Radical Numerics, co-founded by Eric Nguyen, is developing genomic language models (GLMs) that could materially accelerate biological research while also reshaping the biosecurity threat landscape. The company and its founders argue that the same AI capabilities that expand biological possibilities create pressure on defensive systems, producing an arms-race dynamic.
What GLMs do and why they matter
Eric Nguyen recalls that during his time at Stanford the idea of genomic language models struggled for acceptance: many biologists doubted the approach would work, that outputs could be validated, or that it would offer practical advantages over existing methods. Nguyen persisted, helping to develop Evo and contributing to Evo 2 at the Arc Institute. Those early models were later used by a separate Arc/Stanford team to generate entire bacteriophage genomes that were synthesized into functional viruses.
DNA differs from natural language in key ways: its alphabet has only four characters (A, C, T, G), but sequences can be extremely long — an average human gene can be around 60,000 bases, some genes up to 2.3 million, and the whole human genome roughly 3 billion bases. The ability to handle long-context sequences made recent GLM advances possible about three years ago, well before frontier labs began building models with 1M+ context windows.
Thinking in DNA: chain-of-thought and self-optimization
Radical Numerics’ GLMs already generalize across DNA, RNA and protein because genomic sequences contain clear markers for genes and RNAs, and specific genes encode proteins. That lets models generalize to multiple biological “languages” even before training on other modalities such as 3D protein structure, epigenetics, or natural language.
Nguyen describes an experiment on aptamers: the team withheld a subset of the best-performing sequences and presented the model only with lower-performing examples, ranked progressively from worse to better. They then asked the model to continue that progression. The model was able to recapitulate some of the higher-performing sequences it had not been shown — an effect the team likens to chain-of-thought reasoning, here operating in the language of DNA.
The arms race and defensive challenges
Nguyen warns that defensive biosecurity is currently losing ground to rapidly advancing model capabilities. Radical Numerics’ position is that defenders must use frontier models themselves — and keep them open and up to date — to avoid falling behind offensive uses of these tools.
Topics discussed in detail include:
- biosecurity as an arms race and how defense can keep pace;
- viewing the genome as the imprint of the environment on DNA;
- moving toward truly multimodal models;
- how chain-of-thought works when a model “thinks” in DNA.
The team behind Radical Numerics
The founders include Eric Nguyen (Co-founder and CEO, PhD in Bioengineering & AI from Stanford University, who helped develop Evo and Evo 2), Michael Poli (Chief AI Scientist, Stanford PhD, formerly a founding scientist at Liquid AI), Stefano Massaroli (President, a former postdoc with Yoshua Bengio and a founding team member at Liquid AI) and Armin W. Thomas (CTO, a former Stanford postdoc who worked with Chris Ré and previously at Liquid AI).
Why this matters
Advances in GLMs speed up biological capability and introduce new security considerations. Radical Numerics’ approach — leveraging long-context models and DNA-centric reasoning — is promising for research applications but also highlights the need for defensive strategies that can match the pace of offensive development. Without such parity, the balance between enabling innovation and managing risk may tilt toward increasing biosecurity vulnerabilities.
