🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
Radical Numerics co-founder and CEO Eric Nguyen, who holds a PhD in Bioengineering and AI from Stanford University, discussed biosecurity as an AI arms race, according to a Latent Space episode summary. Nguyen previously helped lead development of the genomic language models Evo and contributed to Evo 2 at Arc Institute.
Those models, the summary states, were later used by a separate Arc/Stanford team to generate entire bacteriophage genomes that were synthesized into functional viruses.
The summary frames biosecurity as a parallel concern to cyber-security, citing Anthropic filters that flag both cyber-security and biology. Clem Delangue argued that cyber-warfare defensive capabilities need to be open and keep pace with frontier models' attack capabilities. A quoted post from Adrien Carreira said open models were used during an incident, arguing AI security will not be solved by one company in secret.
The summary describes differences between DNA and natural language: DNA has a four-character alphabet (ACTG) and very long sequences, with an average human gene around 60K, sequences up to 2.3M, and the whole human genome around 3B. It credits long-context model innovation about three years ago with making this work possible.
According to the summary, Radical Numerics' GLMs already perform well with RNA and protein because DNA contains markers for genes and for genes encoding proteins, meaning the models generalize across multiple languages before training on other modalities such as 3D protein structure, epigenetics and natural language.
In one experiment described, the team used a large dataset of aptamers, held out a portion of the best-performing ones, showed the model only lower-scoring RNAs ranked progressively higher, and asked it to continue the pattern. The model recapitulated some higher scores it had not been shown, which the summary calls chain-of-thought thinking in DNA.
Nguyen said defense is currently losing this battle, while Radical Numerics argues to push the frontier harder. The episode also covers the genome as an imprint of the environment on DNA, going truly multi-modal, and how chain-of-thought works when thinking in the language of DNA.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Radical Numerics is using biological chain-of-thought and multimodal perception to keep up with the bio-defense arms race, design new genomes and gain insights into biology itself.