The show
One idea, turned over properly. New most weeks, and shorter when the question deserves twenty minutes rather than an hour.
We open our own machinery. The agents that draft, the editors that cut, the fact-checking pass that catches what the drafts invent, and the places where a human still has to sit down and decide.
The least glamorous idea in applied AI and the one that separates systems that ship from systems that demo. What an evaluation set is, why twenty examples beats zero, and how to build one this afternoon.
There is a version of using AI that makes you sharper and a version that quietly hollows you out. They look identical from the outside and produce nearly identical documents. The difference is what you can do next week.
Splitting a hard problem across specialised agents is genuinely powerful and usually premature. Daniel Okafor on coordination cost, the failure modes nobody warns you about, and the far simpler thing to try first.
A short one. Why fluent, confident, completely invented answers are a feature of how these systems work rather than a bug to be patched — and the two habits that catch almost all of them.
You pasted the contract into a chat window. Now what? We trace the real path of a prompt through consumer tools, enterprise tiers, and API endpoints, and explain what each tier actually promises in writing.
The demo works. Six months later the project is quietly shelved. Priya Raman has shipped and buried enough AI systems to know exactly where the floor gives way, and she walks us through each drop.
No frameworks, no maturity models. A blow-by-blow account of one working week: what we handed to a model, what we refused to, what came back wrong, and the running tally of hours it actually saved.