Latent Space

Eric Nguyen on why the model that writes DNA must also guard it

Eric Nguyen· CEO and co-founder of Radical Numerics at Radical Numerics
·~92 min·English·Latent Space
AI SafetyLLMTrainingAI Company
TL;DR

Radical Numerics' new Omni model shows DNA language models can now beat specialized tools at predicting disease-causing mutations, and Eric Nguyen argues the same generative models must be turned into sequence-level biosecurity defenses.

01Core Mental Model

DNA Is a Language

A genome language model is an LLM trained on DNA instead of words, learning to read and write the letters of life.

a large language model trained on DNA sequences. So very much like natural language and chatbots you see but not train on words or natural language but on the raw fabric of life which is these sequence of letters that make up DNA.

— Eric Nguyen, Latent Space
Key Insight
Calling DNA a language is not a metaphor for Nguyen, it is a commitment: it lets his team import the entire LLM playbook (scale, tokenization, alignment) into biology, which is the reason a general model can start beating tools hand-built for one task.
02The Unlock

From Reading to Writing

The leap from reading DNA to writing it let these models invent a working CRISPR system and, later, the first genome built from scratch by AI.

what you can do with a generative DNA model was to generate the first genome from scratch using AI.

— Eric Nguyen, Latent Space
Key Insight
Generation flips biology from cataloguing what already exists to proposing what does not, and that same power is exactly what forces a safety question the read-only era never had to ask.
03The Thesis

One Model, Two Mandates

The same model that is good at generating DNA is also good at spotting a dangerous sequence, so Radical Numerics builds both the design tool and its defense.

A model that is good at generating turns out is also very good at discriminating or predicting if a sequence is pathogenic or not.

— Eric Nguyen, Latent Space
Key Insight
The dual mandate is a moat as much as a principle: because the generator and the detector are the same model, a lab that leads on design is unusually well placed to lead on defense too, while a defense-only effort is left chasing a frontier it does not itself set.
04How Omni Won

Alignment Made the Difference

Omni passed specialized tools after Radical Numerics added the alignment steps, mid- and post-training, that made chatbots useful.

EVO was just showcasing the potential of that pre-training but Omni is a step of actually making it useful for folks like scientists

— Eric Nguyen, Latent Space
Key Insight
Omni's jump suggests much of the gap in genomics was about training rather than raw capability: the missing ingredient was the alignment stack the language-model field had already built, so advances in ordinary LLM training can keep feeding this work.
05The Mechanism

The Surprise Factor

A DNA model flags a likely disease variant by measuring its own surprise: a rare, unexpected letter scores as more suspicious than a common one.

the raw outputs of the language model which basically you can think of it like a probability for predicting what the next letter is

— Eric Nguyen, Latent Space
Key Insight
Because the base signal is just the model's own surprise, it needs no labeled disease data to reach the vast majority of variants no clinician has ever annotated; a smaller supervised step then sharpens that raw signal into a clinical call.
06Where It Shines

The Dark Genome

Old tools largely focused on the small protein-coding fraction of the genome, but much of human disease hides in the non-coding regions where Omni does best.

turns out many if not most of the diseases are in these non-coding regions and so there's been a real desire to build models that can actually pick up these variants

— Eric Nguyen, Latent Space
Key Insight
The one-and-a-half-to-two-percent figure reframes the whole field: tools optimized for the protein-coding fraction were always at their weakest in the non-coding regions where much of human disease actually lives.
07The Defense

Beyond the Spell-Check

Most biosecurity screening today matches spelling against a list of known pathogens, which misses novel or reworded sequences, so Nguyen wants defenses that read function instead.

you can have a sequence that has the same function like a pathogen but actually look different in terms of the letters.

— Eric Nguyen, Latent Space
Key Insight
Function-aware detection is a direct consequence of better generators: the same models that can reword a pathogen to dodge a spell-checker are also what is needed to catch the rewrite, so defense has to be built on the attacker's own tools.
08The Stakes

An Arms Race You Can't Opt Out Of

Design capability is racing ahead of defense, and Nguyen argues the lagging defensive side has to be pulled back up to par before the gap widens.

the design side's going to get more capable. The defensive side needs to try its best to get ahead.

— Eric Nguyen, Latent Space
Key Insight
The cybersecurity analogy Nguyen reaches for has a limit that surfaced in the conversation: software gets patched, genomes do not. That is the real case for detection and surveillance over prevention, since you cannot assume you will stop every dangerous design, so you invest in catching what gets out.