Seres to Lead Aito as Harmony Intelligent Mobility Confirms Cooperation Shift

On Sept. 14, Xiaomi founder Lei Jun visited Unitree Robotics accompanied by Wang Xingxing and watched a humanoid robot performance, according to STAR Market Daily. Public information did not disclose whether they discussed investment, product cooperation or other specific arrangements. Lei also visited Bilibili's Shanghai headquarters the same day and met Chairman Chen Rui. A large screen displayed a welcome message for Bilibili UP host Lei Jun. Chen Rui joined Kingsoft early in his career as Lei's technical assistant, and the topics of the visit were not made public.

WeChat employee Kecun Xiaojiang responded on Sept. 15 to privacy controversy over the AI assistant Xiaowei, saying the information Xiaowei can access does not exceed what the user can view and that claims of directly reading chat records or peeking at all private data are untrue. Xiaowei reads specified chat records temporarily from the local WeChat client only when a user initiates an Ask Xiaowei request in a chat scenario, for analysis and summarization, and original chat records are not saved, according to the response. Recommendations for public accounts and video accounts may reference follow, browsing and interaction data within user permissions. Xiaowei is developed by the WeChat team and remains in small-scale beta testing; it can query weather, express delivery and food delivery by text or voice and can be invoked in Moments, public accounts, chats and video accounts.

Google has opened Anthropic's Claude to all engineers, according to Business Insider. A Google spokesperson confirmed that engineers can now use the Opus 5 model through the internal development tool Antigravity; previously only some DeepMind and high-priority teams had such access. Claude has a per-user quota and is positioned as a supplementary tool, while Gemini remains the default and main model. Google had generally not allowed most employees to use external AI tools at work. The expansion gives engineers a formal channel to compare programming performance across models and turns Claude from an exception used by a few teams into a supplementary option available to all engineers.

Nvidia CEO Jensen Huang took a speakerphone call from President Trump during the All-in Summit in Los Angeles on Sept. 14. During the call, Trump called AI doomsday scenarios a scam, said robots could never rule the world and criticized calls to slow AI development as politically motivated. Trump also mentioned his uncle John George Trump, who taught at MIT for more than 40 years, and said he inherited a related talent and therefore understood AI; Huang responded that it was no wonder Trump knew AI so well. Separately, according to the Financial Times, Huang expressed a different approach to AI safety and regulation at Salesforce's Dreamforce conference. He disagreed with Anthropic CEO Dario Amodei's call for the industry to slow new model development, as well as with Musk and OpenAI CEO Sam Altman supporting controls on the pace of research. Huang defined AI safety as an engineering problem: pause and handle a product if it goes out of control, continue development once it is confirmed controllable, and embed evaluation and protection in the development process rather than stopping technological evolution. On regulation, he said existing laws and market mechanisms are sufficient to constrain product liability and that no separate body of law is needed for AI.

A First Financial investigation published on Sept. 15 found that some AI simulation short-drama producers use real-person base, fusion face and image padding techniques to mix facial features from multiple faces or alter skin color, age and contours on a real-person base image. Such images sit between similar and not similar, making it harder for ordinary people to prove that an AI character used their likeness. The core standard for courts to find portrait-right infringement is identifiability. Lawyers interviewed said that after facial features are mixed and deeply modified, victims often lack sufficient image samples and technical appraisal support; even if a lawsuit succeeds, compensation may only cover notarization, legal and lost-work costs. Some leading models now block generation requests that are too close to real people, but producers can bypass filters through overseas services or self-deployed open-source models, and platforms mostly remove infringing content after receiving complaints.

Waymo, Nihon Kotsu and the mobility platform GO announced a strategic cooperation on Sept. 15 and plan to launch a Level 4 fully driverless taxi service in Tokyo in 2027. According to the Japanese announcement, this would be Japan's first fully driverless commercial taxi project. The three companies have conducted road tests in seven Tokyo wards since 2025, with trained Nihon Kotsu drivers supervising from inside the vehicles to help the Waymo Driver adapt to narrow streets and crowded areas. After commercialization, passengers will be able to hail rides through the GO or Waymo apps, and the service is planned to operate 24 hours a day.

404 Media reported, based on internal documents and actual prompts, that OpenAI is using a project code-named Project Lily to have hundreds of outsourced workers review real ChatGPT user prompts, with some tasks presenting full conversations. Reviewers rate and comment on model replies to improve response quality, including reducing sycophancy and anthropomorphic descriptions of the model. The materials do not show ChatGPT usernames to reviewers. OpenAI said conversations pass through a Privacy Filter before review, but sensitive personal information may still slip through. Anthropic also confirmed to 404 Media that it uses human review to improve models. OpenAI's data controls show that conversations from personal ChatGPT and Codex can be used to improve models; users can turn off Improve the model for everyone in Settings under Data Controls, and new conversations after that will not be used for training.

Universal Music Group and its Capitol Records sued DistroKid in the U.S. District Court for the District of Delaware on Sept. 15, accusing the music distributor of deceptive business practices and continued distribution of infringing recordings. The complaint distinguishes clearly labeled AI-generated music from content impersonating human artists. UMG said DistroKid let streaming platforms believe its distributed content was created by real musicians with rights, and after admitting it had no right to distribute some recordings, still sent the same content to other services. UMG listed 1,000 specific recordings and said this was only part of the problem. DistroKid says it distributes for more than 4 million musicians, about 40 percent of global new music releases; the lawsuit will further scrutinize music distribution platforms' responsibilities for AI labeling, copyright verification and anti-stream-manipulation measures.

Cornelis, an AI networking company, announced on Sept. 14 that it raised $205 million in a funding round led by IAG Capital Partners and launched Active Compute Fabric, which aims to reduce GPU idle time spent waiting for data transfers. According to TechCrunch, the network allows chips to send information while processing data and uses an open architecture that can work with GPUs and accelerators from different vendors. Cornelis was spun off from Intel in 2020 and plans to launch its next-generation product later this year. Altera confirmed on Sept. 15 that it had confidentially submitted a U.S. initial public offering application. Reuters previously reported that the chip company, backed by Silver Lake and Intel, could list as early as this year and raise more than $2 billion; it makes programmable chips for data centers, communications networks, industrial equipment, AI, aerospace and defense systems. Factory, an AI coding-agent company, said on Sept. 15 that it raised $200 million at a post-money valuation of $5 billion, bringing cumulative funding to more than $400 million. Its products target enterprise software development teams and can be deployed in the cloud, on-premises or in isolated environments, allowing enterprises to control which models are used, how they learn and where they run. The company says Nvidia, Adobe, T-Mobile and Royal Bank of Canada are customers.

Zhu Xiaohu of GSR Ventures said in an interview with Shenwang that office agents are more likely to be led by large companies. He said platforms such as Enterprise WeChat, Feishu and DingTalk already hold organizational relationships, historical chats and enterprise knowledge bases, and these contexts will become underlying barriers for office agents; startup opportunities are more likely in vertical areas deep enough that platforms are unwilling or unable to cover themselves. Zhu also said foundation model capabilities are converging and models will eventually become commoditized like public utilities, with competition in the application layer returning to entry points, customer acquisition, data accumulation and ecosystem collaboration. He advised investors to use revenue growth, recurring revenue quality and cash return cycles to distinguish industrial demand from valuation bubbles, and said revenue sustainability for coding agents is more important than short-term growth because API customers can quickly switch to similar, cheaper models.

Editor's Summary

Harmony Intelligent Mobility confirmed that Seres will lead Aito's product, brand, channel and service businesses, while Huawei Terminal remains an enabler. The same news cycle included Lei Jun's visits to Unitree and Bilibili, Google's broader internal access to Claude, Waymo's 2027 Tokyo robotaxi plan and new legal and funding moves across AI. These developments point to continued shifts in corporate cooperation, AI oversight and autonomous driving commercialization.