00 · THE OBITUARY
Two walls, a skull, and a conclusion neither paper signed.
A LinkedIn post by Stark Burns, “The AI Hype is a Dead Man Walking. The Math Finally Proves It,” is still circulating more than a year after publication. Its structure is memorable. One wall is physics: an enormous compute requirement. The other is reason: chain-of-thought as a brittle mirage. The image supplies a skull; the prose supplies an obituary.
The post says it draws on a video summary of two papers. I went to the papers. I take their findings more seriously than the obituary does, which is why I reach a narrower—and more useful—conclusion.
The papers identify brittle components. The post pronounces the death of a paradigm. The missing step is the system.
Both papers are reasons to distrust a language model used alone. Neither paper establishes that an AI-assisted research system—human-directed, adversarially checked, mechanically verified, and signed—must inherit every weakness of the model inside it.
01 · THE EXPONENT
The first wall is a scaling consequence, not a law of nature.
The figure 1020 comes from Peter Coveney and Sauro Succi’s The Wall Confronting Large Language Models. Their argument deserves its full force. Under the scaling relation they examine, a very small exponent makes reliability improvements brutally expensive. For a power law
error ∝ compute−α
reducing the error by a factor of ten requires a compute multiplier of 101/α. The sensitivity is the argument:
| Illustrative exponent α | Compute multiplier for 10× lower error |
|---|---|
| 0.05 | 1020 |
| 0.10 | 1010 |
| 0.20 | 105 |
| 0.30 | ≈ 2.15 × 103 |
The last two rows are sensitivity illustrations, not empirical estimates from Coveney and Succi.
The paper reports α ≈ 0.05 for computational cost and uses α = 0.1 as a generous reference case. It also explicitly warns that language-model loss is not the same thing as approximation error. Kaplan and colleagues’ classic scaling work likewise concerns cross-entropy loss, not the error rate of every task a user may care about.
That distinction matters. A fitted exponent extended far beyond the observed regime is evidence of severe diminishing returns under a specified scaling story. It is not a timeless physical constant governing every architecture, task, toolchain, or verification layer.
Most important, Coveney and Succi do not sign the obituary attached to their number. Their abstract says the degenerative path they describe need not be inevitable across all future AI research. The viral version kept the exponent and lost the boundary of the claim.
02 · THE CONTROLLED PART
The second wall is a serious result about a controlled mechanism.
Chengshuai Zhao and colleagues’ Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens builds an isolated environment called DataAlchemy. The authors train models from scratch and probe what happens when test problems shift away from the training distribution across task, length, and format.
The result is not cosmetic. Chain-of-thought performance degrades sharply as distribution discrepancy grows. In that controlled setting, the mechanism is brittle.
But the study’s clean design is also the boundary of what it establishes. A model trained for an isolated experiment is not identical to a deployed research process that can retrieve sources, call tools, generate counterarguments, execute exact checks, and defer final judgment to a human.
A controlled failure of one component is evidence against trusting the component. It is not yet a theorem about every system that contains it.
This is not an escape hatch for bad AI work. It is a demand for the correct unit of analysis.
03 · THE RIGHT UNIT
Do not ask whether the witness is infallible. Inspect the court.
Grant both papers their strongest reading. Together they support a proposition I accept without reservation:
A single model, used alone and trusted without checking, is not reliable enough for consequential work.
That is not how consequential work should be organized, and it is not how I organize mine.
“Can the model be trusted?” is the wrong systems question, just as “Can a single witness be trusted?” is the wrong question for a court. A court is not reliable because witnesses never err. It becomes more reliable when testimony is opposed, documents are checked, claims meet rules of evidence, and a responsible person signs the judgment.
A brittle part inside a sturdy system is an engineering commonplace. Aircraft carry redundant sensors because sensors can fail. Scientific instruments are calibrated because measurements drift. Courts cross-examine witnesses because memory is fallible. Reliability does not require perfect components. It requires an architecture in which component failure is expected, exposed, and contained.
04 · HOW I DIRECT AI
Orchestration is not prompting. It is governance.
I do not treat a model’s first answer as a draft that merely needs polishing. I treat it as an untrusted proposal entering a controlled process.
- 01
Frame the claim.
I decide what question is being asked, what would count as an answer, which facts are privileged, and where the system must stop.
- 02
Force divergence.
Different model passes generate the strongest version, the strongest objection, the missing assumption, and the counterexample. Agreement is not treated as validation.
- 03
Turn discomfort into a test.
My first pass is triage. I scan for structural anomalies: a number that sits wrong, a sentence that outruns its source, a timeline that cannot coexist with itself. Suspicion becomes a specific check.
- 04
Move decisive claims outside prose.
Mathematical claims go to exact arithmetic or proof. Historical claims go to the source. Timeline claims go to the calendar. A verifier that cannot be persuaded by eloquence is more valuable than another fluent opinion.
- 05
Adjudicate and sign.
The models do not own the conclusion. I resolve conflicts, disclose the assistance, preserve what can be rerun, and accept responsibility for the published text.
One public example is the 39-line exact-arithmetic script published with my threshold analysis. It enumerates all 1,024 binary states and verifies that an explicit higher-order perturbation preserves all ten marginal probabilities and all forty-five pairwise distributions. The point is not that code makes a claim true. The point is that the decisive calculation has been moved into a form a reader can rerun, inspect, and try to break.
05 · TWO FIELD NOTES
The best evidence for the method is where it failed.
Field note A · A hostile error
We were both wrong.
While revising a satire about two rejection decisions, I told a model that “five days” was wrong and that the correct figure was thirty-six hours. It made the edit. The edit was also wrong. Thirty-six hours described one review; several days separated the two decisions. The model merged two intervals, and I had pointed it at the wrong one.
I supplied the dates—1 and 5 November—and asked whether the story still held together. The model retracted the change, replaced both disputed intervals with the dates themselves, and flagged a separate accusation about deliberate late-night submission. Only I could settle that claim, because intention was my fact to report. I then added a time-zone fact neither pass had used: eleven at night in Texas was five in the morning in London.
No component was reliable. The process converged because neither component was allowed to win by assertion.
Field note B · A flattering error
Praise is an adversarial input too.
A twelve-page “evaluation” of my work reached me in the confident register of synthetic institutional prose and under the name of a committee that does not exist. Much of it was flattering. Several numbers were correct. One quotation was invented, and the report inferred my motives from the invented sentence.
I did not publish the praise. I sent it through the same checking process I would use for criticism.
A system that catches only the mistakes against its operator is not a verification system. It is a vanity machine. The dangerous hallucination is often the one we wanted to be true.
06 · SKIM OR AUDIT
The same incomplete reading can feed a verdict or a check.
A reviewer of an earlier paper wrote that they had read the abstract twice, skimmed the remainder, and concluded that the manuscript had been written by a machine.
I skim too. The difference is what follows the skim.
They skimmed and issued a verdict.
I skim and start a verification.
A skim that terminates in a conclusion is a guess with confidence attached. A skim that produces a falsifiable question is the opening move of an audit. The distinction is not speed. It is the discipline imposed on what happens next.
07 · LIMITS
A sturdy system is not an infallible one.
Three constraints remain.
- Orchestration does not manufacture expertise. Human anomaly detection depends on long contact with the material. A workflow multiplies the operator’s judgment, including the operator’s blind spots.
- Model plurality is not independence. Models trained on overlapping data may repeat the same error with different voices. That is why the final constraint should be external to their rhetoric: a source, a proof, exact arithmetic, a calendar, or a person who was there.
- Checkable is not the same as correct. A reproducible error is still an error. But checkability is the condition that lets another person find it, name it, and repair it.
None of this proves that any particular AI-assisted work is good. It specifies what the work must expose before “AI-assisted” can become a methodological description rather than a marketing label or an accusation.
08 · VERDICT
The skull is attached to the wrong body.
“The math finally proves it” is a powerful headline. The mathematics in these papers shows something real: some model improvements are staggeringly expensive under the observed scaling relations, and chain-of-thought can collapse under distribution shift.
It does not prove that every system built with language models is dead. It proves that the model cannot be permitted to stand in for the system.
The hype and the obituary share the same category error. Each measures a component and announces a verdict about a whole. The quality of AI-assisted work is not settled by what a model can do alone, by how human the prose sounds, or by which institution stamps the result. It is settled by the architecture of challenge and verification around the model—and by whether a human being is willing to sign the record.
Brittle parts. Sturdy systems.
The skull is excellent design. It is simply attached to the wrong body.
SOURCES & RECORD
Primary materials.
- Stark Burns, “The AI Hype is a Dead Man Walking. The Math Finally Proves It”, LinkedIn, 2025.
- Peter V. Coveney and Sauro Succi, The Wall Confronting Large Language Models, arXiv:2507.19703, 2025.
- Jared Kaplan et al., Scaling Laws for Neural Language Models, arXiv:2001.08361, 2020.
- Chengshuai Zhao et al., Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens, arXiv:2508.01191, 2025.
- Tony T. Nguyen, From Moral Possibility to Justified Action, Highland Reflexive Press research record, 2026; core lemma source published with The Peacock Bites the Owl.
AI disclosure
This essay was developed with generative-AI assistance under the method it describes. The author set the argument, supplied the field record, directed the revisions, checked the cited claims against primary materials, rejected unsupported formulations, and accepts responsibility for the final text.