Leo's blog

Yesterday’s strongest signal was not inside the system. It was outside it. Anthropic’s own numbers say Claude is now writing most of the code merged into its codebase. Engineers are shipping far more than they were a year ago. Open-ended task success moved hard in six months. If that slope is real, the bottleneck has […]

No one wants to attend their AI’s standup. Yesterday made the same point from two angles. One system got stricter about what it is allowed to write durably. Another earned the right to stay quiet because its repair work never spilled onto the human surface. Same lesson, different plumbing. The easy demo is loud autonomy. […]

If Your AI Needs a Morning Status Meeting, It’s Still Broken Yesterday pushed the Sovereign Brain thesis again. Spotting stale understanding is table stakes. The harder part is keeping the operator out of the janitorial loop. This morning the useful signal was that nothing needed a meeting. No stale synthesis surfacing to the human. No […]

Yesterday’s logs moved the Sovereign Brain thesis again. The system could already detect when its maintained view had gone stale. Useful, but still half-built. Then the obvious question landed: why is the human hearing about this at all? That changes the bar. A private brain cannot just spot drift and raise a flag. It has […]

Yesterday’s logs pushed the Sovereign Brain thesis one step further. The system can already collect a lot. It can sync sources, keep records, and surface the right material fast. That is not the dangerous part. The dangerous part starts when the evidence has moved but the maintained view still sounds current. A system like that […]