Newsletter Archive

LangChain shares its framework for building targeted evaluations that shape agent behavior: source production-relevant data, design metrics that measure specific behaviors you care about (not just benchmark scores), and run focused experiments over time. The key insight—more evals ≠ better agents—challenges common testing practices.

LangChain · 9 min read Tools

Sam Rose's interactive essay breaks down LLM quantization from first principles, revealing why outlier "super weights" can tank model quality if removed and demonstrating empirically that 16-bit to 8-bit quantization incurs almost no accuracy penalty using perplexity and GPQA benchmarks on Qwen 3.5 9B.

Simon Willison · 2 min read Tools

Mistral AI launched Voxtral TTS, a 4-billion-parameter text-to-speech model supporting 9 languages with emotional expression, low latency, and voice adaptation. The lightweight model is designed for production voice agent workflows and is now available in Mistral Studio.

Mistral AI · 6 min read Industry

Meta is constructing a colossal Hyperion AI data center on 2,250 acres of rural Louisiana farmland, reshaping a community with thousands of workers, round-the-clock construction, and truck convoys. The project offers a rare on-the-ground look at the physical, economic, and human toll of the infrastructure race underpinning the AI boom.

Fortune · 7 min read Hardware

MIT's VibeGen AI model generates novel proteins by specifying desired motion patterns rather than just 3D shapes, enabling design of proteins with custom mechanical dynamics for therapeutics and materials.

MIT AI News · 8 min read Research

Sam Altman told OpenAI staff he was trying to "save" competitor Anthropic during Pentagon contract negotiations, while privately criticizing Anthropic's CEO and ultimately securing the defense deal Anthropic lost. Internal Slack messages expose a calculated play that positioned OpenAI as a peacemaker while it capitalized on Anthropic's collapse.

Axios · 3 min read Industry

A federal judge blocked Trump's Pentagon ban on Anthropic as a supply chain risk, granting a preliminary injunction after the company argued the designation caused irreparable business harm. The ruling provides temporary relief while separate cases proceed on First Amendment and procurement law grounds.

Axios · 2 min read Policy

A federal judge issued a temporary restraining order blocking the Trump administration's ban on Anthropic, one of the largest AI safety companies, signaling legal challenges to the administration's approach to AI governance.

Npr · 4 min read Policy