Anthropic: how AI is beginning to change scientific modeling

Scientific models are how researchers make sense of the universe, whether they are tracking a single neuron firing or calculating what happens when neutron stars collide.

While fitting models to data has gotten faster lately, the initial setup remains a huge bottleneck. Scientists still write their models manually, often having to juggle and compare multiple competing theories that each explain the evidence slightly differently.

Models written and tweaked as code

In a guest piece on Anthropic's Science Blog, researchers Jakob Macke, Daniel Gedon, and Stefan Wahl highlight how modern AI is taking over this tedious loop.

Instead of humans drafting every variation by hand, AI can now write scientific models directly as code, run simulations, fit them to incoming data, and iterate on the fly.

This workflow lets researchers explore a vast array of potential model structures in parallel, quickly showing which theoretical explanations are actually backed up by the data.

Why it matters

We think this points to an important shift in how AI assists science: moving beyond simple data processing and stepping directly into theoretical reasoning.

For researchers, testing dozens of candidate hypotheses by hand is an exhausting grind. Offloading the coding and comparison cycle to an AI agent could significantly shorten the path to experimental breakthroughs.

That said, the AI is still operating within the code-based modeling frameworks researchers give it. Helping refine candidate hypotheses is a huge win, but coming up with entirely new scientific paradigms from scratch is still firmly up to humans.