Article

Why the Standard Model is so hard to put on a computer

For curious readers, an AI-assisted physics proposal explains the weak force’s grid problem—and the gap between a mathematical rule and a simulator.

Editorial illustration for Why the Standard Model is so hard to put on a computer: a geometric block represents a model release. Not documentary evidence.

Physicists already make successful Standard Model calculations. The puzzle is why a straightforward spacetime grid introduces unwanted particle species—and how to preserve the weak force’s different treatment of left and right. In a November 15, 2026 account, Nathaniel Craig explains an AI-assisted proposal for a remaining mathematical consistency rule, not a ready-to-run simulator.

The underlying preprint was submitted on October 6; November brings an explanation of that earlier work. For physics students and curious readers, the surprise is that knowing the laws of nature and having a complete recipe for calculating their consequences are different achievements.

What physicists can already calculate

The Standard Model describes elementary particles and the electromagnetic, weak and strong forces. It has an extensive record of successful predictions; gravity is outside its scope, as CERN’s overview explains. This proposal does not rescue an unusable theory or turn it into a theory of everything.

One familiar calculation method starts with a simpler problem and adds interactions as a series of corrections. That is perturbation theory. It can work remarkably well, but physicists also need methods that do not depend on interactions being small. Those are called nonperturbative methods. Craig’s account explains this distinction.

A lattice replaces continuous spacetime with a grid for the calculation. This is already productive in the strong-force part of the theory: a 2024 Department of Energy account describes lattice calculations of how mass is distributed inside protons and pions. The unresolved task concerns a complete formulation that also preserves the weak force’s distinctive rules.

Why the weak force complicates the grid

Particles such as electrons and quarks have a property called chirality, expressed as left- or right-handedness. The weak force treats these differently. A successful calculation must preserve that distinction—not merely produce numbers. Craig’s introduction sets out why this makes the full theory harder to put on a lattice.

Straightforward grid equations can introduce extra particle species, a problem known as fermion doubling. These are unwanted features of the mathematical description, not newly observed particles. Removing them while keeping the required symmetries is the difficulty; simply asking for more computing power does not supply the missing mathematical rule.

The groundwork predates today’s AI. Earlier approaches, including overlap fermions, addressed the doubling obstacle. Martin Lüscher’s 1998 paper supplied a consistent quantum measure for a simpler class of chiral theories, under specified conditions. That was not the full Standard Model result. Craig’s proposal builds on this programme rather than beginning from a blank page.

What the proposed phase rule changes

A quantum calculation needs a rule for combining contributions from possible field configurations. The mathematical measure is part of that rule. These contributions are not simply positive probabilities: their phases matter. Phase is what determines how quantum contributions reinforce or cancel one another.

In the lattice approach Craig uses, the allowed fermion states depend on the force fields. Choosing how to describe those states leaves a phase ambiguity. The answer cannot depend on arbitrary mathematical choices, and the prescription must remain consistent as the fields change—even around a sequence of changes returning to the starting configuration. The paper describes this measure problem.

The construction compares the same force fields on successively finer grids to build a phase rule for the original grid. Craig proposes that the contributions from those comparisons converge to the required correction, with the right consistency around every allowed loop. The rule also has to be local: distant fields must have only a rapidly diminishing influence. These are claims of the construction, not evidence that a full simulation has been run.

The important distinction is that the finer grids are tools for defining the rule on the original grid. This is not just a higher-resolution picture. As an interpretation of the proposal, more computing power cannot substitute for deciding what mathematical quantity the computer should calculate.

What Craig says AI contributed

Craig, a visiting scientist at OpenAI, says he supplied the refinement strategy. He credits GPT-6 Astra with turning it into a construction and leading the detailed proof work, alongside questions from human experts. That is his first-person account, hosted by OpenAI, not an independent measurement of AI’s contribution or time saved. The October paper already disclosed Astra’s involvement.

The manuscript reports checks using Lean, a mathematical proof checker, alongside a two-dimensional comparison and an alternative construction. Lean checks conclusions against encoded statements and assumptions; it does not by itself establish that those assumptions capture the intended physics. The distinction also matters in AI-assisted proofs for Signal’s encryption code, where the stated scope determines what a proof guarantees.

There is outside scholarly engagement. Edward Witten’s November 3 paper explicitly discusses Craig’s construction and develops a related variant. That is more than an isolated vendor announcement, but a related construction is not a line-by-line audit of Craig’s proof or an assessment of who contributed what.

What still separates the rule from a simulation

Craig’s conclusions describe a proposal at finite grid spacing, restricted to admissible fields—fields satisfying bounds at the grid’s scale. Practical implementation and useful numerical bounds remain work to do. The paper does not itself establish the interacting theory in continuous spacetime; that is a separate mathematical task.

Another open question is reflection positivity: a consistency condition needed to connect the lattice formulation to an appropriate quantum theory. The paper uses Euclidean spacetime, a mathematical reformulation treating time like a spatial direction. Establishing that connection is a physical-consistency issue, not a speed benchmark. Craig leaves it for further work.

Witten identifies a different obstacle for his related approach: it is computable in principle, but its quantum weights lack the positivity needed for straightforward Monte Carlo sampling. Those methods ordinarily exploit probability-like weights to sample a huge space of possibilities. Without that property, a formally available calculation can become impractically slow. This is a limitation of the approach he discusses, not a verdict that every possible simulation method must fail.

The two hurdles should not be collapsed into one: physical consistency and efficient sampling require different evidence. Neither a phase prescription nor a formal proof supplies a demonstrated runtime, hardware budget or end-to-end full-Standard-Model simulator.

For researchers, the next useful step is to examine the proposed rule and its allowed fields before asking how to implement it efficiently. For everyone else, the larger insight is about the kind of contribution AI may make: helping extract consequences from established laws, distinct from assessing new AI-generated scientific explanations. A mathematical recipe, a useful simulator and an experimental discovery are three different milestones. This proposal concerns the first—not new particles or evidence that spacetime is literally a grid.