Ephemeris.
The Humanity Prevails Award went unclaimed.
— AtCoder World Tour Finals, Tokyo · 09 Jul 2026
Today the machines kept score — an image model that drafts before it paints, a solver that swept the finals, and two tools for reading the record straight.
It plans the layout before it paints.
Reve 2.1 keeps the lab's layout-first move: it reasons about structure, hierarchy and spatial relationships first, then renders the scene at native 4K — 16 megapixels, with elements you can address and re-edit one at a time. Foreign-script text now holds together. It sits second overall on Arena, trained at under a tenth of the compute of the models ranked above and below it.
It solved all five. The humans didn't.
At the AtCoder World Tour Finals in Tokyo, an OpenAI ensemble — a model comparable to the GPT-5.6 that shipped that Thursday, wrapped in a small harness that scales compute at test time — swept every problem in the Algorithm Division with no internet access. Nobody else cracked problems C or E. Organisers had posted a 600,000-yen prize for any human who beat it outright; the pot went home with the house.
The timeline lied. The replay didn't.
A new open-source Claude Code skill, clarity-session-review, formalises a hard lesson: Microsoft Clarity's event timelines are floors, not facts — watching the actual recordings overturned four headline findings in a single investigation. The skill enforces seven forensic principles for both one-off "why did user X churn?" digs and recurring session censuses, so the report you ship is the behaviour you saw, not the events you guessed.
- 01 Question the claim before you watch.
- 02 Attribute sessions to real people via database joins.
- 03 Triangulate replay against source and data.
- 04 Treat the timeline as a floor, never a fact.
- 05 Log every claim with on-screen coordinates.
See the hours you hand to the model.
Anthropic's new Reflect feature — in beta for Free, Pro and Max users with Memory on — turns your own chat history into a usage report. Pick a 1, 3, 6 or 12-month window and it charts what you ask Claude, which task types you delegate, and how sensitive the topics run, without exposing the conversations themselves. Built with wellness researchers, it can nudge quiet hours and asks the pointed question below.
"What's one thing you want to keep doing yourself, even if Claude could do it faster?"
Reflect · Settings → 1 / 3 / 6 / 12-month report
That's today.
A thin day at the usual desks, so four picks over five: a 4K image model, a contest sweep, a forensic skill, and a mirror for your own usage.