The Physicist’s Chalk

The word “intelligence” absorbs at least six distinct capacities, and most disagreements about who or what is intelligent are disagreements about which one is being measured.

Memory is storage. Information held so it can be recovered later. It is evaluated against the past, and its virtues are fidelity and capacity.

Retrieval is access. Getting the right item out of storage at the right moment. A system can have excellent memory and poor retrieval — the fact is in there, you cannot reach it — and this is such a common human experience that we have a name for it involving the tip of the tongue.

Reasoning is the manipulation of representations according to rules. Deduction, arithmetic, formal inference. It operates on what is present and is judged by validity.

Prediction is the production of a commitment about something not yet observed. It is judged by what arrives.

Compression is the discovery of structure that lets many observations be represented by less than themselves.

Intelligence, on the thesis of this series, is the last two joined: compression that continues to predict as Actual moves forward.

There is a seventh, which the list above deliberately omitted and which the scene depends on: transcription — getting the intended symbol onto the surface. It is not cognition at all. It is motor output, and it is where the exponent went wrong.

The physicist’s error is a transcription failure, possibly a retrieval one. It has essentially no bearing on the fifth and sixth capacities, which are what he is in the room to do. He is attempting to find a compact structure that accounts for an enormous class of phenomena. The chalk is a transcription device for that work, and transcription devices slip.

Everyone in that seminar knows the difference between the search and the transcription. That is why nobody reaches for a conclusion about his intelligence: they are watching him do the thing that is hard, and he made an error at the thing that is easy.

And here is the part that should make you uncomfortable, because it inverts a common intuition: the easy thing is the one machines have always been better at. Perfect clerical fidelity is what we build computers for. Copying a symbol without altering it is the one capability that has been solved since the 1940s. It is not on the frontier and it never was.

On March 1, 2023, in a personal injury case against Avianca in the Southern District of New York, attorney Peter LoDuca filed an affirmation in opposition drafted by his colleague Steven A. Schwartz. It cited Varghese v. China Southern Airlines, Shaboon v. Egyptair, and Martinez v. Delta Airlines, among others. Six of the cited cases did not exist. A seventh, Zicherman v. Korean Air Lines, was a real Supreme Court decision attached to a fabricated Eleventh Circuit citation. Schwartz had used ChatGPT, and when he asked it whether the cases were real, it said they were. On June 22, 2023, Judge P. Kevin Castel imposed a $5,000 penalty jointly and severally on Schwartz, LoDuca, and their firm.¹

This became the canonical story about AI unreliability, and rightly so — it is a real professional failure with real consequences, and by mid-2026 a public database tracking such incidents in court filings had logged nearly two thousand of them.²

But look carefully at what failed, and in what order.

The register was right. The framing of the tolling question, the shape of the appeal to precedent, the rhythm of American appellate prose — close enough to work product that two attorneys, and briefly a federal court, did not catch it on the surface. What failed first and hardest was the citations: the specific, arbitrary, memorizable strings that anchor an argument to the record.

Only when Judge Castel went and read the fabricated opinions closely did the deeper failure appear. He found Varghese showed “stylistic and reasoning flaws that do not generally appear in decisions issued by United States Courts of Appeals,” and that “its legal analysis is gibberish.” That distinction is worth holding onto rather than smoothing over: the surface form of legal writing was compressed beautifully, the citations were not compressed at all, and the internal logic of a fabricated opinion — a document that had to be invented whole — degraded under scrutiny in between.

That middle layer is the physicist’s exponent, at scale and with worse consequences. The model predicted legal-brief-ness extremely well. It predicted this particular case citation extremely badly. These are not two grades of the same capacity. They are different capacities, and the second one is the clerical one.

The reason is now derivable from everything in the previous four parts, and it is not a mystery requiring special pleading.

Compression works by finding structure across many examples. The structure of a legal argument is present in millions of documents; it is enormously redundant; it compresses beautifully. The citation “Varghese v. China Southern Airlines, 925 F.3d 1339 (11th Cir. 2019)” — if it existed — would appear a handful of times in a corpus, or once, or never. It is incompressible relative to the rest. There is no structure connecting it to anything else. It is a fact that has to be stored, not derived.

And we established in Part I what storage costs. A model has finite capacity — on recent measurements, roughly 3.6 bits per parameter — and the corpus is vastly larger. Something has to go. What goes first is precisely the material with the least structure: isolated proper nouns, dates mentioned once, arbitrary identifiers. What survives is the material with the most: syntax, genre, argument shape, the geometry of how legal reasoning moves.

The recent theoretical work makes this quantitative rather than hand-waving. Kalai and Vempala proved that a calibrated language model must hallucinate at a rate close to the fraction of facts appearing exactly once in the training data — the “monofact” rate — even assuming perfect training data with no errors.³ The follow-up paper generalizes it: if 20% of birthday facts appear exactly once in pretraining, one expects base models to hallucinate on at least 20% of birthday facts.⁴

This is not a bug report. It is a lower bound. Under the stated conditions, hallucination on singleton facts is not a defect to be engineered away; it is a theorem about what compression does to information that appears once.

That is the sharpest possible statement of the thesis in this part:

A hallucination is not necessarily the absence of intelligence. It is evidence that predictive compression has lost resolution at that level of detail.

There is a tempting picture here and I want to inspect it before letting it stand.

The picture: the model learned a curve, the true facts are points near the curve, and a hallucination is the model reading off the curve where the actual point sat slightly off it. Smooth structure, rough reality, plausible-but-wrong output.

As a teaching image for the resolution point, this survives. It captures the essential thing — that a compressed representation reproduces the general shape and loses the individual deviation.

As an explanation of hallucination, it is inadequate, and the literature is emphatic that hallucination is multi-causal. The surveys distinguish intrinsic hallucination (output contradicting the source) from extrinsic (output that cannot be verified from the source at all), and organize causes into three families: those arising from data, from training, and from inference.⁵ The mechanisms include:

• Probabilistic generation. Sampling from a distribution produces low-probability tokens sometimes. That is what sampling means.
• Incomplete or contradictory training evidence. The corpus contains errors, and it contains gaps.
• Uncertainty the model cannot express. A model asked a question it has no basis for answering still has to produce a distribution over tokens, and that distribution has a maximum somewhere.
• Objective mismatch. The training objective is next-token likelihood. It is not truthfulness. These coincide often and not always.
• Inability to distinguish known from unknown. This one is more nuanced than usually reported. Models turn out to have partial access to their own uncertainty — Anthropic researchers found large models are reasonably well-calibrated on multiple-choice and can self-evaluate the probability that their own answer is correct, though poorly zero-shot.⁶ The signal exists. It is not reliably used.
• Incentives to answer rather than abstain. This is the finding that reframes the whole problem.

On that last point: Kalai and colleagues show that hallucination persists after pretraining largely because of how we grade. Under binary scoring — right or wrong, one point or zero — “abstaining is strictly sub-optimal. IDK-type responses are maximally penalized while an overconfident ‘best guess’ is optimal.”⁷ Every major benchmark scores this way. We have spent a decade running an optimization process whose scoring rule punishes admitting ignorance, and then expressed surprise that the resulting systems do not admit ignorance.

Notice that this is exactly the failure Part III warned about. A binary grade is not a proper scoring rule. It does not make honesty optimal. The log score does, and nobody grades on it.

One more thing to flag rather than assert: exposure bias — the mismatch between training on ground-truth prefixes and generating from one’s own outputs — is the standard textbook explanation, and it is genuinely contested. Some work finds it drives hallucination under domain shift; other work directly challenges “the popular belief in exposure bias,” finding distortion limited and models capable of self-recovery.⁸ I include it as a candidate mechanism, not a settled one.

The point of enumerating all this is not completeness. It is to refuse the single-cause story in both directions. Hallucination is not simply “the model is dumb,” and it is not simply “compression lost resolution.” It is a family of failures with a family of causes, some of which are architectural, some statistical, and at least one of which is our own fault for grading badly.

I want to be exact about the boundaries of this argument, because it is easy to convert into an excuse and I am not making one.

It does not license the failures. A fabricated citation in a filed brief is a serious harm. A wrong drug interaction is a worse one. The practical consequences of losing resolution at the level of specific fact are severe, and they scale with how much the specific fact mattered. Nothing in this analysis reduces the obligation to fix them.

It does not claim the model “really knew.” There is no hidden correct answer that was garbled in transmission. If a fact was never compressible and never stored, there is nothing in there to have known.

It does not make hallucination a mark of sophistication. It is a failure. It is simply a failure of a specific capacity, and identifying which capacity is what lets you fix it.

What it does license is a correction to inference. From “this entity produced a false specific fact,” you may conclude that its retrieval and storage are unreliable at that resolution. You may not conclude that its predictive compression is poor, because those are separate capacities and you have measured only one of them.

And the reverse holds with equal force. From “this entity recalled a great many specific facts correctly,” you may not conclude that it is intelligent. That was Part II’s thousand points: perfect performance on the record, silence past the edge.

Once the capacities are separated, the fix becomes obvious and is exactly what the field has done.

If predictive structure and exact retrieval are different capacities, do not demand both from the same mechanism. Give the compressed predictor access to an uncompressed store.

This is retrieval-augmented generation: fetch the actual documents, put them in context, generate against them.⁹ It is also tool use — a calculator for arithmetic, a search index for facts, a database for records. It is the same solution the physicist uses when he checks the exponent against his notes, and the same one the lawyer was supposed to use when he pulled the cases from Westlaw before filing.

The results are instructive precisely because they are partial. Legal research tools built on retrieval hallucinate substantially less than an unaugmented model and still hallucinate: a Stanford study of three commercial systems found rates spanning roughly 17% to 33%, against a considerably higher unaugmented baseline.¹⁰ Retrieval reduces the problem. It does not dissolve it, because the model must still decide what to retrieve, whether the retrieved material is relevant, and how to represent it — and each of those is a predictive judgment that can go wrong.

The architecture that emerges is not one intelligence. It is a division of labor:

• Compressed predictive structure — generates, generalizes, reasons about the shape of things.
• Uncompressed external store — holds the arbitrary specifics that were never compressible.
• Retrieval — moves specifics into the predictor’s working context.
• Verification — checks that the two agree before anything is acted on.

That is also, and not coincidentally, how a competent professional works. The expert does not hold the case law in their head. They hold the structure of the law in their head and know where the case law lives. The skill was never in the storage.

The physicist writes R³, having been corrected, and continues.

What he is doing in that room is trying to find a compression — a compact structure that accounts for an enormous class of phenomena, that will still be accurate when applied to phenomena nobody has measured yet. Success would be a relationship holding across regimes never tested, which is the highest-value output a mind can produce and the only one that cannot be automated by looking things up.

The exponent is a transcription error inside that search. It costs four seconds.

Which one you notice tells you which theory of intelligence you are actually running. And most people, most of the time, are running the wrong one — not because they believe it, but because Part IV explained why: the correct exponents produced no surprise and therefore no experience, and the wrong one produced both.

There is one distinction left to draw. The physicist, if he succeeds, will end up with something he can write on a board — a compact expression another human can read, check, and carry away. The machine that compresses a corpus ends up with something nobody can read at all.

Both are compressions. Both are mathematical. They differ in a way that matters enormously and that is almost always described wrong.