Signal
Foundation models trained on DNA now read and draft genetic code the way language models handle text: one open model designs bacterial-scale genomes and grades mutations, another reads a million letters at a time and predicts how an edit changes gene activity, and a crop startup's model proposes sequence changes that tune a gene for the yield a breeder wants. Drafting is collapsing to minutes at a keyboard. The season it takes to learn whether the plant agrees has not moved.
So what?
Drafting a genome is now a text task. Proving it still takes a season. Biology's bottleneck moves from the idea to the grow-out.
Evidence
Phys.org, 4 March 2026 (study): Evo 2, published in Nature by Arc Institute, NVIDIA and university collaborators, was trained on 9.3 trillion nucleotides from over 128,000 genomes, grades BRCA1 mutations with over 90% accuracy, designs genomes as long as a simple bacterium's, and is fully open source.
Nature, 28 January 2026 (study): Google DeepMind's AlphaGenome takes up to a million base pairs of DNA, predicts thousands of regulatory properties and scores a mutation by comparing edited and unedited predictions. An API preview opened in June 2025 and the paper and code followed in January 2026.
AgFunderNews, 20 July 2026 (startup): UK startup Phytoform is working with Beck's Hybrids and RAGT on corn architecture, designing DNA sequences that change a target gene's activity without foreign DNA, after raising $4 million earlier in 2026 for tomato work.
Worth asking
If a first draft of anything in your work took minutes, what would you spend the saved time proving?
