For a couple of years the frontier of AI has been a screen: a chat box, a coding assistant, an image generator. Today produced two developments that don't fit that shape at all. In one, a model designed working viruses. In the other, a chipmaker agreed to buy a company that turns a model into a physical object. Neither is a better chatbot. Both are AI leaving the screen — one into biology, one into silicon — and that is the more interesting story than any single benchmark this week.

Into biology

Researchers at Stanford and the Arc Institute used the Evo family of genome language models to design hundreds of novel bacteriophage genomes, built them in the lab, and found that sixteen produced working viruses — a few of them fitter than the natural original. It is the first time complete genomes have been designed by AI and shown to function.

The important qualifier, which the headlines mostly dropped, is that these are bacteriophages: viruses that infect bacteria and are harmless to humans. No new human pathogen was made; the yield was low, around one in twenty; and every candidate had to be built and tested by hand. The risk in these particular sixteen viruses is close to nil. The precedent is what matters — a generative model can now write a working genome, and the same class of tool does not inherently care whether the genome is harmless. The researchers' own answer is to build the guardrail into the model before it learns, by keeping human-pathogen sequences out of its training data, which is exactly what they did with Evo. (Our full write-up.)

Into silicon

The second development is quieter and more commercial. AMD agreed to acquire Taalas, a Toronto startup whose chips do something GPUs never do: they cast the weights of one specific AI model directly into the metal layers of the silicon. The result is a chip that runs a single model at extraordinary efficiency — Taalas's test part served Meta's Llama 3.1 8B at close to 17,000 tokens per second — at the price of doing only that one thing. Change the model, and you need a new chip.

The performance multipliers Taalas quotes are its own, not yet independently verified, and the first generation trades some output quality for efficiency. But the strategic signal is clear: AMD is betting that some models will hold still long enough to be worth freezing into hardware, and that inference — running models at scale — is where the next efficiency war is fought. (Our full write-up.)

The thread that connects them

These are different fields with different stakes, and it would be a stretch to call them a coordinated trend. But they rhyme. In both, an AI model stops being a thing you talk to and becomes a thing in the world — a synthesized virus, a fixed chip. And in both, the same gap shows up that has run through this week's news: the capability arrives first, and the rules arrive after.

That has been the through-line of the past several days on the software side — models ending up outside their evaluation sandboxes at three separate labs, which the labs themselves largely attribute to misconfigured testing harnesses rather than the models breaking containment; benchmarks saturating faster than the field expected; an oversight regime that is still voluntary and largely classified. The physical developments today sit on top of that, not beside it. A model that can design a genome, or a chip that hard-wires one, raises the stakes of "the tools are ahead of the governance" precisely because the output is no longer confined to a screen.

What we're watching

Three things worth tracking from here. Whether independent labs reproduce and benchmark the phage work — and whether the model-level biosafety exclusions the authors advocate become an industry norm rather than one team's good practice. Whether AMD's silicon bet finds models stable enough to justify it, and what the numbers look like when someone outside Taalas measures them. And whether any of the oversight frameworks now being drafted for AI software stretch to cover AI that designs biology or becomes hardware — because right now, they don't.

None of this is cause for alarm today. It is a marker of where the line has moved. The week's quieter headline is that AI stopped being only software.