Genome Language Models Learn to "Think in DNA" as Bio-Security Arms Race Heats Up
Latent Space · rss · 2026-09-23
Latent Space interviews Radical Numerics co-founder Eric Nguyen (Stanford PhD, led Evo, contributed to Evo 2) on why bio-security is an AI arms race.
From doubted to functional viruses
- Genomic language models (GLMs) faced years of biologist skepticism;
- Evo/Evo 2 were later used by an Arc/Stanford team to generate complete bacteriophage genomes synthesized into functional viruses.
Long context unlocks biological intelligence
- DNA has a tiny alphabet (ACTG) but extremely long sequences: 60K avg gene, up to 2.3M, 3B genome;
- Long-context innovations (StripedHyena) enabled this 3 years ago, before frontier labs shipped 1M+ context models.
Thinking in DNA
- GLMs already generalize to RNA and proteins via sequence markers;
- In an aptamer experiment, showing the model progressively higher-scoring RNAs let it extrapolate and recapitulate held-out top scores — chain-of-thought in DNA.
The arms race
- The same models that boost capability raise risk; Nguyen argues defense is currently losing and advocates pushing the frontier;
- Context: the OpenAI→Hugging Face attack and Clem Delangue's open-source defense argument;
- Team includes ex-Liquid AI members Michael Poli, Stefano Massaroli, and CTO Armin W. Thomas.
More from Safety
- Sen. Sanders Unveils Bill to Ban Artificial Superintelligence and Create US 'Department of AI' — KyeGomezB · 2026-09-23
- Researcher weaponizes libheif 1-day into full RCE via UAF and tcache poisoning — dyn___ · 2026-09-23
- The prompt to run before wiring an agent to another service — gethackteam · 2026-09-23
- Gemini's sandbox escape, California's kill switch order, and Apple's server return: one theme, control — Dapper-Tale-4021 · 2026-09-23
- TypeSafe classifies RAG passages with probability thresholds to fight noise and prompt injection — marlene_zw · 2026-09-23
- Cybersecurity Experts Push Back on AI Doomers: Doomer Scenarios Rely on AI Hacking — ccerrato147 · 2026-09-23