Anthropic essay: AI must build models for insight
In a new research essay, Anthropic argues that AI systems for science need to show why a model works, and which alternatives fail, rather than just returning a single best fit to the data.
Anthropic has published an essay on how AI can help scientists build, fit and interpret models. It reports no new system, benchmark or release.
The essay starts with the Hodgkin-Huxley model of the squid nerve cell. In 1951, while Cambridge University's computer was out of action, Andrew Huxley spent more than three weeks solving the equations on a hand-cranked calculator. To fit the data, he and Alan Hodgkin had to raise one gating variable to the fourth power. They guessed this meant several charged particles had to move before a gate would open. Decades later, another team found the structure of a related potassium channel protein: four identical pieces arranged around a central pore.
Anthropic's point is that an AI system "could plausibly find an equation like Hodgkin and Huxley's in an afternoon." But a model that only reproduces the data gives no understanding, because many models can do that. What matters is knowing which features a working model shares with other working models, and why simpler versions fail.
The essay splits modeling into three steps: building the model, constraining its parameters, and interpreting what it says about the real system. It says AI has arguably made the most progress on constraining. Using a dinner-bill analogy, it argues that many sets of parameters can fit the same data. That is the case for Bayesian inference, which returns a probability distribution over parameter settings instead of a single best fit.
The excerpt available to us ends at that point. A summary of the full piece says it goes on to cover simulation-based inference, LLM agents that propose models, and returning many candidate models instead of one. It also says that, so far, such systems have mostly rediscovered models scientists already knew.