Remix.run Logo
▲ criemen 3 hours ago

Here I was under the assumption that since GPT-4ish we mostly improved models via RLing, and that pre-training stayed (roughly) constant. Mind blown :)