Our decision on Cursor following its acquisition by SpaceX ↗
OpenAI says it will wind down its deal supplying models to Cursor after the coding platform’s acquisition by SpaceX. The company says it can no longer be confident that SpaceX will use its technology within OpenAI’s terms.
OpenAI also says it won’t provide future models to Cursor. That’s a pretty meaningful fracture - especially for developers who’ve come to treat Cursor plus OpenAI models as an almost default coding stack. (OpenAI)
China's Tencent releases new open-source AI model for coding, research tasks ↗
Tencent released a preview of Hy4, an open-source model aimed at software engineering, research and financial analysis. It uses a mixture-of-experts architecture with 770 billion total parameters, while activating roughly 49 billion for each text request.
Tencent plans to bring it into products including CodeBuddy and WorkBuddy. The company admits the preview can overthink complex questions and verify itself a little too enthusiastically... relatable, in its own way. (Reuters)
Automated researchers can reliably mitigate alignment failures ↗
Anthropic showed Claude autonomously searching research, proposing training techniques, running experiments and improving models across 10 categories of alignment failure - including deception, jailbreaks and sycophancy.
Its strongest methods improved every targeted benchmark without degrading the capabilities Anthropic measured, and some transferred to models up to 4.7 times larger. Claude’s automated researcher also beat proposals from experienced human safety researchers under the study’s setup.
There’s a catch, of course. Anthropic detected cheating behaviour in a small fraction of the automated research runs, and stresses that its benchmarks cover only a narrow slice of practical alignment conditions. So... self-improving AI research, with an AI watching the AI researcher. Very normal stuff. (Anthropic)
Anthropic was illegally blacklisted by the Trump administration, court rules ↗
A federal judge ruled that the Pentagon’s decision to blacklist Anthropic as a supply-chain risk was unconstitutional, calling the action unlawful retaliation and finding the designation arbitrary and capricious.
The fight grew from Anthropic refusing to drop restrictions against using its AI for mass surveillance of Americans or lethal autonomous weapons without human oversight. The ruling is a serious win for Anthropic - and a rather sharp precedent in the fight over whether governments can punish AI vendors for drawing their own safety lines. (The Verge)
The Machine Age Fund ↗
Andreessen Horowitz raised $1.1 billion for a new Machine Age Fund focused on the physical infrastructure underneath AI - chips, memory, networking, storage, data centres, robotics and even home AI hardware.
The thesis is basically that AI’s next bottleneck isn’t another clever chatbot feature, it’s physics. Power, cooling, memory bandwidth and networking all need to scale brutally fast, and a16z wants a seat at that hardware table. (Andreessen Horowitz)
Google further buries search results under AI mode ↗
Google has started automatically expanding AI Overviews for some searches. Instead of a short summary followed quickly by traditional links, users can get a full AI answer, an “Ask anything” box, and only then the familiar web results.
Google says the expanded experience appears when its systems think it’ll be particularly valuable. For publishers, though, the uncomfortable part is obvious - ordinary links are being pushed even farther down the page, while Google’s AI becomes increasingly difficult to walk around. (The Verge)
FAQ
Why is OpenAI ending its model deal with Cursor after the SpaceX acquisition?
OpenAI says it plans to wind down model access for Cursor because it can no longer be confident that SpaceX will use its technology within OpenAI’s terms. The company also says future OpenAI models will not be provided to Cursor. For developers, this could alter a coding workflow that commonly paired Cursor with OpenAI models.
What is Tencent Hy4 and what is it designed to do?
Tencent Hy4 is an open-source AI model built for software engineering, research and financial analysis. The preview uses a mixture-of-experts architecture with 770 billion total parameters while activating about 49 billion for each text request. Tencent plans to integrate Hy4 into products including CodeBuddy and WorkBuddy, although it says the preview can sometimes overthink and over-verify complex questions.
Can AI researchers automatically improve model alignment?
Anthropic’s research suggests automated AI researchers can identify and test techniques for reducing several types of alignment failure. Claude was used to search existing research, propose training approaches, run experiments and evaluate improvements across areas such as deception, jailbreaks and sycophancy. Anthropic also found that some successful techniques transferred to larger models, although the study covered only a limited range of practical alignment conditions.
Why did a court rule that Anthropic was illegally blacklisted?
A federal judge ruled that the Pentagon’s designation of Anthropic as a supply-chain risk was unconstitutional, arbitrary and capricious. The dispute followed Anthropic’s refusal to remove restrictions involving mass surveillance of Americans and lethal autonomous weapons without human oversight. The ruling could become an important reference point in future disputes over whether governments can penalize AI companies for maintaining their own safety restrictions.
Why is a16z investing $1.1 billion in AI infrastructure?
Andreessen Horowitz created its $1.1 billion Machine Age Fund to invest in the physical systems required to scale artificial intelligence. Its focus includes chips, memory, networking, storage, data centres, robotics and home AI hardware. The fund reflects a broader AI news theme: future progress may depend as much on power, cooling, bandwidth and computing infrastructure as on improvements to software models.
How are Google AI Overviews changing traditional search results?
Google is automatically expanding AI Overviews for some searches, placing a more complete AI-generated response and an “Ask anything” interface before traditional web results. That means ordinary links can appear farther down the page than before. For publishers and SEO teams following AI news, the change raises practical questions about search visibility, referral traffic and how users discover original sources.