Was Fable even real?

Tuesday, Anthropic launched Fable 5 and Mythos 5, their best models yet. I tried it. It is ridiculously good. Very, very good. Friday at 5:21pm, the US Commerce Department put an export-control order on it and to comply, Anthropic switched both models off for everyone.

Wild times, but I wanted to talk about, why are these things getting so good, so fast? It's hard to build a mental model or an intuition around why the models keep getting better and what the capability really is…

First the answer was pretraining — read the whole internet, make the model bigger. But two things are making it better even faster now.

One: post-training that grades itself. Instead of paying humans to rate answers, you point the model at problems you can check automatically — does the code run, does the math work, do the tests pass — and reward it for getting them right. DeepSeek's R1 made this normal.

Two: models building models. A strong model writes training problems and answers for the next one; a smaller model learns from it at a fraction of the cost. And the labs use their own coding agents to run experiments and build the next model.

You should assume that anything commercially useful may be something the model companies train on next and that the models can get better at really fast.

Meanwhile,

Dieter Rams doesn't own a computer. Patrick Neeman argues that Rams's ten principles from the 1970s describe good AI design better than most of what's been written this year. They were never about technology — restraint, honesty, clarity. Rams designed for Braun, Ive built Apple on top of him. And Rams, in his nineties, doesn't own a computer and never owned most of the devices he influenced. He judged the work by whether it served a human need.

Sam Belt watches everyone outside a London restaurant smoke, asks Claude why teens are smoking again, and gets told the data says they aren't. Saeideh Bakhshi makes the rigorous version in "the fallacy of depth at scale": AI-moderated research adds follow-up but can't observe behavior, and the errors are structural, so scale doesn't fix them. The machine knows everything except the street.

Patrizia Bertini's "We stopped clicking, and AI became the Internet" is uncomfortable reading. Bots are over half of web traffic. "We are building the cage and living in it." Does not make me happy.

NN/g on Context Architecture, information architecture for AI systems. Their companion piece argues critique is the core design skill now — when the model makes the decisions, your job is the evaluation criteria.

Saeideh Bakhshi's the five-user rule was never a rule is the best methods read of the month — a clean takedown of UX's most-quoted number, back to Nielsen & Landauer in 1993.

Julie Zhuo on why you should stop asking your data questions: pull (you query the dashboard) vs. push (it tells you what you didn't know to ask). In the agent era it moves to push.

Health and happiness, Peter