A Microsoft 'theft' memo surfaces; OpenAI builds Astra for Law

A summary judgment motion unsealed on September 17 in the news publishers' copyright case against OpenAI and Microsoft quotes Microsoft's director of applied science, Brent Hecht, calling the scraping of news for AI training "an astonishing theft of unprecedented proportions" and "the largest theft of labor in human history." The filing also cites Microsoft data showing click-through rates to some publishers falling 83 to 93 percent under Copilot, and an internal OpenAI message from ChatGPT head Nick Turley describing chatbots as an "existential threat" to publishers. The quotes come from the plaintiffs' own brief, and the underlying exhibits remain sealed. Earlier this month Microsoft argued that its chatbot almost never reproduced Times articles; this motion shifts the fight to substitution, the fair-use factor that asks whether a product replaces the work it learned from.

OpenAI released Astra for Law, GPT-6 Astra paired with a legal search index of more than 230 million URLs covering United States case law, statutes, regulations and court rules, with case law supplied by the nonprofit Free Law Project. On 200 questions from Vals AI's Legal Research Bench, OpenAI reports it answered 54 percent correctly against 38.7 percent for GPT-6 Astra with web search alone. Access starts with selected firms through a Trusted Access program in ChatGPT and Codex, and an API version has no date or price yet. Harvey and Legora are named as customers, and 26 partner plugins launched alongside, including Thomson Reuters and iManage.

Pew Research Center surveyed adults in 37 countries between February and June and found that in 34 of them, people are more likely to expect AI to bring fewer jobs than more over the next 20 years. Worry runs highest in wealthy countries: around seven in ten adults in Australia, South Korea and the United States expect job losses, and the American figure is up 7 points in two years. Across all 37 countries, a median of 41 percent feel equally concerned and excited about AI, while 37 percent are mostly concerned. Concern is rising fastest among young adults in several countries.

Researchers from Google and Google DeepMind launched the DeepMind Institute, an essay platform directed by Shane Legg, James Manyika and Demis Hassabis that says its pieces are not Google's official view. The newest essay, by safety researchers Rohin Shah and Anca Dragan, argues that the ability to read a model's step by step reasoning is worth protecting deliberately. They propose measuring it, capping how much hidden sequential computation a model can do before it must reason in readable text, and never training a model's reasoning to merely look safe. The institute also republished the frontier-standards proposal Hassabis first posted in July.

Huawei moved the launch of its next-generation Ascend 960DT AI chip up to the first quarter of 2027, from a previous target in the third quarter, announcing the change at its Huawei Connect conference a week before the September 24 meeting between President Trump and President Xi Jinping. The company says its Atlas 950 SuperCluster can link up to 256,000 accelerator cards. Analyst Rui Ma noted a catch: the Atlas 960 system Huawei described this week holds 4,096 chips, well short of the 15,488 it had originally laid out.

Anthropic redesigned projects in Claude Code around a coordinator: a user sets a goal and connects repositories, and Claude breaks the work into threads, each a cloud session on its own branch, then reviews the output and reports which pull requests should merge first. Threads share a project memory that builds over time, and work continues after the user steps away. The beta reaches selected Pro and Max subscribers who use cloud sessions, widening over the coming week, with Team and Enterprise plans to follow.

Two months after its first 27 billion parameter model, PrismML shipped Ternary Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B that stores each weight as minus one, zero or one, for a total footprint of 5.9 gigabytes, small enough to run on a laptop or in a browser. The company says it keeps 98.2 percent of the full model's aggregate benchmark score while being more than nine times smaller, with a 262,000 token context window and image input. The weights are on Hugging Face under the Apache 2.0 license, the same terms as the base model.

Alibaba's Qwen team released Qwen3.8-Omni-Flash, a model built on Qwen3.8-Flash-Next that takes text, images, audio and video in a context window of one million tokens, and can call tools to work through long recordings. It is available through Qwen Chat and Alibaba's API, with no open weights, at 15 cents per million input tokens and 47 cents per million output tokens on Vercel's AI Gateway. Alibaba says its audio and video performance approaches Google's Gemini 3.8 Flash, a claim no independent evaluation has yet tested.