观点：OpenAI 靠后训练加 Astra 预训练规模取胜

Angel 🌼 @Angaisb_原推发布于 2026.09.21

在 X 上打开原推

I wouldn't call Anthropic the kings of pretraining anymore, it's easy to ruin good pretraining with bad RL and you can see it in how bad Opus 5 and Sonnet 5 are OpenAI mastered post-training, which is the hard part, they just needed to scale pretraining with Astra

这条是引用着另一条写的：「爆料：OpenAI 筹备 GPT-6 Sol/Luna，Astra 6.1 在训」