Blog.

Everyone names benevolence as the only real safeguard for AI, then admits nobody is building it. So let's build it. The moment you try, it stops being philosophy and becomes architecture, because you cannot legislate character into a mind. The question was never how to cage it, but how to raise it.

The most credible pessimist in finance calls AI the biggest investment bubble in history, and relaying the godfather of AI, he lands on a single word as the only way through: benevolence. Everyone names the answer. Almost no one is building it. That's the conversation I want to have.

For the entire history of the internet, there was a human at the end of every transaction. Someone clicked buy. That assumption is quietly breaking: machines have started paying each other, in stablecoins, with no checkout page. The question every business now has to answer is whether an agent can find you, and pay you.

The headlines said Japan built a frontier model to rival Anthropic's best. It didn't. Sakana built a conductor, a small model that reads your task, routes it to the strongest existing models, and verifies the result. There are two ways to win with AI, and Fugu just proved the cheaper one works.

SpaceX just signed a $60 billion all-stock deal to buy Cursor, the AI coding tool millions of developers open every morning. A rocket company bought a code editor. The lesson was never the price. It's that leverage belongs to whoever owns the loop, not whoever rents a seat inside it.

A fifteen-year-old defended his PhD in quantum physics, then flew out the next day to start a second doctorate. The internet made him a curiosity. The real lesson isn't genius, it's leverage: he removed friction, worked across fields, and refused to wait. Unlike genius, leverage is something you can build.

Companies are racing to give AI agents the keys, to inboxes, codebases, and customer data, and almost none have asked the only question that matters once an agent can act: what happens when someone turns it against you? Strip away the mystique and an agent isn't an agent. It's an organized file structure.

A medicine used to take a decade to find. Isomorphic Labs just raised $2.1 billion to collapse that timeline into days. What they built isn't a drug, it's an engine. And the gap between owning an engine and doing the work by hand is what separates the companies that compound from the ones that grind.