Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won't) | Not Hacker News!