Arc Institute — Evo 2 (Genome-Scale Foundation Model)
organizationAI & Compute · Biotech & Transhumanism
An AI that reads and writes DNA a whole genome at a time.
Who they are
The Arc Institute, working with NVIDIA and researchers from Stanford, Berkeley and UCSF.
What they do
Evo 2 is a giant AI trained on 9.3 trillion letters of DNA from 128,000 species. It can flag which gene changes are likely dangerous, and it can write brand-new DNA as long as a simple bacterium's whole genome.
How it works
By learning the statistical patterns of life's code the way a language model learns text — then predicting or generating new sequences.
Why it matters
It's the cheap engine underneath two things at once: ranking people by their genes, and writing genomes from scratch. Powerful, and dual-use by nature.
The engine's record — word for word
**In plain terms:** the genome-scale AI that reads and writes DNA at whole-genome scale — the compute core sitting beneath the Selection Substrate and the genome-writing arm. Evo 2, built by the Arc Institute with NVIDIA (plus Stanford, UC Berkeley and UCSF), was trained on over 9.3 trillion nucleotides from more than 128,000 genomes across all three domains of life[1]. On the breast-cancer gene BRCA1 it reached over 90% accuracy separating benign from potentially pathogenic mutations[2], and it can generate DNA as long as the genomes of simple bacteria, with a context window up to ~1 million nucleotides. Engine placement: the predictor that makes population-scale genetic risk-scoring computationally cheap (feeds genome_as_selection_substrate / heliospect_genomics) AND the writer that operationalizes synthetic_human_genome_project and nucleic_acid_compiler. Held — a foundational dual-use tool, not proof of a selection programme; name no holder. [AI & Bio-Selection Audit — verified at source Sep 2026.]
Follow the trail
Walk this on the live map →