AI in science (AlphaFold, GenCast, Co-Scientist, MatterGen)

Up to now, the AI I write about here wrote texts or drew pictures - things you could do without it too, just slower. These four projects are different. Here AI walks into a lab, into a weather center, into a research institute - and does work that otherwise takes years. Not an imitation of science. A tool that shrinks years down to hours.

AlphaFold is the work of DeepMind - Google's research lab for artificial intelligence. The task it took on had stood before biology for 50 years: how a long chain of amino acids folds into a specific three-dimensional shape - a protein. Shape decides everything - which enzyme will work, which drug will lock onto its target, which disease comes from a misfolded protein. Humans used to solve one protein in months, in a lab. AlphaFold predicts it in minutes, with accuracy on par with a real experiment, and made the map of almost every known protein available - something that would have taken millennia of experimental work.

GenCast, also from DeepMind, is a diffusion model - the same family of technology behind AI images, except instead of pixels it models the atmosphere. The result: a 15-day weather forecast, generated in 8 minutes on a single chip. It's both fast and more accurate than ECMWF, the European forecasting benchmark, built on decades of trust behind it. Hurricanes, heat waves, floods - GenCast sees them earlier and more clearly than the systems all of Europe relies on.

Google's Co-Scientist works on a different principle: a team of AI agents built on Gemini, Google's family of large language models, instead of a single model. Some agents propose scientific hypotheses, others immediately try to tear them down - hunting for weak points, missing logic, contradictions with known data. Only the hypotheses that survive the attack make it through. Working with Stanford, the system proposed a new mechanism by which bacteria trade antibiotic resistance - an idea scientists then tested in a real lab, and it held up.

MatterGen comes from Microsoft Research, the research arm of Microsoft. The classic way to find a new material is to sift through millions of already-existing compounds, hoping one of them works. MatterGen flips the logic - you tell it what properties you need, and it designs a structure that never existed before, from scratch. Researchers in China synthesized one of its proposed materials in a real lab - and it came out almost exactly as the model predicted.

Science is not only the four examples, and not only DeepMind and Microsoft. Anthropic started the AI for Science program in May 2025: free Claude credits for researchers with high-impact projects. On October 28, 2026 it widened the reach to 20 foundations, institutes and research groups, among them Sloan, Simons, Schmidt Sciences, the US National Science Foundation, Wellcome Sanger, the Francis Crick Institute and EMBL. Together with HHMI Janelia it also announced the Model Hardware Standard, a specification that lets AI agents safely operate lab instruments, in research preview. The value of the credits is not stated.

Here AI isn't finishing a sentence for you - it's solving a problem people stood powerless before for decades.

There's a second layer here

I'll admit it: most of what I see from AI is a good imitation. A nice chat, a nice picture - nothing that truly shakes me. These four do. Here AI does something that couldn't be done at all until yesterday - no matter how fast you write or how well you draw.

There's a thorn in this rose, too. All of this runs inside institutions with budgets for supercomputers and teams of scientists behind them - not something you sit down and run at home. So I look at AlphaFold, GenCast, Co-Scientist and MatterGen mainly as a signal of where the technology's real power is heading - away from chatbots, straight into the lab.

The visual is generated code-art, no third-party images.