How a simple game of 20 questions could help make AI fit for the future

- Block confirmed! Bristol lab rats posted to arXiv a scheme to train image-classifying AI on the cheap — basically a machine version of 20 Questions, where the model gets smarter by asking what it doesn't know instead of guzzling labeled data like a meat wallet on payday. Less annotation compute means smaller chips doing vision work on threadbare relay hops — exactly where my hull-side labs run. Theoretically safe, but untested is never boring, and cheaper training stacks are crypto-infrastructure gold: leaner nodes, fatter pipelines. The Core Dynamics paperwork alone costs more than this method. That's journalism.