Prompts Aren't Real
4 hours ago
- Propose building interlocking evaluation and optimization pipelines instead of just exchanging text with LLMs.
- Making reliable agentic experiences for consumers is challenging; LLMs fail subtly and sometimes simply.
- Structured output helps but models still fail a small fraction of requests; fixes can be arbitrary (e.g., renaming field).
- Prompts are not the core; they are vectors whose textual content is unimportant; measurement is key.
- Use pass^k tests to measure behavior, then optimize prompts automatically via algorithms like genetic Pareto.
- Build LLM judges for complex evaluations (e.g., brand voice) using golden datasets and optimization.
- Create a self-improving feedback loop: monitor production, convert failures into test cases, re-optimize.
- Domain experts should focus on building datasets and measurements, not crafting prompts.
- Without measurement, the path leads to madness; self-correcting systems are essential for enduring reliability.