š¬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
Summary
Radical Numerics is betting that genome language models will progress from reading DNA to designing it, with defense built into the same technical approach. Eric Nguyen traces the path from HyenaDNAās million-token context to Evoās generative genomics and Omniās task alignment. Because a model that designs biological function may also recognize it, the company has adopted a ādual mandateā spanning scientific discovery, human health, and biosecurity.
Omniās claimed breakthrough is not merely scale but the post-training that turns a foundation model into a usable genomic system. Evo and Evo 2 showed broad capabilities yet could lose to specialized models; Omni uses task-specific formatting, special tokens, supervised examples, and question-and-answer structures to answer practical questions such as whether a mutation causes disease. Nguyen says this has pushed it toward state-of-the-art performance across variant-effect tasks, especially in long-range, non-coding DNA.
The high-value wedge is the roughly 98% of the human genome outside protein-coding regions, where conventional interpretation remains weakest. Coding regions constitute only about 1.5ā2% of the genome, while many disease-associated variants sit in regulatory regions whose effects depend on context. Omni compares the likelihood of wild-type and mutated sequences, converts the difference into a āsurprise factor,ā and adds aligned training to produce directly usable predictions.
Radicalās most speculative design result treats biological optimization trajectories as a genomic version of chain-of-thought reasoning. Given RNA aptamers ranked from low to progressively higher fitness, the model continued the trajectory and generated plausible sequences with scores beyond those shown during training. Nguyen repeatedly hedges the resultāthe candidates still require wet-lab validationābut argues that the paradigm could extend to antimicrobial phages and selective rare-earth-binding proteins.
Long biological context, rather than isolated protein structure, is the companyās proposed technical advantage. Nguyen describes DNA as an āimprint of the physical worldā: upstream regulatory regions, neighboring genes, organism identity, and evolutionary history can constrain which designs are plausible. That framing supports a unified model spanning DNA, RNA, proteins, epigenomics, metabolomics, and eventually natural language, rather than a collection of single-modality tools.
Biosecurity is both a mission commitment and a potential product category, but the speakers do not pretend the screening problem is solved. Radical divides defense into detection, attribution, countermeasures, and government-level deterrence; its models are intended to recognize functional danger even when a sequence no longer matches a known pathogen database. RJās pushback is material: at the scale of billions of sequences, false positives could obstruct legitimate research, and unlike software vulnerabilities, āwe have fixed genomesā that cannot simply be patched.
Nguyen nevertheless sees an āarms raceā because design capabilities are advancing while biological defense remains significantly behind. The nearer-term danger may be accidental release, although state bioweapons programs remain a wartime concern. His explicitly named bottlenecks are GPUs and a cultural one: scientists need domain expertise while retaining enough imagination to challenge entrenched workflows.
Deep dive
1. Genome language models learn the grammar written into DNA
Nguyen defines a genome language model, or GLM, as a large language model trained on DNA letters rather than words. Radical Numericsā foundational bet is that AI can learn to āread and write DNA,ā exposing rules in the āfabric of lifeā that scientists still understand only partially.
The first goal was functional prediction: given an unprocessed sequence, infer regulatory effects and other biological properties. Because motifs can interact over great distances, a useful model must capture more than local patterns; it must learn the long-range āgrammarā connecting distant parts of a genome.
HyenaDNA addressed that constraint with a convolution-based Hyena operator instead of conventional attention. It processed contexts up to one million tokensāthen an unusually large language-model contextāand demonstrated that DNA models could retain signals across long sequences. Nguyen used ācontext rotā as an analogy for performance degradation in longer chatbot inputs.
The scale mattered because earlier genomic models commonly saw only 1,000ā2,000 bases at once, while even a typical human transcript is around 3,000 bases and the complete human genome contains roughly 3 billion. HyenaDNA established the reading capability; the more consequential question became whether the same machinery could write plausible biology.
2. Evo turned sequence prediction into generative genomics
Evo was designed to move beyond predicting function and generate new DNA. Nguyenās thesis was simple but initially unpopular: if generation transformed natural-language models, introducing generation into genomics might let scientists explore designs that laboratory search and human-written rules could not reach.
DNA offered a shared substrate across modalities because it gives rise to RNA and proteins. Rather than maintain separate generators for each, Evo could model multi-part systems such as CRISPR-Cas, whose operation involves DNA, RNA, and protein components, within one sequence framework.
Asked to produce new CRISPR-Cas systems after learning from natural examples, Evo generated a new candidate that attracted broad attention, including a Science cover and Nguyenās subsequent TED talk, because it made āwriting DNAā concrete rather than merely conceptual.
Scientists later demonstrated what Nguyen describes as the first functional genome created from scratch using AI: a bacteriophage genome. That result became a ākey turning pointāāevidence of design potential and of the potential to create organisms unlike those found in nature, but also evidence that models manipulating lifeās substrate could create unfamiliar biological risks.
3. Omni applies the alignment playbook to biology
Evo and Evo 2 were broadly capable, but specialized systems could still outperform them, particularly in human genetics. That left a fair question from scientists: āWhy use a giant LLM when I can use these smaller, more specialized models?ā Omni is Radicalās answer to that usability and performance gap.
Nguyen distinguishes pretraining from alignment. Pretraining supplies a āraw ability to create embeddingsā through next-token prediction or masked reconstruction; intermediate and post-training teach the model what users actually want, such as comparing a wild-type sequence with a mutation and returning a disease-relevant result.
Omni introduces special tokens, curated task datasets, demonstrations, and question-and-answer formats that identify the requested operation and expected output. The details remain intentionally vague as āsecret sauce,ā but the conceptual shift is clear: scientists should query a prepared model directly rather than extract embeddings, fit another head, and run a separate regression.
Radical experimented with both task-specific heads and a single unified head. Nguyen says either can win in individual cases, but the strategic direction is unification because one model sharing representations across DNA, RNA, proteins, and future modalities should create more opportunities for transfer and generalization.
4. Variant-effect prediction supplies Omniās first practical proof
Most individual DNA changes do nothing, while particular substitutions, deletions, inversions, or larger alterations can cause disease. Since clinicians know the significance of only a small fraction of possible variants across roughly 3 billion letters, patients often receive variants of unknown significance without a causal diagnosis.
A language model can score the reference sequence and its mutation, then compare their likelihoods. Nguyen describes the resulting ratio as a āsurprise factorā: a sequence far from patterns common or conserved in human genetics may be more likely to be pathogenic, although alignment adds task-specific evidence beyond that zero-shot signal.
Omniās strongest claimed advantage appears in non-coding and long-range regulatory regions. Protein-coding DNA represents only around 1.5ā2% of the genome; the remaining sequence helps control when and how strongly genes are expressed, yet its disease effects have historically been much harder to predict.
The benchmark set included Evo 2 and Borzoi, a supervised model mapping DNA directly to functional tracks such as chromatin accessibility and gene expression. Radical also targeted Caduceus, which Nguyen describes as the longstanding ākitchen sinkā benchmark because it combines leading methods with an SVM; the internal goal was to beat the aggregate, not select a convenient weaker comparison.
5. Better benchmarks require both supervision and leakage control
Evo 2ās raw likelihood scoring represents the zero-shot baseline. Other researchers had already shown that fine-tuning Evo-derived outputs could reach state-of-the-art performance; Omni goes another step by embedding the benchmarkās structure and requested answer format into intermediate and post-training.
RJ presses on a central credibility risk: benchmark variants or nearly identical sequences may already exist in massive genomic pretraining corpora. Nguyen says Radical explicitly reviews and curates the data, removes duplicates and potentially similar benchmark sequences, and applies bioinformatics quality control, although he offers no quantitative contamination audit in the conversation.
The commercial pull is already visible in requests from hospital systems and nonprofits holding patient genomes. They may know a patientās phenotype but cannot identify which DNA alteration caused it; Nguyen presents unresolved variants of unknown significance as one of Omniās clearest potential diagnostic applications.
6. Biological chain-of-thought means learning an optimization trajectory
Natural-language chain-of-thought works by āshowing your workā: intermediate tokens expose the steps leading to an answer and give the model more room to reason. Biology has no equivalent words, so Radical tested whether progressively improving sequences could serve as the intermediate steps.
Its experiment used RNA aptamers paired with measured fitness. The model saw sequences ordered from lower to progressively higher scores, while a portion of the best examples was withheld, and was then asked to continue the trajectory. It generated plausible aptamers that received some higher scores than those shown during training.
Nguyen is careful about the boundary of the result: āWe are currently checking those sequences in the laboratory.ā Until those wet-lab tests are complete, the experiment shows that the model can extrapolate the scoring pattern, not that it has necessarily produced a better physical molecule.
The underlying selection process offers a reusable template. Researchers generate enormous librariesāRJ estimates that one such experiment involved on the order of 10^11 sequencesāmeasure binding or another phenotype, retain better candidates, mutate them, and repeat; recording those intermediate rounds could let a model learn the same gradual optimization in silico.
7. Context is the proposed advantage across therapeutics and industrial biology
RJās challenge is sharp: designing a single protein often looks like a structural problem where protein-specific models should have the advantage. Nguyen agrees that isolated structure design is not Radicalās natural edge; its advantage appears when organism, genomic neighborhood, regulatory sequence, and evolutionary provenance materially constrain the answer.
For rare-earth extraction, Radical is working with a national laboratory on proteins that bind one target rare-earth element while avoiding others. The model can use the source microorganism and non-coding regions upstream of a relevant gene as context, then generate plausible variants whose selectivity can be tested comparatively.
Nguyen calls DNA an āimprint of the physical world.ā The desired protein is only one component; surrounding sequence records something about where it arose, how it is regulated, and which biological function it performs, allowing genome models to narrow design space in ways an isolated structure may not.
A second application is antimicrobial resistance. Designed bacteriophages might selectively kill bacterial strains rather than act like conventional antibiotics, addressing a problem Nguyen associates with roughly 2 million deaths from bacterial infections worldwide; progressively stronger phage genomes could also fit the chain-of-thought optimization format.
8. A unified biological model must extend beyond virtual cells
Radicalās public Omni work emphasizes humans because Evo and Evo 2 were perceived as weaker on the human genomeās noisy, repetitive, complicated grammar. Nguyen nevertheless describes the ambition as covering āall areas of life,ā including prokaryotes and viruses relevant to therapeutics and biosecurity.
DNA is the preferred foundation because RNA, proteins, and other molecular states ultimately connect back to it. The aptamer experiment is suggestive: the model was pretrained on genomes and DNA, received RNA only for that task, yet appeared to transfer implicit structural information into a new modality.
Nguyen argues that many āvirtual cellā efforts over-focus on RNA transcripts. A serious cellular model should combine DNA, transcriptomics, proteins, metabolomics, epigenomics, chromatin accessibility, and methylation; Radical does not want to stop at the cell, but to integrate āall these sensorsā across biology.
Natural language is not yet one of the modelās modalities, though Nguyen expects it to become one. He considers text-image and text-video integration comparatively understood; the more difficult recipe is joining heterogeneous biological signals without losing their shared structure.
9. Long context grew from a research obsession into the companyās origin
RJ asks whether synthetic-biology design now requires a context length of roughly 2 million, still far below a 3-billion-base human genome. Nguyen treats that gap as an active research frontier rather than a solved scaling exercise: larger organisms require architectural, systems, and hardware innovation, not merely another open-source model.
Long context has been Nguyenās focus since his Stanford doctorate. His teamās enthusiasm reaches GPU implementation detailsāāDid someone say kernels?ā is his caricature of how quickly the researchers become absorbed in making long-sequence operations faster.
The first AI-generated bacteriophage genome did not require genome-scale human context; its genome was only about 6,000 base pairs. Viruses are āincredibly efficient,ā but the example also limits what it proves: success on one exceptionally compact genome does not resolve whole-organism design at bacterial or human scale.
Nguyen spent six months asking Stanford scientists what they would do with generated DNA and heard that the idea was useless, unverifiable, or impossible because biologyās rules were too noisy. People at Arc nevertheless agreed to spend GPUs on the experiment; Evoās first checkpoint then proved competitive on ProteinGym despite never being explicitly told what proteins were.
10. Interpretability could turn model internals into biological discovery
Evoās unexpected cross-modal results led NVIDIA to support the larger Evo 2 effort, while Greg Brockman took a four-month break from OpenAI and joined late-night debugging sessionsāsometimes at 3 a.m. The origin story is less a planned product roadmap than a sequence of surprising capabilities that demanded practical validation.
Radical now treats mechanistic interpretability as a potential branch of biology. A model compresses recurring structure into weights, activations, and embeddings; inspecting that compression may reveal which motifs and long-range relationships it used, including patterns scientists have not yet articulated.
The early targets move from simple to complex: GC content, repetitive sequence, transcription-factor motifs, and representations of disease diversity. Transcription factors bind DNA and alter gene expression, while also being encoded and regulated by genes themselves, creating a combinatorial system too large to map manually.
Nguyen sees the disease ālandscapeā inside model representations as a starting point, not a finished result. Extending interpretability across dozens of modalities could convert predictive models into instruments for mechanistic discoveryāāall green spaceāābut the episode offers visualizations and hypotheses rather than validated new biological laws.
11. Biosecurity turns the same models into an explicitly defensive stack
Radicalās ādual mandateā begins with a symmetry: a model that generates functional sequences should also be good at distinguishing whether a sequence is pathogenic. Nguyen argues that natural-language refusalsāsuch as a chatbot declining to discuss virusesāare useful but cannot replace inspection of the biological substrate itself.
He divides biosecurity into four layers: detection and environmental surveillance; attribution of whether a threat is natural, accidental, or engineered; countermeasures such as antivirals or antimicrobials; and deterrence, largely a government responsibility. Radical intends to concentrate on the first three.
Traditional screening compares an ordered sequence with databases of known pathogens. That approach can miss something novel or deliberately altered: different letters may preserve the same structure or function, as illustrated by Microsoft work Nguyen calls biological āparaphrasing.ā Function-aware models are intended to flag the capability even when spelling-level similarity disappears.
The immediate deployment surface includes DNA-synthesis companies, which can manufacture a submitted digital design and return physical DNA ālike an Amazon package.ā Existing providers already screen orders, but Nguyen argues that their tools are generally not AI-based and are likely unreliable because they mostly compare sequences with known samples.
RJās objection survives the pitch: a screen operating at the scale of billions of sequences needs an extraordinary ROC curve, or false positives will obstruct ordinary science. Nguyen does not claim perfect prevention; āthe bar is improvementāāgiving manufacturers, licensed laboratories, regulators, and monitoring systems a stronger starting tool than exact sequence matching.
The cybersecurity analogy also breaks down because humans have āfixed genomesā that cannot be patched after every exploit, while turning printed DNA into a successful pathogen remains difficult. Nguyen concedes those differences but still sees āessentially an arms raceā: design capability advances, attackers adapt, and defense must continually move forward.
His threat model includes state programs, especially during wartime, as well as accidental release by researchers manipulating biological function without adequate containment. He considers unintended incidents potentially more likely in the near term because AI lowers required expertise and accelerates experimentation, increasing the volume of attempts.
Even so, Nguyen says the prospect does not keep him awake and remains āmuch more optimistic.ā He believes health and discovery benefits significantly outweigh potential harm, while the research and policy communities can mobilize; Radicalās purpose is to narrow the defensive gap before capabilities scale further.
Asked what he would remove as a bottleneck, Nguyen gives the standard frontier-lab answerāGPUsāthen a cultural one. Deep expertise can produce pessimism about every new approach; Radical wants domain experts who retain the imagination to change established scientific workflows.
His closing message is that AI researchers ādonāt need to chooseā between frontier technology and work that advances humanity. Radical Numerics is intended to make biology itself a frontier-AI problem, combining architecture, hardware-aware optimization, multimodal modeling, and wet-lab relevance rather than consigning leading talent to another chatbot.