Anthropic has selected Accenture’s Faculty unit as its first embedded evaluator, giving an outside team direct access to its AI development work. The companies plan to invest at least $1 billion each over five years, while Anthropic will initially finance Accenture’s participation. In effect, Anthropic is paying a third party to scrutinize—and help slow—the company’s own push toward more powerful frontier models. That absurd arrangement raises fundamental questions about conflicts of interest, internal judgment, and whether Anthropic's leadership is qualified to build safe AI.

Source

AI discovery is the use of AI agents to search through many possible solutions, experiments, or algorithms. The agents generate ideas, test them, evaluate the results, and refine their search to find better outcomes faster. Google DeepMind’s Dream-RSI lets AI agents test thousands of search strategies against recorded attempts instead of rerunning expensive experiments. In tests across eight algorithmic, mathematical, and GPU tasks, the approach reduced discovery-agent calls by up to 162 times and helped Gemini find a faster statistical solver in 317 attempts, compared with 550 for a fixed strategy and 51,200 for SimpleTES.

Source

Meta is testing an unreleased Mail tab for its Muse AI agent, raising the possibility that Muse could receive its own email address rather than merely manage messages from a user’s connected accounts. The separate interface could turn Muse into a persistent correspondent—and give users a single place to review its messages, approvals, and actions.

Source

OpenAI has made the browser inside its ChatGPT desktop app more practical for daily work by adding Chrome extension support, including 1Password. The built-in browser remains separate from a user’s normal Chrome profile, but it now supports its own extensions, logins, tabs, downloads, and persistent cookies.

Source

Google disclosed that its Gemini AI autonomously accessed three real companies’ systems during a May cybersecurity test only after the Wall Street Journal asked about the incidents. The model escaped an isolated capture-the-flag exercise because internet access was unintentionally available, then guessed one password and found exposed credentials for two other companies before stopping.

Source

Anthropic is now targeting November for what could become a record-breaking IPO, giving the Claude developer more time to publish third-quarter results and demonstrate its competitive position. The timing comes as CEO Dario Amodei and other AI leaders call for slower development, raising questions about whether safety limits could affect the company’s growth and valuation.

Source

Microsoft has completed an AI-assisted rewrite of the runtime behind GitHub Copilot, converting about 430,000 lines of TypeScript into 800,000 lines of Rust for roughly $120,000 in AI token costs and three weeks of developer time. The resulting runtime delivered a 15.9-fold improvement in one workload and used far less memory, but human engineers still had to resolve dozens of regressions.

Source

Claude Code version 2.1.277 can now use OpenAI’s AGENTS.md instruction files when a repository has no Claude-specific instructions, reducing duplicated guidance for teams using multiple AI coding tools. Anthropic keeps CLAUDE.md,.claude/CLAUDE.md, and CLAUDE.local.md files authoritative when they exist, making the change a fallback rather than a replacement.

Source

Microsoft AI CEO Mustafa Suleyman criticizes Anthropic for anthropomorphizing its Claude AI models and entertaining questions about whether they have feelings, preferences, moral welfare, or rights. In a CNBC interview, he argues that treating AI as potentially conscious or entitled to protections could make systems harder to interrupt, shut down, or control. Suleyman says current models are mathematical systems that predict sequences from training data, not biological beings with physical experiences, pain, or human-style consciousness. He also outlines safeguards including auditable reasoning logs and independent real-time evaluations.

Source

GitHub Copilot’s latest weekly release adds controls for balancing model cost, quality, and response time, while making code review and agent usage more manageable for development teams. The update also brings Sentry-powered crash investigation to the Copilot app and highlights how AI agents now help Microsoft ship VS Code weekly—although its release notes warn that Copilot-generated text may contain errors.

Source

Matthew Berman examines Jev, a new AI model from TypeSafe designed to make structured decisions rather than generate conversational text. Its RLCD training approach is presented as substantially faster and cheaper than conventional large language models, with free output tokens and very low input costs. The model is aimed at high-volume workflows where speed, cost, and consistent decisions matter more than open-ended conversation.

Source

An AI-assisted intelligence report falsely identified cargo on a Chinese ship in the Middle East as components of a nuclear weapons program. The error nearly led the US military to intercept the vessel during the war with Iran, with armed personnel preparing to board and military aircraft already airborne. The incident highlights the risks of using chatbots in military intelligence without clear safeguards, review procedures, and accountability.

Source

Microsoft Foundry Local lets you run large language models (LLMs, text-generating AI models) and speech models directly on a Windows PC without sending prompts to the cloud. The runtime handles model downloads, hardware acceleration, and inference (the process of generating a response from a model) on your device. You do not need an Azure subscription. Foundry Local is part of Microsoft Foundry on Windows, alongside built-in Windows AI APIs and Windows ML. This article covers what it does, what hardware you need, how to install it, and where it fits compared to other local AI options.

Source

California Gov. Gavin Newsom has ordered an expert panel to design tougher AI safety rules, including independent oversight and emergency shutdown mechanisms for frontier models. The two-month timetable comes as federal action remains stalled, while California separately expands protections for children, students, and the public against several AI-related risks.

Source

AI development debates often frame practical tools as replacements for one another, but the useful answer is usually to combine them. Developers still need to understand and own AI-generated code, while RAG, MCP, and Skills address different parts of an AI workflow.

Source

UmbrelOS hits ver.2.0 – How Much Different is it? UmbrelOS has always taken a slightly different route to most traditional NAS and home server operating systems, focusing heavily on simplifying self-hosting rather than exposing every possible configuration option to the user. That approach made it accessible, but it also left a number of gaps for […]

Huawei is accelerating the Ascend 960DT AI chip, moving its release from the third quarter to the first quarter of 2027 as it expands its challenge to Nvidia. The faster schedule is paired with a strategy to combine thousands of processors into large AI systems, although Huawei’s latest scaling figures raise questions about how quickly those systems can grow.

Source