White House officials were expected to raise concerns with Xi Jinping at the summit that Chinese companies used distillationDistillationTraining a smaller or cheaper model on the outputs of a larger one so it inherits much of the larger model's ability. It is a standard technique, but it has become a trade dispute because it lets a lab reproduce a rival's capabilities without matching its compute spend., training a smaller model on the outputs of a stronger one, to reproduce what top American AI models can do, The New York Times reported. The technique is common and not illicit in itself; the dispute centers on claims that Chinese firms tapped American models without authorization, per Newsmax. Anthropic has accused DeepSeek and Moonshot AI of routing user requests to its Claude models, according to AP. Chinese officials have rejected the allegations, and some analysts say distillation is not the main reason Chinese models have improved. Beijing has in turn raised concerns about risks to its critical infrastructure from American AI, including Anthropic's Claude Mythos.
Read at The New York Times ↗ • Read at AP ↗ • Read at Newsmax ↗
Installers for a desktop version of DeepSeek Harness, the Hangzhou company's AI agentAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. workspace, began circulating among developers before any official announcement, per 36Kr. The Mac build carries Apple's verified developer signature registered to Hangzhou DeepSeek Artificial Intelligence Co., Ltd., and its download links point to DeepSeek's own servers. The app reuses the existing web interface and its agent, tool and plugin systems, with separate modes for office work and for coding. Unlike DeepSeek's free chatbot, the app requires users without credit to top up an account and complete real name identity verification, or connect their own API key, a paid developer access code. Neither DeepSeek's website nor its GitHub releases page had posted a download link as of publication. The preview carries version number 0.1.7.
Read at 36Kr ↗
Xirang Kaiwu, a physical AI startup founded by two former Huawei AI executives, has closed seed and angel rounds worth hundreds of millions of yuan (at least $14 million) at a $500 million valuation, QbitAI reported. Dunhong Asset led the financing. Chief Executive Li Yin was chief technology officer for Huawei Cloud's Pangu large language models, and co-founder Zhang Hanwang, a Nanyang Technological University professor, was Huawei's chief multimodal scientist. The company is building what it calls a large physics model, pretrained on video, robot movement records and real world interaction so one model can carry over across different robots and tasks. Xirang said its model ranked first for trajectory accuracy on one track of the WorldArena 2.0 benchmark, though the model has not yet finished pretraining.
Read at QbitAI ↗
Mifeng Technology, a data subsidiary of Shanghai humanoid robot maker AgiBot, launched Mifeng Pai on Sept. 23, a crowdsourcing service that pays ordinary people to wear its recording gear during daily work and chores, per QbitAI. Users claim tasks in an app, record with a panoramic headset and wrist camera or a handheld gripper, and are paid by valid recorded hours after review. The company grades the footage and sells it to developers training robot models. It said 20,000 users registered during a monthlong closed test and submitted 13,000 tasks, with one earning more than 5,000 yuan ($700) in a month. Mifeng has produced 20,000 recording kits and said it holds about 1 million hours of footage across 22 scene categories.
Read at QbitAI ↗
Moore Threads, a Chinese designer of AI chips, released test data for its MTT S5000 accelerator on robot learning workloads, Pandaily reported. The company said training curves on the chip tracked those of unnamed mainstream foreign GPUs with a 0.976 correlation under controlled conditions, and claimed speeds on core AI calculations 1.6 to 2 times those GPUs in the tested setups. The figures come from Moore Threads and have not been independently checked. RLinf, an open source framework for training robots by scoring their own attempts, has supported the chip since version 0.3. Moore Threads said real robots trained on the chip completed a dual arm assembly task about 92% of the time.
Read at Pandaily ↗