AI-found antibiotics win a Pritzker as agents reach payments and robots

The most interesting AI story today has nothing to do with chatbots. MIT's Regina Barzilay and James Collins just won the Pritzker Prize for AI and Science Research Excellence for using machine learning to discover new antibiotics. I find this far more persuasive than most product launches, because it answers the question every sceptic should ask: what has this technology actually produced that we couldn't have had otherwise? Antibiotic resistance is a slow-motion crisis, and the economics of conventional drug discovery have failed it for decades. If AI can search chemical space faster and more cheaply than human labs, the second-order effect is not just new drugs but a change in which diseases are commercially worth fighting.

Meanwhile, the agent story keeps moving from demos toward plumbing, and plumbing is where the money tends to follow. Microsoft is opening the technology behind its Copilot agents to developers through Foundry, giving enterprises a managed environment to build and run their own. Mastercard, for its part, has struck partnerships with OpenAI and Tencent to widen access to agentic commerce and cross-border payments. Put those together and you can see the outline of the next fight: not who has the smartest model, but who controls the rails an agent uses to act and to pay. By the way, the Tencent link is worth noticing. Agentic payments are being built on both sides of the US-China divide at once, and Mastercard intends to sit between them.

That raises an uncomfortable question. If an agent spends your money, how do you know it did what it claimed? One small but important item today is progress on zkSNARKs, a class of cryptographic proof well suited to demonstrating that a model's inference was run correctly. It is not glamorous, and it is not production-ready for large models yet, but I think verifiable inference will matter more than most people expect once agents handle expense reports and payments. An essay making the rounds argues that OpenAI's Codex, paired with computer use, is the closest thing yet to a real-life JARVIS, automating coding, SSH sessions, video editing and expense reports. I'd read that with some caution, since a single impressive workflow is not a reliable assistant. Still, the direction is clear, and verification is the missing piece.

Robotics had a quietly strong day as well. ETH Zurich built a single controller that lets humanoid robots climb ladders, vault obstacles, crawl and get back up after a fall. Separately, researchers showed a humanoid system that learns walking and object handling from separately collected data and combines them into multi-stage tasks without ever seeing a full demonstration. The second result strikes me as the more important one. Collecting complete task demonstrations is the bottleneck for embodied AI, and composing skills instead of demonstrating them removes much of that cost.

So here is what I'm watching: when agents can move, pay and prove their work, who decides what they are allowed to do?