爆料：Anthropic 与 OpenAI 内部新模型已超越 Astra
有人根据长期潜水观察称，两家实验室的未公开预训练模型已超过 GPT-6 Astra 和 Fable，内部团队正用 Lean 规格加数千并行智能体冲击数学难题。
so in short what is happening inside labs right now both Anthropic and OpenAI have new pre-trained models that leap Astra and Fable, and they believe these capabilities are such that millennium problems can be solved with a combination of: 1. a correct Lean spec 2. a fuckton of compute (1000's parallel agents) 3. humans telling the AI "you can do it!" so that's what they've been doing. these internal teams are literally tasked with, and competing to, solve open math problems, and whoever solves it gets credited and then they do robot dances to celebrate (ok 3 was a bit of a joke but is it that far from truth?) OAI's model will probably be public in 1-2 months no idea if A\ has plans to launch and may keep for them also the pacing Ai stuff is about new training runs, for models even smarter than these. they're more concerned about misbehavior during training, than after the model is completed I have no leaks that's just what I collected / inferred from 24/7 lurking so this post can range from wrong or obvious