> China’s rapid AI compute buildout is straining local grids, particularly in coastal urban areasnot even Chinese grid is omnipotent, wow. Datacenters are already enough to…
I don't get why Meituan is not admitting it's Ascend 910Cs and instead pointedly says "ASICs". Or if it's somehow not 910C, same question. Are they…
cries in poorLotto: That’s actually crazy A new frontier is found imohttps://www.localmaxxing.com/en/runs/cmqznfqaa02rfoe01tuis95dv
I think in this case it's just that RL is currently easier logistically than pretraining on Ascends. I suspect it's just not a very good pretrain,…
oh brotherChujie Zheng: This guy seems to have left our team after the first-gen Qwen model. Our team also did not have a “head of RL”…
RT αιamblichusWhenever people on this website try to tell you that the EU (a $23T economy) should give up on building a SOTA LLM because "even…
RT Xiuyu LiStepFun Foundation Team @StepFun_ai 诚招 26 / 27 届校招和实习生，岗位包括- Pretrain / Post-train / 多模态 / 语音- 算法 / Infra / 数据完整 JD 可查看 https://app.mokahr.com/social-recruitment/step/94904#/欢迎发送简历和方向至…
RT eliethe new sparse attention method introduced with this model is basically a combination of components from existing ones. let's go over each sparse attention method…
RT thebes"Assistant! My dividend is half the size it should be!".oO ( Thinking... ) You are correct, User. Your dividend of the Beyond was $25,514,376.30. Using…
RT changyi yangWeight sync in RL post-training looks RDMA-bound. But >99% of BF16 weights are byte-identical step to step — ship only the delta: ~100–1000× less…