Anthropic published a new research report that goes in an unexpected direction: teaching Claude chemistry. Real chemistry. Not the atoms-and-electrons stuff from high school, but the analytical work that chemists do every day.
What it’s about
Anthropic chemist David Kamber examined how Claude handles NMR spectra — the single most important analytical tool in chemistry. NMR spectroscopy (Nuclear Magnetic Resonance) reveals how atoms are arranged within a molecule. For chemists, reading NMR spectra is as fundamental as reading code is for developers.
The key challenge: chemists constantly switch between different representations — hand-drawn structures on a whiteboard, instrument readouts, database query strings, technical notations from patents and publications. Each representation encodes the same underlying chemistry but demands a different kind of fluency.
Why this is exciting
Frontier models like Claude are multimodal and capable of explicit reasoning. That means Claude can read a chemical structure directly from a journal figure or a hand sketch, parse the experimental detail of a methods section as it’s actually published, and show its reasoning step by step — so a chemist can audit the outputs.
Anthropic’s stated goal is straightforward: working chemists should know where Claude can save them time and where they still need to rely on their own expertise. No promise that Claude replaces the chemist — just that it becomes a useful tool.
My take
I find this approach compelling because it shows how Anthropic tackles specialized domains: systematically, with real experts, and without overselling. Instead of claiming Claude can do everything, they’re honestly investigating what works and what doesn’t. That’s the right way to build trust in science.
Sources: Anthropic Research