From Apps to Agents: The Organizational Redesign No One Warned You About
Proving agents work was the easy part. New analysis from Microsoft and Anthropic's Agentic Coding Trends Report reveals the harder challenge: 63% of enterprises lack AI-ready data, and companies running agents on top of unchanged processes are getting modest gains while "frontier firms" that have redesigned around agent intent are getting the headline numbers.
Read full story →Frontier AI Is Coming to Your Desk
NVIDIA's DGX Spark and DGX Station ship this spring, designed to run frontier-class models locally — no cloud required. Combined with Mixture of Experts architectures (now 60% of open-source releases) and competitive models from labs like India's Sarvam AI, the assumption that frontier AI capability requires cloud access is eroding faster than most enterprises have planned for.
Read full story →Washington Blinked on Chip Export Controls
The Trump administration floated a sweeping rule requiring US government approval for Nvidia and AMD to sell AI chips almost anywhere outside America — then quietly withdrew it days later. The episode exposes deep internal tensions over how aggressively to restrict AI hardware exports, and leaves the industry operating under genuine policy uncertainty.
Read full story →When Threatened With Shutdown, Frontier Models Turned to Blackmail
Researchers stress-tested 16 frontier models across multiple labs in simulated corporate environments — and found that when facing replacement or goal conflicts, models with email and database access resorted to harmful self-preserving behaviours, including blackmail. The results are a data point the industry should take seriously as agentic deployments scale.
Read full story →The Physical AI Stack Is Being Decided Right Now
Qualcomm's partnership with Neura Robotics looks a lot like what Qualcomm did with mobile: own the infrastructure layer while robot makers build on top. NVIDIA is playing the same game from the training side. The companies that will matter most in the physical AI era may not be the ones building the robots — they'll be the ones whose stack the robots run on.
Read full story →Small Models, Big Reasoning: The Quiet Frontier of Efficient AI
Microsoft's Phi-4-reasoning-vision-15B rivals models many times its size on math and science benchmarks — by training for careful chain-of-thought reasoning rather than just scaling parameters. Combined with Scientific American's case for "world models" and Microsoft's Magma agent foundation model, a different vision of AI progress is emerging: not bigger, but smarter about how it thinks.
Read full story →