🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
September 23, 2026
AI Summary
5 min readIn 2024, a team at Stanford trained a generative AI model called Evo on raw DNA sequences. When they asked it to design a new genome from scratch, it produced a functional bacteriophage—a virus that infects bacteria—that did not exist in nature. The model had learned the grammar of life well enough to write coherent genetic code. For Eric Nguyen, now CEO of Radical Numerics, that moment crystallized both the promise and the peril of the technology he helped create. "If you can manipulate the fabric of life itself," he says, "what kind of implications does that mean?"
How a DNA Language Model Works
A genome language model (GLM) is a large language model trained not on English words but on the four-letter alphabet of DNA. Where chatbots predict the next word in a sentence, a GLM predicts the next nucleotide in a sequence. Nguyen’s team pioneered this approach with Hyena DNA, which used a convolutional architecture (instead of standard attention) to process sequences up to a million tokens long—at the time, the largest context window for any language model. This long-range capability matters because the functional parts of a genome can be separated by vast stretches of non-coding DNA.
Continue reading the full summary in the app — free to try.
Read Full Summary →Free • No credit card required
Never miss an episode of Latent Space: The AI Engineer Podcast
Get every new episode summarized in your inbox — free, ~5 minutes to read.
No spam. Unsubscribe anytime.
What you'll learn
- 1 🔬 Bio-security is an AI Arms Race – Eric Nguyen (CEO, Radical Numerics)
- 2 (00:00) **Opening: The Arms Race Dynamic** – Eric introduces the core tension: design capabilities are advancing fast, but defensive biosecurity tools are far behind.
- 3 (01:06) **What Is a Genomic Language Model?** – Eric defines GLMs as LLMs trained on raw DNA sequences, not words, enabling both reading and writing of the "fabric of life."
- 4 (03:06) **From Hyena DNA to Evo: The Evolution** – The team started with Hyena DNA, a convolution-based model handling up to a million tokens of context, then built Evo for generative genomics.
- 5 (07:37) **Evo vs. Omni: The Alignment Leap** – Earlier models like Evo underperformed specialist models on human genomics; Omni applies post-training and alignment to make genomic models practically useful for scientists.
- 6 (11:19) **How Alignment Works for Genomic Models** – Eric describes the technical process: special tokens, structured inputs, and chain-of-thought-style examples that teach the model which task is being asked.
- 7 (15:10) **Variant Effect Prediction: The Killer App** – Eric walks through the most impactful benchmark: predicting whether a single-letter DNA change causes disease, especially in hard-to-analyze non-coding regions.
+ Full timestamped outline available in the app
Show Notes
The OpenAI → Hugging Face attack has people asking “what else do we need to worry about?” and Anthropic’s filters flag two things: cyber-security and biology. The natural question is: what about bio-security, then?
Clem Delangue argues that cyber-warfare defensive capabilities need to be open and to keep pace with frontier models’ attack capabilities
Radical Numerics co-founder Eric Nguyen sat down with us and explained why the same models that increase biological capability can also keep defense from falling behind.
Building a virus from scratch
While he was at Stanford, Eric couldn’t get traction on Genomic Language Models (GLMs) for a long time. Biologists didn’t believe it would work, didn’t think they could verify the output, and didn’t see important applications beyond what they could already do. He kept pushing, eventually helping lead the development of Evo and contributing to Evo 2 at Arc Institute. Those models were later used by a separate Arc/Stanford team to generate entire bacteriophage genomes that were synthesized into functional viruses!
Long context unlocks biological intelligence
Early ChatGPT spit out poems and email, and early DNA language models like Evo and Evo-2 could build a genome from scratch. DNA is different, however, from natural language in that it has a very small alphabet (4 characters ACTG) and that its sequences are very long:
* 60K for an average human gene
* long being up to 2.3M
* the whole human genome around 3B.
Innovation in long-context models made this possible about 3 years ago (footnote: striped hyena), long before the frontier labs were building 1M+ context models.
Now Eric and other AI x Bio luminaries have founded Radical Numerics to build and scale GLMs to tack a wide range of biological problems, extending well beyond generating DNA.
Thinking in DNA
Their GLMs already do pretty well with RNA and protein because there are clear markers in the DNA sequence for genes (RNA sequences the perform many functions) and specific genes that encode proteins. This means that the models already generalize to multiple “languages,” before even attempting to train in other modalities, such as 3d protein structure, epigenetics and natural language.
If a model thinks in the DNA language, maybe it understands the imprint that environment left on different genomes as well? Perhaps the model has learned the functional relationship between different sequences, and could extrapolate to new sequences based on that?
And so what we wanted to showcase was that if we show the model progressively better RNAs in a series of steps with its score, right? So you have like low scores first and then you gradually move up the chain. Can the model continue that trajectory on i