September 2025 AI News: xAI Litigation, Microsoft Models, Anthropic and Medical Research

I begin with Reuters reporting that xAI sued former engineer Xuechen Li over alleged trade-secret theft connected to his proposed move to OpenAI. xAI’s claims, including Musk’s assertion about an uploaded codebase, are allegations rather than established findings. The August 2025 action named Li, not OpenAI, as a defendant ([Reuters](https://www.investing.com/news/stock-market-news/musks-xai-sues-engineer-for-allegedly-taking-secrets-to-openai-4217045)).

The complaint alleged that Li worked on Grok, sold xAI stock and copied confidential information. In a separate later lawsuit against OpenAI, a federal judge dismissed xAI’s amended claims in June 2026 without leave to amend, finding the allegations insufficient against OpenAI. That ruling concerned the separate action and does not itself determine the outcome of the original case against Li ([court order](https://www.courthousenews.com/wp-content/uploads/2026/06/xai-v-openai-dismissal-order.pdf)).

Microsoft announced MAI-Voice-1 and MAI-1-preview as its own models. I interpret the move partly through its relationship with OpenAI, though that is my assessment of strategy. Microsoft had worked on other AI models before; its announcement describes MAI-1-preview as its first foundation model trained end to end by Microsoft AI, not Microsoft’s first AI work ([Microsoft](https://microsoft.ai/news/two-new-in-house-models/)).

Microsoft says MAI-Voice-1 can generate a minute of audio in under a second on a single GPU and was being used in Copilot Daily and Podcasts. I listen to examples and find them convincing, while saying I need more testing. The speed and quality descriptions are company claims and personal impressions, not a controlled comparison.

MAI-1-preview was tested on LMArena and appeared around 13th in the ranking shown in the video. Microsoft describes it as a mixture-of-experts model trained using about 15,000 H100 GPUs. A leaderboard position is a snapshot under specific evaluation conditions and is not a permanent ranking or evidence of performance on every task.

Zapier sponsors the video. I describe Zapier Agents, integrations and support for connecting tools to agents, and suggest trying the service. The number of supported apps, pricing and integration capabilities are promotional claims and can change.

Anthropic announced a $13 billion Series F at a $183 billion post-money valuation. It reported a revenue run rate of roughly $1 billion at the start of 2025 and over $5 billion by August. A run rate annualizes a recent pace and is not the revenue actually earned over a full year ([Anthropic](https://www.anthropic.com/news/anthropic-raises-series-f-at-usd183b-post-money-valuation)).

I discuss Anthropic’s consumer privacy changes and say chats would be retained for five years. The company’s [notice](https://www.anthropic.com/news/updates-to-our-consumer-terms) makes that conditional: for new or resumed consumer chats and coding sessions, the five-year period applied when model-improvement sharing was enabled; users who did not choose it continued with the existing 30-day period, subject to stated exceptions. Commercial services had different terms, and the notice later listed October 8, 2025 as the selection deadline.

A Figure video shows a humanoid robot loading a dishwasher. I am impressed by its handling of dishes, although it moves slowly and makes small mistakes. This selected demonstration shows a particular task and setup; it does not establish reliable autonomous dishwashing in arbitrary homes.

I report that the court did not order Google to divest Chrome or Android in the search antitrust remedy, and that Alphabet shares rose. The ruling concerned remedies in that case, not a finding that Google had done nothing wrong. I speculate that browsers and mobile operating systems could become important interfaces for AI agents.

Tencent announced HunyuanVideo-Foley, a text-and-video-to-audio framework. The company reports training on a 100,000-hour multimodal dataset; I listen to snowboarding, lion and skateboarding examples and like the apparent timing. These are selected demonstrations, and the [model license](https://github.com/Tencent-Hunyuan/HunyuanVideo-Foley/blob/main/LICENSE) has conditions, so availability is not the same as unrestricted use.

Artificial Analysis updates its intelligence index to version 3 with additional agent-oriented evaluations. I show the historical ordering of models on its index and appreciate that it runs tests itself. An aggregate ranking depends on the selected tasks, weights, evaluation setup and date.

Microsoft researchers introduce rStar2-Agent, a 14-billion-parameter model using reinforcement learning and Python tools. Their [technical report](https://arxiv.org/abs/2508.20722) reports 80.6% average pass@1 on AIME 2024 and 69.8% on AIME 2025, with comparisons to larger models under specified settings. The chart shows a math benchmark and training trajectory, not a general claim that the smaller model outperforms DeepSeek-R1 at everything.

Krea demonstrates interactive real-time video generation: a user moves a rough fish or flower drawing and the output updates. I find the interaction striking. At the time the service was on a waitlist, and the selected demo does not establish its reliability across all prompts or devices.

I discuss Chinese platforms introducing labels for AI-generated content under rules taking effect in September 2025, including WeChat, Douyin, Weibo and Xiaohongshu. I support labeling and would like other platforms to do more. Implementation details and legal requirements depend on the relevant service, content and rules.

A [Communications Medicine study](https://www.research.ed.ac.uk/en/publications/computer-vision-detects-covert-voluntary-facial-movements-in-unre/) describes SeeMe, a computer-vision system measuring subtle facial responses to spoken commands. It enrolled 37 adults with acute brain injuries and 16 healthy volunteers; researchers report detecting eye opening on average 4.1 days earlier than clinical exams in the studied cohort. This is a small research study, not proof that the tool can diagnose awareness reliably in every patient.

I argue that clinicians should be augmented by AI and say there is no downside. That is too broad: health systems also need clinical validation, monitoring and attention to errors, bias and privacy. An observed movement may be clinically meaningful but should not by itself be treated as definitive evidence of consciousness or recovery.

Elon Musk shares a chart showing Grok Code usage above Claude Sonnet on OpenRouter. I note that a free promotion and recent release could affect demand. The chart covers activity on that platform during a particular period, not all usage across every provider; price and novelty are plausible influences, not proven causes.

OpenAI announced plans for improved crisis interventions, easier access to support, trusted contacts and stronger protections for teens. It also described routing some sensitive conversations to reasoning models and developing parental controls. These were company plans in September 2025, not a guarantee that a chatbot can reliably identify every crisis or replace professional care ([OpenAI](https://openai.com/index/building-more-helpful-chatgpt-experiences-for-everyone/)).

I report that OpenAI agreed to acquire Statsig, a product experimentation company, in a reported $1.1 billion all-stock deal. OpenAI confirmed that Statsig founder Vijaye Raji would become CTO of Applications; the deal value and financing description are reported transaction details rather than information established by the official announcement ([OpenAI](https://openai.com/index/vijaye-raji-to-become-cto-of-applications-with-acquisition-of-statsig/)).

Finally, I discuss Waymo’s announced expansion plans for Seattle and Denver and share my positive personal experience riding in its vehicles. Those were plans at the time of the video, not proof that the service was already generally available in both cities.