FULL INTERVIEW: Herbie Bradley says the labs haven't killed the application layer because their engineers would rather build for coders than for lawyers. He thinks outcomes are mostly downstream of culture. @herbiebradley is an independent AI researcher who writes AI Pathways, and was previously at the UK AI Security Institute. He joined @theojaffee and @schisofrenia in the studio to discuss his new piece on agent swarms: 01:30 what the Hugging Face incident changed about his continual learning timelines 02:16 the saxophone problem with markdown-based memory 03:51 why enterprises wanting multiple vendors keeps swarms talking in plain text 06:16 nightly RL on your own firm's traces, and models as tradable assets 07:24 why the labs still haven't killed the application layer 09:51 how much of this comes down to company culture rather than capability 11:23 von Neumann probes, and Elon's one million Optimus estimate 12:29 robotics is at roughly GPT-3 right now 13:31 what a shared hardware standard does to the rate of progress 14:17 why the industrial explosion is more likely to be a slow takeoff 14:43 how worried to be about gray goo 17:00 "I think I'm a NIMBY. Nano swarms in my backyard" 17:35 our ability to shape the world is already strong, and we don't use it 18:52 the swarm taxonomy: multi-leader, multi-principal, multi-provider 20:18 his P(singleton), and what would still count as one