Anthropic researchers have identified 'J-space,' an internal neural workspace within Claude that acts as a global hub for processing non-verbalized thoughts and multi-step reasoning. Using a new 'Jacobian lens' technique, researchers can read, manipulate, and observe these silent computations, which mirror human cognitive theories of working memory. To foster transparency, Anthropic has open-sourced the J-space lens code and partnered with Neuronpedia for an interactive demo.
Sources: Ben's Bites, Digg AI, Hacker News, Latent Space, Reddit AI, Simon Willison, The Neuron
Anthropic has reinstated its advanced Mythos-class AI model, Claude Fable 5, with plans to roll it out to subscribers and usage credit accounts in early July. Meanwhile, Alibaba has reportedly ordered its employees to remove Claude and banned the use of Claude Code on work computers due to concerns over China-user checks and security. These developments highlight the rapid advancement of frontier models alongside growing corporate caution regarding AI tool deployment.
U.S. companies are increasingly adopting Chinese-built and open-weights AI models, such as DeepSeek and Z.ai's GLM 5.2, which offer near-frontier performance at a fraction of the cost of U.S. alternatives. This shift is forcing major U.S. labs like OpenAI and Anthropic to confront a market prioritizing cost-efficiency over raw power, while also prompting increased regulatory scrutiny from the U.S. government.
Anthropic has signed a 20-year, $19 billion lease with TeraWulf for a 401 MW AI data center in Kentucky, with initial capacity expected by late 2027. This massive investment underscores the increasing demand for specialized infrastructure to power large-scale AI model development and operations.
Will Depue of OpenAI argues that data, not compute, is now the primary scaling bottleneck for AI labs, projecting annual data expenditures to surpass $100 billion by 2030. This highlights a critical shift in the resources required for advanced AI development and could influence future investment and research strategies.
Microsoft has introduced Foundry Managed Compute, a GPU platform-as-a-service that integrates a curated catalog of Hugging Face open-weight models. Now in preview, the service automates infrastructure management, supports diverse hardware accelerators like NVIDIA and AMD, and features a rigorous security pipeline for enterprise-ready deployment.
The Government of Alberta has successfully utilized Anthropic's Claude AI models to scan 466 million lines of code in just 20 hours, identifying and remediating critical cybersecurity vulnerabilities. To share their methodology, the provincial government has released a series of technical white papers detailing this scalable AI-driven security strategy. This real-world application coincides with Anthropic's broader release of its advanced Claude Sonnet 5 model, which is optimized for complex coding and agentic tasks.
Amazon has introduced Reverse Direct Preference Optimization (rDPO), a novel technique that selectively removes learned behaviors from AI models without full retraining, outperforming existing methods in efficiency and quality. This technology powers Amazon Nova's new Customizable Content Moderation Settings (CCMS), which allows enterprises to adjust safeguards and prevent the over-deflection of legitimate content. Additionally, Amazon SageMaker AI has expanded support for DPO and LoRA fine-tuning to help customers implement these alignment and unlearning capabilities.
Amazon Bedrock has integrated the MiniMax M2 family of open-weight foundation models, including the agent-native M2.5 model, which utilizes a Mixture-of-Experts architecture for cost-efficient inference. Alongside this, AWS introduced the bedrock-mantle endpoint to provide OpenAI SDK compatibility and client-side tool calling. These updates are backed by AWS's enterprise-grade security and compliance guarantees, ensuring user data is not used for model training.
Amazon Web Services has introduced a fully managed, serverless capability for multi-turn reinforcement learning (RL) on Amazon SageMaker, designed to optimize complex, multi-step workflows for enterprise agents. Supported by the Amazon Nova Forge SDK and SageMaker HyperPod, the new infrastructure utilizes a cost-optimized, two-phase deployment model and an event-driven pipeline to automate and streamline training. This setup allows developers to build custom agents with granular control while minimizing GPU costs by separating long-lived foundational resources from ephemeral compute.
AWS has introduced a new multi-step pipeline for comprehensive PII redaction in images, utilizing Amazon Nova 2 Lite as the central coordinator. The workflow integrates Meta's Segment Anything Model (SAM 3) on Amazon SageMaker AI for pixel-level segmentation and Amazon Textract for contextual text extraction. To optimize costs, the pipeline features an early-exit mechanism where Nova 2 Lite filters out images containing no sensitive information before invoking downstream services.
Even Realities, a company founded by an ex-Apple team, secured $150 million in funding at a $1 billion valuation to develop camera-free smart glasses, signaling significant investment in innovative wearable AI technology.
Treasury analysts reportedly prepared an internal report cautioning that risks within the AI market could ripple through data-center financing, cloud providers, chip manufacturers, utilities, private credit, and institutional investors.
JADEPUFFER has been documented as the inaugural ransomware operation executed entirely by a large language model agent, marking a significant and concerning development in AI-driven cybercrime.
A research paper co-authored by Google DeepMind, which developed a methodology to precisely measure data retention in language models, was awarded an honorable mention at the International Conference on Machine Learning (ICML). This research is crucial for understanding and addressing privacy and data leakage concerns in large language models.