Kafka on the Shore
5.5K posts


以前车圈有BBA, 现在电车能称为 "BBA" 的除了特斯拉还有谁? 小米,蔚来? #tesla



活久见系列,有人贴心的帮 CC Switch 做了一个官方网站,并且链接到我的 GitHub 仓库,但是你这个定价是几个意思?在此声明一下,CC Switch 目前没有官方网站,唯一的下载地址是 GitHub,并且开源免费,凡是收费的都是盗版,谢谢大家!



@bboczeng 你的意思是量产车改造的增幅比原型车还大?



The DeepSeek narrative was always “2,000 H800 chips, $5.6 million, and Chinese ingenuity beat $100 billion in American spending.” That story just collapsed in 72 hours. On Sunday, Reuters confirmed via a senior Trump administration official that DeepSeek’s upcoming V4 model was trained on smuggled Blackwell chips clustered in an Inner Mongolia data center. On Monday, Anthropic published detailed evidence that DeepSeek ran 150,000+ exchanges through 24,000 fraudulent accounts to distill Claude’s reasoning capabilities. Google followed the same day, reporting 100,000+ prompts targeting Gemini’s reasoning traces. And today, Reuters reports DeepSeek is withholding V4 from Nvidia and AMD entirely, giving Huawei weeks of early access instead, because showing the model to American chipmakers would expose which hardware actually produced it. That sequence tells you everything. DeepSeek plans to launch V4 claiming it was built on Huawei chips. The U.S. government is saying, on the record, that’s a lie they’re watching DeepSeek prepare to tell. Count the dependencies. Smuggled Nvidia Blackwell chips they can’t legally possess. Distilled reasoning from OpenAI, Anthropic, Google, and xAI they can’t legally access. And now a cover story attributing the results to Huawei hardware that didn’t produce them. The “$5.6 million training run” was always a mirage. SemiAnalysis estimated DeepSeek’s parent company spent $500M+ on Nvidia GPUs. The founder admitted in 2023 to stockpiling 10,000 A100s before the export ban hit. The real V3 training cost was $5.87 million for the final run alone, excluding all prior R&D, ablation experiments, and data costs. And those numbers assumed legally obtained hardware. What the last 72 hours reveal is a company that simultaneously needs banned American chips to train, banned American models to teach, and a fabricated origin story to ship. Anthropic made the point explicitly: “Without visibility into these attacks, the apparently rapid advancements made by these labs are incorrectly taken as evidence that export controls are ineffective.” The proof that chip bans don’t work was itself built on chip ban violations. The V4 launch will be fascinating to watch. DeepSeek will claim Huawei. The U.S. government has already told you that’s false. And every benchmark that follows will carry an asterisk the size of an Inner Mongolia data center.



















