
OpenAI's pretraining problems are apparently solved. According to a new report from SemiAnalysis, the lab has moved past its previous bottleneck and is now pushing forward with a far larger pretraining run — one reportedly codenamed Doug.
Leaker ChrisGPT adds some color: Doug isn't GPT-6. Instead, it's OpenAI's biggest pretraining effort yet, and it could arrive by November. He even suggests Doug will make Anthropic's own model, Fable, look "primitive."
So where does that leave GPT-6? ChrisGPT thinks GPT-6 is likely Astra, the model currently going through safety testing. That would mean OpenAI is running two tracks at once: ship Astra/GPT-6 first, then roll out the larger Doug afterward.
https://twitter.com/ChrisGPT/status/2086220662264250764
