BIG STORY #1: Delaware Weighs Letting AI Agents Run Companies
Ars Technica reports that a draft bill in Delaware, where a huge number of U.S. companies are legally registered, would let AI agents manage a new kind of business called an "AIC." 🏢 The plan includes a 30-month regulatory "sandbox," a supervised trial period to test the idea before anything spreads more widely.
Why should you care? It's only a draft, so nothing has changed yet. But it raises a question that could touch anyone who works, shops, or invests: if software is making the business decisions, who is responsible when something goes wrong? Watching how this trial is designed will tell us a lot about how seriously lawmakers take that question.
BIG STORY #2: AI Designs Tiny "Molecular Scissors" With Atomic Precision
Here's some good news. SingularityHub reports that researchers in David Baker's lab at the University of Washington built an AI model, RFdiffusion3, that designs proteins at the level of individual atoms. 🧬 They used AI tools to design enzymes called proteases, which act like molecular scissors that cut proteins at very specific spots. One of them cut amyloid-beta, a protein linked to Alzheimer's.
It's early days: some designs were less active than the natural enzymes they were based on, and everything still has to be tested in cells. Still, designer enzymes could one day lead to new medicines for diseases that have been hard to treat, which is a very different side of AI than chatbots and headlines.
IN PLAIN TERMS: What is a "regulatory sandbox"?
Picture a restaurant that wants to try a bold new dish. 🍳 Before putting it on the menu for everyone, the chef serves it to a few tables while the owner watches closely and can pull it off the menu at any time. A regulatory sandbox works the same way: a limited, supervised test where a new idea can run under relaxed rules, with regulators looking over its shoulder. 👊
QUICK HITS
🔓 Mistral's big open model: Wired reports Mistral says its trillion-parameter, open-weight model matches or beats leading Chinese models at coding and cybersecurity.
🧑💻 Copilot gets a team of agents: InfoWorld reports GitHub's new "Copilot Projects" lets several AI agents work in parallel on bigger goals, plus a multi-agent code review feature.
🛰️ A data center in orbit: TechCrunch reports Google's prototype space data center, carrying four AI chips, is online and talking to the ground.
This week's thread is "who's in charge, and who's checking?" Whether it's an AI running a company or an AI designing medicine, the exciting part comes with a very human need for careful testing. Stay curious, and we'll keep translating the headlines. 🌟