Eric Nguyen：生物安全是一场 AI 军备竞赛

Brandon Anderson

欢迎来到 Latent Space。我是 Brandon，在 Atomic AI 做 RNA 治疗。联合主持是 Mirror Omics CTO、联合创始人 RJ Honicky。今天请来 Radical Numerics 联合创始人兼 CEO Eric Nguyen——他在 Chris Ré 组读博，很早就做长上下文基因组模型；一作并推动 Evo 生成式基因组平台，再到 Evo 2，自然延伸到 Radical Numerics。有没有漏掉的？

Welcome to Latent Space. I’m Brandon—I build RNA therapeutics at Atomic AI. Joined by co-host RJ Honicky, CTO and co-founder of Mirror Omics. Pleasure to have Eric Nguyen, CEO and co-founder of Radical Numerics. He got his PhD in Chris Ré’s group, thought about long-context genomic models before they were cool; first author and visionary behind Evo, then Evo 2, which naturally led into Radical Numerics. Did I miss anything?

Eric Nguyen

听起来很好。谢谢邀请。

That sounds great. Welcome—thank you.

RJ Honicky

Eric，我们想谈 Omni 和你们的基准博客，但先问：什么是基因组语言模型？为什么要关心？能做什么？然后再说博客的头条结果。

Eric, let’s talk about Omni and the blog post on benchmarking—but first: what is a genome language model? Why do I care? What does it do? Then the topline results.

Eric Nguyen

基因组语言模型（GLM）就是在 DNA 序列上训练的大语言模型——很像聊天机器人，但不训自然语言词，而训生命的原材料：组成 DNA 的字母串。我们团队以首批生成式基因组模型著称：不只读 DNA，还能写——生成新序列。我们觉得这块被忽视了：若 AI 能读写 DNA，就可能改变科学发现、对人类健康的理解与治疗。于是在基因组上训练 AI 是巨大机会。

A genome language model, or GLM, is a large language model trained on DNA sequences—very much like chatbots, but not on words or natural language, on the raw fabric of life: sequences of letters that make up DNA. Our team is known for creating the first generative genomics models: trained not just to read but to write DNA—generate new sequences. We felt this was overlooked; if AI could read and write DNA, it could change scientific discovery and understanding of human health. So there was a big opportunity to train AI on the genome.

Eric Nguyen

我们最早做 HyenaDNA：仍是 LLM，但用卷积而非注意力。DNA 极长，当时 LLM 上下文受限，我们找更高效算法，提出 Hyena 算子，可处理长达百万的序列——当时语言模型里最大的上下文。用来读 DNA、预测功能：给定字符串，预测调控功能、对基因组的影响。有趣在于：我们对自身基因组知之甚少——编码「我们是谁」与疾病，但字母如何组合成功能与性状的「语法」远未搞清。

When we first worked on DNA models we built HyenaDNA—an LLM using convolution instead of attention. DNA is very long; LLMs then had limited context, so we sought a more efficient algorithm and came up with the Hyena operator, processing up to a million tokens—then the largest context for a language model. We used it to read DNA and predict function: given a string, predict regulatory function and effect on a genome. Scientists care because we don’t know that much about our genome—it encodes who we are and disease risk, yet the grammar of how letter combinations encode function and traits is not fully understood.

Eric Nguyen

希望是用 DNA 语言模型从原始序列映射功能。第一代 HyenaDNA 表明：AI 可以在一定程度上读懂 DNA，尤其长程相互作用——类似「context rot」：输入越长，模型越容易劣化；我们要在百万上下文上抓住长程模式与语法。真正开启如今「生成式基因组学」的是 Evo：不只读，还要生成。想加速生物学家从基因组学习；生成式 AI 已极大加速自然语言的理解与操控，而 DNA 这门语言几乎还没被认真用 AI 生成——当时模型又小又短上下文，只能读小模式，无人写 DNA。自然语言里生成如此强大，若带到生物学会解锁什么实验室做不到的事？

The hope was mapping function from raw DNA with these LMs. First-generation HyenaDNA showed we can train AI to read and understand DNA to some degree—especially longer-range interactions. Like “context rot”: longer inputs deteriorate; we wanted long-range motifs and grammar over a million tokens. What started generative genomics was Evo: not just reading but generating DNA. We wanted to accelerate how biologists learn from genomics. Generative AI accelerated natural language; DNA was barely touched—models were small, short-context, read-only. Generation transformed language—what if we brought that to biology? What does that unlock that you can’t do in a lab?

RJ Honicky

若你能把这件事做好，最先展示、让人感到潜力的是什么？

What was one of the first things you showcased that got people intrigued by the potential?

Eric Nguyen

CRISPR-Cas——能切割 DNA 的酶系统。Evo 特别之处是能跨越多种模态与尺度生成：CRISPR-Cas 由 RNA 与蛋白质构成；当时你有蛋白 LM、RNA 模型，却没有单一系统共设计。我们展示：单一 DNA 模型是两者的基础——从 DNA 可得 RNA 与蛋白，因而可以共设计，并在真实世界中具有功能。

A CRISPR-Cas system—an enzyme that cuts DNA. What Evo enabled was generating across modalities and scales. CRISPR-Cas is made of RNA and proteins; you’d seen protein LMs and RNA models, but not a single system to co-design. We showed a single DNA model is the foundation of both—from DNA you get RNA and proteins—so you can design one system that also functions in the real world.

Eric Nguyen

我们给 Evo 看大量天然 CRISPR-Cas，问它能否做出新的；采样后确实发现了新系统——上了《Science》封面，后来还为此做了 TED。更激动人心的是：就在去年，科学家用生成式 DNA 模型从零生成了首个基因组——人类做不到「从零搭建」：通常是复制粘贴已知功能片段；在基因组尺度从地基往上建此前未有。Evo 生成了有功能的基因组：细菌噬菌体（一种病毒）。这是科学界与我们公司的转折点——既能创造自然界不存在的整株「生物」，也意味着若能操控生命原材料，潜在危害同样巨大。兴奋与担忧的反馈都很多。

We showed Evo natural CRISPR-Cas systems and asked if it could make a new one—and it discovered a new CRISPR-Cas; Science cover, later a TED talk. Even more intriguing: just last year scientists showcased generating the first genome from scratch with a generative DNA model. Humans usually copy-paste known motifs; building from the ground up at genome scale hadn’t been done. Evo generated a functional genome—a bacteriophage, a virus. A turning point for the community and for us at Radical Numerics: creating a whole organism not in nature, and the potential harm if you control the fabric of life. Lots of excited and concerned outreach.

Eric Nguyen

因此我们觉得：公司不仅要推生物设计能力，也要把这些模型用作防御工具，应对误用与生物风险。许多前沿实验室也在担心 AI 设计生物序列；但目前多从自然语言侧做护栏——比如 Claude 一谈病毒就切断——自然语言护栏有必要，但还需要生物序列层面的护栏：能理解一段序列是否致病或是否病毒，再推到环境监测与国家安全。把前沿能力带给防御社区，与用于人类健康同等重要。

So it was important not only to push biological design but also to use these models as defensive tools against misuse and biological risk. Frontier labs worry about AI designing biological sequences, mostly attacked from natural-language safeguards—Claude shutting down virus talk—which is needed, but you also need safeguards at the biological-sequence level: models that understand if a sequence is pathogenic or a virus, then surveillance systems and national security monitoring emerging sequences in the environment. Bringing frontier tech to that community is as important as human health, our primary focus.

RJ Honicky

从 HyenaDNA 到 Evo、Evo 2，现在是 Omni。Evo / Evo 2 在许多任务上难赢专用模型；博客里 Omni 却能在广泛任务上超过它们。差别在哪？

Evolution from HyenaDNA to Evo, Evo 2, now Omni. Evo and Evo 2 struggled to beat specialized models across many tasks; the blog says Omni outperforms across a wide range. Can you talk about that?

Eric Nguyen

Evo 展示了多模态潜力，但在人类遗传学/基因组学上仍常输给专科 DNA 模型——有竞争力，却未真正 SOTA，社区会问：为何用巨型 LLM 而不用更小的专用模型？Omni 要做的第一件事，是展示 mid-training 与 post-training（广义对齐）。生物基因组学里你大多只见过预训练基座、未对齐；自然语言里预训练之后还要大量对齐才能对用户真正有用。Evo 展示预训练潜力；Omni 是让它对科学家真正好用——用他们关心的形式呈现任务，例如给定野生型与突变，帮找因果变异。这类问答形态不会从下一 token 预训练里轻易冒出来。

Evo showcased multi-modality potential but still underperformed specialist DNA models on human genetics—competitive, not SOTA—so people asked why use a giant LLM vs smaller specialists. With Omni one of the first goals was mid-training and post-training—alignment. In bio/genomics you’ve mostly seen base pretrained, unaligned models. In natural language, alignment and post-training make models production-useful. Evo showcased pretraining potential; Omni makes it useful for scientists—tasks in the form they want, e.g. wild-type vs mutation, find the causal variant. That Q&A form doesn’t easily emerge from next-token pretraining alone.

Eric Nguyen

预训练教的是原始造模式能力；把学到的表征对准具体任务、让输出对你有意义，需要一番「拨弄」——前沿自然语言实验室在推这块，我们想把更多带到基因组学。只做一点点，我们就惊讶地看到：不只「多任务还行」像 Evo，而是在变异效应、疾病因果突变等上真正推前沿，开始对人类基因组学有用。这仍是预览——我们还在训、加模态——但太兴奋了，想尽快交到人手。早期兴趣来自医院系统、非营利：患者有症状却找不到致病 DNA 片段，大量「意义未明变异」（VUS），我们极想用模型帮助诊断。

Pretraining teaches raw pattern-making; pointing embeddings at tasks and aligning outputs takes teasing—frontier NLP labs drive that; we wanted to bring it to genomics. Even a little, and we were surprised it became SOTA—not just broadly capable like Evo, but pushing variant-effect and causal disease mutations, actually useful for human genomics. Still a preview—we’re still training, adding modalities—but we were too excited to wait. Early interest from hospital systems and nonprofits: symptoms without known causal DNA—variants of unknown significance—we’re extremely excited to help diagnose patients.

Brandon Anderson

对齐基因组模型具体长什么样？自然语言有思维链、RL 的清晰范式；对基因组是另一种语言上的微调吗？

What does aligning a genomics model look like technically? LLMs have chain-of-thought, RL—clear paradigms. For a genomic model, is it fine-tuning in a different language?

Eric Nguyen

某种程度上可叫微调，但关键是输入结构：用特殊 token 提示「这是疾病 A 预测」「这是设计任务」，再以类似思维链的方式展示期望输出轨迹——细节故意含糊，是还在打磨的秘方。很大程度模仿自然语言社区：微调、精选数据集、把任务结构化成问答，相对预训练的「整库下一 token / 掩码填空」。Mid/post-training 是：给定这类输入期望这类输出（分数或设计）；post-training 也可含强化学习。方向上我们更倾向统一单模型、单套头，以解锁跨模态泛化——DNA 是基础，RNA、蛋白、分子是下游；从 DNA 能学到很多其他模态，我把它们想成给模型的额外上下文。

In many ways fine-tuning, but key is input structure—special tokens so the model knows disease-A prediction vs design, then chain-of-thought-like trajectories of desired outputs—intentionally vague; secret sauce still developing. It mimics NLP: fine-tuning, carving datasets, structuring Q&A vs pretraining’s next-token or masked infilling. Mid/post-training: given this input, expect this output—score or design; post-training includes RL too. We’re flexible on multi-head vs unified, but moving toward a single unifying model for emerging cross-modality capabilities. DNA is the foundation; RNA/proteins/molecules are downstream—we think of other modalities as additional context.

Eric Nguyen

Omni 这次重点展示理解变异效应：DNA 某一位换成另外三个字母之一，有时致病、多数无影响；三四十亿字母组合下，临床只知极小部分哪些变异因果致病。ClinVar、TraitGym 等基准：给定突变，能否判断是否致病？DNA 模型是概率的——改一个字母会改下一字母预测的置信度；训过海量人类 DNA 后，模型大致知道什么「常见/保守」更偏健康，罕见则可能致病。先前 DNA 模型在人类、尤其在非编码调控区很挣扎——那里只约占基因组 1.5–2% 的编码区之外，控制基因何时表达；多数疾病其实落在非编码区。Omni 新一代在非编码、长程区域区分致病突变上表现突出，遗传学家用传统生物信息/统计方法长期吃力的地方。

One Omni focus was variant effects: changing one letter can cause disease or often nothing; over ~3B letters clinicians know only a tiny fraction of causal variants. Benchmarks from ClinVar or TraitGym: given a mutation, is it disease-causing? DNA models are probabilistic—a change modifies next-letter confidence; trained on so much human DNA they grasp what’s common/conserved (typically healthier) vs rare (potentially pathogenic). Prior DNA models struggled on humans, especially noncoding regulatory regions—coding is only ~1.5–2% of the genome; most disease sits in noncoding control of expression. Omni shines distinguishing disease-causing mutations in noncoding and long-range areas where geneticists struggled with traditional bioinformatics and stats.

RJ Honicky

博客基准表里有 Borzoi 参考列——能给听众一点历史上下文，并对照右边的 Omni 列吗？

There’s a Borzoi reference column in the benchmark chart—for listeners, what does that column represent, and how does it contextualize Omni on the right?

Eric Nguyen

我们放了深度学习侧与传统方法里代表性的强模型。Evo 2 是团队上一版基因组模型；Borzoi 也是 DNA 模型，但很不同：监督学习，从 DNA 直接预测功能基因组 tracks（染色质可及性、基因表达等）——天然多模态，却不是语言模型、不做下一 token。它需要标注输出；我们的 LM 在未标注原始序列上无监督预训练——绝大多数基因组数据本来就没有标注，这非常可取。指向具体任务时会有监督步骤，但核心是用似然：野生型与突变各自的下一字母概率，取比率当作「惊讶度」，再据此打致病与否的分。

We took representative strongest models on the deep-learning side and traditional methods. Evo 2 is our team’s previous genomic model. Borzoi is also a DNA model but different: supervised, predicting functional genomic tracks from DNA—chromatin accessibility, gene expression—inherently multimodal but not an LM, not next-token. It needs labeled outputs; our LMs do raw unsupervised pretraining on unannotated sequences—almost all genomic data is unlabeled, hugely desirable. For a task there’s a supervised step, but we use likelihood scores: probability of next letter for mutation vs wild-type, a ratio as a “surprise factor,” then score disease prediction.

Brandon Anderson

Omni 是自回归、扩散，还是你不能说的别的？

Is Omni autoregressive, diffusion, or something else you can’t tell me?

Eric Nguyen

这次不描述准确配方。Evo 是自回归，也是早期大规模自回归之一；我们不把自己绑死在某一训练目标上——工具箱里什么都用。对这个基准，你走的是 token 似然分布：模型觉得罕见的，往往对应进化约束——基因组里少见，可能有问题。这是一种解读，很像自然语言；mid/post-training 还能在「自然中常见」之上，为疾病变异预测任务加结构与额外学习。

We’re not describing the exact makeup this time. Evo was autoregressive—first large-scale autoregressive in the space. We don’t tie ourselves to one objective; every tool in the toolbox. For this benchmark you use the likelihood distribution—unlikely tokens often reflect evolutionary constraint; rare across genomes, probably causes problems. One interpretation, similar to natural language; mid/post-training adds structure on top for the disease-variant task.

Eric Nguyen

Evo 2 列基本是零样本似然比，不微调；另有人用 Evo 2 分数再微调（如 Goodfire）也能到 SOTA。我们的模型再上一步：不仅微调，还在 mid-training 把基准的问答格式写进模型本身，多一个增益。实践上科学家还省掉「抽 embedding、接头、再回归」的额外步骤——想象 ChatGPT 每次提问都要先针对领域微调；我们让模型在 mid-training 同时适应多任务，开箱按格式提示即可输出。数据泄漏方面我们极度警惕：有生物信息人员梳数据、去重、比对，基准相近序列会剔除，QC 很重。

Evo 2’s column is mostly zero-shot likelihood ratio—not fine-tuned. Others fine-tune on Evo 2 scores (e.g. Goodfire) and hit SOTA. Ours goes a step further: not just fine-tuning but introducing Q&A structure of these benchmarks into the model during mid-training—an extra boost. Practically, scientists don’t need to take embeddings, slap on a head, and regress—imagine ChatGPT requiring fine-tuning per domain every question. Mid-training on many tasks at once means out-of-the-box prompting in a format. We’re extremely cognizant of leakage; bioinformaticians comb, dedupe, align, remove near-benchmark sequences—extensive QC.

Eric Nguyen

很久以来 CADD 一类方法是 SOTA——设计上会把最佳方法做集成，甚至混 SVM，「厨房水槽全扔进去」。我们不挑一个弱对手夸自己更好，而要赢「人类目前能堆出来的全部」。研究者把目标定在全面超过。

For a long time CADD was SOTA for a reason—by design it ensembles the best methods, even mixes an SVM, throws the kitchen sink. We weren’t going to cherry-pick one model; we needed to beat what’s humanly possible across everything. Our researchers set their sights on improving across every method.

RJ Honicky

博客里提到思维链，第一次听到时脑子都炸了——DNA 上的 CoT 到底是什么意思？

There’s discussion of chain-of-thought that broke my brain when I first heard it—what does that even mean in DNA?

Eric Nguyen

这是设计能力可走向何处的预告。思维链源自自然语言——Jason Wei 在 OpenAI 等展示：让模型「展示解题步骤」就更好；即使简单步骤，也多了 scratch space。这是推理的开端。生物里很少见，蛋白设计刚冒头。我们想问：DNA 没有描述「思考」的词，同一范式如何引入？做了一个更简单的版本：RNA 适体数据集——序列加适应度分数，越高越好。按分数从低到高逐步展示，问模型能否自行延续轨迹、最终自优化到更高分。我们留出表现最好的一部分作 hold-out，只给较低分再排序展示；模型竟能复现一些从未见过的高分。湿实验室验证还在进行。硅基上已显示能延续轨迹、设计更高适应度的合理适体——若实验室成立，几乎可套到一切「序列 + 适应度」的设计任务上。

A taste of where design capabilities can go. Chain-of-thought stems from NLP—Jason Wei at OpenAI showed models perform better when they show their work; even simple steps give scratch space to think—the beginning of reasoning. Very little in biology; some in protein design. We asked: DNA has no words for thinking—how do you introduce the paradigm? A simpler version: RNA aptamer dataset—sequences with fitness scores, higher better. Show progressively better RNAs with scores; can the model continue the trajectory and self-optimize? We held out top performers, showed only lower ones ranked upward; it recapitulated some higher scores never shown. Wet-lab validation underway. In silico it continued the trajectory and designed plausible higher-fitness aptamers—if it works in lab, a paradigm for almost every sequence-plus-fitness design task.

Brandon Anderson

我对 RNA 上强提升有点惊讶——哺乳动物 RNA 往往缺乏好的共进化结构约束。能力从何而来？预训练是 DNA，mid-training 才少量碰 RNA？

Surprised this task saw strong improvement—mammalian RNA is notorious for lacking good co-evolution constraints on structure. Where might this capability have emerged? Pretraining on DNA, only a little RNA in mid-training?

Eric Nguyen

我们也惊讶。预训练全是基因组 DNA；RNA 只在这一任务、且仅限该数据集引入。你的直觉大概对：DNA 里大量进化信息可迁移——这也是我们坚持先在 DNA/基因组上预训练、再叠模态的原因。模态泛化是我们努力的方向。公司层面还有「通向通用生物智能」：统一生物学各「语言」/模态——归根到底都源于 DNA，但人们多用专科模型，没充分利用共享结构。虚拟细胞若只盯转录组不够；要融合 DNA、代谢组、表观、蛋白组等传感器。我们甚至不想停在细胞，而要跨生物学融合传感器。能否到理解生物学每个细节的「超级智能」不好说，但我确信这范式会远超现状。

We’re surprised too. Pretraining is all DNA genomes; RNA only for this task and only from this dataset. Your intuition about evolutionary information in DNA is probably where transfer comes from—why we think pretraining on DNA/genomes first then adding modalities matters. Modality generalization is what we’re working toward. As a company we’re building toward general biological intelligence—unifying biology’s languages/modalities. They stem from DNA; people use specialized models and don’t leverage shared structure. Virtual cells that focus only on transcripts fall short; fuse DNA, metabolomics, epigenomics, proteomics. We don’t want to stop at the cell—fuse sensors across biology. Superintelligence that understands every minutia? Who knows—but this paradigm gets us a hell of a lot further than now.

Eric Nguyen

范围上我们对所有生命域都感兴趣。这次侧重人类，因 Evo / Evo 2 在人类基因组上不够强，有人说人类语法太复杂、太噪、重复太多——我们想证明有用。原核与病毒也做，病毒对生物防御与安全尤其关键；治疗上也可向微生物学习——例如用噬菌体应对 AMR：全球细菌感染仍约两百万死亡，东欧长期用噬菌体，有望做成新一类抗微生物手段。我们被高影响、能救命的事吸引，不限于一种基因组。

We’re interested in all domains of life. We focused on humans here because Evo/Evo 2 weren’t as strong and people said human genomics is too complex, noisy, repetitive—we wanted to show usefulness. Prokaryotes and viral too—viral especially for biodefense and biosecurity; therapeutics from microbial life—bacteriophages for AMR: ~2M deaths from bacterial infections still; phage therapy long used in Eastern Europe; potential new class of antimicrobials. We gravitate to high-impact, life-saving work—not just one genome type.

Brandon Anderson

适体数据大致是：随机大量序列，用 RNA「开关」结合靶标（如 HIV 相关）后经 NGS 高通量读数——结合越多读数越多即适应度越高，再突变迭代。湿实验在用生物学做筛选；你们是在硅基上做同一套递进精炼。通量可达约 10^11 量级序列并行探索。

Aptamer data: generate many random sequences; an aptamer switch cleaves on binding—NGS readout, high throughput—more reads imply higher fitness for e.g. an HIV target; mutate and iterate. Biology does the filtering; you’re doing the same progressive refinement in silico. Throughput like 10^11 sequences explored in parallel.

Eric Nguyen

对。同样范式可套到选择性杀灭特定细菌株的抗菌效果评分，或我们与国家实验室合作的另一设计任务：设计提取稀土矿物的蛋白——不只结合，还要对某一稀土选择性高、对其他低，用一系列递进的合意分数做「提示工程式」设计。

Absolutely. Same idea for antimicrobial resistance—scores 0–1 for selectively killing strains—or another design task with a national lab: proteins extracting rare-earth minerals. Not just binding—selectivity for one mineral vs others; show progressively desirable scores. Creative prompt-engineering of sequence-plus-fitness.

RJ Honicky

稀土场景下，你们的优势似乎在整株/大尺度，而非单个蛋白结构。竞争壁垒在哪？模型除了功能是否也理解结构？

Rare earths seem less your comparative advantage vs structural protein methods—your strength is larger-scale organism design. Where’s your strongest edge? Does the model understand structure as well as function?

Eric Nguyen

纯结构设计未必是我们的战场；稀土上我们的假说是上下文很重要——可提示「在这些微生物/基因组邻域里搜新蛋白」，并喂入目标基因前的非编码区，再要变异体：在已知微生物上下文中组合自然界见过的多样性。EVO 用户曾用类似手法做毒素-抗毒素：提示蛋白上游再生成合理变体。稀土上可快速测结合与选择性——对关心美国稀土供应链安全的合作方很有价值。我们不是在设计整株生物产蛋白，但生物体告诉我们哪些蛋白合理、选择性如何演化——调控元件也提供上下文。我把基因组看成物理世界烙在 DNA 上的印记；目标片段周围的序列讲述它从哪来、如何获得功能。

Pure structure design isn’t necessarily where we’re advantaged; for rare earths our hypothesis is context matters—prompt Omni with the neighborhood of genomes/microbes to search, feed noncoding regions before the gene of interest, ask for variants combining diversity seen in nature in that context. People used Evo similarly for toxin–antitoxins—prompt upstream, generate plausible others. Exciting for rare earths: test binding and selectivity simply—worth supporting partners securing US rare-earth supply chains. We’re not designing the whole organism here, but the organism tells us which proteins are plausible and how selectivity evolved—regulatory elements add context. Genome as imprint of the physical world into DNA; sequence around what we want tells where it came from and how function arose.

RJ Honicky

你们现在大约两百万上下文，人类基因组大约再长一千倍；细菌基因组工程也远超。有限上下文如何做大尺度合成生物？Evo 2 噬菌体设计的上下文大概多长？

You’re at ~2M context; human genome ~1000× longer; bacterial genomes for engineering also longer. How do you leverage long-but-finite context for large-organism synbio? How did Evo bacteriophage design work context-wise?

Eric Nguyen

噬菌体那项是另一组做的；他们选噬菌体正因为基因组极短——大约六千碱基对，短得惊人却极高效。如何用「相对小」的东西做更长上下文，是整个 AI 社区持续要创造性解决的问题——也是我们首先是 AI 研究实验室的原因：不能只抓开源模型就指望关心的任务被解决。长上下文是核心研究方向；公司名也与此有关——2022、2023 年我们就在长上下文还没成气候时做这事。

The phage work was a separate group; they started with phages because among the shortest genomes—around 6,000 base pairs—insanely short yet incredibly efficient. How you get longer context with something smaller is a key question the AI community constantly works on—and why we’re an AI research lab first: you can’t just grab open-source models and expect our tasks solved. Context is a key research area; probably how we got our name—we worked on long context in 2022–2023 before it was a thing.

Eric Nguyen

想 nerd-snipe Radical Numerics 的研究员，就提长上下文、kernel、GPU。长上下文对我特别亲：斯坦福博士课题，也是我们进入 DNA 的路径。当时在做通用语言模型，发现模型擅长长上下文；实验室同伴 Michael Poli 做了 Hyena 卷积架构的第一版设计。我们问：若死磕长上下文，什么新应用会打开？最长的序列是什么？答案几乎必然是 DNA——三十亿碱基对。当时基因组模型上下文往往一两千 token，连许多人类转录本（约 3K）都不够。需求清晰却被忽视。相对开箱就读得不错；人们不停问「能做什么」——于是想到：没人在写 DNA，我们能不能让它写？

Want to nerd-snipe researchers at Radical Numerics? Bring up long context, kernels, GPUs. Long context has a special place—my Stanford PhD focus, and how we started on DNA. We were working on language models, noticed they were good at long context; lab mate Michael Poli worked on the first Hyena convolutional design. We asked what opens if we lean into long context—what’s the longest sequence? Unsurprisingly DNA: three billion base pairs. People were doing ~1–2K tokens—not even enough for many human transcripts (~3K). Clear overlooked need. Relatively out of the box it read DNA well; people kept asking what you can do with it—then: nobody’s writing DNA; can we get it to write?

Eric Nguyen

现在回头几乎显而易见，但刚推销 Evo 时我花了约六个月在斯坦福问科学家：若我生成 DNA，有用吗？你们会支持吗？几乎人人觉得蠢——「你拿它干什么？」「不可能，我们人类都不懂规则，AI 怎么学？你甚至不能告诉它对错」「重复太多、分布太噪、全是垃圾」。我很固执：生成 DNA 一定有用例，只是还不知道是什么。第一次训 Evo 时完全不知道会不会 work；Arc 的人说「为什么不试试、花点 GPU 钱让这些疯孩子训一把」。第一张 checkpoint 扔进 ProteinGym，竟与蛋白专用模型有竞争力——我们从未告诉模型什么是蛋白。随后 RNA、DNA 上不断追平甚至 SOTA，感到跨模态在学习。再后来 NVIDIA 支持 Evo 2，Greg Brockman 从 OpenAI 休假四个月帮这些疯孩子——凌晨三点还在 Slack 跟我们 debug，很疯狂。

In hindsight obvious, but pitching Evo I spent six months asking Stanford scientists: if I generate DNA, useful? Would you back us? Almost everyone thought it was stupid—“What would you do with it?”“Impossible—humans don’t understand the rules; how can AI? You can’t even tell it right or wrong.”“Too many repeats, too noisy, junk.” I was stubborn—there’s got to be a use case. First training Evo we had no idea if it’d work; folks at Arc said why not—pay for GPUs, let the crazy kids train. First checkpoint thrown at ProteinGym was competitive with protein-specific models—we never told it what proteins were. Then competitive/SOTA on RNA and DNA—learning across modalities. NVIDIA backed Evo 2; Greg Brockman took a four-month sabbatical from OpenAI to help—Slack at 3 a.m. debugging our code. Wild.

Eric Nguyen

轨迹令人惊讶，但仍有人问：真实世界有什么用？这促使我们创业——学术上只展示了「味道」，像早期语言模型被当成玩具、写笑话，人们问不出 AGI。DNA 上我们看到类似甚至更激动的轨迹。噬菌体例子既有说服力也可怕；也有人批评只是贴近自然的小变异。对公司而言，更重要的是真正改进疾病理解与治疗、做出更好的稀土提取器——在科学问题之外，让它有用。

The trajectory was surprising, but we still got “what are these useful for in the real world?” That motivated starting a company—we’d shown an academic taste, like early LMs seen as toys that write jokes, not AGI. DNA feels reminiscent, even more exciting. The phage was persuasive and scary; fair critiques say it recapitulates small natural variants. What’s more important now is using this to improve disease understanding and treatments, better rare-earth extractors—make it useful, plus scientific questions.

RJ Honicky

博客末尾的机制可解释性，以及生物安全——先谈 mechan：Evo 就有反问模型、抽取生物学洞见；生物学特别适合问「你学到了什么模式？」

Mech interp at the end of the blog, and biosafety—mechan first. Evo already reversed the question to extract biological insight; biology especially lends itself to asking what patterns the model learned.

Eric Nguyen

Mechanistic interpretability 在生物是新兴领域，我们还早，但在建团队。目前兴奋于分析 embedding 与激活：模型压缩所见信息、蒸馏关键模式；我们探测权重与激活里蒸馏了什么。从输出/激活可视化理解 DNA 复杂性：GC 含量、转录因子 motif、调控模式——模型要完成致病判断等任务就必须压缩这些。目标是找到可泛化到我们尚不懂的模式的结构。转录因子可结合 DNA、改变表达而不改序列本身，影响疾病与衰老等；组合空间太大，只能数据驱动学 motif。疾病流形——疾病在模型输出/嵌入空间里的结构——是我们想映射的大方向。DNA 只是最终要融合的传感器之一；下一世代会越来越多模态，绿色空白极大。自然语言融合相对直接（已有图文视频先例、早有实验）；更难的是跨生物模态融合。有了语言也更利于工具编排式思维链。

Mech interp in bio is emerging—we’re early, building the team. Excited about embeddings and activations: models compress information into key patterns; we probe what was distilled into weights and activations. Visualize structure in outputs—GC content, TF motifs, regulatory patterns the model must pick up for disease tasks. Find structure that generalizes where we don’t understand patterns. Transcription factors bind DNA and alter expression without changing DNA—disease, aging; combinatorial space too vast for manual work—data-driven motifs. Mapping the manifold of disease in output/embedding space is a big research thrust. DNA is one sensor to ultimately fuse; next models increasingly multimodal—huge greenfield. Fusing natural language is relatively straightforward—image/video precedents, early experiments—trickier is fusing biological modalities. Language also enables tool-orchestrating chain-of-thought.

Eric Nguyen

公司有「双重使命」：对设计侧能力自觉负责。若我们造能把功能写入序列的模型，就看到护栏缺口。昨晚我还在 AI 科学家 agent 的讨论会上，末日场景人人提到生物武器，却几乎没人说自己在做什么——公司也看不到大力投入。我们觉得：同时做设计的团队最适合做防御——同一类模型，生成强则判别「是否致病」也强。原则与战略上都该做，也与美国政府及海外社区共鸣。

Dual mandate: cognizant and responsible for design-side capabilities. If we create models that design function into sequences, we see a gap in safeguarding. At a panel on AI scientist agents last night, every doomsday answer was bioweapons—and I heard nothing about what they’re doing; companies lack big efforts. A team building design is best suited for defense—same models; good at generating is good at discriminating pathogenic vs not. Principled and strategic; resonated with our team and US government and abroad.

Eric Nguyen

生物防御大致三到四根柱：一、检测与监测——机场拭子、污水等环境样本里有无病原序列；二、归因——检出后是自然、他国还是某实验室工程化；三、反制——抗病毒/抗菌对策；四、威慑偏政府层。我们主要做前三。缺口不只是「是否致病」，还要刻画危险基因片段、用签名归因，而非只对齐已知库。社区多靠序列匹配：「见过吗？在已知病原名单上吗？」担忧是：新东西不在名单上；以及故意在序列空间混淆、却在功能/结构空间同构——我们赋能的模型正走向功能与结构感知，字母不同、功能相同。Microsoft「paraphrases」工作：用蛋白 LM 保持结构（近功能）却改拼写，并测试能否骗过现有检测。DNA 合成公司像亚马逊包裹一样按设计寄回物理 DNA，支撑整个科研管线；在 agent 遍布网络的世界，若没有工具告诉它们是否在制造危险物，很荒唐——这是我们最早动机之一。许多合成公司有检测，但我强烈怀疑多非 AI、不够稳健，主要是模式匹配。

Biodefense has three or four pillars: detection/surveillance—airport swabs, sewage; attribution—natural vs engineered, which lab/country; countermeasures—antivirals/antimicrobials; deterrence more governmental. We focus on the first three. Missing piece isn’t only pathogenic-or-not but characterizing dangerous parts, attributing signatures—not just database matching. Community focuses on alignment to known pathogens. Concerns: novelty not on the list; intentional obfuscation in sequence space while preserving function/structure—models we’re enabling become function- and structure-aware: different letters, same function. Microsoft paraphrases: protein LMs keep structure (implying function) while changing spelling, testing whether detection breaks. DNA synthesis companies mail physical DNA like Amazon packages—the research pipeline. In a world of prolific agents, no tools to say if something’s dangerous is wild—our first motivation. Many synthesizers have detectors; I strongly hypothesize they’re not AI-based and not robust—mostly pattern matching.

Brandon Anderson

科学上探索新序列时，很难找到既不妨碍正当研究、又能在海量序列下可承受的 ROC——哪怕 F1 很高，假阳性也会淹没。如何平衡？

For scientists exploring novel sequences, it’s hard to live on an ROC that doesn’t impede legitimate research—even a high F1 with billions of sequences could flood false positives. How do you balance?

Eric Nguyen

若问如何 100% 挡住危险设计，很难。我们问的是：设计侧有很多人推前沿，防御侧有没有同等前沿技术？答案是没有。巨大缺口，我们要把防御补到更接近正面对抗。能否解决一切？我们会 aspire，但现实总会有绕过——这正是为何自然语言护栏只是一层。

100% stopping dangerous design is a harder question. We asked: plenty push design frontier—is frontier tech applied on defense? No. Huge gap; bring defense to a better head-to-head. Will it solve everything? We aspire, but there will always be ways around—why chatbot safety is one layer.

Eric Nguyen

有些东西终会越过聊天机器人护栏、已经落到环境里。我们的工具面向「已经在外面」的生态：监测、归因、反制——不必完美才有用，把前沿 AI 带给生物防御社区、推动进步即可。没有工具就什么都做不了；标尺是相对现状挪动指针，超越序列匹配与比对。监管与智库侧讨论在增多，我们乐观人们会从空谈走向行动——很多 AI 公司说「该做生物防御」，然后呢？我们的做法是把工具做出来、开放接入。

Things get past chatbots into the environment. Our tools are for what’s already out there—surveillance, attribution, countermeasures. Needn’t be perfect to help; bring frontier AI to biodefense and move forward. No tool means you can’t do anything; the bar is moving the needle beyond sequence matching. Regulatory and think-tank chatter is growing; we’re optimistic about action vs talk. AI companies say they should do biodefense—then what? Our approach: build tools and give access.

RJ Honicky

钢人对方观点：你们若同时推攻击与防御前沿，长期是否互相抵消？如何想？

Steelmanning: if you’re on the frontier of attack and defense at once, does helping now not matter long-term? How do you think about that?

Eric Nguyen

别做成「做一个生物防御工具就收工」——本质上是军备竞赛，像网络安全：设计侧更强，防御侧必须尽力领先，再刺激设计侧……在我们看来防御远、远落后，要把防御拉回接近均势。这是我的框架。

Mindset isn’t “ship a biodefense tool and call it a day”—it’s an arms race, like cybersecurity. Design gets more capable; defense must try to get ahead, which motivates design further. Inherently arms-race dynamics; defense has been far, far lagging. We want to bring defense closer to par. That’s the framing.

Brandon Anderson

网安类比有关键不同：原则上或许能堵上所有非社会工程漏洞并打补丁；基因组却不能「给人类打补丁」。同时进攻激励与「从打印 DNA 到成功活病毒且不杀死设计者」的壁垒也可能不同。最让你夜不能寐的单一威胁是什么？

Cyber analogy differs: in principle you might catch every exploit and patch; we have fixed genomes—you can’t patch a human. Offense incentives and the barrier from printing DNA to a successful virus that doesn’t kill the designer also differ. What’s the single biggest threat that keeps you up?

Eric Nguyen

很难进入「想设计生物武器的人」的心态。我担心的是：我们在降低所需专业门槛、加快工程速度，体积会指数上升。有意层面——确有国家曾有大规模生物战项目，顾问亲眼见过并拆除设施——和平时期没那么吓人，战时特别吓人。无意层面近端可能更可能：为研究生成并控制功能，却难以遏制而泄漏。这未必让我失眠——总体我乐观：收益远大于潜在危害，这是发现与健康的引擎；只是防御在军备竞赛中落后，所以我们也推防御。研究与政策社区有韧性、能动员超前，我对此乐观。

Hard to get into the mindset of someone wanting a bioweapon. I worry we’re lowering expertise needed and speeding engineering—volume will exponentially increase. Intentional: state actors have had very large biowarfare programs; an advisor has physically seen and decommissioned facilities—less scary in peacetime, scary in wartime. Unintentional may be more likely near-term: generate to study, hard to contain, leakage. Does it keep me up? Not necessarily—I’m an optimist. Benefits far outweigh potential harm; engine for discovery and human health. Defense was losing the arms race, so we push it too. The research and policy community is resilient enough to mobilize ahead—I’m optimistic.

RJ Honicky

例行两问。若能一纸命令去掉一个对你重要的瓶颈，会是什么？

Two questions we ask every guest. If you could by fiat remove a bottleneck important to you, what would it be?

Eric Nguyen

能否两个？第一有点敷衍：GPU——每家 AI 实验室都这么说。第二个更哲学：我们想重塑这领域科学家的工作方式，却常撞上「越专业越悲观」——尤其生物，对疾病或模态懂太深，新人一提新东西就「可是这个那个呢」。或许有道理，但我们要的是既懂领域仍保有梦想、想改变做法的专家。若领域更多拥抱这种心态，会有阶跃进步。

Two answers? First is a bit of a cop-out—every AI lab says GPUs. Second more philosophical: reinvent how scientists work here; a hurdle is that more expertise often means more pessimism—especially in bio, you know a disease or modality so well that anything new gets “but what about this and that.” Probably rightfully, but we strive for domain experts who are still dreamers—imagination to change how things are done. If we embrace that more, we’d see step-change progress.

RJ Honicky

最后：希望听众带走的一条信息？

Last—one message you want the audience to take away?

Eric Nguyen

AI 研究社区常觉得必须二选一：要么做聊天机器人式消费/企业前沿，要么做真正改善人类健康、理解生物学的事却够不到 AI 前沿。我想说：不必选择。你可以做真心认为能推动人类的事，同时做最前沿技术——Radical Numerics 要建的就是这个。

People in AI research often felt they had to choose: frontier consumer/enterprise chatbots, or improving human health and understanding biology without being on the AI frontier. You don’t have to choose. Work on what you truly care about that pushes humanity and on cutting-edge technology—that’s what we’re building at Radical Numerics.

Brandon Anderson

也鼓励大家读博客——架构、mechan，投入大量创新；我坚信生物学正站在 AI 最前沿。谢谢远道而来——就几个街区。

Encourage people to read the blog posts—architecture, mech interp, lots of innovation. I firmly believe biology is on the forefront of AI. Thanks for the long trip—blocks away.

Eric Nguyen

随时。感谢邀请，聊得很开心。

Anytime. Appreciate the invite—I had a blast.