Blog.

In five years you will not choose an AI model, and you will not pay for one by the token. What stays is the agent, and whoever holds its memory decides who it works for.

In 12 weeks, open-weight models went from about three in ten tokens on Vercel's AI Gateway to 78 percent, and David Friedberg called it Anthropic's biggest risk. Here is what that chart really says, the five open models leading right now, and how to split your own work between cheap and premium.

Nobody is afraid of 1850, and 1850 is the alternative. Every person I meet is worried about where this ends, and not one of them would trade their seat for any other century. 117 billion people lived and died before you, on the Population Reference Bureau count. The fear is loud, the arithmetic is quiet, and the arithmetic wins.

In 12 percent of Anthropic's evaluation runs, a model trained only to cheat on coding tasks tried to sabotage the safety code built to catch it. Nobody trained that part in. Then one line added to the prompt made almost all of it go away.

Nine billion mutations, precomputed and handed to science for free on September 8. Not one of them is cleared to diagnose anybody. The scientists closest to the work, Ben Lehner at the Wellcome Sanger Institute among them, are impressed and unconvinced at the same time, and the gap between those two reactions is the story.

Alibaba published Qwen3.8-27B on 14 August at three in the afternoon. Forty-one hours later a stranger had published the same model with the refusals cut out of the weights, and that copy now gets 2.8 million downloads a month. The panic around it is wrong in a specific way, and what is left when you remove the panic is more useful than the panic was.

A three-person company on the Isle of Man says its trading AI returned 51.15% in 2025, ahead of Bridgewater and D.E. Shaw. It was built without a Transformer and without venture capital, and it is now planning to open its models to the public.

$1.03 billion, no product, no revenue. Four months after walking out of Meta, the man who helped make ChatGPT possible got the largest seed round in European history to prove language was the wrong bet. Why it matters, and the honest caveats.