Newsletter Archive

Google released Gemma 4, a family of four open-weight models (2B to 31B parameters) claiming unprecedented intelligence-per-parameter efficiency. The 31B and 26B variants rank 3rd and 6th on Arena AI's leaderboard while outperforming models 20× their size; all support vision/audio processing for edge and server deployment.

Engadget · 2 min read Industry

Anthropic launched Claude Code, an agentic system that autonomously handles multi-file coding tasks, debugging, and project work. Developers can now hand off complex engineering problems to Claude rather than iterating step-by-step, marking a significant shift toward AI agents that operate independently on real codebases.

Anthropic · 4 min read Industry

Google DeepMind's Gemma 4 multimodal models are now available on Hugging Face with Apache 2 licenses, supporting audio alongside text/vision and deploying on everything from cloud to edge devices. The release includes implementations across transformers, llama.cpp, MLX, WebGPU, and Rust with strong arena benchmarks.

Hugging Face · 22 min read Tools

Kilo launched KiloClaw, an enterprise governance platform designed to detect and manage autonomous agents deployed by employees outside official procurement ("shadow AI" or BYOAI). Employees bypassing IT are routing agents through personal infrastructure and API keys to automate workflows, exposing corporate data in Slack, Jira, and code repos — KiloClaw provides centralized visibility and control without blocking productivity.

AI News · 4 min read Tools

Moonlake AI is building interactive, multiplayer world models bootstrapped from game engines—enabling indefinite simulation lifetimes and physics-accurate interactions, contrasting sharply with single-player limitations in competitors like Google's Genie 3. Chris Manning and Fan-Yun Sun explain how their approach scales to long-horizon planning and multi-agent environments.

Latent Space · 58 min read Research

LangChain's evals show open-weight models like GLM-5 and MiniMax M2.7 now match closed frontier models on core agentic tasks (file ops, tool use, instruction following) while costing 8-10x less and running faster—making them viable production alternatives for real-world workflows.

LangChain · 7 min read Tools

Meta's Ranking Engineer Agent now includes KernelEvolve, an autonomous system that generates and optimizes GPU/CPU kernels for different hardware without manual engineering. The agent automatically tunes kernels for Meta's mix of NVIDIA, AMD, MTIA custom chips, and CPUs—addressing the exponential scaling problem of hand-tuning across model variants and hardware generations.

Meta AI · 15 min read Tools

Simon Willison argues that November 2024 marked an inflection point where GPT-5.1 and Claude Opus 4.5 crossed a threshold: code now works correctly most of the time rather than requiring constant debugging. He covers implications for software engineering roles, dark factories, testing bottlenecks, and why AI coding agents are suddenly viable for real security work.

Simon Willison · 16 min read Community