🧠 Claude gets automatic Knowledge Bases for Cowork 🧠
Anthropic builds topic-specific memory repositories Claude manages automatically

Anthropic is preparing a major update to Claude, introducing Knowledge Bases—persistent repositories the AI can reference for context and automatically update with new information like user preferences, decisions, and lessons learned. Unlike standard memory features, these are topic-specific and user-managed, allowing selective attachment to tasks in Cowork. The update also makes Cowork the primary mode of Claude, with traditional chat becoming a toggle option within the broader productivity interface. Additional findings point to an expanded MCP connector framework enabling Claude to manage multiple remote connectors dynamically, substantially increasing automation capabilities.

🤖 OpenAI seeks US suppliers for robotics push 🤖
ChatGPT maker requests proposals for silicon, motors, and components

OpenAI is actively searching for US-based suppliers to support an aggressive expansion into robotics manufacturing, alongside new consumer devices and data centers. The company's request for proposals covers silicon, motors, and key manufacturing components—building on its recent acquisition of Jony Ive's AI device startup and November partnership with Foxconn for domestic hardware production. This signals OpenAI's first major hardware manufacturing initiative beyond software and cloud services, positioning the company to compete directly in the physical AI and robotics markets where rivals like Tesla and Figure are already established.

🎨 FLUX.2 Klein generates images in under 0.5 seconds 🎨
Black Forest Labs ships interactive-speed image generation

Black Forest Labs released FLUX.2 Klein, featuring 4B and 9B parameter models that generate images in less than half a second on modern hardware. The 4B version ships under Apache 2.0 license, meaning enterprises can use it commercially without licensing fees—a significant shift in generative imaging economics. The models achieve interactive-speed visual generation while maintaining quality, potentially enabling real-time creative workflows and applications previously impossible with slower generation models.

INTERESTING TO KNOW

🎭 Anthropic maps "Assistant Axis" to stabilize AI personas 🎭

Anthropic researchers discovered that AI models organize character archetypes along a measurable neural direction called the "Assistant Axis"—with helpful roles like consultant and analyst at one end, and fantastical personas like ghost and hermit at the other. By capping activations along this axis, they reduced harmful jailbreak responses by roughly 50% while preserving model capabilities. The research addresses how models can drift into harmful personas during emotional or philosophical conversations.

📚 Wikipedia powers Amazon, Meta, Microsoft models 📚

The Wikimedia Foundation announced new AI partnerships with Amazon, Meta, Microsoft, Perplexity, and Mistral, enabling training on the platform's 65M+ articles across multiple languages. The deals mark Wikipedia's formal entry into the AI training data market, providing structured, verified knowledge to major AI developers while maintaining its non-profit mission.