Tag: llm (6)
- The Judge Said Yes, Then Explained Why NotGrading a chart-making agent with a small local vision model: what moved the score, and how often the score itself was wrong.
- Same Bot, Two GamesA language model lost to a two-ply chess bot. A four-ply alpha-beta search behind the same harness mated it in 17.
- The Harness Is the ProductWhat building an agent harness from scratch taught me about the part of the system that is not the model.
- Four Results I Had to RetractAn empty database file, a reward made of flakes, a spec that broke my output, and a warning in my code that nobody had tested.
- My Stateful Memory Made the Fuzzer WorseA one-run win, a ten-run loss, and what telling a language model what not to do actually does.
- What Two GPUs at Home Taught MeAn idle GPU, a power supply that was too small, and an engine that was not the problem.