Assertions, not assumptions. We’ve known small models can do wild shit for years now. Google themselves proved this by beating their world-news Go bot with a much smaller model that studied no human games, and then beating that with an even smaller model that also played other board games.

American AI companies keep scaling up because that prevents competition. They insist it takes big data because they love an excuse to slurp up your data.

The endgame has always been local models. And since their scale is constrained, and you can always just train them some more… no company can lock it all down and make their model the model. It’s gonna stay weird.