Newsletter Archive

A Baltimore teen was tackled by police after an AI surveillance camera misidentified a Doritos bag as a gun; a Tennessee grandmother spent five months in jail due to faulty facial recognition. These cases expose how law enforcement treats probabilistic AI outputs as certainties, creating serious harms that demand scrutiny of AI deployment in criminal justice.

The Conversation · 7 min read Policy

OpenAI announced a $4 billion company focused on accelerating enterprise AI adoption, representing a significant expansion beyond consumer API access into dedicated infrastructure and support for large organizations.

PYMNTS · 3 min read Industry

You can now execute LLM prompts directly in shell script shebangs, turning plain English files into executable programs with tool calls and function definitions—e.g., `#!/usr/bin/env -S llm -f Generate an SVG` or YAML-based calculator templates that chain LLM operations.

Simon Willison · 2 min read Tools

Microsoft is targeting a $92 billion return from its early OpenAI investments, exposing the scale of financial commitments underwriting the AI boom and signaling massive valuations ahead.

Bloomberg Tech · 1 min read Market

OpenAI launched Daybreak, an automated vulnerability detection and patching initiative powered by the Codex Security agent, which analyzes code to identify and patch security risks before attackers exploit them. The move directly competes with Anthropic's recently announced Claude Mythos, a security-focused model Anthropic restricted to private deployment.

The Verge AI · 2 min read Industry

AWS and Hugging Face outline how modern foundation model training and inference now requires coordinated infrastructure across three scaling regimes: pre-training, post-training (SFT/RL), and test-time compute. The piece breaks down distributed compute, networking, and storage requirements for efficiently managing the full model lifecycle at scale.

Hugging Face · 20 min read Tools

Thinking Machines released TML-Interaction-Small, a 276B-parameter MoE model with 12B active parameters that advances realtime voice AI with 200ms interaction latency, encoder-free early fusion, and 80–96% cache hit rates—positioning it as a practical competitor to GPT-4o's 'her' demo while multiple quantization studies question recent TurboQuant hype.

Latent Space · 10 min read Research

Paris startup White Circle raised $11M to build real-time controls for deployed AI models, backed by executives from OpenAI, Anthropic, DeepMind, Mistral, and Hugging Face. Founded by Denis Shilov (who discovered a universal jailbreak prompt that went viral), the company targets the gap between model safety in testing and actual workplace behavior when users interact with deployed systems.

Fortune · 5 min read Industry