Date: July 22nd, 2026 9:08 PM Author: Jared Baumeister
and oddly enough it's the same thing I've been saying: you shouldn't need more than a few billion parameters to do most shit. "Distillation" is a fake term that obscures what people are actually doing, which is making models that require fewer parameters. They only need to think, not memorize
(http://www.autoadmit.com/thread.php?thread_id=5885302&forum_id=2,5#50018056) |