Teaching AI to build scientific models: check “mostly rediscovery, sparse evidence on unknowns,” not “AI already delivers understanding”

In 1951, Hodgkin and Huxley wrote equations for the squid nerve cell’s action potential. When the Cambridge computer was down for six months, Huxley spent more than three weeks on a hand-cranked calculator. They raised one variable to the fourth power because simpler forms fit worse; decades later, a related potassium channel protein turned out to have four identical pieces around a central pore.

Anthropic Science’s guest post on 6 November 2026 uses that story to draw a line: an AI system might find a closer-fitting equation in an afternoon, but handing over a model that fits is not the same as handing over understanding. Authors Jakob Macke, Daniel Gedon, and Stefan Wahl (ELLIS Institute Tübingen / University of Tübingen; Anthropic AI for Science grantees) put the limit in checkable prose: current systems have mostly rediscovered known models, and evidence on unknown problems is sparse.

Hard facts (https://www.anthropic.com/research/teaching-ai-to-build-scientific-models; published Nov 6, 2026; accessed the same calendar day in Asia/Shanghai):

• Form: Anthropic Science guest post by AI for Science grantees Jakob Macke, Daniel Gedon, and Stefan Wahl (ELLIS Institute Tübingen and the University of Tübingen).
• Three steps: build (what the model is made of and what form its equations take); constrain (tune parameters to measurements); interpret (what the model says about the real system, such as a fourth power suggesting several moving parts in a gate). In practice the steps loop.
• Constraining: standard Bayesian inference is often impractical for complex models; simulation-based inference (SBI) runs many forward simulations, trains a network, and returns a distribution of parameters compatible with an observed output. The post cites neutron-star merger localization fast enough to point telescopes, and applications to nerve-cell models including Hodgkin-Huxley.
• Building: LLM agents write candidate model code from data and a natural-language description, then constrain, compare, and revise. Several such systems appeared in the past two years (footnotes include Romera-Paredes et al., 2024; Shojaee et al., 2025; Ma et al., 2024; Novikov et al., 2025; Xia et al., 2026; Ziaei Bideh and Gryak, 2026). The authors argue for a distribution of plausible models rather than a single best fit (see also Wahl et al., 2026, and related footnotes).
• Interpreting: they call this the hardest step to hand to AI. Fit is measurable; “understanding” is hard to write as an objective. Proposals include checking predictions after intervening on the system, or asking whether outputs help human scientists read a design principle (a quantum-experiment program generator is one example).
• Scale: complete fruit-fly brain wiring diagrams now exist (footnotes to Dorkenwald / Schlegel / Azevedo et al., 2024), enabling large models on top. Climate ensembles and collider “which theory versions remain open” are the analogy: shared properties of data-compatible models matter more than one true equation.
• Authors’ own limit, in substance: so far these systems have mostly been tested on rediscovering models scientists already know; evidence that they work on problems with no known answer is sparse.

Evidence strength: this is a funded researchers’ argument and method map, not a new leaderboard. The Hodgkin-Huxley history, the SBI workflow, the fly-connectome citations, and the footnote paper list can be checked against the post and the public literature. The hardest falsifiable sentence is the authors’ own limit: mostly rediscovery so far, sparse evidence on unknowns.

How to check: verify the three-step definitions and the SBI section; open the footnotes for the named model-building agents and distributional work such as Wahl et al.; confirm the post still says evidence on problems with no known answer is sparse. If later independent work stably publishes reproducible, interpretable model families on several unknowns, raise how far “building” can be automated.

Steelmanning the piece: insisting on a distribution of models plus failed alternatives blocks reading a lucky fit as a mechanism. The fourth power carried interpretive weight because lower powers were not good enough, not because an optimizer emitted a four.

Strongest steelman: treating model building as inference over a set of candidates is the same honesty as returning a parameter distribution when constraining. Shared features support stronger claims; disagreements point to the next experiment. Fly-scale brains and climate ensembles show that once systems get large, “the one true model” was never the working target.

Steelman risks: the post is argument and literature map, with almost no new quantitative board. Readers should weigh the guest authors’ Anthropic grant relationship. Interpretation still lacks an operational automatic objective; “helps humans understand” is hard to turn into one score. Rediscovery tests do not clear the bar for open scientific problems.

The Hamming line that “the purpose of computing is insight, not numbers” is a normative claim, not a new experimental result in this piece.

For scientific software and funders, the next six months are about whether model-building agents start publishing checkable model families and failed alternatives on unknowns, whether SBI and LLM search loops share public benches, and whether interpretation gets a harder test than “reads like insight.” For readers, the check is the authors’ self-limit and the footnote trail, not a claim that AI already finds Nobel-grade mechanisms.

What to check in this guest post is the three-step split and the authors’ stated limit: current AI model-building systems have mostly rediscovered known models, and evidence on problems with no known answer is sparse. What not to check is whether AI already delivers scientific understanding. If the demand for distributions over single fits holds, this is a checkable methods stance. Reading “finds an equation in an afternoon” as already understanding the system leaves the source behind.

【No.2】