EU just blocked U.S. AI tools over security fears

The EU Parliament just threw down the gauntlet on U.S. tech giants, blocking AI tools over cybersecurity and privacy concerns (yikes), while Meta's doubling down on the chip arms race with a massive Nvidia expansion that includes Grace CPUs. Meanwhile, Anthropic quietly dropped Claude Sonnet 4.6, which somehow rivals Opus at a lower price point with wild improvements to computer use, and a new open-web simulator called WebWorld is training agents on 1M+ trajectories to match frontier models. Oh, and there's a helpful guide reminding us that in the agentic era, harness and app matter as much as the model itself. So here's the real question: are you ready to switch from picking models to picking agents?

Image via Philipp Dubach

The EU Parliament banned AI tools from lawmakers' devices over fears that sensitive data could be exposed to U.S. authorities and used for AI model training, highlighting the fundamental tension between data protection regulations and AI development in a geopolitically fragmented tech landscape.

regulationprivacydata-protectioncybersecurity

Open-Web Simulator for Agent Training

arXiv

WebWorld enables efficient web agent training through large-scale simulation, allowing smaller models to achieve frontier performance without expensive real-world interactions. This addresses a critical bottleneck in scaling AI agents across web automation, code, and GUI tasks.

agentsworld-modelsweb-automationsynthetic-data

Meta Expands Nvidia Deal to Use Millions of AI Chips in Data Center Build-out, Including Standalone CPUs

CNBC

Meta's major expanded Nvidia deal (tens of billions in value) locks in millions of AI chips including first-ever standalone Grace CPUs, reflecting both companies' commitment to comprehensive AI infrastructure dominance as Meta pursues its $600 billion spending plan.

nvidiametachipsdata-centers

Claude Sonnet 4.6

Anthropic

Claude Sonnet 4.6 delivers frontier-level reasoning and capabilities at mid-tier pricing, with users frequently preferring it over the more expensive Opus 4.5 model, marking a significant shift in the performance-to-cost ratio of AI development.

llmanthropicclaudecomputer-use

A Guide to Which AI to Use in the Agentic Era

One Useful Thing

AI's evolution from chatbots to autonomous agents fundamentally changes how users should evaluate and deploy AI tools, requiring attention to models, apps, and harnesses rather than just conversation quality. Learning to manage AI agents doing real work—not just talking about it—is now the critical skill for leveraging modern AI effectively.

ai-agentsllmopenaianthropic