Field notes on making AI actually work
The interesting problem of this decade isn't making models smarter — the labs are handling that. It's the layer between human intent and model capability: context, memory, prompts as contracts, verification, orchestration. That layer is where outcomes are decided, and it is nowhere near solved. This is where we publish what we learn while building in it.
positions taken · claims sourced · counted in tokens, not vibes · wrong in public when wrong
FIELD NOTE 01The model is not the product
The application layer — not the model — decides what people actually get from AI. On what a harness owes the model, what it owes the human, and why the iteration loop is the real interface.
read → FIELD NOTE 02The burden of proof is on splitting
Multi-agent systems look like org charts and cost like them — about 15× in tokens, plus coupling. Why one excellent prompt is the right default, and the one split that is almost always worth paying for.
read → FIELD NOTE 03The training shows through
How a language model is actually made — pretraining, tuning, and the side effects installed along the way — and why every prompt practice that works is a patch for a specific scar in that process. With the full ledger, traced.
read → FIELD NOTE 04The rules file that made things worse
Two 2026 studies measured the same instruction file and disagreed — one found agents 28.6% faster with it, the other that generated ones cost 23% more and slightly reduced success. The experiment that reconciles them, and the line-by-line test it gives you.
read → FIELD NOTE 05Your check is already an eval
A 13B model with custom rubrics matches GPT-4 at judging; agreement with humans swings from 0.17 to 0.86 depending only on how specifically you asked. The judge is a commodity — the rubric is the asset, and you probably wrote one already.
read →