Anthropic Calls for a Pause in AI Development, Warns of Self-Improvement Risk
On June 4, Anthropic issued a major statement via a WSJ report. The company is calling on major AI labs to consider slowing the pace of frontier AI development.
Behind the move is concern over "recursive self-improvement" — the stage at which AI systems can improve themselves without human intervention may arrive sooner than expected. In a blog post authored by Marina Favaro, head of Anthropic's internal research organization, and Jack Clark, head of policy, the company argues that model progress is accelerating rapidly and that it is desirable to have options for letting social structures and alignment research keep up with technological progress.
While noting that self-improvement has not yet happened and is not inevitable, the post warns that it "could arrive sooner than most institutions are prepared for." It proposes forging a global consensus, along with mechanisms to verify that competitors are actually adhering to any agreement.
Around the same time, Anthropic completed a funding round at a valuation of roughly $1 trillion and is making preparations to go public. The company's annualized revenue is projected to reach $50 billion by the end of this month. Rival OpenAI is also said to be preparing for an IPO.
First AI Agent "Poke" Approved for Apple Messages for Business
A startup called Poke has become the first AI agent approved on Apple's Messages for Business platform. Poke is a service that lets users access AI agents through simple text messages, whereas Messages for Business had previously assumed a human representative on the other end. This is the first time Apple has officially recognized an AI agent on its business messaging platform, marking a major milestone for future AI agent integrations across platforms.
Meta Targets Cost Savings with Tent-Style Data Center Construction
To cut data center construction costs, Meta is adopting an approach pioneered by Tesla: temporary data centers with tent-like structures. According to TechCrunch, Meta has begun building tent-style facilities as one way to rein in its enormous data center expenses. With AI compute demand surging, this method — which enables faster and cheaper construction than conventional large-scale data centers — offers a new perspective on infrastructure investment.
Stanford Releases 28-Trillion-Pixel Open Image Dataset "GPIC"
A research team at Stanford University has released GPIC (Giant Permissive Image Corpus), a massive image dataset of roughly 28 trillion pixels. It consists of 100M training, 200K validation, and 1M test images, all captioned by state-of-the-art vision-language models.
Its biggest feature is that every image is offered under permissive licenses that allow both research and commercial use. The dataset has been safety-filtered and deduplicated, and is centrally hosted on Hugging Face. A benchmark protocol for image generation models is also provided, aiming to support reproducible research in large-scale visual generative modeling.
Nvidia Acquires Kumo AI
Fortune reports that Nvidia has completed its acquisition of Kumo AI. While the purchase price has not been disclosed, Nvidia has been pursuing acquisition after acquisition to expand its AI infrastructure ecosystem. Kumo AI specializes in predictive AI built on graph neural networks and is seen as complementing Nvidia's data center and enterprise portfolio.
Science Journal Warns That the Window for Understanding AI Is Closing
An editorial titled "A narrowing window to understand AI" has been published in the journal Science. As AI systems grow more complex and black-boxed, the editorial argues that the "window" currently open for humans to understand and control AI's inner workings is gradually closing. Within the research community, calls are growing to accelerate investment in model interpretability research.
Anthropic Releases an Open-Source Vulnerability Discovery Framework
Anthropic has released an open-source framework for AI-powered vulnerability discovery, "defending-code-reference-harness," on GitHub. It has picked up 50 points on Hacker News and is drawing attention from the security community. As AI-driven security testing tools proliferate, it is notable that a major AI company is providing security research infrastructure of its own.