Three weeks ago a logistic regression beat eight LLMs at reading job postings. I gave it a rematch against a model built for tables. TabPFN-3.5 beat the lion, beat GPT-6 Astra and Claude Fable on the same postings, ran in 12 seconds per thousand, cost nothing, and fits in 876 MB.
For decades you could read a manager's rank off the number of assistants in the hallway. The executive floor managed exceptions; everyone else managed errands. This month the assistant arrived for everyone. The judgment about what to hand it did not.
I tested eleven technologies and eight LLMs on one triage problem over a weekend. A logistic regression won. The bigger finding: a test that used to take a quarter now takes a weekend, and that changes what an architect is for, especially where the worst case is a missed diagnosis.
A relative lost his job to an agent, so I built him one. Choosing which model should interview him cost a day, $22, and one reversal I did not enjoy. The model was never the problem. Six other things were, and finding them is the eval nobody runs.
A coding agent asked a product question at 10:40 on a Tuesday. The PM was in a review, the designer was on another project, the developer was at lunch. The agent took the default, the tests passed, and the product decided for itself. Who answers the agent? Everyone, by topic, with a default.
Watching is not a job. It is the transitional state between two jobs, and it is the most dangerous place in the building, because it looks like supervision and it is not. I hired a team on Saturday. By Sunday it did not need me, and I was there anyway, pressing the green button.
Two frontier AI models, an external reviewer, thirty-nine signed rulings, and not one of them ever said stop. I did not fire my agent team for being wrong. Every answer was right. I fired them because no answer ever included the sentence that mattered.
Ten runs of one boring question about fruit teach more about AI than any slide deck. Probabilistic is not a definition; it is something you watch happen to your own keyboard.
Physicians are building their own AI tools, and patient data cannot leave the building. So which local model deserves to be in the room with a chart? Not the biggest one. I measured.
A PM who has never watched an agent work cannot write its spec, its evals, its boundaries, or its intent. The pod model assumes PM, UX, and engineering share an understanding of the product's behavior. For agentic products, only one of the three has it.
Healthcare asks who is accountable for a machine's mistake before it asks what the machine can do, because it has no choice. On 18 August a federal regulator applied that question to generative AI. The framework it produced is not new. It is clinical governance, written down for everyone else.
A new paper rebuilds the DIKW pyramid for medicine and shows where AI actually breaks. Strong on the arrows nobody worried about, weak on the one the structure depends on. The pyramid always assumed you could tell which layer you were standing on. That assumption is gone.