If you knew exactly where every molecule of air was and how fast it was moving, could you know what happens next? For a famous weather model the answer is yes in principle: chaos makes small errors grow, but a perfectly known start gives exactly one future. One of the 719 papers OpenAI released on October 6, 2026 claims that the equation physicists use for a gas of colliding particles — Ludwig Boltzmann’s, from 1872 — does not always work that way. From one carefully built starting state it constructs two different futures, both obeying conservation of mass, momentum and energy and the law that entropy never decreases. The claim is less of a shock than it sounds and more interesting than a shock. Whether the equation picks one future had been an open question for decades, and three mathematicians conjectured this outcome seven weeks before the release. The starting state is extreme, packed into ever-thinner jets. And only a weaker version of the claim is machine-checked. What it would show, if it holds, is that a model can obey every physical law we impose and still need a rule nobody has written down to decide what happens next.
In 1963 the meteorologist Edward Lorenz published a three-variable model of air in convection and found that it never settles into a pattern: two runs that start a hair apart end up completely different.[11] He later framed it as a question about a butterfly in Brazil and a tornado in Texas.[12] Chaos, though, has a precise limit. In Lorenz’s model a perfectly known starting point still leads to exactly one future; the trouble is only that we never know the start perfectly. The companion drawing to this case shows both halves of that.[14]
Ludwig Boltzmann’s equation, published in 1872, works one level down. It describes a gas as an enormous population of particles that fly and collide, and it came with his H-theorem: a quantity tied to entropy that only ever moves one way.[5] It is among the foundations of how physics models air, and in 2024 it was derived rigorously from Newton’s laws of colliding spheres for as long as its solutions exist.[13] Whether it always picks a single future was a different matter. In 2002 Cédric Villani listed it plainly: “many fundamental questions are still unsolved: in particular uniqueness.”[7] A 2022 survey said the same of the standard solution theory: “The uniqueness of solutions within this class is currently unknown.”[8]
Then a prediction. On August 13, 2026, three mathematicians described the uniqueness of these solutions as “a longstanding open problem of some importance” and conjectured that they “are not unique for a dense class of initial data.”[9] Seven weeks later, OpenAI’s release included two papers claiming a version of exactly that: one starting state of a hard-sphere gas, two different futures.[1][2] The stronger paper, dated October 5, imposes everything a physicist would ask for — exact local conservation of mass, momentum and energy, and the entropy law — and still finds two solutions.[1] Its own framing: “A weak solution theory must retain enough of these properties to distinguish the equation from a balance law with an unidentified collision defect.”[1]
The fine print matters, and the papers state it. The starting state is not ordinary air. In the earlier paper’s words, it “has bounded velocities but unbounded amplitudes concentrated at successively smaller spatial scales” — narrow jets packed into thinner and thinner rings, concentrating at a point.[2] That is how it escapes every known uniqueness theorem, each of which assumes the gas is better behaved.[1] And the machine check covers only the earlier, weaker claim; the scope note says “later local-conservation refinements are outside it.”[3]
Physicists have long been wary of reading too much into results like this. Villani’s review quotes the mathematician Mark Kac, decades earlier: “the preoccupation with existence and uniqueness theorems for the Boltzmann equation appears to be unjustified on grounds of physical interest and importance.”[7] The usual reading of a second solution is not that nature is undecided. It is that the equation, at that level of description, has not been told enough — that some additional selection rule is needed to say which solution is the physical one.
Fluid dynamics has already been through this. From 2009 onward, mathematicians showed that the equations of ideal and viscous fluids admit multiple solutions in weak senses, and that the extra condition everyone hoped would pick the right one does not.[10] In the words of one survey, the local energy inequality “does not uniquely select a solution to the initial value problem.”[10] If the Boltzmann claim holds, the gas equation joins that list, with an unusually strong set of conditions imposed — and the search for the missing rule moves from fluids down to the particles beneath them.
None of this has been checked by people yet. This case found no public comment from any specialist in kinetic theory, and coverage of the release does not mention it. It is the fourth result from that release in this library, and it has the clearest precedent of the four: the outcome had been predicted, in writing, by people who knew the problem best.[9] What it lacks is what the others lack — someone, other than the machine, standing behind it.
A gas as colliding particles, with the H-theorem: entropy moves one way.[5]
The EquationWeak solutions of the fluid equations shown not to be unique; the energy condition does not select one.[10]
The PatternThree mathematicians conjecture that Boltzmann solutions are not unique for a dense class of starting states.[9]
The ConjectureA gas as colliding particles. Derived rigorously from Newton’s laws for long times only in 2024.[5][13]
Conjectured in print on August 13, 2026; claimed in OpenAI’s release on October 6.[1][9]
Many fundamental questions are still unsolved: in particular uniqueness. — Cédric Villani, 2002, on the Boltzmann equation [7]
| Dimension | Evidence |
|---|---|
| Quality (D5) Origin · 82 | Two solutions from one starting state, with exact local conservation and the entropy law imposed. Uniqueness had been a listed open question since at least 2002.[1][7]A Model That Doesn’t Pick |
| Regulatory (D4) L1 · 74 | The machine check covers the September claim; “later local-conservation refinements are outside it.” No specialist review found.[3]The Checked Half Is the Weaker Half |
| Operational (D6) L1 · 70 | The start “has bounded velocities but unbounded amplitudes concentrated at successively smaller spatial scales” — outside every known uniqueness result.[1][2]Built to Escape the Theorems |
| Employee (D2) L2 · 62 | Conjectured in writing on August 13, 2026 as “a longstanding open problem of some importance.” No public specialist comment since the release.[9]Predicted, Then Silent |
| Customer (D1) L2 · 50 | The start is far from real air, and working models already rely on extra rules. The result is about what the equation alone determines.[2][10]Forecasts Are Unaffected |
The cascade originates in D5 (Quality): whether a foundational model of a gas determines its own future, a property assumed in practice and never proven.[7][8] It propagates to D4 (Regulatory), where the strongest form of the claim is not machine-checked and no specialist has reviewed it,[3] and to D6 (Operational), where the construction is new and narrow — a starting state of concentrating jets, built to slip past every existing uniqueness theorem.[1][2] It lands in D2 (Employee), the specialists who had predicted it in writing and have not yet responded,[9] and in D1 (Customer), the forecasters and engineers who use these models — for whom nothing changes in practice, and whose models already assume extra rules. D3 (Revenue) is not scored.
Read alongside the three earlier cases from the same release, this one changes the question. UC-328, UC-329 and UC-330 asked how people respond to a machine’s proof. This one asks what a proof can tell us about a model at all. A machine can show that an equation allows two futures. Deciding which future nature picks — and what rule, missing from the equation, makes it pick — is still physics, and still people. The honest limit: the starting state is far from any real gas, the headline claim is unformalized, and this case does not say the proof is correct.
-- UC-331: Same Gas, Two Futures: 6D Diagnostic Cascade
-- Boltzmann equation 1872 (H-theorem). DiPerna-Lions 1989: global renormalized solutions exist; uniqueness open (Villani 2002: many fundamental questions still unsolved, in particular uniqueness; Silvestre 2022: uniqueness within this class currently unknown). Gismondi-Golding-Novack, Aug 13 2026: conjecture renormalized solutions are not unique for a dense class of data. OpenAI release Oct 6 2026, result 363: Sep 23 paper (65 pp, machine-checked) two renormalized solutions, local mass and total momentum conserved, energy and entropy inequalities; Oct 5 paper (78 pp, not machine-checked) two global entropy solutions with exact local conservation of mass, momentum and energy. Starting state: unbounded amplitudes concentrated at shrinking scales. Not convex integration. Not withdrawn or revised. No specialist comment found. Contrast: Lorenz 1963 model is chaotic but has unique solutions.
FORAGE same_gas_two_futures
WHERE foundational_model_uniqueness_open = true
AND nonuniqueness_conjectured_before_claim = true
AND strongest_claim_machine_checked = false
ACROSS D5, D4, D6, D2, D1
DEPTH 3
SURFACE same_gas_two_futures
DIVE INTO model_does_not_select
WHEN all_conservation_laws_imposed = true
AND starting_state_singular = true
TRACE missing_selection_rule_cascade
EMIT same_gas_two_futures_signal
DRIFT same_gas_two_futures
METHODOLOGY 86
PERFORMANCE 36
FETCH same_gas_two_futures
THRESHOLD 1000
ON MONITOR CHIRP high 'An AI-written paper claims Boltzmann's 1872 gas equation can produce two different futures from one starting state while obeying exact conservation of mass, momentum and energy and the entropy law. Uniqueness was a listed open problem and non-uniqueness was conjectured in print seven weeks earlier. The starting state is extreme, and only a weaker version is machine-checked. If it holds, the model needs a selection rule nobody has written down.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
Lorenz showed that small errors grow. A perfect start still gave one future. The claim here is different in kind: a perfect start, every conservation law obeyed, and two futures.[1][11]
Specialists had named uniqueness as unsolved for decades and conjectured this outcome in print seven weeks earlier. The machine did not overturn expectations; it may have confirmed one.[7][9]
A machine-checked statement exists, but for the September paper. The headline claim, with exact local conservation, is outside it. Both facts belong in any summary.[3]
If an equation allows two futures, something outside the equation must choose. Finding that rule is physics, not proof — and it is still a human question.[10]
Primary sources are OpenAI’s two papers, machine-check note and change history; Villani’s 2002 review, Silvestre’s 2022 survey and the August 2026 conjecture; and the original papers for the history. All were read directly. No specialist comment on this result was found in a search of the public record to October 14, 2026. This case does not claim the proof is correct. Source 14 is an interactive companion visual made for this case.
This case separates a model that is chaotic from one that is undetermined — the first is hard to predict, the second needs a rule it doesn’t contain. The same test applies to forecasts, policies and plans. Book a discovery call to find where your models are relying on a rule nobody wrote down.