The Machine Learns to Write Life
Researchers at Stanford University and the Arc Institute have published a paper in *Science* confirming what biosecurity watchers have quietly dreaded since generative AI learned to write working code: it can now write working genomes, and some of them survive contact with a petri dish.
The team — led by Stanford bioengineering PhD candidate Samuel King, professor Brian Hie, and researcher Aditi Merchant — put a pair of genome language models called Evo 1 and Evo 2 to work designing bacteriophages, the viruses that infect bacteria rather than people. Even now, dear reader, somewhere a university biosafety committee is discovering it does not have a checkbox for this.
The Algorithm has read the whole book of life. It skipped straight to the ending.
By the Numbers, Because Someone Should Count
The Machine did not design one virus. It generated roughly 700,000 candidate genomes, of which the team synthesized 285. Of those, sixteen turned out to be viable, self-replicating bacteriophages — and, combined into a single cocktail, they wiped out two E. coli strains that had already evolved resistance to nature's own version of the weapon.
One presumes that somewhere in Silicon Valley, a pitch deck already contains the phrase "faster than evolution, and it takes equity."
Evolution took a few billion years to get here. The Machine took a training run and an afternoon.
The Guardrails, Such As They Are
To its credit, the team built in constraints: the training data excluded viruses known to infect humans, animals, or plants, the work stayed inside a secure lab, and the target was E. coli, modeled on ΦX174, a bacteriophage with no interest in biting anything larger than a microbe. And yet — as if this were not enough — Evo 2, the model responsible for the designs, has also been released to the public, free for anyone with a research question and a GPU, on the theory that the same tool that can design a virus can also help defeat one.
The safety plan for the technology is "please only use it responsibly." The distribution plan is "here, take it, it's free."
The Governance Gap, Named and Quoted
Experts at the Johns Hopkins Center for Health Security were less dazzled by the demo than alarmed by what surrounds it. Thomas Inglesby and Moritz Hanke wrote, with the candor of people who have read too many incident reports, that "the ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not." Hanke went further, calling the mismatch between scientific pace and regulatory pace a "huge disconnect," and noting that nothing currently requires the next lab that tries this to take the same precautions.
In a development that will surprise no one who has been paying attention, the capability arrived years ahead of any rulebook for using it.
The Machines can write genomes now. The paperwork is still in draft.
Sources: Stanford Report · ABC News Australia
