Your model's output is a sample, not a return value
Twenty identical calls, ten different answers. What changes when a language model enters your code path, and why chasing determinism is the wrong first week.
writing
Everything here comes from real work: implementing AI for businesses, teaching people to build, shipping my own products, and short pieces adapted from my book, AI Engineering, Measured. One rule: it has to be useful.
Twenty identical calls, ten different answers. What changes when a language model enters your code path, and why chasing determinism is the wrong first week.
The same twelve-turn conversation replayed four ways. Cutting history to save tokens made the agent do the one thing it had been told not to do.
Moving retrieval into Postgres changed no ranking and cost about a millisecond and a half. The vector index, on the other hand, was never used.
No hype, no jargon — five specific, affordable automations that pay for themselves, ranked by effort versus payoff.
Connect with me on LinkedIn, or tell me what you'd like me to write about next.