这是什么
从 2022 年底美国禁止英伟达向中国出口顶级 AI 芯片开始,这场「技术脱钩」已经在悄悄推进。最新动向是:华盛顿不再满足于单点制裁,而是要求盟友在制度层面表态——是否加入美国的 AI 出口管制阵营。 受影响最大的三类对象:- 芯片制造国:荷兰(ASML)、日本(Tokyo Electron)、韩国(三星、SK 海力士)
- 算力部署国:阿联酋、沙特——这些中东国家原本是中国 AI 公司出海的重要跳板
- 模型与云服务国:东南亚、印度——本地数据中心选址正在重新评估
Not available in English yet
More on #DeepSeek
On two NVIDIA DGX Sparks, GLM-5.3 Flash beat DeepSeek V4 Flash on HumanEval (97% vs 94.5%). GLM ran 30% slower with one-quarter the context.
Reddit user hit 67-84 token/s on DeepSeek V4 Flash with a 1M-token context window on two ~$9K NVIDIA DGX Sparks. The local-LLM cost barrier is collaps
Reddit users spotted speculative decoding working on local GPUs—AI instantly outputting phrases via MTP. The local inference cost curve is being quiet
A Reddit developer tested DeepSeek/Qwen on four DGX Sparks: 'deep thinking' mode lowers scores and doubles runtime — a direct cost warning for AI infe
DeepSeek Agent framework has 5 event-dispatch modes decoded in an 8,000-word Chinese source tutorial - a shift from API calls to source-reading.
Tencent's Hy4 open-weights preview: 770B total, 49B active, 1M context, text-only. 2.6x scale in 6 weeks — but 49B active sets real compute cost.