Researchers just crossed a line that, frankly, we all knew was coming. A team taught a machine learning model to draft viral genomes from scratch, no template needed, and some of those sequences actually proved viable when tested in the lab. The virus wasn't copied from nature. It was generated from the ground up by a model that learned the rules of what makes a virus work.

Stop and think about what that actually means for a second. We've watched generative models produce text, images, code, protein structures. I've covered AlphaFold and the whole protein folding revolution pretty closely here. But going from predicting how a protein folds to writing a functional viral genome is a different category of problem. It's the difference between reading someone else's codebase and shipping your own app. You have to understand not just individual components but how the entire system hangs together.

The model learned the underlying grammar of viral genetics. Which nucleotide sequences produce functional genes. How those genes need to be organized. What regulatory elements matter. Then it started writing its own drafts from a blank slate, and the output wasn't just plausible-looking genetic code. Some of it actually worked in living cells.

That's wild.

Now, the immediate reflex for most people reading this is probably concern, and that's a fair instinct. I spend my days inside machine learning systems, and even I feel the weight of this one. The dual-use conversation around AI in biology has been brewing for years. If a model can draft a viable virus, what stops someone from drafting something dangerous?

But context matters here. The researchers were working under strict lab safety protocols, and the viruses they generated weren't pathogens designed to cause harm. They were proofs of concept. The goal was to understand whether AI could bridge the gap between theoretical genome design and biological reality. Turns out it can, at least at this scale.

And the upside is real. Being able to design viruses from scratch opens doors in gene therapy, vaccine development, and targeted cancer treatments. Engineered viruses already deliver corrective genes into patients with inherited disorders. If AI can optimize those delivery vehicles faster and more precisely than manual design, we're looking at a genuine acceleration in how quickly new therapies reach clinical trials. Imagine iterating on a vaccine candidate the way my team iterates on a build pipeline. Days instead of months.

The tension is that the same capability that lets you design a therapeutic vector also lets you design something nastier. This is the classic security problem in software engineering, except the stakes are biological. Open source code democratized development but also made it easier to find vulnerabilities. Synthetic biology faces a similar tradeoff, except the vulnerabilities are living things.

What strikes me most is how this connects to a thread I keep pulling on in this column. We keep watching AI cross thresholds that felt theoretical just months earlier. AlphaEvolve solving decades-old math problems. AlphaFold cracking protein structures. Now generative viral genomics. Each one on its own is a breakthrough. Stacked together, they suggest we're entering a phase where AI stops being a tool that helps us understand biology and starts being a tool that writes it.

I don't think panic is the right response. But neither is indifference. The research community needs robust frameworks for governing generative models in biology the same way we're starting to build them for large language models. And those frameworks need to come before the capability outpaces our ability to contain it.

As someone who works in code every day, I can tell you that once a system can generate functional output at this level, it rarely gets worse at it. It only improves. The question isn't whether AI-designed biology is coming. It's whether we'll be ready for it...