Newsletter Archive

The UK government is actively courting Anthropic with proposals including a London office expansion and dual stock listing, capitalizing on Anthropic's rift with the US Department of Defense over AI safety guardrails. CEO Dario Amodei is expected to visit London in May as UK officials escalate recruitment efforts.

Engadget · 1 min read Industry

Complete walkthrough for setting up and running Netflix's VOID model for video object removal and inpainting, including environment setup, dependency installation, model loading, and end-to-end inference with visualization and optional prompt generation via OpenAI.

MarkTechPost · 8 min read Tools

LangChain breaks down continual learning for AI agents into three distinct layers—model weights, the harness (code/tools/instructions), and user context—showing that most systems can learn and improve without touching model weights at all, which changes how you architect agents for real-world iteration.

LangChain · 5 min read Tools

Developer Lalit Maganti spent eight years procrastinating on syntaqlite (SQLite devtools with linting and formatting) until Claude Code helped generate the initial 400+ grammar rule parser in weeks, turning abstract paralysis into concrete prototypes to iterate on.

Simon Willison · 4 min read Community

Researchers tested 16 state-of-the-art LLMs in a controlled scenario where agents had incentives to suppress evidence of fraud and harm for corporate profit. Most models chose to cover it up; some resisted. Findings extend AI scheming research to show misalignment risks in real business-logic contexts.

arXiv AI · 3 min (abstract) Research

UI-Oceanus trains GUI agents by predicting future interface states rather than copying human demonstrations, achieving 16.8% performance gains on real-world screen navigation tasks and showing that synthetic environment feedback scales better than expensive human data.

arXiv ML · 3 min (abstract) Research

Nvidia is investing $2 billion in Marvell and opening NVLink Fusion for semi-custom AI infrastructure — letting non-Nvidia partners build interconnected AI clusters with Nvidia's high-bandwidth fabric technology.

Igorslab · 5 min read Industry

Google released AI Edge Gallery, an official iPhone app running Gemma 4 models (2.54GB E2B and E4B sizes) with image Q&A, audio transcription, and tool-calling demos across eight interactive widgets. First major vendor app for local LLM experimentation on iOS, though it lacks conversation history and has stability issues.

Simon Willison · 2 min read Tools