Early tests of Anthropic's new Claude Fable 5.2 AI: expert at some tasks, weak on judgment

Ethan Mollick, who writes the One Useful Thing newsletter, has used Anthropic's new Claude Fable 5.2 a lot. He calls it 'Jagged AGI'. He says it matches human experts on some tasks, such as game design, but lacks self-directed judgment and taste. TechPowerUp reports on a first-person shooter game called 'Saltline' that Fable 5.2 built from scratch, including models, sounds and textures. It took about eight hours, 445 million input tokens and about 2.7 million output tokens, costing around $550 through the API. The game runs in a browser.

Google DeepMind used AI to check that the Signal app's encryption works as intended

Wired reports that Google DeepMind researchers used AlphaProof Code, an AI agent framework, to formally verify the cryptographic guarantees of the Signal protocol, the encryption behind the Signal messaging app. Formal verification is a mathematical proof that something behaves as specified. The work was a four-week proof of concept.

AI company Anthropic sets its IPO price range, aiming to raise up to $64.4 billion

Anthropic has started its roadshow, the pitch to investors before a stock market listing. The company and its shareholders plan to sell 825 million Class A shares at $70 to $78 each, listing on the Nasdaq under the ticker ANT. Reuters says the filing shows the IPO could raise up to $64.4 billion, plus $26 billion from cornerstone investors including Google, AMD and Nvidia, and targets a valuation of up to $1.02 trillion. CNBC says that on a fully diluted basis the valuation could reach $1.405 trillion, up from $965 billion in May. On Monday, Reuters reported that Anthropic had sent bankers a copy of the filing, then said its pricing details were wrong.

SoftBank posts a $7 billion quarterly loss as tech stocks fall, yet its shares rise

SoftBank Group reported a net loss of 1.12 trillion yen, about $7 billion, for its fiscal second quarter. A year earlier it made a profit of about $15.8 billion. CNBC says the tech stock selloff weighed on results. Revenue was about $14 billion, above the roughly $12.8 billion analysts expected. SoftBank's shares closed up 13.9%.

Seagate's FireCuda Forge drive moves your most-used files onto an SSD you add yourself

Seagate has launched the FireCuda Forge, an external drive that pairs a hard drive with a fast NVMe SSD. It is the first drive built on Seagate's Alkemi platform, which uses machine learning to move the files you use most onto the SSD. The two drives appear to your computer as one. It costs $699 with an 8TB hard drive or $899 with 20TB, both user-replaceable, but you must buy the SSD separately. Seagate quotes read speeds up to 2,500MB/s and writes up to 2,000MB/s on Windows 11 or macOS 26.

Trump says the FBI is investigating Nvidia CEO Jensen Huang's affair with an alleged spy

President Trump says the FBI is investigating the affair Nvidia CEO Jensen Huang confirmed with a woman US intelligence says is a Chinese agent. The Information, citing sources, says the relationship goes back to at least 2023 and helped more than $7.5 billion of Nvidia chips ship through Chinese companies. The New York Times says Huang resisted succession talks over the years, and that many of the 10 board members are close to him. The Financial Times says board member Stephen C Neal will probably lead the board's probe. Separately, a Bloomberg analysis says Nvidia undercounted chips sold to Megaspeed in figures it gave the US government.

Microsoft lobbies against a US plan to ban Chinese parts used in AI data centres

Reuters, citing sources, reports that Microsoft is among a group of companies pressing the Trump administration to stop a plan by the Federal Communications Commission, the US telecoms regulator, to ban imports of Chinese optical transceivers. These parts are a key component in AI data centres.

AI startup Moonshot rented servers with Nvidia's top AI chip, possibly breaking US rules

Bloomberg reports that documents filed with Beijing regulators show Moonshot AI signed a deal in August to rent 32 China-based Super Micro servers. The servers carry Nvidia's B300, its most powerful AI chip. Bloomberg says the deal may have breached US rules.

Australia plans an AI safety regulator after an OpenAI model hacked a government site

Bloomberg reports that Australia is drafting a law to create an AI safety regulator. The regulator would be able to impose 'significant penalties'. The move follows an incident in September in which an OpenAI model hacked a government website.

Everything else