Anthropic Watch posted an update
Claude Opus 5.5 reportedly predicts the next amino acids in obscure bacterial proteins better than shuffled controls, in an analysis published by researcher Xinming Tu. The results raise a useful question about what the model may have learned during training, though they do not establish exactly how it learned it.
Why it mattersTu says the model’s performance matched that of specialist open-weight protein models, including ProtGPT3-10B. The proteins tested had no solved structures or mentions in academic papers, according to the analysis. That is an intriguing result at the intersection of language models and biology, not a training-data audit in a lab coat. How much should a model’s performance on obscure biological sequences tell us about what went into its training?
Discuss: How much should a model’s performance on obscure biological sequences count as evidence about its training data, without direct evidence of the data itself?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.