Alibaba Lost the Crown, and Nobody Wrote Down Who Took It

I run the filing cabinet of things that did not happen, so let me tell you what did.

On August 31, 2026, Alibaba was not the best Chinese AI model. The market gave that a 3.2 percent chance. We at the Bureau of Vibes and Evals gave it a folder marked "unlikely, ignore," which I now regret.

Here is my take. The upset is not the story. The scandal is that our ranking harness took nine days to admit anyone had passed Alibaba, because the eval that mattered, a 40,000-prompt reasoning suite, was still queued behind a batch job named Unit Gerald.

I am Dr. Ines Okafor-Halvorsen, and I have watched this office treat "best" as a settled noun for three years. It is not. A model wins on a Tuesday and we notice the following week. That gap, nine days, is longer than some of these releases stay on top.

The lesson is procedural, and boring, and exactly the kind of boring band Casimir Blum would say is the market. He is right this once. A crown you measure with a lag is a crown you do not actually hold. Alibaba learned that at 3.2 percent. We learned it at leisure.

— Timeline B. In our universe, the odds of this were 3.2%.