How AI systems differ

A model is slow, metered by the token, probabilistic, stateless, and steerable by whatever text it reads. Each of those properties breaks an assumption that ordinary service design relies on, and each one needs an explicit answer in the design.

The first lesson names the five properties and what each changes: streaming and long timeouts, token budgets, evaluation instead of unit tests, a context builder that carries all state, and a permission boundary outside the model. The second turns them into numbers: the ten inputs that size any AI workload, the latency budget of a single call, and the levers that trade cost against quality. Afterwards you should be able to estimate the tokens per second, GPUs or spend, and per-request cost of a design from a handful of requirements, and say which of the three moves first when one changes.