The startups rethinking what AI can make
1 Atomic Machines comes out of stealth with "vibe manufacturing"
After six years out of view, Jeff Holden, who was Amazon's 10th engineer, unveiled Atomic Machines. Its Matter Compiler is an AI-driven factory system that turns code into tiny physical machines without building new tooling for each design. The first product is the PrimeSwitch relay, and the company has raised $250M. If it works, making hardware could start to feel more like shipping software.
2 TypeSafe AI reaches a $7.5B valuation one month after launch
Co-founded by ChatGPT co-creator Diogo Almeida, TypeSafe raised $870M for Jev, a model designed for AI agents, and says a third of the Fortune 500 already use it. Almeida joked about the exhausting month in a launch-recap thread. Competition is arriving fast: Microsoft released its own fast decision model last week.
Satya Nadella's 7 rules for using AI without losing control
Why it matters. Microsoft's CEO published seven principles for any business putting AI to work, and the post has passed 10M views. They read less like a policy paper and more like a checklist you can apply this week.
Never let AI grade itself. No single model should be your only dependency, and no model should check its own work. Use a second model, a set of rules, or a person to verify important outputs. Validation should always be independent of the model being tested.
Leave a paper trail. Every meaningful action an AI takes should leave a record that cannot be altered and that a human can read. Test the whole system, including how it fails, how it handles attacks, and what it does with unusual inputs.
Decide access yourself. Your organization, not the model, should decide what an AI can see and do. Assume any model could be compromised and set limits from day one, the "emergency brakes" that stop a problem before it spreads.
Share what goes wrong. The last principle asks companies to disclose AI failures quickly and share the lessons across the industry. Most businesses still treat AI mistakes as something to hide. The question is whether anyone follows Microsoft's lead, or waits for a regulator to make them.
The money pouring into AI-driven science
A $1.8B virtual cell. Biohub, the Zuckerberg-backed nonprofit, has raised $1.8B to build a "universal virtual cell" with the US Department of Energy, the NIH, DeepMind, and Isomorphic Labs. The goal is a computer model of a cell detailed enough to test ideas before they reach a lab bench.
Bigger bets on drug discovery. Alphabet's Isomorphic Labs is reportedly raising money at a valuation of up to $40B. Startups such as Chai Discovery, Lila Sciences, and Enveda are chasing the same goal.
Unexpected entrants. Midjourney, best known for AI images, is building a full-body ultrasound scanner. Google veteran Jeff Dean has also left to co-found Discovery Loop, which plans to run large-scale experiments with AI.
👀 Shopping Demo Backfires. After Elon Musk showed Grok Bot ordering Lego Star Wars sets, one user replied that he canceled an order Grok placed and warned people not to share personal details with it.
🍿 Weekend Update. SNL took aim at Kevin O'Leary and his Utah data center in a three-minute sketch.
🤦 A Sign of a Sign. An AI-generated sign outside a spa showed, well, a sign (50K likes).
👓️ Smart Glasses on the Job. A video of a worker using Apple Vision Pro on installation jobs has people wondering who will bring this to the trades first (1.9M views).
🩻 DIY X-Rays. One user says ChatGPT produced an X-ray of their sore hand. It is very likely a joke, and AI images are no substitute for a real scan (5M views).
