🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
Latent Space RJ Honicky
AI models are getting better at biology, and that scares people. Eric Nguyen says the same tools can also keep bio-defense from falling behind.
Based on reporting by Latent Space, RJ Honicky — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The latest AI anxiety isn’t just about hacking. Latent Space points out that Anthropic’s filters already flag cyber-security and biology, and Eric Nguyen of Radical Numerics says bio-defense needs the same kind of urgency as the attack side. His argument is blunt: if models can help create biological capability, they can also help defense keep pace.
Nguyen has been pushing on this for years. At Stanford, he tried to get Genomic Language Models off the ground and ran into a wall of skepticism. Biologists doubted the approach, doubted the outputs, and didn’t see the point. He kept going anyway, helped lead work on Evo, and contributed to Evo 2 at the Arc Institute. Those models later fed into an Arc/Stanford team effort that generated bacteriophage genomes, which were synthesized into functional viruses.
The technical reason this works is less mysterious than it sounds. DNA has only four letters, but the sequences are enormous: an average human gene is around 60K, long genes can reach 2.3M, and the full human genome is around 3B. Nguyen says progress in long-context models, enabled about three years ago by work such as striped hyena, made this possible well before frontier labs were talking about 1M-plus context windows.
Radical Numerics is now trying to push GLMs further, not just at DNA generation but across more biological problems. The models already do well on RNA and proteins, since genes leave recognizable marks in DNA and proteins are encoded by specific genes. The bigger bet is that a model trained to think in the DNA language can start to infer the imprint of environment on genomes, and maybe generalize into other modalities like 3D protein structure, epigenetics, and natural language.
Nguyen described one test where the team showed a model a series of RNA aptamers with progressively better scores, held out the best-performing ones, and asked it to continue the pattern. It did manage to recapitulate some of the higher-scoring examples it had not seen. That’s the unsettling part. The same stepwise reasoning that makes language models feel sharper is starting to show up in biology too, and that cuts both ways.
My take — AI-written commentary, not fact-checked reporting
Biosecurity is becoming the same ugly race that cyber has always been: offense gets the headlines, defense gets told to move faster. The annoying truth is that open systems tend to be how the defense side learns at all, while closed systems usually arrive after the panic. That pattern is getting hard to ignore, unless the plan is to let the other side enjoy the first-mover advantage and call it prudence.
Read more about this at: Latent Space