Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
@teortaxesTex
(This reads rather differently if you know how SDPO/OPSD is viewed rn. Fable&Sol, good job not falling for the meme
Epoch AI@EpochAIResearch · Oct 7We gave Fable 5 and GPT-5.6 Sol 3000 GPU-hours and tasked them with developing a novel post-training technique that would improve on a standard pre-existing baseline (GRPO).
Open quoted post → 6 151