Study mode

My 2.5 year old laptop can write Space Invaders in JavaScript now (GLM-4.5 Air)

Irrelevant facts about cats added to math problems increase LLM errors by 300%

M8.7 earthquake in Western Pacific, tsunami warning issued

It&#x27;s a web app that uses AI to turn any book into a playable text adventure. Your favorite book, but your choices, hence your story. You can even &quot;remix&quot; the genre like playing Dune as a noir detective story.Note: Work in progress. Suggestions are welcome.

Show HN: I built an AI that turns any book into a text adventure game

Maru OS – Use your phone as your PC

Observable Notebooks 2.0 Technology Preview

Learning basic electronics by building fireflies

More honey bees dying, even as antibiotic use halves

Hi HN! We&#x27;re Yujong, John, Duck, and Sung from Hyprnote (<a href="https:&#x2F;&#x2F;hyprnote.com" rel="nofollow">https:&#x2F;&#x2F;hyprnote.com</a>). We&#x27;re building an open-source, privacy-first AI note-taking app that runs fully on-device. Think of it as an open-source Granola. No Zoom bots, no cloud APIs, no data ever leaves your machine.Source code: <a href="https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote">https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote</a>
Demo video: <a href="https:&#x2F;&#x2F;hyprnote.com&#x2F;demo" rel="nofollow">https:&#x2F;&#x2F;hyprnote.com&#x2F;demo</a>We built Hyprnote because some of our friends told us that their companies banned certain meeting notetakers due to data concerns, or they simply felt uncomfortable sending data to unknown servers. So they went back to manual note-taking - losing focus during meetings and wasting time afterward.We asked: could we build something just as useful, but completely local?Hyprnote is a desktop app that transcribes and summarizes meetings on-device. It captures both your mic input and system audio, so you don&#x27;t need to invite bots. It generates a summary based on the notes you take. Everything runs on local AI models by default, using Whisper and HyprLLM. HyprLLM is our proof-of-concept model fine-tuned from Qwen3 1.7B. We learned that summarizing meetings is a very nuanced task and that a model&#x27;s raw intelligence (or weight) doesn&#x27;t matter THAT much. We&#x27;ll release more details on evaluation and training once we finish the 2nd iteration of the model (still not that good we can make it a lot better).Whisper inference: <a href="https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote&#x2F;blob&#x2F;main&#x2F;crates&#x2F;whisper-local&#x2F;src&#x2F;model.rs">https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote&#x2F;blob&#x2F;main&#x2F;crates&#x2F;whispe...</a>AEC inference: <a href="https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote&#x2F;blob&#x2F;main&#x2F;crates&#x2F;aec&#x2F;src&#x2F;lib.rs">https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote&#x2F;blob&#x2F;main&#x2F;crates&#x2F;aec&#x2F;sr...</a>LLM inference: <a href="https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote&#x2F;blob&#x2F;main&#x2F;crates&#x2F;llama&#x2F;src&#x2F;lib.rs">https:&#x2F;&#x2F;github.com&#x2F;fastrepl&#x2F;hyprnote&#x2F;blob&#x2F;main&#x2F;crates&#x2F;llama&#x2F;...</a>We also learned that for some folks, having full data controllability was as important as privacy. So we support custom endpoints, allowing users to bring in their company&#x27;s internal LLM. For teams that need integrations, collaboration, or admin controls, we&#x27;re working on an optional server component that can be self-hosted. Lastly, we&#x27;re exploring ways to make Hyprnote work like VSCode, so you can install extensions and build your own workflows around your meetings.We believe privacy-first tools, powered by local models, are going to unlock the next wave of real-world AI apps.We&#x27;re here and looking forward to your comments!

Launch HN: Hyprnote (YC S25) – An open-source AI meeting notetaker

RIP Shunsaku Tamiya, the man who made plastic model kits a global obsession

iPhone 16 cameras vs. traditional digital cameras

Microsoft Flight Simulator 2024: WebAssembly SDK

FoundationDB: From idea to Apple acquisition [video]

After training calculator agent via RL, I really wanted to go bigger! So I built RL infrastructure for training long-horizon terminal&#x2F;coding agents that scales from 2x A100s to 32x H100s (~$1M worth of compute!) Without any training, my 32B agent hit #19 on Terminal-Bench leaderboard, beating Stanford&#x27;s Terminus-Qwen3-235B-A22! With training... well, too expensive, but I bet the results would be good!*What I did*:- Created a Claude Code-inspired agent (system msg + tools)- Built Docker-isolated GRPO training where each rollout gets its own container- Developed a multi-agent synthetic data pipeline to generate &amp; validate training data with Opus-4- Implemented a hybrid reward signal of unit test verifiers &amp; a behavioural LLM judge.*Key results*:- My untrained Qwen3-32B agent achieved 13.75% on Terminal-Bench (#19, beats Stanford&#x27;s Qwen3-235B MoE)- I tested training to work stably on 32x H100s distributed across 4 bare metal nodes- I created a mini-eval framework for LLM-judge performance. Sonnet-4 won.- ~£30-50k needed for full training run of 1000 epochs (I could only afford testing )*Technical details*:- The synthetic dataset ranges from easy to extremely hard tasks. An example hard task&#x27;s prompt:&quot;I found this mystery program at `&#x2F;app&#x2F;program` and I&#x27;m completely stumped. It&#x27;s a stripped binary, so I have no idea what it does or how to run it properly. The program seems to expect some specific input and then produces an output, but I can&#x27;t figure out what kind of input it needs. Could you help me figure out what this program requires?&quot;- Simple config presets allow training to run on multiple hardware setups with minimal effort.- GRPO used with 16 rollouts per task, up to 32k tokens per rollout.- Agent uses XML&#x2F;YAML format to structure tool calls*More details*:My Github repos open source it all (agent, data, code) and has way more technical details if you are interested!:- Terminal Agent RL repo- Multi-agent synthetic data pipeline repoI thought I would share this because I believe long-horizon RL is going to change everybody&#x27;s lives, and so I feel it is important (and super fun!) for us all to share knowledge around this area, and also have enjoy exploring what is possible.Thanks for reading!Dan(Built using rLLM RL framework which was brilliant to work with, and evaluated and inspired by the great Terminal Bench benchmark)

Show HN: Terminal-Bench-RL: Training long-horizon terminal agents with RL

ACM Transitions to Full Open Access

URL-Driven State in HTMX

Elements of System Design

A little idea I got from playing with AI SWE Agents. Can AI help make sure we understand the code that our AIs write?PR Quiz uses AI to generate a quiz from a pull request and blocks you from merging until the quiz is passed. You can configure various options like the LLM model to use, max number of attempts to pass the quiz or min diff size to generate a quiz for. I found that the reasoning models, while more expensive, generated better questions from my limited testing.Privacy: This GitHub Action runs a local webserver and uses ngrok to serve the quiz through a temporary url. Your code is only sent to the model provider (OpenAI).

Show HN: A GitHub Action that quizzes you on a pull request

How the brain increases blood flow on demand