A language model wrote the genome of a working virus, and the lab that built it is the one that set the safety rules.
AspireAI community · Quartz
**A language model wrote the genome of a working virus, and the lab that built it is the one that set the safety rules.**
Stanford and the Arc Institute used the genome language models Evo 1 and Evo 2 to generate complete bacteriophage genomes from scratch. The team synthesised 285 of 302 designs; 16 inhibited the growth of E. coli, 12 of 227 sequences from Evo 1 and 4 of 58 from Evo 2. Viability tracked similarity to natural genomes, reaching 46.2 percent among the closest designs. A cocktail of the AI-designed phages cleared bacterial strains that had evolved resistance to the natural template phage, which natural phages could not.
The part that argues this is not recombination: cryo-electron microscopy at 2.9 angstrom showed the designed phage Evo-Phi36 carries a truncated DNA packaging protein J from a distant relative, 25 residues in place of 38, a swap reported as nonviable in the wild type.
The honest limit: the target was PhiX174, a 5,400-letter phage infecting only harmless gut bacteria. Half the working viruses picked up mutations, pointing to natural evolution finishing the job. The team excluded human-infecting viruses from training, but Johns Hopkins argues the next team may not.
https://qz.com/scientists-ai-designed-viruses-phage-genome-biosecurity-101926
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.