3 posts
The tried and true methods of root cause analysis assume a failure has one cause and will happen again. AI failures break both assumptions. What people smarter than me are saying about it, a test from my own series that needs revising, and why the useful question after an incident is which change to make, not which cause to name.
I went looking for how to do evals better and found an argument about whether they work at all. What I've read, what I've heard, and why the most important discipline in AI product management has quietly turned into a performance.
Everyone agrees AI is transforming product management. Nobody agrees on which skills survive. The debate over strategy, taste, and the 'editing function' reveals a profession in the middle of an identity crisis.