Elizabeth Barnes, CEO of METR, announced that AI experts lack sufficient control over frontier AI risks, based on a pilot Frontier Risk Report that evaluated AI agents at Anthropic, Google, Meta, and OpenAI. This report is significant as METR had direct access to these major AI labs, highlighting critical concerns about the safety and governance of advanced AI development.
Sources: Digg AI
Anthropic's initial Project Glasswing report reveals its unreleased Mythos Preview models identified over 10,000 vulnerabilities and blocked a $1.5 million fraud attempt, showcasing performance superior to GPT-5.5 on exploit benchmarks. Anthropic plans to expand access to these models to US and allied government partners first, emphasizing their role in critical security applications.
Sources: Digg AI
Key figures like OpenAI's Greg Brockman, along with companies like AI21 and DeepSeek, are increasingly focusing on building AI agents and comprehensive systems rather than just standalone large language models. This trend signifies a shift in product strategy, where model quality alone is no longer sufficient, emphasizing integrated solutions that include harnesses, workflows, UI, and memory.
Sources: Latent Space
The Trump administration has ended a policy that allowed O-1 visa holders, including advanced AI researchers at companies like OpenAI, to remain in the US during green card applications. This change will require applicants to pursue consular processing abroad, potentially leading to years-long waits and impacting the retention of top AI talent in the US.
Sources: Digg AI
An ArXiv paper titled 'From Entropy to Epiplexity' has initiated a discussion on X about the limitations of information theory in providing explanatory power and testable predictions for AI systems, with replies highlighting practical ML barriers like noise and unreliable estimators. This discussion is significant for understanding theoretical challenges in AI development.
Sources: Digg AI
NVIDIA has launched Nemotron-Labs Diffusion, a new family of diffusion language models (DLMs) that generate text by generating multiple tokens in parallel and iteratively refining them. This approach offers significant runtime performance benefits, achieving up to 6.4 times higher tokens per forward pass compared to autoregressive models, and allows for revision of generated tokens.
Sources: Hugging Face Blog
Project Glasswing, a collaborative effort with 50 partners, has used Claude Mythos Preview to discover more than 10,000 high- or critical-severity software vulnerabilities, significantly accelerating bug finding in critical software and shifting the cybersecurity bottleneck from discovery to patching.
Sources: Anthropic News
Google's Josh Woodward, Koray Kavukcuoglu, Liz Reid, and Jeff Dean explored how proactive AI agents are fundamentally changing and enhancing productivity across various domains.
Sources: Google AI Blog
Major AI providers including Anthropic, OpenAI, and Google have significantly increased service prices, leading to a 20-37% jump in American AI software costs. This surge has caused enterprises like Microsoft to cancel internal licenses and Uber to deplete its 2026 AI budget in four months, prompting a widespread shift towards usage-based billing models across the industry.
Sources: Keep Bookmarks
MIT CSAIL PhD student Ryan Bahlous-Boldi introduced Vector Policy Optimization (VPO), a new reinforcement learning method that maximizes vector-valued rewards to preserve distinct objectives during LLM post-training, demonstrating improved pass@k scores on LiveCodeBench over GRPO baselines.
Sources: Digg AI
Hyperagent has launched a cloud platform designed for autonomous AI agents that can accumulate skills over time, offering $10 million in credits to 500 founders. This platform is already utilized by 80% of Fortune 100 companies, indicating significant enterprise adoption.
Sources: Digg AI
Thomas Dimson, who played a significant role in the development of OpenAI's Sora video generation model and previously authored Instagram's core recommendation algorithms, has announced his departure from OpenAI after 1,000 days. This executive change is notable given his involvement with a prominent AI product.
Sources: Digg AI
Meta has shifted more than 1,000 software engineers from its advertising teams to data labeling roles for AI model training, signaling a significant internal strategic pivot towards strengthening its AI capabilities.
Sources: Digg AI
Polsia, a company operating businesses autonomously through AI, secured $30 million in funding at a $250 million valuation, demonstrating significant investor confidence in its zero-employee, AI-first business model and its rapid growth towards a $10 million annual run rate.
Sources: Digg AI
An interactive tool and Substack analysis found that a Granta prize-winning story contained numerous n-grams, rare phrases, and unique segments traceable to online fanfiction, suggesting it was assembled from memorized online passages. This raises questions about originality and potential AI influence or plagiarism in creative writing.
Sources: Digg AI