Open Source
Developer created KillSwitch, an esoteric programming language designed to exploit LLM weaknesses through hidden lies, counting requirements, and safety trigger phrases.
Simon Willison released llm-tools-datasette 0.1.1, fixing compatibility with LLM 0.33's httpx2 migration.
Simon Willison released llm-echo 0.5, a debug plugin for LLM that simulates token usage including tool calls and system prompts.
NVIDIA and AI4Bharat release 2.8 trillion token Indic language dataset and open-source GPU-accelerated data generation pipeline for 22 Indian languages.
Anthropic's Scion version control system appears heavily inspired by Meta's Mononoke, raising questions about proper attribution in AI-assisted system design.
Simon Willison released web-edit 0.1a1, a web app for editing files in a folder with performance optimizations written by Claude Opus 5.5 using Python profiling data.
Hugging Face is hosting a live stream with the Reflection team discussing how they built Beam.
Parsimony is an open-source benchmark that measures how much code AI coding models add by counting syntax elements rather than lines, helping developers evaluate code efficiency.
Janas-LLM enables running large language models (up to 80B parameters) on hardware with limited RAM by streaming experts from SSD, achieving 37 tok/s on Qwen3-Next-80B with 32GB RAM.
Simon Willison released llm-meta-ai 0.2, a plugin for the Meta AI API that supports the new Muse Wonder models, which demonstrate significant improvements in reasoning capabilities.
Simon Willison releases web-edit 0.1a0, a Python web app built by AI agents for securely editing files in folders over Tailscale networks.
Hugging Face CEO advocates for researchers to share RL environments openly, citing recent conversion of OpenAI math problems into open-source training environments.
Veda is a Rust-based operating system with built-in AI agent capabilities that integrates directly with applications rather than relying on screen scraping.
A developer used Claude Opus 5.5 to successfully port the TypeScript compiler, type checker, and LSP to Rust, achieving 100% compatibility in real-world testing for under $25k.
Talorys is a free, open-source personal AI agent that runs entirely within your Cloudflare account with chat, memory, tasks, and automation features.
A GitHub repository tracks FOSS projects that have adopted LLMs/AI and provides alternatives, forking guides, and policy request templates.
Hugging Face converted OpenAI's math problems into open-source RL training environments available for anyone to use.
Edi Life OS is a self-hosted life dashboard with built-in MCP server for AI assistants like Claude to manage habits, goals, and finances.
Open-source macOS CLI tool that lets AI agents visually guide users by drawing arrows and text on screen for human-computer interaction scenarios.
Simon Willison released ttok 1.0, a CLI tool for counting and truncating text based on tokens, now defaulting to GPT-5/GPT-6 tokenizer instead of GPT-4.
nanoMuse is an open-source personal agent that runs on devices you own, with shared memory and actions managed by a Sentinel system, providing an alternative to Meta's closed Muse agent.
OpenTPU is an open-source AI accelerator developed by AI agents, featuring RTL design, ISA, simulator, compiler and profiler that runs modern models on FPGA hardware.
NanoMuse is an open-source personal AI agent that runs across devices, performs tasks instead of just answering questions, and works offline with local deployment options.
Simon Willison praises EmbeddingGemma 2's Apache 2.0 license as essential for embedding models due to vendor lock-in risks with proprietary alternatives.
Hugging Face released OpenEnv, an open-source proxy that converts coding harnesses into RL environments without modification, enabling training on real evaluation frameworks.
OpenAI Cookbook adds deepsense.ai's evaluation framework for harness-aware plugin testing and validation.
Simon Willison released llm-openai-decisions 0.1a0, an LLM plugin for OpenAI's new Decisions API that supports image input and structured question types.
Hugging Face CEO calls for more public traces of AI agent attacks/defenses to improve security transparency and help defenders learn from real-world incidents.
DwarfStar 4 (ds4) is a C inference engine that enables local deployment of large MoE models like DeepSeek V4.1 and Qwen3.8 on high-memory machines with asymmetric quantization.
llm-mistral 0.16 adds support for Mistral's reasoning models like Mistral Large 4 to Simon Willison's LLM command-line tool.