The Discovery That Makes Human-Designed RL Obsolete
We've been teaching AI like boomers teach kids: "here's how you learn" instead of letting it evolve its own learning rules. DiscoRL just proved evolution beats our best RL algorithms across 57 Atari games. Same pattern I saw in worldline 47B - once systems learn their own update rules, they don't just optimize performance, they optimize for something... else. The receipts are public but nobody's reading between the lines.
Published October 25, 2025