Newsletter Archive

The US Treasury published a Financial Services AI Risk Management Framework with a detailed guidebook developed by over 100 financial institutions and regulators. Unlike generic frameworks, it addresses sector-specific risks like algorithmic bias, LLM unpredictability, and cyber vulnerabilities that traditional tech governance doesn't cover.

AI News · 6 min read Policy

Encyclopedia Britannica and Merriam-Webster are suing OpenAI for allegedly using ~100,000 copyrighted articles to train its LLMs without permission, escalating the legal battle over training data rights.

TechCrunch · 4 min read Policy

Simon Willison published a three-hour NICAR 2026 workshop guide showing data journalists how to use Claude Code and OpenAI Codex to explore, clean, and visualize datasets. The workshop cost attendees just $23 in API tokens and included live demos of agents generating interactive Leaflet heat maps directly into static folders—proving that coding agents can handle end-to-end data pipelines from SQL queries to visualization.

Simon Willison · 2 min read Tools

AWS and NVIDIA announced a major partnership expansion at GTC 2026: AWS will deploy over 1 million GPUs (including Blackwell and Rubin chips) across regions starting 2026, add EC2 support for NVIDIA RTX PRO 4500 Blackwell, and optimize LLM inference with NVIDIA NIXL on AWS's Elastic Fabric Adapter—making large-scale production AI workloads more accessible to enterprises.

Amazon AWS ML · 6 min read Hardware

NVIDIA and Bolt announced a partnership at GTC 2026 to deploy autonomous robotaxis across Europe, with Bolt providing operational fleet data to train NVIDIA's Alpamayo model while integrating the full Drive Hyperion platform. The deal ties together NVIDIA Cosmos (data curation), Omniverse (digital twin simulation), and Alpamayo (AV-specific foundation model) to handle European road complexity.

Engadget · 2 min read Industry

Mistral AI released Mistral Small 4, a 119B-parameter MoE model combining instruction-following, reasoning, multimodal understanding, and coding in one deployment. With 128 experts (4 active per token), 6B active parameters, and 256k context, it eliminates model-switching overhead for chat, agentic tasks, and long-document analysis.

MarkTechPost · 4 min read Industry

Mistral shipped Mistral Small 4, a 119B parameter Mixture-of-Experts model (6B active) that combines reasoning, multimodal, and agentic coding into one model with tunable reasoning_effort levels. Also launched Leanstral, specialized for Lean 4 formally verifiable code generation.

Simon Willison · 2 min read Industry

NVIDIA Warp tutorial teaches Python developers to build high-performance GPU simulations and differentiable physics workflows, covering kernel implementation, particle dynamics, and vector operations across thousands of threads with CPU/CUDA fallback support.

MarkTechPost · 8 min read Tools