Ok I was wrong here. My confusion comes from one line in the OpenAI post: "Since August 28 we have been training a new internal model." If you read "training" as pretraining from scratch, the timeline is impossible. You cannot pretrain a capable super-frontier model in a week on any cluster on earth. But OpenAI's own Sept 1 post explains the date. in "PAth to Asta", they say "On August 28th, we restarted the large frontier RL run that was previously paused after the new safety and security…

Peter Wildeford🇺🇸🚀@peterwildefordSep 8

I genuinely didn't know you could train a model for only 11 days and have it be useful for solving Millennium problems