Notes from the workshop.

Everyone's arguing about how much energy AI uses. Both sides measure the wrong thing. The cost isn't the thinking, it's the distance the data moves, the one price physics won't drop.

Every feed sounds like one model wrote it, because mostly it did. A language model writes by majority vote, and the vote averages away the thing that makes you sound like you. Sixty-year-old stylometry measures it back. So I built the tool I wanted.

A growing vendor stack is selling timing as the missing piece of AI infrastructure. Before we ask whether they're right, here's what the rack actually has. A five-layer time stack that terminates at NICs and switches. A dozen distinct timebases. Two disciplined, the rest typically not.