Schneider uses the Navier-Stokes finite-time blow-up proof to argue that AI predictions can be trusted only inside an auditable causal chain

The article separates scientific activity into episteme (explanatory understanding, where understanding is the goal itself) and techne (the craft of predicting and making), and argues that whether the two can be separated depends on how easily predictions can be checked. Relative to discussions that treat AI capability as broadly progressive or threatening, this distinction supplies a criterion for whether AI can be trusted in a given task: black-box models can stand alone in daily-checkable weather forecasting, but not where verification is too slow, costly, or dangerous. An argumentative article, illustrated by AlphaFold predictions checkable against experimentally determined structures and by AI weather forecasts checkable daily; it is a conceptual framework rather than an empirical study.

The article notes that although the finite-time blow-up proof for the forced Navier-Stokes equation is supported by Lean formal verification, mathematicians are still working to digest it, and the result says little about modeling, predicting, or understanding turbulence. It separates 'the proof is correct' from 'the reasoning advances collective understanding of methods and structures,' and explains that this singular solution differs from turbulence, in which viscosity dissipates energy at a small but finite Kolmogorov scale and which is the ubiquitously realized physical phenomenon. Cites Terry Tao, Timothy Gowers, and Yehuda Rav, and notes that the unforced version of the problem remains open; no proof details are given, so this is a commentary judgment.

The article argues that end-to-end AI weather models are unsuited to decadal climate projection because greenhouse gas concentrations are typically not among their inputs, and even if they were, no observations exist to learn the response; such models also do not enforce conservation laws such as energy, so long integrations can drift by accumulating errors. It sharply distinguishes 'forecasts checkable daily' from 'climate projection,' the task of predicting how weather statistics change over decades in response to a forcing, and notes that the former lack a stability, consistency, and convergence theory (no analog of von Neumann stability analysis or the Lax equivalence theorem) and that how an initial condition becomes a forecast is difficult to audit. An argument grounded in established physical modeling and numerical analysis theory, without new numerical experiments or statistical results.

The article proposes embedding AI in auditable scaffolds: use error-controlled numerical methods for the resolved large scales, learn closure models for unresolved subgrid scales from data, and let AI agents run the closure search in a closed loop, proposing symbolic closures, testing them against high-resolution simulations, observations, or experiments, and revising them. Against end-to-end alternatives, this route grounds trust in individually testable links, and notes that large scales are where climate change moves the system out of the current distribution, whereas small-scale physics obeys the same local laws in a warmer or colder climate as in today's. Draws on decades of practice in which climate models and CFD codes are built from resolved dynamics with embedded closures, and mentions that neural network closures can be individually tested and partly interpreted and that symbolic regression (as in SINDy) can sometimes produce more easily interpretable closures.

The article addresses settings where predictions must be trusted before they can be verified, such as decadal climate projection and certification of a novel aircraft; in these settings the author argues for placing AI at links of an auditable chain, using error-controlled numerical methods for resolved large scales and learning individually testable subgrid closures from data. For short-range weather forecasting that can be checked daily, the author explicitly holds that black-box end-to-end models can stand on their own, so the argument does not seek to restrict such uses. For researchers and practitioners seeking to bring AI into Earth-system or engineering prediction, the article offers a framework for how to use it and where trust comes from.

The article's discussion of the forced Navier-Stokes finite-time blow-up proof rests on a public announcement and formal verification, and the author himself notes mathematicians are still digesting the result, so its mathematical significance remains for the community to assess. On the limits of end-to-end AI climate projection, the article offers an argument from physical and numerical-analysis theory rather than quantitative tests of specific models. In addition, the text read here is an incomplete version lacking figures and reference details, so readers needing to verify specific citations or numbers should return to the original. The author also notes that closure models may face renewed extrapolation problems where available data do not span the conditions of intended prediction, with local high-resolution simulations for offline calibration and uncertainty quantification as mitigation.