Eric Nguyen: Bio-security is an AI Arms Race

Against the backdrop of biosecurity being increasingly viewed as part of the AI arms race, Radical Numerics co-founder Eric Nguyen gave a Latent Space interview and put forward a counterintuitive proposition: the models that enhance biological capabilities can equally be used to keep defense from falling behind. This formulation itself constitutes the core tension of the entire discussion—capability and defense are not two independent tracks, but share the same technological foundation.

But Nguyen also offered a less optimistic judgment: by his account, defenders are currently at a disadvantage in this arms race, and Radical Numerics' position is to push the frontier harder. Two levels need to be distinguished here: the proposition of dual-use models is the core argument of the interview, while "defense is losing" is Nguyen's personal judgment, and the source did not provide quantitative evidence to support it.

This judgment did not emerge from nowhere. The article mentions the "OpenAI → Hugging Face attack" incident and states that Anthropic's filters flag both cybersecurity and biology content. It is precisely such incidents that make "what else should we worry about" a public question, and biosecurity is thereby pushed to the forefront. It should be noted that the specific nature, timing, and impact of that attack incident are only mentioned in the source and not elaborated in detail.

To understand Nguyen's position, one needs to return to his personal technical path. According to the interview, during his time at Stanford he long promoted genomic language models without gaining recognition—biologists did not believe this path could work, did not think outputs could be verified, and saw no application scenarios beyond existing capabilities. He kept pushing, later helped lead the development of Evo, and contributed to Arc Institute's Evo 2.

The key validation node appeared later: Evo and Evo 2 were used by another Arc/Stanford team to generate complete phage genomes and synthesize functional viruses. The significance of this fact is that it moved genomic language models from the debate over "whether they can generate" to the empirical level of "whether what is generated is functional." However, the source did not elaborate on the specific details, verification methods, or reproducibility of that experiment.

It was based on this path that Nguyen, together with other AI×Bio figures, co-founded Radical Numerics, with the goal of building and scaling genomic language models to address a broad range of biological problems, far beyond generating DNA. The team composition includes CEO Eric Nguyen (PhD in bioengineering and AI from Stanford), Chief AI Scientist Michael Poli, President Stefano Massaroli, and CTO Armin W. Thomas.

The reason genomic language models have long been difficult to advance lies in the fundamental differences between DNA and natural language. DNA's alphabet is very small, with only four characters—A, C, T, G—but sequences are very long. The magnitudes given in the interview are: the average human gene is about 60,000, long ones can reach 2.3 million, and the entire human genome is about 3 billion.

This structure of "an extremely small alphabet and extremely long sequences" determines that the modeling difficulty lies not in vocabulary size but in context length. According to the interview, long-context model innovations made this kind of genomic modeling possible about three years ago, far earlier than frontier labs building models with contexts above 1 million. This time gap is the premise for understanding Radical Numerics' technical path: the capability they bet on was already feasible before mainstream frontier labs turned to long context.

From a reasoning perspective, this also explains why the capability leap of genomic language models and the long-context breakthrough of natural language large models are not synchronized in time—the former relied on infrastructure that matured earlier, but application validation came later.

The core experiment Radical Numerics presented was to transfer the chain-of-thought concept from natural language models to DNA language. According to the interview, their genomic language model already performs quite well on RNA and proteins, because DNA sequences contain clear markers of genes and protein-coding genes, and the model generalized to multiple "languages" before being trained on other modalities.

The specific experimental design was: they used a large-scale aptamer dataset, showed the model only lower-scoring RNA aptamers and gradually raised the scores, then asked the model to continue along this pattern. The result was that the model could reproduce some higher-scoring sequences that were not shown to it. The authors called this process "thinking in DNA" chain-of-thought, meaning the model can self-optimize along a score trajectory.

It must be made clear that the specific dataset size, scoring metrics, and magnitude of improvement in this experiment were not disclosed, and "performs quite well" also lacks specific benchmarks or numerical support. Therefore, the conclusion of this experiment should be regarded as a demonstrative claim by the source, not a capability conclusion verified by a third party.

A key inferential argument in the interview is: just as long-context reasoning, chain-of-thought, and multimodal perception unlocked complex reasoning in natural language large models, these capabilities are also giving rise to increasingly complex "biological intelligence" in genomic language models and the greater danger that comes with it. This judgment is labeled by the source as an inference, based on the analogical logic of capability transfer rather than direct experimental evidence.

This inference forms a logical loop with Nguyen's judgment that "defense is losing": if capability upgrades and danger upgrades share the same origin, then for defenders to keep up, they must likewise master these capabilities. This is precisely the argumentative basis for Radical Numerics' position of "pushing the frontier harder"—in their view, slowing the frontier cannot give defenders an advantage and may instead widen the gap.

But the weak point in this argumentative chain is equally obvious: whether capability upgrades necessarily lead to danger upgrades, and whether defenders mastering equivalent capabilities is sufficient to reverse the disadvantage, are both unsupported by quantitative evidence in the source. This is a positional claim, and readers should treat it separately from confirmed technical facts.

This discussion was not confined to Radical Numerics itself. The interview cited Clem Delangue's argument: cyberwarfare defense capabilities must be open in order to keep up with frontier models' attack capabilities. This position echoes the open-weights debate in the biosecurity field—if attack capabilities can be obtained, should defense capabilities also be made public to maintain balance?

The article also cited the claim that Zai_org shared GLM5.2 with the world as open weights (free) and that it became a key part of its security response. This statement comes from a tweet paraphrase and lacks a first-party source, so it can only be regarded as a source claim and cannot be treated as confirmed fact.

The open-weights path is especially sensitive in the biosecurity context: the dual-use nature of genomic language models makes the boundary between "openness to strengthen defense" and "openness to lower the threshold for misuse" extremely difficult to draw. The interview did not offer a solution to this tension, but presented it as part of the arms race backdrop.

Taken together, the verifiable information provided by this interview is concentrated at the level of technical facts: the DNA alphabet and the magnitudes of sequence lengths, the timeline of long-context technology, the phage synthesis application of Evo/Evo 2, and the team composition and research direction of Radical Numerics. These constitute a reliable basis for understanding the current state of genomic language models.

The parts involving judgments and claims, however, need to be treated with caution: "defense is losing this race" is Eric Nguyen's personal judgment, and the source provided no quantitative evidence; the specific scale and magnitude of improvement of the aptamer chain-of-thought experiment were not disclosed; the details and reproducibility of Evo/Evo 2 generating phage genomes were not elaborated; GLM5.2's open weights and its role in the security incident come from a tweet paraphrase; and the specific nature of the "OpenAI → Hugging Face attack" is only mentioned.

This means that Nguyen's proposition that "the same models can both enhance capabilities and strengthen defense" is a proposition supported by a technical basis, but the conclusions that "defense is losing" and that we should "push the frontier harder" remain positional claims requiring more evidence to test. In the high-risk field of biosecurity, distinguishing technical facts from strategic judgments is more important than accepting either side's narrative.