Why the CNN skills hit a wall, how sharing a single VLM backbone across tiny expert heads changes the economics, and the learned router experiment we are about to run.

Wiring Claude Sonnet to hypertuned CNNs, building a self-improving agent, and what happens when your system starts training its own replacements.

Five SFT runs that scored zero, a failed RL attempt, switching models, and learning to decompose the problem.

Decompiling the benchmark, designing a three thread architecture, and learning how robots think about vision.

Procedural data generation, testing existing models, and discovering the connection to robotics.

What I learned about latency, failure modes, and why pipelining intelligence does not work.

A 30 step web challenge and the start of something bigger.