy principle proposes that living systems persist by minimizing exactly this quantity: perception is bound-surprise minimization, action is choosing the observations that are least surprising. The principle is contested and, to be honest, its strong claims overreach — but the core mechanism is real and formal: approximate inference is exact inference plus a gap, and the gap has a price.

That price is the quiet thread through everything below.

Occam's razor is a budget constraint

"Occam's razor" usually reads like a stylistic preference: all else being equal, prefer the simple explanation. The formal work — Solomonoff induction, minimum description length — shows it's something colder. Solomonoff's prior assigns a hypothesis a weight proportional to 2 to the negative of the length of the shortest program that generates it. Minimum description length, Rissanen's program, makes model selection compression: the best model of your data is the shortest description that still explains it.

In both frameworks, simplicity wins not because simple things are truer but because simple things are cheaper to verify, cheaper to update, and overfit least. The razor is a budget line drawn through the space of explanations. It is also, on its own, a heuristic — "shortest program" is only cheaply approximated, never computed — which means the razor itself sits inside the same trade we're about to examine.

Two algorithms that say the same thing

Every objective function encodes a model of the world. The objective is where the assumption lives. Two standard examples make this concrete.

Ordinary least squares. We teach OLS in every statistics course, and it's easy to describe as a convenience — the math is clean, the closed form is pretty. But that's not the story. Under the assumption that errors are Gaussian, OLS is the maximum-likelihood estimator, and by the Gauss–Markov theorem it's the best linear unbiased estimator, full stop. The "easy math" and the "most accurate" are the same thing — inside the assumption.

And the assumption is doing all the work. OLS's famous failure mode — a single outlier drags the entire fit — is that Gaussian assumption leaking out. Your data has a heavy tail, one sensor glitched, one typo in the column, and the mathematically optimal estimator is doing exactly what the math says: giving your outlier the weight a Gaussian world would assign it. The robust alternatives — L1, Huber, LAD — cost you more analysis, more computation, more care. You don't get robustness for free, because robustness is a different model of the world, and modeling costs.

Cosine similarity. In vector search we measure distance with cosine similarity rather than raw Euclidean distance in Cartesian space. The precise statement: embeddings live on (or near) the unit sphere, and cosine is Euclidean distance after normalizing away the norm. Meaning in a good embedding is directional; magnitude is mostly scale noise. So cosine throws a dimension away for free — the norm — and buys scale invariance, which is exactly what retrieval needs.

But cosine is wrong when the magnitude is the signal. The choice between cosine and L2 is not a style choice; it's an answer to the question "what is noise in this feature space?" and that answer is a heuristic — a cheap, usually-right guess about which degrees of freedom carry information.

Both algorithms are the same move: a proxy that is optimal conditional on an assumption, where the assumption is cheaper to state than to test.

Which is why heuristics buy you knowledge and bias

This is where the honest part of the argument lives, because the proxy always leaks.

In the energy picture, a heuristic is a coarsened landscape: you've merged many microstates into fewer, wider basins. You move faster, and you can afford more of the world. The cost is that your basins are no longer the true ground state's — you can settle into a local minimum that the exact system would never occupy. That's bias. In a language model, that's hallucination: the next-token distribution is a Boltzmann-like distribution over plausible continuations, and "plausible" is not "true." The model has minimized surprise with respect to the training corpus, not with respect to the world, and the difference is the hallucination rate.

But here's the thing that's easy to miss from the other side: a system with a zero hallucination rate and zero speed is a lookup table. Accurate, and only inside the table. The error rate is the price of generality, and the price is what you're paying for the ability to be right about things you've never seen. Herbert Simon called it satisficing — find the good-enough option, spend the rest of the budget on something else. Gigerenzer's fast-and-frugal heuristics showed the same thing in human judgment: single-attribute rules that beat elaborate models in the right domains, not because humans are clever, but because elaborate models overfit the noise in the domains we can actually test them in.

Every decision you make is a heap of these heuristics. That's not a limitation of decision-making; that's the mechanism of decision-making. The exact answer is a theoretical object; the decision is what you can afford to compute.

The crystal and the sorted file

A correction, because the naive version of this argument is wrong, and it's worth knowing where it's wrong.

The naive claim: "knowledge is low energy." But a crystal is low free energy and boring in an information-theoretic sense. A perfect crystal is maximally redundant — every atom is exactly where the symmetry predicts, so there's no surprise in the state. Shannon entropy near zero. The naive claim predicts a crystal is a deep knowledge. It isn't.

The minimal counterexample is a file with all its 1s first and all its 0s later — sorted binary. Super organized, zero information:

Low free energy — trivially settled, zero work to arrive
Zero Shannon information — the distribution is deterministic, no contingency
Low Kolmogov complexity — "sort it" is the whole program

Pure noise is the mirror image: maximum entropy, high Kolmogorov complexity (you need the bits themselves as the program), no stable structure to settle into. Both are dumb, and in different ways. You can't get to "knowledge" by going to either end.

This curve has a name. Murray Gell-Mann and John Searle, in the Santa Fe Institute research program (The Sciences of the Complex, 1990) and Gell-Mann's The Quark and the Jaguar, defined effective complexity: the information content of the most compressible model of a system. It's zero for a perfect crystal (order: the model is the symmetry itself), zero for pure noise (nothing modelable), and peaking at intermediate "structured complexity." Knowledge sits at or near the peak — your sorted file and your noise file are the two ends of the same curve.

So what separates knowledge from both dumb ends? Not low energy. It's which alternatives were excluded. A sorted file excluded only permutations of the same content — no real alternatives were in play. A noise state excluded nothing — everything is still alive. A known fact excluded real, structural alternatives: different hypotheses about the world, and you picked one, and the picking is what did the work.

That's the refined thesis: knowledge is low free energy achieved by suppressing real, structural alternatives — the work of settling into a specific hypothesis among alternatives that were actually alive, not among permutations of a fixed content. Low F achieved by suppressing permutation entropy (redundancy) is just order. Low F achieved by suppressing real entropy is knowledge.

Local minima can't tell you they're local

The word for the failure mode is metastability. Ice below 0°C is metastable: a local minimum, stable relative to its current context (the container, the lack of a nucleation site), but not the global minimum — liquid water is lower. It's "right" relative to the constraints that got it there, and "wrong" relative to everything else.

That's the hallucination mechanism, precisely. The model settled into a basin where the contextual constraints were relaxed enough that "plausible continuation" and "true continuation" live in the same basin. From inside the basin, both look equally settled. The difference — which basin is deeper, which one contains the ground state — is a global property of the landscape, and it's not visible from inside. This is also distribution shift, epistemic scope, and the Gettier problem, all the same thing from different angles.

And this is now measured, not just intuited. HalluShift (Dasgupta et al., arXiv 2504.09482) shows LLMs drift away from factual accuracy in subtle parts of a response — the model's internal state space shifts toward misinformation in a way you can detect from the trajectory, beating existing baselines. That's the local-basin idea with sensors attached: the detection works by watching the trajectory, i.e., by looking at more of the landscape than the local output.

The cleanest single-sentence version of the whole hallucination problem: a local minimum can't tell you it's local. Knowing you have accuracy is a global, expensive question — it costs the same as looking at the whole landscape, which is exactly the cost of verification. That's why accuracy is not free. Not because accuracy is expensive, but because certainty of accuracy is a different problem, and the local state doesn't show it to you.

There's one more piece, and it's where the crystal is actually doing work for you. A metastable crystal encodes the path that trapped it — the temperature history, the nucleation event, the way it fell. The information content of a low-energy state isn't in the state; it's in the trajectory that got it there. A perfect crystal and a random string have the same information in their states (both boring), but their histories differ radically — one is the result of a long, constrained settling, the other of no settling at all.

Knowledge as stored work — flagged as a synthesis

Here's the synthesis, and I want to be honest that I couldn't find it written down anywhere: I checked, and the unification below is a weld on published parts, not a published result. The pieces are all there — Landauer's price of a bit, Solomonoff's price of a hypothesis, Gell-Mann–Searle's curve — but I don't think anyone has written the sentence. So: treat this as one person's reading of the pieces, not a theorem.

Knowledge is pre-paid structure — a bit you no longer have to pay for.

Landauer says erasing a bit costs heat: that's the thermodynamic price of a bit. A known fact is a bit that's already been paid for — already settled, stored in a low-energy state. You don't re-derive, re-measure, or re-search it. So:

A protein: the work of finding the fold was paid by evolution shaping the landscape; the molecule just falls
A trained model: the work of finding the weights was paid by gradient descent; inference is cheap
A law: the work of finding the regularity was paid by the experiment; applying it is cheap

A low-energy state is, in every case, a configuration where the work has already been done. That's why knowledge looks like a ground state — because, operationally, it is one. And that's why the hallucination problem and the knowledge problem are the same problem from the same side of the table: settling is how you get structure, and settling is how you end up in a basin you can't see the bottom of from inside.

Design for total efficiency, not accuracy

If decisions are heaps of heuristics, then system design is not about maximizing accuracy. It's about allocating compute to match the cost of error. The accuracy you need is the accuracy the decision can survive losing, and the compute you spend is the cost of buying the rest.

This is where the interesting engineering is, and it's worth being precise about the cases:

Cascades and routing. A cheap model answers first; a hard model fires only when the cheap model is uncertain. You're paying for deep analysis only where the cost of error justifies it.
Speculative decoding (Leviathan, Kalman, Matias, arXiv 2211.17192). A cheap draft model proposes continuations; an expensive model verifies them in parallel, and rejection sampling keeps the output distribution exactly identical to the expensive model alone. This is the counterexample that sharpens the whole argument: the efficiency/accuracy trade is strictly beatable when you have an exact verifier. The reason it's not free is that verification is itself compute — and in most real problems, verifying a claim costs more than searching for one, which is exactly the regime where the hallucination rate lives.
Early exit, mixture-of-experts, adaptive test-time compute. All the same knob: how much analysis does this input deserve?

And the knob has a name. The sampling temperature in the Boltzmann distribution, P(state) ∝ exp(−E/T), is the depth-of-analysis dial, literal and unironic. Low temperature: deep, expensive, close to the ground state. High temperature: fast, cheap, approximate. The same equation that prices a protein's fold prices your model's willingness to settle.

Where this leaves AI

The current instinct — throw more parameters, more data, more compute at accuracy — is the right instinct inside a fixed model of what the problem is. But the next-order problem isn't accuracy. It's calibrated depth: a system that knows how much analysis a given problem deserves and spends exactly that. A router that knows when the cheap model is right. A verifier that knows when it can check for free and when it can't. A temperature that's set by the cost of the mistake, not by default.

That's not a new architecture. It's the design principle the rest of this post has been building: accuracy is not free, the price is information, and the goal is never the exact answer — it's the right answer at the right cost.

Evolution already solved this. It selected for organisms that compute the right approximation fast, not the exact answer slowly. The organisms that got the exact answer eventually, precisely, and expensively are not here to be designed.

Design for the budget. The budget is the thing.