Intel / article

🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

Source: latent.space
Author: RJ Honicky
Date: Published Sep 23, 2026

Context

Radical Numerics, a company building genome language models, describes an experiment in which a model continues a sequence of progressively better examples. In a Latent Space interview write-up, cofounder Eric Nguyen says the team showed a model RNA sequences paired with performance scores, ordered from lower to higher, while holding back some of the best-scoring ones. He says the model continued the pattern and recovered some of the higher scores it had not been shown. That is what the account means by for biology: extending a trajectory of improving examples. It suggests a direction for molecular design, but the account supplies neither enough experimental detail nor defensive results to establish a biosecurity benefit.

Key takeaways · AI-distilled

Evo and Evo 2, genomic language models Eric Nguyen helped build at Stanford and Arc Institute, were later used by a separate Arc/Stanford team to generate whole bacteriophage genomes that were synthesized into functional viruses.
DNA has a four-letter alphabet but very long sequences (about 60K bases for an average human gene, roughly 3B for the whole genome), so the post credits long- architectures such as StripedHyena, about three years ago, with making genome models practical.
In a Radical Numerics experiment, the model saw RNA aptamers in ascending score order with the best ones held out; asked to continue the pattern, it recapitulated some of the higher-scoring held-out sequences, which the post frames as chain-of-thought in DNA.
Nguyen says models trained on DNA already do well on RNA and protein because genes and protein-coding regions are marked in the sequence. He also says biosecurity defense is currently losing, and Radical Numerics argues for pushing the frontier harder.

Terms in this piece · Glossary

chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source www.latent.space