Anthropic grantees on teaching AI to think like scientists, not curve fitters

Back in 1951, Cambridge University's computer broke down for six months. Physiologist Andrew Huxley was forced to spend over three weeks grinding through a hand-cranked mechanical calculator to solve equations describing squid nerve cell firing.

To make the curve fit their real-world measurements, Huxley and Alan Hodgkin took an educated guess: they raised a key variable to the fourth power. They didn't know why, only that simpler versions failed. Decades later, structural biologists finally mapped the relevant potassium channel and found four identical protein subunits surrounding a central pore. The math had predicted physical reality.

Today, modern AI could fit those squid curves in an afternoon. But on November 6, Anthropic AI for Science grantees Jakob Macke, Daniel Gedon, and Stefan Wahl from the ELLIS Institute Tübingen and the University of Tübingen argued that curve fitting is nowhere near enough. What science actually needs is genuine insight.

The $120 dinner tab

Traditional scientific modeling rests on three pillars: building the equations, constraining the parameters, and interpreting what the model says about the world. For decades, constraining complex models was a massive bottleneck.

The researchers use a restaurant analogy. Imagine getting a group dinner bill of $120 with no itemized receipt. Everyone could have ordered an average burger, or one diner had steak and lobster while everyone else sipped tap water and soup. Biological systems behave the same way: wildly different combinations of cellular parameters can produce virtually identical electrical spikes.

If an AI hands you a single best-fit parameter set, it essentially hides every other plausible mechanism. Standard Bayesian inference can map out these possibilities, but complex simulators are often impossible to run in reverse.

To fix this, the researchers rely on simulation-based inference (SBI). Instead of calculating exact likelihoods with rigid formulas, they run a forward simulator hundreds of thousands or millions of times with random inputs, then train a neural net on the pairs. Once trained, the network can ingest new experimental data and return the full distribution of plausible parameters in seconds. The technique has already proven its worth in cosmology, pinning down neutron star collisions fast enough to redirect optical telescopes.

Generating a distribution of models

Now, LLM agents are creeping into the very first step of science: writing the candidate equations themselves. Researchers can hand an LLM agent some data and a system description, and the agent will write code, evaluate fits, and revise the model structure iteratively.

Yet the exact same trap appears one level up. Run an agent twice on the identical dataset, and it might output two completely different models that both fit cleanly. Picking just one is just as deceptive as picking a single parameter set.

The Tübingen researchers argue that AI must provide a whole distribution of plausible models. By having agents generate many candidates, weighting them by parsimony and fit, and seeing what surviving models share in common, scientists can pinpoint which mathematical features are fundamental versus merely incidental.

Why it matters

For researchers and computational scientists, this framework offers a principled roadmap to move past black-box predictions. Simply matching experimental curves has never been the goal of science; understanding underlying mechanisms is. Tools combining SBI with LLM reasoning give scientists an automated partner that highlights uncertainty and suggests where the next lab experiment is actually needed.

Across the AI landscape, this highlights Anthropic's push into rigorous AI for Science. Rather than treating frontier models solely as coding copilots or end-to-end data predictors, there is growing momentum to weave foundation models into formal scientific inference pipelines.

We think the Tübingen team hits the nail on the head. As Richard Hamming famously observed, the purpose of computing is insight, not numbers. Brute-force neural networks can approximate almost any function, but an AI that hides competing explanations is a scientific liability. The lingering catch is compute: running millions of forward simulations alongside iterative LLM loops remains prohibitively heavy for many labs. Making these SBI workflows lightweight will be the real test for widespread adoption.