Working with AI: Delegation & Oversight
What to hand to a machine, what to keep, and how to tell whether the oversight you have is real.
6 chapters · all free during the beta · ~92 min total
Syllabus in review — not final. These chapters are outlined, not written. Their checkpoint scenarios are live now, and they are the part you can actually train on.
- 1What to Hand Over
Whether a task should go to an AI system is a sum with three terms, and the one people leave out, how long the output takes to check, usually decides it. Which part of a task can be handed over, and why bad output is more often a bad brief than a weak model.
Free13 min1 checkpoint - 2Oversight That Catches Things
Why "a human reviews it" is a hope rather than a design, why review slides into rubber-stamping when agreeing is one click, and the four things a review has to specify before it can catch anything.
Free16 min2 checkpoints - 3Judging the Output
Why a model's wrong answers read exactly like its right ones, why a clean random sample is weaker evidence than it looks, and how an eval turns "it's been working well" into a number that can go down.
Free15 min1 checkpoint - 4How Long It Can Run Alone
Why an agent that is reliable on short tasks can fail most long ones, why "handles 80%" means nothing without a task length and a success rate, and how to cut a long run into steps that get checked before the next one builds on them.
Free in beta16 min1 checkpoint - 5Access Is Authority
Why giving an agent access to code, mail, money or customers is a grant of authority, how to bound what it can do before you let it act alone, and why anything it reads can try to give it orders.
Free in beta16 min0 checkpoints - 6What Lasts
The long-run costs a delegation business case leaves out: the people left with only the hard part, the skill to judge the work fading once nobody does it, and engineering built around today's model that expires with it.
Free in beta16 min0 checkpoints