A notebook, not a brochure
Neural networks in 2001, to see how they learn. AI systems now, to see what they did and why - and what it takes for them to hold up under inspection. What I find ends up here, in talks, and in the reading list at the bottom.
Taking apart
My master's thesis: self-organizing networks with every step of the learning drawn on screen, so you could see the map form instead of trusting the numbers. Rebuilt for the browser with verification notes. The 2001 look was kept on purpose.
Twenty-five years later the question is the same, only the networks are bigger and nobody draws the learning any more.
876 rows, each traced to the article it comes from. What is left when you do that is not a checklist. It is three questions, in the order the Act itself asks them: know and govern · build and prove · run and respond. Everything else is a row under one of the three.
What surprised me most: how much of the work is deciding, per system, which rows apply at all - and how little of it is the rows themselves.
Triage sorts by how bad it would be if it failed - no likelihood, no controls, minutes. Assessment asks how likely, what to do about it, and what is left afterwards - days. Assurance asks whether what you did actually works - tested, not estimated.
Most arguments about "AI risk" are three people using one word for three different things. Name the level first; the argument usually ends.
“Nobody has AI under control. The work is not to chase reliability but to manage unreliability. The way we already do with people: checks, second readings, recertification.”
Work
What I read, and why
Five sources, on purpose. The output of reading them is a note here or a change in a talk - not a folder.