Genomic language models have progressed from predicting biological sequences to generating complete bacteriophage genomes that researchers synthesized into functional viruses. Radical Numerics cofounder Eric Nguyen argues that this capability creates a biosecurity race: models can expand biological design, while defenders need equally capable tools to evaluate what those designs might do.
Unlike ordinary language, DNA uses only four characters but stretches across enormous sequences. A typical human gene can span around 60,000 bases, unusually long genes reach millions, and the full genome contains roughly three billion. Recent long-context architectures made it practical to model enough sequence at once to capture relationships that shorter systems miss.
Nguyen previously helped develop the Evo genomic models. He described an experiment in which a model saw RNA aptamers ordered from low to progressively higher scores, then generated candidates resembling withheld high performers. The result suggests sequence models can learn optimization trajectories, although that does not establish broad biological understanding or safe real-world performance.
Radical Numerics plans to combine DNA with protein structure, epigenetic information and natural language. Its claim that defense should push the same frontier is a strategic position, not a demonstrated safety solution. Better models may help identify threats, but they also increase the importance of access controls, laboratory validation and independent risk assessment.