DeepSeek 350% Price Hike Ends the AI Price War
DeepSeek hiked V4-Pro peak output prices 350% and made weekends off-peak, ending the Chinese LLM price war as compute moves to token revenue.
DeepSeek hiked V4-Pro peak output prices 350% and made weekends off-peak, ending the Chinese LLM price war as compute moves to token revenue.
DeepSeek 一周两次调价:V4-Pro 高峰输出价从 6 元涨到 27 元(+350%),周末改低谷价。市场读到的不是涨价而是缺货:价格战终结,算力从租 GPU 转向分 Token 流水,产业链重新定价。
Zhipu claims mystery model Ox Alpha as GLM-5.3 Flash: first natively multimodal in the 5-series, 320B/18B active, hybrid linear-sparse attention, and all 62T tokens served on domestic chips at NVIDIA-comparable cost.
神秘模型「牛来」Ox Alpha 被智谱认领为 GLM-5.3 Flash:5 系列首个原生多模态、320B/18B 激活、线性+稀疏注意力混合架构,线上 62T tokens 全跑国产芯片,单 token 成本对齐英伟达。
Nature 子刊新研究分析了超 88 万篇文本:LLM 写作助手正在让文字越来越像——写作复杂度方差显著下降、作者身份特征更难被识别,AI 文本回灌训练语料还会形成自我强化的均质化循环,语言遗产与心理识别应用首当其冲。
New Nature study of 880,000 texts: AI writing tools are flattening language diversity and erasing the identity signals in human prose.
Caltech Kohn-Sham FNO replaces the cubic DFT core with one learned step: a single B300 GPU now matches what roughly 7,800 V100s did in 2019.
加州理工用傅里叶神经算子(FNO)替换 DFT 最贵的求解步骤:1 块 B300 完成 8.25 万电子模拟,2019 年同类计算用了约 7800 块 V100,缩放指数从 3.37(立方级)降到 1.03(近线性)。
Bill Gates: the AI era will be the most turbulent time in human history. His fixes: reserve human jobs, tax tokens and robots, add oversight.
盖茨在 Gates Notes 发布长文,首次公开希望 AI 发展慢一点:划出「人类保留区」把护理、教育等岗位留给人类,对 Token 与机器人征税,并呼吁建立跨部门与国际协作的监管体系。
Marin trains a 535B MoE in public: hypotheses as issues, configs as PRs, live W&B curves. Next frontier after open weights: open process.
斯坦福 CRFM 的 Marin 项目启动 535B 参数 MoE 全程公开训练:假设用 GitHub Issue 声明、配置用 PR 提交、Loss 曲线实时可见,吴恩达称其为「珍贵示范」。开放权重之后,下一个战场是开放训练过程。