Daily AI Catchup
Research

Claude models show signs of training on raw protein data

An outside researcher found that recent Claude models can predict the next amino acid in obscure, little-studied bacterial proteins, something that only works if the model saw real protein sequence data during training. The ability disappears if the protein sequence is shuffled, and Anthropic's Fable 5.2 reportedly beats every open protein language model tested.

Why it matters

๐Ÿ’ป Developer ยท If true, this hints at what data mixes top labs are using beyond public text, useful context if you're building bio or science tools on these models.

๐Ÿ“ฆ Product ยท A model that understands protein sequences natively could open new science and biotech product angles without a separate specialized model.

๐ŸŽจ Design ยท Not directly design-relevant, but it's a reminder that 'general' chat models increasingly carry hidden specialist capabilities.

๐Ÿ“ˆ Business ยท Anthropic may have a quiet edge in biotech and pharma use cases that competitors can't easily match.

๐Ÿค” Just Curious ยท Claude apparently picked up biology skills nobody announced, discovered by an outsider poking at it rather than Anthropic itself.

Sources: Claude was trained on protein sequences at scale