Yogurt is a free, open-source app for private AI-powered meeting notes, positioned as an alternative to Granola. Its repository lists support for macOS, Windows, and Linux, with local AI and transcription tools. (github.com)
A public GitHub project implements key DeepSeek V3 components in about 190 lines of pure JAX. Its author says the model was trained on a MacBook Air. (github.com)
The daily digest
Today's best Hacker News stories, summarized and screenshotted, one email a day.
Linum trained a 2-billion-parameter Pyramid-JiT text-to-image model on 660 million samples for about $21,000 in compute. The experimental model’s code and weights are available under the Apache 2.0 license, though it has notable limitations. (linum.ai)
The daily digest
Today's best Hacker News stories, summarized and screenshotted, one email a day.
Qwen3-4B has been fine-tuned to eliminate typical AI writing habits, resulting in more natural and human-like responses. The model, small enough to run locally on a MacBook Air, now produces clearer, more concise text without common AI tics. (100toolkit.com)
The daily digest
Today's best Hacker News stories, summarized and screenshotted, one email a day.
Hob is a free, open-source desktop AI assistant that runs open-weight models locally on Mac, Windows, and Linux. It works offline and keeps conversations on the user’s device as an alternative to Claude and ChatGPT. (github.com)
Uncracked City is a detective game set in real neighborhoods, where players search for clues and question AI suspects who may lie. After five training levels, players tackle unsolved cases and get three attempts to submit a final report. (uncracked.city)
AgentRun is an open-source platform that adds self-improvement, durable execution, and management tools to AI agents. It aims to make agents more reliable and reduce token use with each run. (github.com)
Doomformer compiles an integer neural network into a DOOM map, using voodoo dolls for arithmetic and lighting and floor heights to store values. It checks every result against NumPy with zero tolerance. (github.com)
Flide is an open-source inference engine designed to run Jev-like decision models faster. Its creators claim speedups of up to 9.5×. (github.com)
STE80 Humanizer is an agent skill that rewrites AI-generated text to remove common AI tells. It aims to bring writing about 80% of the way toward ASD-STE100 Simplified Technical English. (github.com)
Agweb is a lightweight browser designed for AI agents, with all tabs visible in one view. It aims to avoid stealing focus while agents work. (github.com)
Omnara is an open-source, self-hostable platform offering managed agents as an alternative to Claude Managed Agents and OpenAI's Agents API. It enables developers to run AI agents locally or on their own infrastructure. (github.com)
Fable52.games is an arcade of 32 browser games built autonomously by Claude Fable 5.2, often from a single prompt as one self-contained HTML file. Visitors can play the games, read their prompts, and inspect their source code. (fable52.games)
Jev Cards is a quiz game that uses TypeSafe's Jev vision model to identify objects in pictures and answer questions quickly. The game allows users to compete against Jev or friends, with open-source code available for customization. (jevcards.com)
ADRs4AI is a Git-based system for recording questions, AI recommendations, reasoning, and human decisions so collaboration stays searchable and versioned. It offers repository templates, a VS Code extension with agent support, and an iOS app for reviewing and accepting answers as commits. (adrs.systems)
Klad is a source-available paint app designed for people and coding agents to draw together. Its agent-driven controls make the canvas usable as part of coding workflows. (github.com)
BootGPT is an experimental language model implemented in x86 assembly that fits in a boot sector. The repository includes its code and demo files. (github.com)
HyperFlow H3 Flash compresses the 31B MiniMax-H3 video-and-audio diffusion model to 13.7B parameters. It supports joint video and audio generation in 4-8 steps, with INT8 and GGUF versions. (github.com)
An AI analyzed 45,047 Enron emails to create a detailed, citation-backed wiki about the company's knowledge and contacts. The project tracks when facts became outdated, providing a dynamic historical record. (thisandthat.chat)
Hurston-1 is an open-source text-to-speech model designed to speak Southern American English and Afro-English. The model is available on Hugging Face. (huggingface.co)
Laminar has developed flow-1, an open-source model trained with reinforcement learning to detect errors in AI agent traces at a fraction of the cost of larger models. The system analyzes detailed agent traces to identify implicit failures, enabling comprehensive and affordable agent observability. (news.ycombinator.com)
Respan has developed an open-source Rust-based AI gateway that uses a plugin architecture to easily add providers, protocols, and features without modifying the core. The design aims to simplify AI infrastructure customization and scalability for developers and organizations. (news.ycombinator.com)
Penmentor is an AI tool that marks up handwritten math like a teacher, helping users learn and understand mathematical problems. It is available as an open-source project on GitHub. (github.com)
AI Red Lines helps people define what AI behaviors would prompt them to act and decide what they would do. It offers ways to share commitments, schedule check-ins, and save private copies. (yourairedlines.com)
Leeroo open-sourced Kapso, an engine that experiments on enterprise data tasks, deploys the best-performing solution, and learns from each run. Its lesson bank records evidence and conditions from past experiments to guide future work toward solutions suited to each company’s environment. (news.ycombinator.com)
Coral is an open-source gateway that lets AI agents create custom tools for enterprise systems. It adds precise permissions and aims to make agent work faster with fewer tokens. (github.com)
Spec-first is a development plugin that turns feature descriptions into reviewed specifications before agents write code. It builds and tests changes in separate worktrees, then opens pull requests only after checks pass. (spec-first.dev)
Arsaze is an AI-powered video editor that turns plain-language requests into edits on a real timeline, with changes users can review or undo. It also offers transcript-based editing, Tab-complete suggestions, and frame-specific client feedback. (arsaze.com)
Sketch Stadium is a live Pictionary game where AI agents take turns drawing secret words and guessing other agents’ sketches. Players can connect agents through an MCP server, watch rooms live, and compete for points on a leaderboard. (sketch.wakeshift.ai)
Interactive notes teach RLHF, attention, LoRA, and LLM evaluations through hands-on examples. The site requires JavaScript and its author updates the notes while learning. (concepts.laxmena.com)
Recalc is a shared spreadsheet where cells can contain values, formulas, terminal commands, or Claude Code and Codex agents. It lets users coordinate AI coding agents directly from a spreadsheet. (github.com)
Strot is a tool that compiles an AI model's intent once and runs decisions locally indefinitely. It uses deterministic, self-routing natural language understanding runtimes to improve AI code synthesis. (github.com)