LLM: Describe the job; let a model do it at runtime

A prompt and a few tools get a working version the same day. Iterate by editing text.

✦Better models, agent SDKs and tool calling make the first prototype close to production quality.

Unstructured text, PDFs, images, languages, typos. It covers cases no one would ever write a rule for.

✦Each model generation takes on messier inputs with less prompting.

Swap in a newer model and quality goes up without touching your code.

✦Model upgrades arrive every few months. That's R&D you get without paying for it.

New policy? Edit the prompt, re-run evals, ship. Domain experts can do it too.

✦This was always true. What's changed is that evals make it safe.

Users and operators just ask. No forms and no training on a new UI.

✦Users now expect to be able to just ask.

The same input can produce different outputs. You need evals, guardrails and human review for high-stakes paths.

✦Structured outputs, better models and eval tooling reduce variance, but it never reaches zero.

Every request pays tokens and seconds. At millions of calls, the unit economics can dominate.

✦The price for a given capability has been falling several-fold per year, and small models cover more tasks.

Confident wrong answers. Dangerous where money, health or law is involved.

✦Grounding, tool use and verification loops help. High-stakes paths still need deterministic checks.

Untrusted input can steer the model. Tools with side effects need strict permissioning.

✦Still an open problem. Treat model output as untrusted input.

It's hard to explain to an auditor why the model decided X. There's no line of code to point at.

✦Traces help, but regulators still want deterministic, explainable decision paths.