Meta's Lucas Beyer argues 29 million agentic sessions provide massive post-training scale, while Susan Zhang demands concrete token counts

@giffmana @Alibaba_Qwen "a good chunk" is not a numerical amount i'm able to parse :)

@giffmana @suchenzang @Alibaba_Qwen depends on what "an exchange" is. one turn? tool calls included? might be very few. and you need to jailbreak claude in a way that gets you the full, not the summarized cot, which might push it OOD

(actually interesting research direction here, which no one will do :()