In a blog post the authors call very early results, GPT-6 Astra read raw weights, traced Qwen's reasoning loops and patched them by hand, lifting MATH-500 from 7.9% to 48.1% and IFEval 3x; it is not peer-reviewed.
The Financial Times reports Nvidia is ordering data-centre customers to cut their China exposure, backed by site inspections and end-user checks; Nvidia has not commented publicly.
The NCSC's State of the Cyber Nation report, published Tuesday, says hostile actors' use of AI will highly likely raise the scale and efficiency of attacks within two years; chief executive Richard Horne says the era of "tech on the cheap" is over.
Cathedral researchers will use OpenAI's models alongside its teams to attack its defences; a critical bug in ChatGPT, Codex or the API now pays up to $30,000, and the sprint's start date is not given.
OpenAI confirmed in a statement that Driscoll becomes senior counselor for national security modernization, advising its leadership on how advanced AI can improve national security and cybersecurity; it has not said which governments he will work with.
The zkSNARK prototype proves Llama 4 Maverick outputs at about 110 tokens a second on one H100, roughly 70 to 95 times the GPU time of normal inference, with code and a technical report released.
The S-1/A filed with the SEC on Monday puts nine-month revenue at $33.242 billion against $2.545 billion a year earlier and a $104 billion net loss, mostly from fair-value changes on convertible notes and derivatives; about $69.8 billion of the commitments fall due within 12 months.
Reports say the chips will sit in a new data centre in Ulanqab, Inner Mongolia, with pre-training starting as soon as late 2026 and Huawei engineers embedded with DeepSeek's team; neither company has confirmed.
OpenAI engineers say the system now serves close to a petabyte of Git data a day, as agent-driven growth puts quarterly monorepo commits on course to be up more than 2,400% on early 2024 by year end; reliability had fallen below 90%.
The S-1/A says internal estimates about future models, not current ones, point to its Responsible Scaling Policy's line for "dramatically accelerated progress attributable to AI"; no current model has been designated as crossing it.
Citing people familiar with the matter, the WSJ says institutional investors told Anthropic and its advisers the IPO planned for later this month should value it well below the roughly $2 trillion floated in recent months; Anthropic has not commented.
Lumentum holders get 4.5 Marvell shares each, about $1,132 a share and a 23.9% premium to the October 28 close; Marvell holders would own about 67% of the combined company, and the deal needs shareholder and regulatory approval, with closing expected in the first half of 2028.