Radical Numerics CEO Calls for DNA-Level Safety Checks in New Interview

In a new Latent Space interview, Radical Numerics CEO Eric Nguyen called for AI safeguards that examine DNA sequences for danger, rather than relying on chatbot refusals alone. His company builds models that can generate biological sequences. He argues their ability to assess those sequences should advance alongside their ability to design them.

A check on the DNA, not just the request

A genome language model learns patterns in DNA rather than words. Some such models predict what a sequence might do; generative models can produce new sequences. Nguyen says a model good at generation can also help judge whether a sequence is pathogenic, or capable of causing disease.

That would be a different check from refusing a chatbot request about viruses. One responds to what a person asks; the other examines the biological sequence itself. Nguyen says both are needed. Radical Numerics, he added, has a dual mandate to develop design tools and defenses against their potential misuse. He argues that teams building the design models are well placed to build the defenses because the model capabilities are closely related.

The defensive side needs to try to get ahead.

Eric Nguyen, speaking to Latent Space

What Evo showed him

Nguyen pointed to earlier work with Evo, a generative DNA model. In one demonstration, he said, researchers showed it natural CRISPR-Cas systems, then generated a new one that functioned. Because CRISPR-Cas involves RNA and protein components, he presented the result as an example of designing a working system with more than one kind of biological component.

A second example had greater weight for him: Nguyen said Evo generated a functional bacteriophage genome. A bacteriophage is a virus that infects bacteria. He described the result as a turning point that drew both excitement and concern, and helped push Radical Numerics toward defensive work. The concern he raised was about what more capable design tools might eventually enable, not a claim that this work produced a harmful human pathogen.

A Nature Biotechnology research highlight adds detail to the phage example. Researchers used Evo 1 and Evo 2 to design complete genomes using a known phage as a template. They further trained the models on roughly 15,000 related sequences and filtered generated genomes computationally, leaving 302 candidates. The account makes clear that selecting promising outputs was part of the work; it does not say every generated genome functioned.

Design evidence is not a screening test

The CRISPR-Cas and phage examples explain why Nguyen wants defenses to keep pace with design. They do not show how reliably a model could recognize pathogenicity in unfamiliar sequences. That is the unresolved test of his proposal: whether developers can demonstrate that a sequence-level safeguard catches the risks their design tools might help create.