The biology team at Anthropic mapped the first 3D structure of a vault particle from a bacteriophage, and the most useful part of their write-up is the list of what Claude got wrong.

Vaults are hollow protein barrels found in nearly all the cells in our bodies. They were discovered in 1986 by Leonard Rome and Nancy Kedersha, and 40 years later nobody knows what they do. The team used Claude at every step: searching about 15 billion predicted proteins, designing the DNA, optimizing how to isolate the particles, and processing the microscope data. A purification that usually takes three days took about 90 minutes, and a first structural model was ready in 34 hours.

Then they wrote down the failures. A protocol Claude wrote called for a growth additive that interferes with getting DNA into cells, and that cost two weeks. Early in the microscopy processing it skipped a standard step and produced a map blended with its mirror image. The team’s instruction was “Please let’s do normal things unless they don’t work.”

They also say that in areas they knew well they caught mistakes quickly, and in areas they didn’t, the mistakes cost them time. To me that is the honest version of AI speeding up science. The speed-up is real, it comes in stages, and it depends on someone in the room who can tell when a step is wrong. What vaults actually do is still unknown, and the team says so.

Source: https://www.anthropic.com/research/investigating-vault-particles-with-claude