The most confident optimization suggestions were not always the useful ones.
An 8 by 8 register block sounded better than 4 by 4 because it offered more reuse. It ran slower. Packed f16 storage sounded additive with tiling. It was not. A vectorized kernel won its standalone benchmark and initially broke the real training loop.
The productive workflow was simple: take the suggestion, define the right baseline, implement the smallest test, measure it, and keep the negative result. The benchmark had authority over the explanation.
The negative results became more reusable than many wins. They recorded which limit an idea actually attacked and why the next attempt should differ. End-to-end parity tests caught errors that standalone kernel tests could not see.
AI was useful for generating hypotheses and tracing failures. It was unreliable as the judge of whether its own proposal worked. That role stayed with frozen inputs, reproducible commands, and measured output.
The complete pair-programming notes are at https://posttrainllm.com/devlog.