// AI Briefing
March 19, 2026
Mistral just made GitHub Copilot sweat, the EU started handing out real fines, and a Berkeley paper might let you run 70B models on your laptop.
Mistral unveiled Codex 2, a 32B parameter code generation model released under the Apache 2.0 license. The model matches GPT-4o on HumanEval benchmarks while running on a single A100 GPU. This matters because it gives companies a production-grade code assistant they can self-host without vendor lock-in, directly challenging GitHub Copilot's dominance.
Google revealed that Project Astra, its multimodal AI assistant, will be deeply integrated into Android Studio by Q3 2026. Developers will be able to point their camera at a whiteboard sketch and have Astra generate the corresponding Jetpack Compose UI code. The integration signals Google's strategy to make AI-first development the default workflow, not an add-on.
The first enforcement wave of the EU AI Act took effect today, requiring companies deploying high-risk AI systems in healthcare, law enforcement, and employment to register their systems and complete conformity assessments. Non-compliance carries fines up to 3% of global revenue. This marks the world's first binding AI regulation with real enforcement teeth.
A team from UC Berkeley published a paper demonstrating a new speculative decoding technique that achieves 10x faster inference on 70B+ parameter models without quality degradation. The technique uses a cascade of draft models with adaptive verification, making real-time LLM responses feasible on consumer hardware. The research is already being integrated into vLLM and llama.cpp.
Anthropic announced a partnership with Thomson Reuters to build AI tools for legal professionals using Claude. The collaboration will embed Claude into Westlaw and Practical Law, enabling lawyers to research case law, draft documents, and analyze contracts with AI assistance. This is notable as one of the largest enterprise AI deployments in a regulated industry.
Test Your Understanding
Quiz
1 / 7
Why is Mistral's Codex 2 release significant for enterprises?
Summary
A federal judge just told the Pentagon it cannot punish an AI lab for refusing to hand over its models, which is the first time a court has drawn a hard line around what safety policies cost you in Washington. Elsewhere the theme is AI learning by watching rather than being told: a robot foundation model that copies a ten-minute task from one video, and Claude grinding through algebra that stumped a lab for eighteen months. Plus Meta patching the dumbest privacy hole in its glasses, and Pew's number on how many Americans now ask a chatbot about their symptoms.
Top Stories
A Judge Just Ruled the Pentagon Illegally Punished Anthropic for Saying No to Mass Surveillance
Skild's Robot Model Watches One Video of You Making Pancakes and Then Makes Pancakes
Claude Did Eighteen Months of Algebra in Five Weeks, and Got Some of It Wrong Along the Way
Meta Is Finally Killing the Trick Where You Start Recording Then Cover the Light
A Third of Americans Now Ask a Chatbot About Their Symptoms, and Most Won't Share Their Health Data With It