Anthropic — Teaching AI to Build Scientific Modelsauto-ingested

A guest post on Anthropic's research blog by three AI for Science grantees. They argue that AI can now build, constrain and partly interpret scientific models, but a single best-fitting model gives "an answer, but… not understanding." What AI should deliver is a [[Distributions over Best Fits|distribution of plausible models]] plus an account of which alternatives failed. Subscribed-feed capture (Anthropic Research), auto-fetched and auto-ingested 2026-11-08; awaiting human review.

The anchoring story: Hodgkin–Huxley (1951–52)

• Hodgkin and Huxley modeled the squid nerve-cell membrane as an electrical circuit. Huxley solved the equations on a hand-cranked calculator for over three weeks because Cambridge's computer was down.
• One term was an empirical guess: a gating variable raised to the fourth power, because simpler versions fit worse. They hypothesized that four charged particles had to move before a gate opened. Decades later, potassium channel structures showed four identical subunits around a central pore. The model predicted a structure it was never fit to.
• The authors' point: "simply knowing the fourth power did the job was not enough; what mattered was knowing that a simpler version of the model did not work." Hamming (1962): "The purpose of computing is insight, not numbers."
• Hodgkin and Huxley noted a third or fourth power was needed, and that higher powers might fit slightly better, but the "improvement was not considered to be worth the additional complication."

A three-step program

Most scientific modeling repeats three steps, revising when fit or interpretation fails:

1. Build: choose the model's parts and the shape of its equations.
2. Constrain: tune parameters to fit measurements.
3. Interpret: say what the model means about the real system.

AI is involved in all three; it has progressed furthest on constraining and least on interpreting.

Constrain: report every parameter set that works

• Non-identifiability ("splitting the bill"): many parameter combinations produce near-identical output, as many combinations of orders can sum to the same $120 dinner bill. One best fit hides the others, and they may predict differently.
• Bayesian inference returns the full distribution, but needs the likelihood of the data under every parameter set, which is intractable for complex models.
• Simulation-based inference (SBI) sidesteps that. Run the simulator with random parameters (~100,000 runs for Hodgkin–Huxley, millions for larger models), train a neural network on the results, and it returns the plausible parameter range for new data in seconds. Any "parameters in, data out" program counts as a simulator ("Huxley, cranking through the equations by hand, was running one").
• Uses: locating colliding neutron stars fast enough to point telescopes; cosmology, structural biology, geophysics, nerve-cell models. The distribution shows which parameters the data pins down, where several explanations remain, and where the next experiment is needed.

Build: search the space of models, not one model

• LLM-based agents now write candidate model code from data plus a natural-language description, fit it, compare to data and revise the equations in a loop. Several such systems appeared in the past two years (FunSearch, Romera-Paredes et al. 2024; LLM-SR, Shojaee et al. 2025; others). Model design "was thought to be something only humans could do."
• The non-identifiability problem recurs one level up: many differently shaped models fit equally well. "Run an agent twice on the same data, and it might return two different, and equally good, models."
• Their proposal (Wahl et al. 2026 and others): the AI proposes many candidates; each is weighted by its support from the data, and models that need lots of flexibility to fit get less credit; weak ones drop out and the AI varies the strong ones. "We keep both likely and unlikely models, as all models can carry insight." Shared features across survivors are what the data favors; disagreement shows where evidence is thin.
• Their own caveat: these systems "have mostly been tested on rediscovering models that scientists already know. Evidence that they work on problems with no known answer is sparse."

Interpret: the hardest step to hand over

• The four-particle claim "only carried weight because of what they had ruled out."
• A theory is intelligible (philosophy of science) if you can work out its consequences without running it. Hodgkin–Huxley passes. Large models might pass a weaker test: can they be boiled down to a few mechanistic ideas sufficient to rebuild them?
• Fit is measurable; understanding isn't, and "it's not clear how to tell an AI system to make 'understanding' its objective." Two proposals: judge understanding by predictions under perturbation, or judge whether the output helps human scientists understand. Example: an AI for quantum-experiment design writes a short program generating a whole family of setups, so a physicist can read the design principle off the program.
• A single model can't say which parts the data constrains, "so it's easy to read meaning into parts that are not."

Cells and scales

• Famous models are compact because humans had to build, constrain and interpret them. As AI does more of the work, the size limit lifts, at a cost in readability.
• The fruit-fly connectome (>100,000 neurons; 124,635 cells rendered) now supports whole-brain models. At that scale "there is no such thing" as a single correct model (Box: "All models are wrong but some are useful"). The question becomes which properties all data-compatible models share.
• The most useful experiments rule out the most remaining models, "even if they are not always the ones a human scientist finds the most interesting."
• Climate ensembles and LHC exclusion maps already work this way.

The closing claim

"Much of the insight that a model provides comes from the process of creating it; Hodgkin and Huxley understood what the model was telling them because they had built it themselves, and they knew what hadn't worked."

To offer understanding, an AI system must deliver "not just the model that fits, but an account of which parts are pinned down by the data, which are still open, which alternatives failed along the way, and what new puzzles were uncovered." "A scientific model may provide an answer, but it is the process of constructing it that provides the insight."

Key claims

• Many parameter sets, and many model structures, fit the same data; returning one best fit hides the rest.
• SBI makes full parameter distributions practical for any runnable simulator (~100,000 simulations for Hodgkin–Huxley; inference in seconds thereafter).
• LLM agents can now propose and revise model structure, but evidence beyond rediscovering known models is sparse.
• Interpretation is the hardest step to automate, because "understanding" has no objective function; distributions of models are what make interpretation possible.
• Insight comes from the record of failed alternatives, which is what building a model yourself gives you and what an AI must now hand over explicitly.

Connections

• Distributions over Best Fits — the concept this source opens.
• AI-Accelerated Science — adds the build/constrain/interpret decomposition and a third Anthropic-feed view: acceleration of fitting is not acceleration of understanding.
• Falsification Is Asymmetric — interpretive weight comes from ruled-out alternatives; the best experiment eliminates the most models.
• Verification Requires Domain Knowledge — the "account of what failed" is a checkable artifact, a partial substitute for having built the model yourself.
• The Bottleneck Shift — model building moved to AI; interpretation and choosing experiments that matter remain the hard-to-delegate part.
• AI Use Tracks AI Capability — fit is measurable, so AI is strong there; understanding isn't, so AI is weak there.
• Critical Thinking Is Domain-Bound — Hodgkin and Huxley's interpretation drew on biological insight and the experience of the failed fits.
• Method Reflection — T4 sharpened: if insight comes from the process, auto-ingest hands the human the product without the process.
• Anthropic — AI for Science grant program; guest post on the research blog.

Source

• Original: raw/2026-11-06 Anthropic — Teaching AI to build scientific models.md (anthropic.com/research/teaching-ai-to-build-scientific-models). Figures (interactive likelihood map, model-space clusters, connectome renders) were omitted in the capture. Reading leads not ingested: Wahl et al. 2026 (PMLR v306) on model-space inference; Cranmer et al. 2020 (PNAS) on SBI; the original Hodgkin & Huxley 1952 paper.