Delphi Intelligence AMA

Teknium × Tommy Shaughnessy

Nous says 90%+ of its data is synthetic, and Atropos generates reasoning data across more than 1,200 tasks.Teknium sees post-training RL as barely explored and argues open research can pool domain expertise large companies cannot staff internally.He claims Hermes 4 leads Meta’s instruct variant by “something like 60 points” on MATH and 21 points on MMLU-Pro, while urging downstream testing as Psyche’s RL infrastructure is still being built.