Sep 23, 2026 · 1h 32m
AI models turn genomic design into a biosecurity arms race
🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
The same systems that can design biological sequences may also help detect dangerous designs, making defensive capability central to responsible progress.
- 1Genome language models now read and generate long biological sequences across genomic, RNA, and protein-related tasks.
- 2Omni combines long-context modeling and multimodal signals to improve variant prediction and biological design.
- 3Radical Numerics treats biosecurity as a dual mandate: advance biological discovery while building defenses against misuse.
Don't miss
Eric Gwynn explains why chain-of-thought-like training can transfer to biological design through progressively scored experimental sequences.
The brief
Eric Gwynn, Radical Numerics’ CEO and co-founder, describes a widening gap between biological design capabilities and the tools available to detect dangerous designs.
Genome language models apply ideas from natural-language AI to DNA, learning to read, generate, and reason over long sequences rather than merely predict isolated variants.
The discussion traces Omni’s effort to unify genomic tasks, improve variant-effect prediction in non-coding regions, and use unsupervised sequence data beyond annotated benchmarks.
A striking turn comes when chain-of-thought-like methods meet biological design: models can learn from intermediate fitness measurements for RNA aptamers and related experiments.
Gwynn frames biosecurity as an arms race, arguing that better design systems must be paired with equally capable defenses without blocking legitimate research.