Significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview
这是 DeepSeek 官方 Change Log 里的一手数据，直接证实了 EP.97 故事线 B 的核心事实：V4-Flash 在 Terminal Bench、Cybergym 等九项基准上大幅反超自家旗舰 V4-Pro-Preview。一个"轻量版"模型靠后训练反超"旗舰版"，说明模型能力的边际提升正在越来越多地来自后训练配方，而不是参数规模或架构本身。
It is best understood as transferring selected capabilities into a cheaper, locally controlled system, not achieving independence from frontier AI.
这是对"蒸馏能不能让中国AI实现独立自主"这个问题最精确的限定回答——不是独立，而是把前沿模型的部分能力搬进一个更便宜、可控的本地系统。说这话的 Trevor Koverko 是 AI 数据公司 Sapien（https://sapien.io/，专注 AI 训练数据质量验证/Proof of Quality）联合创始人，这句话给 EP.97 故事线 B 提供了一个必要的降温视角：蒸馏管用，但不是万能钥匙。
Teaching a model the right answer is one thing but teaching it the reasoning behind the answer is much harder
这句话点出了本篇 Reuters 独家报道的技术核心：中国军方关联研究者不是在抄答案，而是在系统性提取美国前沿模型"如何推理"这件事本身——这正是 EP.97 故事线 B 的论点起点：出口管制卡得住芯片和权重，卡不住模型输出里蕴含的推理路径。说这句话的 Sunny Cheung 来自 Jamestown Foundation（华盛顿智库，长期研究中国军事与科技政策），本次分析了 60 余篇相关论文。
Ultimately, by inducing a highly aligned cross-embodiment representation, UniT offers a scalable path to distill vast human knowledge into general-purpose humanoid capabilities.
这一值得深入思考的声明提出了UniT在将人类知识转化为通用人形机器人能力方面的潜力，为未来人形机器人技术的发展提供了新的方向。
Anthropic, who accused DeepSeek, Moonshot, and MiniMax of distilling from Claude's outputs.
Anthropic 公开指控 DeepSeek、月之暗面和 MiniMax 从 Claude 的输出中蒸馏数据——这是一个令人震惊的商业伦理事件。更深层的含义是：这些中国公司被迫采用「寄生式追赶」策略，以 Claude 为「免费教师」压缩训练成本。这既是技术现实的写照，也暗示了「无算力优势」下的竞争逻辑：当你无法花钱训练更好的模型，就借用别人训练好的。
method and madness by [[Alan Jacobs]]
via In which I describe my writing “methods." by [[Alan Jacobs]]
reply:
@ayjay Thanks for sharing this. My method is often very much like yours. Lots of internal distillation, slowly over time. I remember hearing a story that Mozart wrote music "like a cow pees" (in one giant and immediate flood and then done). I feel like large works of writing, composing, etc. springing, as if fully formed from the head of Zeus is more common than is acknowledged. Cory Doctorow hints at a similar sort of method in his own work in The Memex Method. I'm also reminded of bits of what neuroscientist Barbara Oakley calls "diffuse thinking" or a more internalized version of Michael Ondaatje's "thinkering" described in The English Patient.
It is important to learn as much and at the sametime as little as possible.J
By abstracting and concatenating portions of material, one can more efficiently learn material that would otherwise take more time.
Instead of dumping the humidity outdoors, why not have an option to condense it to a water holding tank, thus offsetting the need to purchase water? This sounds like it might even be able to distill water, which would further save money for those who need distilled water for apnea devices and other uses.
On Zettel 9/8a2 he called the Zettelkasten "eine Klärgrube" or a "septic tank;" (perhaps even "cesspool"). Waste goes in, and gets separated from the clearer stuff.
Niklas Luhmann analogized his zettelkasten to a septic tank. You put in a lot of material, a lot of seemingly waste, and it allows a process of settling and filtering to allow the waste to be separated and distill into something useful.
Dataset Distillation
一篇知识蒸馏（Distillation）的新应用。Distillation 本身的应用目标是为了“模型压缩”从而知识转化实现小模型上的快速训练，而此文是对“数据压缩”以实现不错的蒸馏效果。