AI Doing Real Science: Exciting and Hard to Verify
Anthropic's Claude reportedly discovered a novel enzyme system with CRISPR-like repeats, and the thread surfaced two competing reactions. The first was genuine excitement about AI as a scientific discovery tool. The second was sharp skepticism: commenters wondered whether the absence of a Nature letter means this is hype, whether another lab already found the same thing, and whether the framing inflates what Claude actually contributed.
Separately, a thread on automated optimization of a molecular simulation program showed Claude producing meaningful performance improvements in a hard scientific computing context. The pattern is consistent: AI systems are doing real work in science and engineering, but the provenance and credit for discoveries are genuinely murky.
One commenter in the enzyme thread put the stakes plainly: in a year or two, stories like this will either look like artifacts of peak hype or evidence of the beginning of something historically significant. There is no clean way to know which it is right now, and that uncertainty is itself a feature of this moment.
So what?
If you are building tools for researchers or scientific computing, the current moment is a genuine opportunity. Researchers are actively experimenting with AI-assisted discovery, and the tooling is still immature. The credibility question around AI attribution in science is also a real problem to solve, not just a PR issue.