In Nov/Dec 2025, just prior to the training cutoff for the first model to show this capability (Opus 4.7) I was working on solving arc-agi using claude code and was exploring using protein data and other adjacent data to improve performance.
In January it seemed very clear the subscription models behind claude code became less competent, there was a whole series of posts about it. I stopped working on it as a result.
Claude just blocks from anything in this field. Can't even work with my own DNA without triggering opus 5.5 50% of the time
It would be cool if Claude were not explicitly trained on protein sequences at scale, and this were an emergent property from better fundamental understanding of chemical structure.
Does shuffling still result in a viable protein? I'm no biochemist, so this is a naive question. Do proteins have some (chemical) constraints on how the "letters" can be arranged?
Is this a valuable skill ? Sorry, bit of a noob question ig but have been seeing a lot of buzz around protein sequences in general
Curious: how can there be a correct prediction for the next element of a "shuffled" protein? The shuffled proteins don't occur in real life, so what makes one answer correct and another incorrect?
Gemini should be good at naming the next sugar in a polysaccharide. :)
You, however, will be banned for "biosecurity" research. Or, as Amodei puts it, "you can't have opensource models because with them you can't ban users"
It's hard to see what the point of this is?
When would one want to predict a full-length unknown protein sequence from only a fragment? Any fragment of any useful length can be back-translated to the DNA sequence and that can be tested against genomic databases (a process called tblastn). And if your protein/gene is not found there, you have no way of knowing whether or not the predicted sequence has any basis in reality.
Testing the model against shuffled sequences is also kinda cheating, as these are no longer random pool of amino acids, but one whose source already exists in the databases.
What's more surprising is that the other models aren't
No doubt that all of them, except maybe Gemini, surely were trained on SciHub and all books + data they could download from torrents and other sources, as well. That could include leaked personal data and proprietary copyrighted material, from anywhere maybe even including Darknet. Add to that public patent registry and the book buyout and scanning campaigns that occur.
I am certain that all these companies have stored and backed up all they could find, so that it can't be taken away from them.
DNA is a language, or at least a protocol. Just a very difficult one. All meaning is locked up in the shape of the Lennard-Jones 12-6 magnetic flux isosurface that is generated when you string the letters together from amino-acids. Probably justified to call that a bit more indirect than English.
And electrostatic interactions.
And some methylation.
And some phosphorisation (but now we're getting deep into somewhat rare grammar)
Yeah, DNA is a language that is way too complicated for human minds to understand, although plenty of humans "get the gist" so to speak after looking at it for a few hours, but ... here's some bad news: it's only somewhat too complicated for LLMs. So why wouldn't you train LLMs on them? Frankly, it's the one language we have "texts" from that is actually difficult to understand for LLMs and ... well, it's definitely improving their understanding.
(also we have quite a bit of DNA text that isn't from the internet or books, it's freely downloadable and unlike everything else LLM companies steal, actually not covered by copyright)
What I find most interesting here isn’t necessarily that Claude can predict protein sequences, but what this tells us about how knowledge is acquired by large models.
The shuffled-sequence control is particularly compelling. It suggests that the models aren’t simply recalling amino acid frequencies or reproducing patterns from scientific literature. They’re capturing structural information embedded in the sequences themselves.
There’s an important distinction, though, between learning statistical representations of protein families and actually understanding the underlying biology. Next-amino-acid prediction is a fascinating proxy, but it doesn’t necessarily demonstrate that the model has learned the mechanisms governing protein folding or function.
I’d be particularly interested in seeing whether these capabilities transfer to genuinely novel protein families, rather than previously unseen members of known families. The post-cutoff experiment is a good step in that direction, although it still tests generalization within existing families.
Either way, this raises an interesting question about foundation models: how much latent scientific knowledge are we inadvertently creating simply by exposing them to sufficiently large and diverse datasets?
Perhaps we’re underestimating what can emerge from learning the statistical structure of the natural world.
So the american "government", "apparently", goes full steam in to develop biological weapons (of which, as usually, they accuse others).
In "Empire of AI", the author mentions that when OpenAI included public GitHub repositories in its training data, the resulting model achieved better results even in fields unrelated to programming.
Aside from the obvious applications in biochemistry, could protein sequences also be used as a new training data source to improve overall model quality? If that's true, it's quite fascinating that we can take highly complex data we don't fully understand and use it to improve LLM training, simply because it contains so many significant structures and relationships.