gpt-6POST
质疑「两周训出超越 Astra 的模型」说法
作者反驳某说法，认为所谓「训练」大概只是在早已准备好的基座上恢复 RL，谈不上两周内实现跨越式提升。
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
@teortaxesTex2026.09.14收录原推发布于 2026.09.14
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex原推发布于 2026.09.14
@TheStalwart > that it didn’t even take OpenAI two weeks to train a model that’s a step change above Astra this is of course BS we don't really know what they mean by "train" but I guess "resumed RL on a base that's been ready for a while"
x.com/teortaxestex/status/2099183493712380266采集于 2026.09.14 16:04