Anthropic introduced Project Glasswing, a cybersecurity coalition with major tech partners including AWS, Apple, Google, Microsoft, and Nvidia, centered around its new unreleased frontier AI, Claude Mythos Preview. Mythos, which demonstrated superior capabilities by flagging thousands of security flaws, will not be publicly released but will be deployed to partners for defensive security applications.
Sources: Hacker News, Keep Bookmarks, Latent Space, Simon Willison, The Neuron, The Rundown AI
AI company Anthropic has finalized a multi-gigawatt compute deal with Google and Broadcom, securing 3.5GW of next-generation TPU capacity by 2027, primarily in the US. This significant partnership aims to support Anthropic's rapid growth and the development of its expanding AI models, following a reported tripling of its run-rate revenue.
OpenAI has published a 13-page policy paper, 'Industrial Policy for the Intelligence Age,' suggesting ambitious reforms like shifting tax bases, piloting a 32-hour workweek, and building new AI infrastructure and oversight to prepare for the impact of superintelligence. The paper signals OpenAI's belief that AI's rapid advancement will necessitate a fundamental redesign of social and economic structures.
Google has released AI Edge Eloquent, a free iOS and Mac dictation application that processes speech into polished text entirely offline. Powered by the new Gemma 4 AI model, this development emphasizes user privacy through on-device processing. The update also includes official support for Gemma 4 in the Google AI Edge Gallery, enabling advanced offline AI capabilities directly on mobile devices.
Netflix has open-sourced VOID, a physics-aware AI tool designed to enhance video editing workflows. This model allows for advanced manipulation of video content, including the realistic removal of objects and background filling, making powerful AI more accessible to developers for post-production.
Researchers have developed MegaTrain, a memory-centric system that allows full precision training of large language models with over 100 billion parameters on a single GPU. This innovation significantly improves efficiency and throughput compared to existing methods like DeepSpeed ZeRO-3, marking a notable advancement in AI model training capabilities.
Sam Altman, CEO of OpenAI, has published a 13-page policy document outlining ideas like taxing AI-driven profits, creating a national wealth fund, and implementing a 4-day workweek to prepare society for the advent of superintelligence. This proposal highlights the urgency Altman sees in addressing the economic and societal impacts of advanced AI, suggesting a fundamental shift in how society operates.
The Wall Street Journal reports that leading AI companies OpenAI and Anthropic are accelerating plans for initial public offerings, even as both firms continue to incur substantial costs for AI model training and experience heavy cash burn. This move signals a push for public market capital amidst the intense financial demands of frontier AI development.
IBM Research has introduced ALTK-Evolve, a novel AI agent learning system designed to enable agents to generalize principles from experience, significantly boosting reliability in complex, multi-step tasks. ALTK-Evolve offers versatile integration options, from no-code plugins for popular LLM clients like Claude Code to pro-code integration with CUGA, facilitating easy testing and continuous improvement. This launch is supported by IBM and UC Berkeley's research into enterprise AI agent failures and the new AssetOpsBench benchmark, aiming to bridge the gap between theoretical and industrial AI applications.
OpenAI's Frontier team successfully built and shipped an internal beta product, Symphony, with over a million lines of code, entirely generated by Codex agents without any human-written or reviewed code, demonstrating a new paradigm in AI-native software development.
Researchers have developed 'Just Pass Twice' (JPT), a novel method enabling causal large language models to perform discriminative token classification for zero-shot Named Entity Recognition (NER) with full bidirectional context. This approach achieves state-of-the-art results, surpassing previous methods by +7.9 F1 and operating over 20 times faster than comparable generative techniques.
Amazon SageMaker AI has introduced a new serverless model customization feature, leveraging Reinforcement Learning with Verifiable Rewards (RLVR), to significantly improve the reliability and accuracy of AI agents in calling tools, thereby addressing issues like hallucination and incorrect parameter usage in production. This feature allows developers to fine-tune models like Qwen 2.5 7B Instruct without managing infrastructure.
Nia Docs has introduced a new tool that allows AI agents to interact with web documentation sites as if they were a Unix filesystem, enabling them to `grep`, `cat`, and `cd` through content. This innovation aims to combat code hallucinations by providing agents with up-to-date information directly, bypassing stale training data and leveraging agents' inherent understanding of filesystem commands.
Chinese AI lab Zhipu AI launched GLM-5.1, a new open-source coding model that achieved a leading score on SWE-Bench Pro, outperforming GPT-5.4 and Opus 4.6. The model also exhibits strong long-horizon agentic capabilities, demonstrated by building a functional Linux desktop as a web app over eight hours without human intervention.
OpenAI, Google, and Anthropic are collaborating through the "Frontier Model Forum" non-profit to identify and limit Chinese rivals from distilling their advanced AI systems. This cooperation aims to protect their proprietary technologies and maintain a competitive edge.