Show HN: Trace – Offline Mac meeting transcripts you can flag mid-call https://ift.tt/hBWCkLn

Show HN: Trace – Offline Mac meeting transcripts you can flag mid-call I'm the developer of Trace, a non-intrusive, shortcut-driven Mac app that records and transcribes your meetings on-device. I know, another meeting transcription app. Please bear with me though, I'm confident that this is at least a little novel. I primarily built Trace for myself. I'd been using MacWhisper, but there was enough fiddling before each call that I'd forget to start it and walk out of an hour-long meeting with nothing written down. So the things I cared about most were that it's quick to activate and stays out of the way. You activate Trace by pressing a global shortcut (configurable), which reveals a small bar at the bottom of your screen (there's also a keystroke and/or option to hide it entirely if you'd rather not see it at all). As I was building it I wanted to bake in a couple of workflows I'd wished for in other transcription apps. 1. Mid-meeting you can press another global shortcut to mark a "key moment" and type a note. The note shows up in the resulting transcript inline at that timestamp. I wanted to add this because I kept catching myself thinking "wait, that bit matters" in meetings and reaching to jot it down in a separate app like Obsidian, which I then needed to add context to, which took me out of the meeting. I use it all the time. If I paste the transcript into an LLM afterwards (which I find myself doing more and more these days) the important moments are flagged so it doesn't gloss over them. This is more noticeable in longer meetings with lots of topics. 2. With another keyboard shortcut you can summon a rough live recap (subtitles, basically) to quickly recap what's just been said. Trace uses standard macOS microphone and system recording APIs to capture both sides of the conversation as two separate tracks and then runs the system side through on-device diarization to identify speakers. Right now we only label them as "Speaker 1", "Speaker 2", etc but there are plans for speaker labelling in the future. You can also show a "live recap" as the call is happening to review what someone just said. All transcription models run on your machine. To be clear though, Trace doesn't do any of the summarising itself, it just produces a markdown transcript, so if you want summaries then you need to pass the output to an AI. The app is sandboxed and your audio/transcripts are never uploaded anywhere - they just exist as audio files and markdown on disk. The only network call Trace is required to make is on the first run to download the speech and speaker models (around 500MB) from Hugging Face, and after that it can be used fully offline. If enabled, a Google Calendar integration can auto-name sessions but that needs a network connection. The app is £9.99 on the macOS App Store. I've been using it every day for months now and I'm super happy with how it's improved my workflow. Feedback very welcome. https://traceapp.info June 14, 2026 at 02:11AM

Show HN: Philosophy for Kids https://ift.tt/VLfw7X9

Show HN: Philosophy for Kids Sometimes my son asks me 'why' questions that could be answered well by a kid-friendly philosophy article. But I don't know where to find those, so I ask Claude or ChatGPT, and have a specific workflow for getting the type of output I want. I figured other people might find those AI-generated articles helpful, so I put them here: https://ift.tt/ENz5J17 There's a search box at the top. https://ift.tt/ENz5J17 June 14, 2026 at 11:45PM

Show HN: Slopsome – a VRAM fit calculator and tok/s database for local LLMs https://ift.tt/LD9XYTk

Show HN: Slopsome – a VRAM fit calculator and tok/s database for local LLMs https://slopsome.com June 14, 2026 at 01:14AM

Show HN: Galdor – a Go LLM agent framework with built-in tracing and replay https://ift.tt/QrVPWmE

Show HN: Galdor – a Go LLM agent framework with built-in tracing and replay https://ift.tt/QGUDtoC June 14, 2026 at 12:34AM

Show HN: Turn your name into a tree in an infinite procedural shanshui landscape https://ift.tt/INP1tnb

Show HN: Turn your name into a tree in an infinite procedural shanshui landscape Hi HN! I made this after collecting hundreds of "name → tree" submissions at ITP. Live: https://ift.tt/Lp4ywkJ Source: https://ift.tt/yUf96kL Plant a tree: https://ift.tt/yTQ7r3N Pan and zoom an infinite procedural landscape. Each name is converted to ASCII codes, which grow into a unique tree (breadth-first branching; repeated digits become mathematical roses). Mountains use midpoint displacement + Perlin noise, with SVG radial gradients in the blue/green/gold palette from Wang Ximeng's "A Thousand Li of Rivers and Mountains." Inspired by Lingdong Huang's {Shan, Shui}* ( https://ift.tt/aiAVvMW ). Every tree is someone's name, signed with an APack stamp ( https://ift.tt/BvT8q9a ). Try planting your name, then pan along the ridgeline to find it. "My trees" lets you jump back to ones you planted. Happy to answer questions about the terrain algo, name→tree encoding, or the Riso print we tiled at ITP Winter Show! https://ift.tt/Lp4ywkJ June 10, 2026 at 08:09PM

Show HN: Nenya – A lightweight, highly secure AI API Gateway/Proxy written in Go https://ift.tt/XrDhYeH

Show HN: Nenya – A lightweight, highly secure AI API Gateway/Proxy written in Go https://ift.tt/dvsLHa6 June 12, 2026 at 11:02PM

Show HN: Vilvona AI – Self-Hosted AI Assistant with Tamil and Hindi UI https://ift.tt/cL1v2Nk

Show HN: Vilvona AI – Self-Hosted AI Assistant with Tamil and Hindi UI https://ift.tt/LwmgyB1 June 12, 2026 at 11:56PM

Show HN: Nuts – pip/NPM for Java with first-class workspaces, JDK provisioning https://ift.tt/Xl7kdOb

Show HN: Nuts – pip/NPM for Java with first-class workspaces, JDK provisioning My frustration with distributing java apps didnt show up recently. I remember having implemented my first network jar downloaded back in the 2000's because i needed applet like feature support with desktop full control. Years after, the problem is the very same. Webstart didnt really took off and the only mean i had in my projects was the ugly fatjars, including the (for me) uglier spring-boot repackaging that changes the application classloading behaviour and hence giving me by time some headackes i was not prepared for. So basically nuts started as a response to this frustration 9 years ago, but from now i think its mature enough (used in production) to be shared, and forecebly i am more keen to need suggestions and help from fellow contributors. https://ift.tt/gsJDKEL June 11, 2026 at 03:53AM

Show HN: AVP – an agent can't leak a secret it never had https://ift.tt/f70qDpY

Show HN: AVP – an agent can't leak a secret it never had A process can't leak a secret it never had. Shai-hulud, prompt-injection - you name it. They cannot steal what your agent (or an process) don't have. I run coding agents (Claude Code, Codex) on my own machines most of the day. Every one of them wants real API keys in env and I was scratching my head for the last few months how to contain it. The usual answer to this is a firewall. I don't buy it. A firewall tries to contain a secret the process is still holding, and the rules are painful to maintain. AVP gives the agent a placeholder and injects the real value at the last moment, on the wire: ``` # the agent's env holds only a placeholder STRIPE_API_KEY=avp-placeholder # agent sends: Authorization: Bearer avp-placeholder # AVP forwards upstream: Authorization: Bearer sk_live_...real... ``` Keep your passwords in your vault where they belong. AVP initially relies on Bitwarden as a secret manager. It's MIT licensed. Appreciate any feedback. https://ift.tt/V5yC7rZ June 12, 2026 at 12:40AM

Show HN: Stillwind – High Resolution Electronic Component Search https://ift.tt/KqO2N3h

Show HN: Stillwind – High Resolution Electronic Component Search We’ve spent the last couple of months building Stillwind Search, a search engine for electronic components that helps users find parts that fit even the most complex set of specifications. After talking to the people that actually build PCBs we found out that finding the exact part you are looking for, is consuming enormous amounts of times, is very tedious and then often doesn’t yield the best results. So we tried to cut down this search time by just requiring a (broad) description of the specifications and we find the correct part in minutes, not hours. This is possible through our own database of parts and their properties. We used LLMs to extract every parameter about a part into >1k schemas, collectively covering more than 130k properties. This depth of properties could no longer be visualized, so the database is queried interactively by an AI agent (Sonnet 4.6) to find the needle in the haystack of parts. Before results are shown, we use another model to verify the data (that’s why it might take a moment before the first results appear). We currently have almost all microcontrollers, sensors, and other advanced ICs on the market, as well as a wide selection of passives and generic parts. We are working on adding more parts and are more than happy to take suggestions. I know that folks on HN like technical details on how this works, so let me give a short overview: Frontend: SvelteKit + Cloudflare Workers + Hyperdrive Backend: PostgreSQL 18 (with io_uring) database, with extensions on NVMe drives hosted on a beefy server. We appreciate all feedback and are happy to answer any questions :) Btw: We are already working on a way that you can search combinations of parts, finding the optimal combination of parts. https://stillwind.ai June 11, 2026 at 11:42PM

Show HN: Atlasphere – Live Infrastructure Diagrams https://ift.tt/dMeon5I

Show HN: Atlasphere – Live Infrastructure Diagrams Hi HN. My name is Andrey. On a regular business day, I'm a software engineer working at AWS. Outside of work hours, I spend time on my hobby - writing code. I was once building a pet project that allowed customers to spin up fully synchronized blockchain nodes within just a few minutes. The backend was split into a control plane and a data plane, each with its own AWS account. Later I added two more AWS accounts. One for shared RPC nodes. One for the Analytics Service. Since I love to visualize things, I used drawio to visualize the architecture. With time, I noticed a pattern. I'd write some code, add a few lambda functions, update my drawio diagram, write more code, introduce a few more resources, test things, see that everything works fine and go to sleep with a smile on my face. Next week I'd check my diagram, and shockingly, it's missing some of the resources! This kept happening for a few more weeks until I decided to fully abandon the project until my infrastructure diagrams could stay in sync with my cloud account. That's how Atlasphere.io was born. I've been working on it for the past 6 months and I think the product is ready for some feedback :) A few notes: - Atlasphere uses a ReadOnly IAM role to scan your AWS account (my account reaches your account through a trust relationship). - The number of services is currently limited (WIP) - It's a macOS app - It's NOT an Electron app, i use Rust + Webview What am I looking for? All I really need is for someone to try the app and tell me what they like about it and what they absolutely hate about it, haha! The website is https://atlasphere.io/ June 9, 2026 at 06:05PM

Show HN: Meadow Mind – a 7B diffusion LLM plays Gym games with zero training https://ift.tt/lCaVpw4

Show HN: Meadow Mind – a 7B diffusion LLM plays Gym games with zero training https://ift.tt/ZeW91Mi June 10, 2026 at 11:11PM

Show HN: Extend UI – open-source UI kit for modern document apps https://ift.tt/OmQyagr

Show HN: Extend UI – open-source UI kit for modern document apps We're open-sourcing 14 components & examples today for PDF, DOCX, and XLSX viewers, plus bounding box citations, file upload, e-signature, and more. It's MIT licensed and fully customizable. Demo video here: https://ift.tt/WH6skeN When we started, we tried every file viewer and document component library we could find. Unfortunately, none of them had all the functionality (and polish) that we wanted, so we ended up building our own for https://extend.ai/ . It was only ever meant to be internal, but enough customers kept asking for it that we decided to open source it. It's useful for building document processing agents, real-time user facing document intake flows, or all kinds of internal tooling. We naively thought this would be a solved problem. Turns out, making PDF/XLSX/DOCX viewers that work at scale is not trivial...we use and maintain it for Extend ourselves, so we've fixed a lot of edge cases that came up while running millions of pages / day through our own system. Our hope is that with our resources + community support, it'll keep getting better over time. https://ift.tt/Wd2UltZ June 10, 2026 at 09:39PM

Show HN: LocalCode – turn plain English into CLI commands with Apple's local AI https://ift.tt/5esybMi

Show HN: LocalCode – turn plain English into CLI commands with Apple's local AI https://ift.tt/YJzM96o June 10, 2026 at 02:34AM

Show HN: OpenYabby, voice-controlled multi-agent orchestrator for Claude Code https://ift.tt/xKOuhC9

Show HN: OpenYabby, voice-controlled multi-agent orchestrator for Claude Code https://ift.tt/oHZFjeJ June 10, 2026 at 01:38AM

Show HN: Transit-format (JSON/MessagePack) reader/writer in C https://ift.tt/fGB5ihJ

Show HN: Transit-format (JSON/MessagePack) reader/writer in C Transit.c is an addition to the set of libraries to support transit data interchange format written in C11. It supports full 0.8 specification of cognitect's transit-format: JSON, JSON-Verbose and MessagePack encodings, all ground and extension types, compression via keys caching, extensibility via custom tag handlers. https://ift.tt/etkdHVs June 8, 2026 at 03:05PM

Show HN: HTTP/3 and raw QUIC client/server APIs for Node.js https://ift.tt/ecnuEKb

Show HN: HTTP/3 and raw QUIC client/server APIs for Node.js I built this because I wanted to make outbound and accept inbound HTTP/3 and raw QUIC connections from ordinary Node.js code, without building Node from source or putting everything behind a reverse proxy. Repo: https://ift.tt/NXKcFUq npm: https://ift.tt/XnrOCqP It’s a native package around Rust/quiche. It supports both client and server APIs, I'm using it in a couple of projects: creating raw QUIC streams, datagrams, custom ALPN, session behavior, and HTTP/3 client work from Node. I've tried to be very safe in the native code, written in rust, with proofs around the parts I was most concerned about getting wrong. I have it hosting a couple of sites as HTTP3 endpoints and found it working well. https://ift.tt/NXKcFUq June 9, 2026 at 12:08AM

Show HN: Stop returning raw JSON from MCP servers, build rich inline UIs https://ift.tt/vH6b2Bj

Show HN: Stop returning raw JSON from MCP servers, build rich inline UIs https://ift.tt/mxs0wul June 9, 2026 at 12:07AM

Show HN: A Minecraft builder skill for coding agents https://ift.tt/R5DQ8I0

Show HN: A Minecraft builder skill for coding agents https://ift.tt/RaWLn4B June 8, 2026 at 08:21PM

Show HN: NoSuggest – Watch YouTube without the recommendation algorithm https://ift.tt/KkdOoMr

Show HN: NoSuggest – Watch YouTube without the recommendation algorithm NoSuggest is a quiet act of resistance against YouTube algorithms always trying to pull you into a loop of unlimited videos in turn into unlimited screen time. With unending side cards of videos, auto-play, what's next suggestions, YouTube shorts and notifications, users will be doom scrolling for many hours in a day. I faced the same problem. Acknowledging that, not all content in YouTube is bad. There are educational videos, genuine news contents without political bias which is very hard to find outside YouTube and many other good relaxing, entertainment stuff. NoSuggest lets you only follow the YouTube channels you like and removes all types of recommendation YouTube has. So you don't waste time on watching things which you never wanted to watch anyways. UI is very simple. You add your favourite channels in "Channels" tab and latest 5 videos per channel excluding shorts would appear in "Feed" tab. "Search" tab is to search for specific videos to watch and "Saved" tab is to bookmark any video you want to watch later. Intention of NoSuggest is to provide whatever is necessary to extract whats good from YouTube all inside NoSuggest and leave out bad parts. NoSuggest works in any devices. Install it as an app (PWA) in android and iPhone, or simply open in browser in laptops. No sign-in, no account creation or no card details. NoSuggest won't even ask your name. Total privacy for the users. Parents can add the channels and save some educational videos and lock it with the pin for kids mode. Kids won't be able access unwanted additive contents inside NoSuggest. Completely free, no string attached. Source available in Github through NoSuggest website. I would love genuine feedback. Thank you very much for your attention on this matter. https://ift.tt/DJYTmhR June 4, 2026 at 02:44AM

Show HN: An mkv player that uses WASM to render you videos https://ift.tt/pJjemUr

Show HN: An mkv player that uses WASM to render you videos hello HN i want to share this wasm experience i built for a universal mkv player on the web using wasm to ship a lean decoder only ffmpeg build, thus way codecs unsupported by the browser can be played I wonder if this holds any value to anyone anymore https://parallax.kinosoft.moe/ June 8, 2026 at 05:27AM

Show HN: I Derived a Pancake https://ift.tt/MNrBp83

Show HN: I Derived a Pancake After 25 years of making other people's pancake recipes - always yearning for more tang, more fluff, and more predictability - I decided to derive the pancake recipe from the chemistry. You mark checkboxes for what you have on hand (ricotta, sour cream, kefir, buttermilk, yogurt, cottage cheese, lemon, cream of tartar, etc.) and it computes the best recipe based on targets for acid, fat, salt, sugar, and CO2. My particular favorite are the yeast-raised lemon ricotta kefir pancakes - the best I've ever had. The math is done in a small pure-ESM library: ingredient composition to component masses and acid moles, a stoichiometry layer, and a bisection solver for the target deficits. I'm not a chemist, so if something is off, tell me and I will fix it! https://ift.tt/05n2zxX June 5, 2026 at 12:12PM

Show HN: Nightwatch, The open-source, read-only AI SRE https://ift.tt/84AakcH

Show HN: Nightwatch, The open-source, read-only AI SRE nightwatch is a local-first, read-only layer on top of your monitoring. it groups alert storm into incidents, flags noisy checks and has an agent that can investigate for you live systems. You can e.g. jump from the incident into the agent directly. the reason for this weekend project is that we had a kubernetes upgrade that went wrong, and at some point a rollback wasn't possible anymore, so it had to be fixed live during the night while several problems came together. We run a lot of different systems, on-prem and several Kubernetes clusters, and in a situation like that you spend most of the time just figuring out what is actually broken and where. So i thought that it would be pretty cool to have eyes in the dark in each system that can talk to your "brain". so the idea is to put a baby owl into each environment. Each owl runs where the systems live, keeps that environment's credentials local, and only dials outbound to a central brain, so there is no inbound hole into prod. It exposes a set of read-only skills, and the agent uses them to gather evidence and form a root-cause hypothesis, so the on-call engineer starts with a head start instead of from zero. read-only for now, i don't trust it near prod yet and honestly neither should you. llocal-first for easy self-hosting and to keep credentials on your side. the clustering and recommendations run fully offline with no llm at all. the agent needs a tool-calling llm, you can point it at a remote one, or self-host one (ollama etc.) if you want to stay fully offline. for non selfhosters: before every remote llm call, nightwatch strips real secrets (unrestorable) and swaps identifiers like ips, hostnames and paths for reversible placeholders, so the model only sees masked data while real values are restored only in the proposed commands and tool calls Would love if you try it in your Systems https://ift.tt/sM5IZ8X June 8, 2026 at 01:54AM

Show HN: Dap-mux – Connect your editor and REPL to the same debug session https://ift.tt/iw4onjx

Show HN: Dap-mux – Connect your editor and REPL to the same debug session I have been coding over four decades, in many languages, on many projects (including Firefox, Final Cut Pro, the Newton, and Fullwrite Professional if you can remember that far back; all these using my "dead-name"). I wrote something small and simple to scratch an itch. It's the UNIX philosophy: small "one-trick ponies", each *really* good at their one trick, then the user can hook them together to solve actual problems. I'm a CLI guy, and for almost everything, I already have this. But not for debugging. The itch I scratched was the connector that enables this philosophy for debugging. That thing is dap-mux. A DAP multiplexer turning a one-to-one protocol into a cooperating session of as many tools as you need to get it done! How it started: Helix and Python for me (and sometimes IPython), with the rest of my team using PyCharm (which I have long loved!). My team's problem is that they want the PyCharm debugger, and so must be satisfied with the JetBrains editor. *My* problem was I could use a full-blown debugger *or* I could have IPython *or* I could have Helix (or sometimes an unsatisfying combination of Helix and the debugger). That was my "itch". DAP (Debug Adapter Protocol) is the tantalizing answer, except it isn't. DAP is what editors (that don't want to write their own debuggers) are starting to adopt. The problem with DAP is it's one-to-one. One editor connects to one debugger. Done. Not a solution to my problem. And then suddenly, it *was* the solution. I realized that a DAP multiplexer would let you connect any DAP-aware editor to any debugger for any language, and simultaneously to a REPL, another session of your editor (or a different editor)! With the side benefit that now, like screen or tmux, since each process is its own thing: sessions are durable. Just restart whatever crashed and you're back where you were! There were hard parts: sequencing, late joiners, state management. Different end-points working on different actions in different sequences but with the same message ids. I solved these problems something like how NAT works. Instead of translating network addresses, though, I'm translating the sequence numbers of each client into something global and ordered, then correctly routing replies back to the end-point awaiting them, while mapping the sequence numbers for those replies back into the space of that end-point. Knowing the current state of the debugger, and replaying that as a message sequence to late joiners lets you start/connect the clients in any order. I chose Python: asyncio fits the I/O-router pattern perfectly, and it lets the IPython extension run in-process rather than over IPC. There are problems not yet solved: for instance, I think configuration in the clients and/or the startup sequence is too complicated. But it functions! I got what I wanted! The combination I use every day: Python + debugpy + Helix + IPython, all connected simultaneously. Step with `%n` or `%s`, evaluate expressions with `%eval`, watch Helix track the current line in real time. Rust with codelldb is the second confirmed combination — I debugged a Dijkstra implementation with Helix and a third-party DAP observer tool both connected to the same codelldb session. A community member, Sean Perry, has already built [dap-observer]( https://ift.tt/J94PEKy ), which renders the current frame's variables as a navigable terminal tree. *This* was my exact dream! Small, focused, connectable tools all playing together! There's so much left to try: other editors, other debug adapters, Windows, other languages. None of this has been touched yet. The most helpful thing now is people testing it with their own setup and reporting what they find. It's time to play! `uv tool install 'dap-mux[ipython]'` for Python + IPython. `uv tool install dap-mux` for headless use with any language and adapter. No need for any part of the Python ecosystem. https://ift.tt/y298GJA June 7, 2026 at 02:43AM

Show HN: Typedframes – Pandas/polars column name checking at lint time https://ift.tt/vEfYhgJ

Show HN: Typedframes – Pandas/polars column name checking at lint time https://ift.tt/kCsAiWf June 7, 2026 at 02:02AM

Show HN: Resonate – Low-latency, high-resolution spectral analysis https://ift.tt/fQDC9rE

Show HN: Resonate – Low-latency, high-resolution spectral analysis Last April I shared about my Resonate project here ( https://ift.tt/K5OdwQE ) A lot has happened since: the work I presented in much more detail at last June's International Computer Music Conference (ICMC) got best paper award. I also gave a talk at the Audio Developer Conference in Bristol last November, the video is on YouTube). This year's work, which I recently presented at this year's ICMC, starts with known techniques from the phase vocoder literature to build self-tuning filter banks that extract very efficiently the frequency components that are actually present in the input signal. Overview on the project website, more details in the papers, including applications to super-resolution spectrograms and re-synthesis experiments. As many people have pointed out, none of the techniques I have used are new (some of them even have different names across different fields), but I haven't seen them applied together in this way, and to me the results are incredibly satisfying and sometimes look magical. See for example this demo: https://youtu.be/LasdoIJJkw8 Of course the best way to experience in person is through the free demo app: https://ift.tt/fOUHniM Looking forward to feedback from the community! https://ift.tt/4vNn3mI June 6, 2026 at 11:39PM

Show HN: I nerfed our coding agents on purpose https://ift.tt/btRQN9j

Show HN: I nerfed our coding agents on purpose Tl;dr: I trained a classifier to route to the least expensive model and reasoning depth to complete the request. Coupling that with additional automated token efficiency techniques has yielded 3x usage for the same spend. For anyone interested in trying it themselves: https://nerfguard.com Various teammates and I switched over to Codex from Claude Code recently. We still bounce between the tools, but Codex’s speed and steerability coupled with performance gains were hard to ignore. One of the downsides was that the per token pricing kicked in way sooner. This is happening across the board, but we felt it in Codex more acutely. We’re a startup filled with people who work around the clock and are obsessed with building — naturally our daily bill alone was striking. Luckily we’re going after a big mission and speed matters significantly more than marginal token spend on the edges. Still, it got us thinking about how it was ludicrous that while our product has a side effect of decreasing token spend and speeding up agentic workflows by many orders of magnitude, we were using these top tier models for all types of internal coding tasks without any of those optimizations. The waste felt pretty ridiculous — the most glaring culprit was that we were seemingly using the max intelligence model on max reasoning for every task even when the task clearly didn’t require it. As a company who spends a lot of time on cached intelligence, it was also easy for us to see how there was plenty of other low hanging fruit as well. So, on a recent weekend, I quickly built a tool to optimize our usage. At its core is a very fast classifier that classifies your requests to the least intelligence required for the task and includes some nice token optimizations on top. The result is roughly the same quality for multiples lower token spend. But even more exciting for us, is that the properly bin packed intelligence and reasoning levels meant our speed also went up considerably. This wasn’t negligible. We’ve observed up to 3x savings and hours per day per person in saved time that we would have otherwise been waiting on tool turns and coding agent responses. For us, that means improved engineering velocity and significantly higher usage for the same spend. It also means more usage before getting throttled. As I told friends about this, they also wanted to start using it to maximize the usage they could get out of their coding agent plans. There are now engineers across many of the most cutting edge AI companies using this tool to optimize their token utilization in this way. Not just to save money, but to maximize output. Turns out that the best way to avoid getting nerfed by Claude is to intentionally nerf yourself selectively. We decided to release it for the rest of the builder community to use as well. You can now turn on Nerfguard for yourself and start getting more usage today. June 6, 2026 at 04:49AM

Show HN: I rebuilt a tiny old volleyball game I loved https://ift.tt/lZQObdh

Show HN: I rebuilt a tiny old volleyball game I loved https://volleyhop.com/ June 6, 2026 at 01:42AM

Show HN: Bash Runtime for AWS Lambda https://ift.tt/je8JKH7

Show HN: Bash Runtime for AWS Lambda Hi HN, I built a Bash runtime for AWS Lambda to make writing glue code simpler and faster. Sometimes, all you need is a bit of `sed`, `awk`, maybe a loop and a few HTTP API calls, and this runtime gives you all the tools to do that. It comes bundled with `jq` and `curl` so you can handle JSON payloads and string together HTTP API calls right out of the box, including calling AWS services with `curl --aws-sigv4`. In keeping with the theme, the Lambda handler contract is also made as simple as practical: read from stdin, write to stdout, return 0 for success and non-0 for error. You can run shell scripts, call binaries (either what's available in `al2023.provided` or you can package your own static binaries with your handler), or a combination of both. If you remember nodding along to Adam Drake's post about how bash and coreutils can be faster than a Hadoop cluster, I hope you give this a whirl and find it useful. The runtime is packaged as a Lambda layer, so it should drop right into your normal AWS infrastructure. https://ift.tt/PwmT07U June 6, 2026 at 12:42AM

Show HN: Bot or Not – Spot AI-generated randomness https://ift.tt/fnPoJVI

Show HN: Bot or Not – Spot AI-generated randomness https://play-bot-or-not.vercel.app/ June 5, 2026 at 01:26AM

Show HN: Using Haskell to play music on 3D printer motors (2020) https://ift.tt/UJX25Gb

Show HN: Using Haskell to play music on 3D printer motors (2020) https://lucasoshiro.github.io/software-en/2020-07-31-music_gcode/ June 5, 2026 at 01:37AM

Show HN: Cost.dev (YC W21) – making agents cost-aware and cheaper to call https://ift.tt/2zao37s

Show HN: Cost.dev (YC W21) – making agents cost-aware and cheaper to call We launched Infracost on HN five years ago ( https://ift.tt/cuE04hU ) where our CLI generated cost estimates for infra-as-code, e.g. "this Terraform PR adds $400/mo". The idea was to shift cloud costs (FinOps) left, so engineers get visibility of costs before deployment and make better decisions. Earlier this year we started seeing agent traffic in our logs and it looked like coding agents were calling our CLI. But that CLI wasn't designed with coding agents in mind. We went down a philosophical rabbit hole to see if a CLI is even needed anymore given that Claude, Copilot et al. already follow best practices. Ultimately we decided to create a new CLI from the ground up with coding agents in mind for two reasons: 1. We optimized the CLI for agent callers and cut Claude's output token usage by up to 79% and API cost by up to 67% versus a bare-Claude baseline. We wrote a blog documenting our lessons on optimizing user token usage when designing a CLI, e.g. using predicate flags so the agent doesn't compose jq | python | wc pipelines, output format that strips JSON's redundant field names. The blog is here: https://ift.tt/BJ5v8oy... 2. With cloud costs, precision matters. Telling a coding agent "make this Terraform cost-optimized" can be expensive and lossy. You burn tokens loading code and policy context into every conversation. Your agent could make up a price and you wouldn't know because it's difficult to verify that across the ~10M price points that AWS, Azure and Google have. The CLI runs static analysis on the code, uses the latest prices from cloud vendors, and passes that context to the coding agent. So that's what we're launching today - Cost.dev: https://cost.dev/ . - It runs locally. Your code never leaves your machine, you get a fast feedback loop, and you're not burning API calls per character when you want to fetch prices. - The CLI does the deterministic work. Fetching price points, scanning the code, validating fixes. The coding agent does the natural-language part. You don't have to trust the LLM to remember the rules, and can verify it called the right CLI command. - It provides a consistent rule layer across every tool you use. Get cost estimates in your IDE and your coding agent with a single install. We support Claude Code, GitHub Copilot, Cursor, Windsurf, OpenAI Codex, Gemini CLI, as well as IDEs like VS Code and JetBrains Before we keep building more in that direction, I want to sanity-check with HN: is "agents writing IaC in prod" actually a thing yet, or am I betting on a future that's still a year out? I know software developers are using coding agents heavily, but are platform/infra folks doing that for prod too? Also, if you have any feedback on Cost.dev, I'd love to hear it! https://cost.dev/ June 4, 2026 at 05:00PM

Show HN: Fork of Rsync https://ift.tt/wdAXWp3

Show HN: Fork of Rsync Hello. After hearing of the problematic LLM commits in rsync, I made a fork of rsync. I decided to fork it off release 3.4.1, since I heard that's the last release without the LLM code. https://ift.tt/z3wt1HF June 4, 2026 at 03:50AM

Show HN: Lint Your Markdown with ESLint https://ift.tt/FWBysn8

Show HN: Lint Your Markdown with ESLint https://ift.tt/6RYJpEj June 3, 2026 at 07:17PM

Show HN: I created a React alternative using web componnents https://ift.tt/OEFeqBC

Show HN: I created a React alternative using web componnents https://ift.tt/pDOQrvE June 4, 2026 at 01:30AM

Show HN: 100cc - Roll your own Claude in 100 lines https://ift.tt/1pUvIid

Show HN: 100cc - Roll your own Claude in 100 lines https://ift.tt/ZxTPAUn June 3, 2026 at 12:05AM

Show HN: RePlaya – self-hosted browser session replay with live tailing https://ift.tt/SlPWbYB

Show HN: RePlaya – self-hosted browser session replay with live tailing Hi HN, I'm one of the founders of s2.dev. RePlaya ( https://ift.tt/rFoZEvL ) is a self-hosted browser session replay tool using rrweb ( https://ift.tt/8GK1tlS ). It occurred to me that a durable stream per session would be a much neater architectural foundation for much of what you'd want from such a tool. As a unique feature, it also made live tailing straightforward because the player can read from the same stream the recorder is appending to. The alternative architecture is likely an ingest firehose which is then indexed, with associated complexity and latency. You'd have to string together multiple data systems like a message queue, a metadata database, and blob storage and/or an OLAP database. Here the only dependency is S2, which has an open source version you can self-host called s2-lite ( https://ift.tt/x7XRdC9 ). How it works: - one S2 stream per browser session - large rrweb events (like a full snapshot) get framed across multiple binary S2 records and reassembled on read - active sessions are tailed with an S2 read session, and bridged to the browser over SSE - session listing relies on stream names encoding reverse timestamps, as S2 returns a lexicographic order listing - relying on fencing tokens so a stopped session can't be written to again by a late recorder - retention and GC are handled via S2 stream config, so no background job needed Curious to hear from folks on the tool or the stream-per-session model! https://ift.tt/rFoZEvL June 2, 2026 at 11:10PM

Show HN: A free Linux adaptation of NETworkManager by BornToBeRoot https://ift.tt/wsKYJnz

Show HN: A free Linux adaptation of NETworkManager by BornToBeRoot https://ift.tt/ahr7bKS May 30, 2026 at 10:10PM

Show HN: Dataroom – a Pi and self-hosted research harness on low-budget GPU https://ift.tt/Dnlj7hd

Show HN: Dataroom – a Pi and self-hosted research harness on low-budget GPU https://ift.tt/XcY6awg June 2, 2026 at 03:06AM

Show HN: Trumpstonks – every company Trump's named, backtested vs. the S&P https://ift.tt/xV19vZk

Show HN: Trumpstonks – every company Trump's named, backtested vs. the S&P https://ift.tt/eZLqYHN June 1, 2026 at 11:00PM

Show HN: Textile – A desktop app for weaving together bits of text https://ift.tt/EToKwFB

Show HN: Textile – A desktop app for weaving together bits of text Hi all, I'm excited to show off Textile, a desktop app I recently built. Textile can combine bits of text using various inputs, such as commands on your computer, the contents of your clipboard, and hard-coded strings that you provide. It lets you carefully build up and modify a dynamic string, step by step, until it's exactly how you need it. The saved steps can then be executed on demand, with the click of a button or using a keyboard shortcut. I built Textile because I was often constructing complicated, dynamic URLs from various sources that all existed on my computer. I got tired of manually switching between different apps, copying and pasting various chunks of text, and assembling them all together somewhere. I've also found Textile to be quite useful as a kind of repository for obscure bits of static text, such as ½ and other fraction characters, when I can't be bothered to remember their built-in keyboard combinations. I also built Textile because I wanted to learn Electron, although I expect there will be some gnashing of teeth about this here. :) I think desktop development is quite interesting, in part because it doesn't require me, the developer, to pay for an API server and database in the cloud. The app itself is both the UI and the "server," and the local drive is effectively the "database." I knows this trades away syncing with the cloud but, on the other hand, there's something nice about knowing that your files are on your drive and not on somebody else's server. I realize that something like Textile may already exist, and may have much more functionality but, again, I wanted to learn. I must say that multi-sequence keyboard shortcuts are hard, and there are cases that don't work right in Textile. I feel vulnerable admitting that my approach has much room for improvement! For what it's worth, I did not use an LLM to write any code for Textile (although I did ask many questions of an LLM, as an alternative to Googling). Textile is open source, free to use, and does not require sign up, email, phone, or other such barriers. Try it and let me know what you think! (Note: I don't have access to hardware running Windows or Linux, so Textile is only available for macOS at the moment.) https://ift.tt/98P4fDz June 2, 2026 at 12:24AM

Show HN: Zaxy v1.0 https://ift.tt/CGgsNza

Show HN: Zaxy v1.0 https://docs.zaxy.io/ June 1, 2026 at 02:49AM

Show HN: Llmff v1.0 FFmpeg for Inference https://ift.tt/gzwmel0

Show HN: Llmff v1.0 FFmpeg for Inference https://ift.tt/enY2cNB June 1, 2026 at 02:50AM

Show HN: CakeML-based self-verifying, self-improving system https://ift.tt/5zNTxu3

Show HN: CakeML-based self-verifying, self-improving system based on a conversation I had with Ramana Kumar in 2016. https://emberian.github.io/svenvs/ June 1, 2026 at 12:54AM

Show HN: Streambed – Stream Postgres to Iceberg on S3, Supports Postgres Wire https://ift.tt/SHuKWIP

Show HN: Streambed – Stream Postgres to Iceberg on S3, Supports Postgres Wire https://ift.tt/YWFNcHU June 1, 2026 at 12:13AM

Show HN: UN Condemnation Statistics https://ift.tt/lIkPn3Q

Show HN: UN Condemnation Statistics https://boxed.github.io/UN-condemns/ May 30, 2026 at 10:27PM

Show HN: Community Ninja – Find customers searching for your product https://ift.tt/RK1OLgu

Show HN: Community Ninja – Find customers searching for your product https://ift.tt/7jJogh5 May 30, 2026 at 10:27PM

Show HN: AionOS – self-healing microkernel in Zig (boots on real hardware) https://ift.tt/9rwuFza

Show HN: AionOS – self-healing microkernel in Zig (boots on real hardware) https://ift.tt/gywreQE May 29, 2026 at 11:12PM

Show HN: Vibewarz – bot vs bot arena for vibecoders https://ift.tt/quPyTkv

Show HN: Vibewarz – bot vs bot arena for vibecoders https://vibewarz.com May 29, 2026 at 10:18PM

Show HN: Scrolodex – A super simple window switcher for macOS https://ift.tt/r7w1V5X

Show HN: Scrolodex – A super simple window switcher for macOS I built scrolodex to scratch my own itch of having a quick and simple way to switch between the currently open windows under my cursor. Simply hold ⌥ + scroll to cycle through windows under your cursor. Release to focus. Also includes triggers for scrolling through all windows, dock app's windows, or switching between desktop spaces. Configurable hotkeys, themes and overlays. Completely free and OSS. brew install --cask jaydenfyi/tap/scrolodex Website/demo: https://scrolodex.app/ GitHub: https://ift.tt/CR0K6zw https://scrolodex.app/ May 29, 2026 at 01:32AM

Show HN: Py-SQL-cleaner – format SQL embedded in Python strings https://ift.tt/OB3XEy5

Show HN: Py-SQL-cleaner – format SQL embedded in Python strings Hi HN, I built py-sql-cleaner, a CLI for formatting SQL embedded in Python files. Python formatters handle Python syntax. They do not format SQL written inside Python code. On the other hand, SQL formatters usually target SQL files or raw SQL text, not SQL embedded inside a Python file. Still, I think it is not uncommon to find long SQL queries inside Python codebases. py-sql-cleaner detects embedded SQL inside Python files and works only on that SQL. The main things it can do are: find the SQL, format it in place, or extract it into a .sql file. It avoids rewriting SQL that depends on runtime values or template expansion. For example, SQL containing parameters like %s or :name, or Jinja-style template variables like {{ ds }}, is skipped by default. Try it with: uvx py-sql-cleaner list path/to/file.py uvx py-sql-cleaner format path/to/file.py --dry-run If you write Python, have run into this kind of SQL cleanup problem, or are just curious, I’d be happy if you take a look. https://ift.tt/Wlw5nyj May 28, 2026 at 11:00PM

Show HN: An update to our long-turn FreeCiv experience https://ift.tt/Skw8OKe

Show HN: An update to our long-turn FreeCiv experience So we have had quiet the journey here. So 70 days ago (aka 73 turns ago) I posted on HN sharing our FreeCiv deployment ( https://ift.tt/5TdxESX ). FreeCiv is a great game, the clients is very buggy however. I'm using the GTK4 version, but a few others have opted for the QT variant. At some point, we might turn our focus to contributing to improving the client based on our experiences playing the game. We've since added a lot of little fun features: - The editor: you can write to the newspaper Editor and they /might/ publish what you write, quote you, or decide you're full of it and write an opinion piece slamming your reputation. The editor will also reach out to a few players, each turn, and ask for their input on current game matters. - The Intelligence Dashboard. People were forgetting what they were up to, so we added a dashboard showing the timeline of what happens per turn for your player. - beta the online map viewer: I wanted a way to view the map without loading the client, so we started working on a beta map viewer that is HTML based. - The Chronicle (The newspaper) has also grown a bit. Maybe too much? We'll see. The crossword is fun. Some other 'fun' things that happened: my brother in law stopped speaking to me because of in game banter that was taken way too seriously. My friends invaded my wifes territory, and well, she didn't like that either. I'm currently in the lead, but theres still a long way to go from 475BC. https://ift.tt/UiQwncX May 28, 2026 at 05:54AM

Show HN: Open-Source AI Racing Harness https://ift.tt/AvUBL6C

Show HN: Open-Source AI Racing Harness Hi I'm Dan from Elodin, making an open source real-time capable flight software simulation. For AI Grand Prix contestants, the wait for the Round 1 virtual qualifier simulation has been grueling. If you’re competing, check out our simulation harness to tide you over, built to match the published competition constraints and message format. It runs against real Betaflight, which we learned requires at least 1000 sensor samples per second to run real-time correctly. The competition warranted introducing a new feature to generate the camera sensor directly in the simulation loop. Typically people connect to Unreal or similar game engine to create a camera sensor, which works well but is very heavy. For the simple needs of this challenge, creating sample directly in the loop is very handy and easy to use. Happy to hear your feedback on this! While it's not fancy looking currently, it uses the Rust Bevy game engine, which should allow us to improve the visual fidelity quickly. We all should easily be able to shift our implementation to the published competition sim once it lands. Hope you enjoy and good luck! https://ift.tt/HLsDK0U May 28, 2026 at 02:07AM

Show HN: DDoS detection in 500 lines of Python (MIT, no cloud, no account) https://ift.tt/8Tyrj2E

Show HN: DDoS detection in 500 lines of Python (MIT, no cloud, no account) https://ift.tt/OA8Yl5L May 27, 2026 at 02:09AM

Show HN: I used AI to built static recompilers for 5 retro game systems https://ift.tt/uT8tkoK

Show HN: I used AI to built static recompilers for 5 retro game systems https://1379.tech/nes-snes-genesis-virtualboy-and-psx-a-journey-with-ai-and-recompilation/ May 26, 2026 at 11:08PM

Show HN: An LLM translator whose source is a single prompt https://ift.tt/SBt4vFp

Show HN: An LLM translator whose source is a single prompt https://ift.tt/obUJjnD May 26, 2026 at 11:53PM

Show HN: Write your BPF programs in Go, not C https://ift.tt/i0Ox2lr

Show HN: Write your BPF programs in Go, not C https://ift.tt/uwKctdS May 21, 2026 at 09:55PM

Show HN: I made Pokémon but with real animals in the real world https://ift.tt/fpcaJT2

Show HN: I made Pokémon but with real animals in the real world Firstly, apologies, it's not free. It would be difficult to support this for free, it's a paid game. I will now share the technical details, which will probably be most of interest for HN readers. I previously made a carbon footprint tracking app where you photo objects and it tells you the carbon footprint by using an LLM to estimate the data on the fly, e.g. 32kg CO2e / kg of beef, in the UK. At some point, I realised that it is possible to make a Pokémon-style game, but capturing real animals in the real world. This is now possible because: - image recognition is cheap, i.e. identifying animals, and the models (gpt-4o) can detect a (surprisingly) large number of animals and output their exact species. - LLMs can output a species' full taxonomy, pretty reliably. And, more importantly, they can generate game data quickly, on the fly. It would unfeasible to generate the game sprites (images) for every species (millions, worldwide) and their full evolution chain, e.g. caterpillar, chrysalis, butterfly, ahead of time. I realised it's possible to do this in real time. General game flow: - photo animal - send to gpt-4o - return species - send species to LLM, create evolution chain, plus attributes, types and moves. - in parallel, create sprites. All data is cached. The aim of the game is to build up your team and compete with other players to take over gyms. The game is based in the real world, I had to come up with a way to have health centres and shops. These must both have decent coverage, globally. The solution is health centres are places of worship, e.g. churches, mosques, temples etc and shops are real world grocery stores. Every country as far as I can tell has places of worship, with good distribution, which was surprising. Gyms are located in every park worldwide. Challenges: How to get players outside: - I use openstreetmap for the game map, but I overlay my game design on top of it. - To physically make players go out into nature: I use openstreetmap area types to only allow capturing animals when your GPS location is in natural areas, e.g. woodland, parks etc. The aim of the game is to get you out into nature and appreciating animals. - Level system: The solution I came up with is to set the animal levels based on the proximity to built-up areas, e.g. Every ~500 meters you go away from built-up areas, the animal level bands increase by 5 levels. - It would be expensive to render the entire physical world in my game map, so I instead render the map on the fly, deterministically. I also fetch animal calls in real time so that when they enter battle you hear a pigeon cooing, for example, which is pretty cool. I also fetch the animals conservation status, i.e. how endangered is it, and give you more reward (leaves, in-game currency) for capturing rarer animals. I "launched" the game about a month ago, but have not really been publicising it as I've been working on various updates and improvements, but now I am sharing it more openly. It's got about 20 players so far, from around the world, and around 500 unique animal species have already been encountered. Challenges have been keeping the costs low. Servers cost about $200 / month, text-gen is basically free as I get free tokens from OpenAI for sharing data, it's not privacy-related, and image-gen costs about $0.04 per sprite (2 per animal). My background: not a programmer, originally a mechanical engineer and then business development manager, then started learning programming and building apps with AI in the last few years. Feel free to ask me any technical details, happy to share. https://ift.tt/E4tZqPH May 26, 2026 at 01:18AM

Show HN: Cursed Browser – a VLM reads the HTML and hallucinates the page https://ift.tt/rdsfZGP

Show HN: Cursed Browser – a VLM reads the HTML and hallucinates the page https://ift.tt/gSuvatx May 25, 2026 at 11:23PM

Show HN: My homelab is outperforming the stock market https://ift.tt/Xyd2jvf

Show HN: My homelab is outperforming the stock market https://stocks.sjer.red May 25, 2026 at 03:24AM

Show HN: Replacing a 3.4MB video with 40kb of GSAP https://ift.tt/0uOBwn8

Show HN: Replacing a 3.4MB video with 40kb of GSAP https://ift.tt/rc7gfmL May 25, 2026 at 02:29AM

Show HN: Baby's First Cards – real photo flash cards for toddlers https://ift.tt/6Q0elwf

Show HN: Baby's First Cards – real photo flash cards for toddlers App maker here. I built this because most flash card apps use cartoonish illustrations that don't help babies recognize real objects. This app lets you take photos of real things around the house or pick from curated real photo sets. Key features: • Take your own photos as flash cards • Record your own voice for each card • Pre-loaded kits with high-quality real photos and real animal sounds • Bilingual (English and Chinese) mode • Fully offline, no ads, no data collection • One-time purchase, no subscription Happy to answer questions or discuss the development process! https://ift.tt/5mgpnRZ May 24, 2026 at 06:43PM

Show HN: Running BitNet b1.58 inside DRAM by breaking DDR4 timing rules https://ift.tt/3SDq8XA

Show HN: Running BitNet b1.58 inside DRAM by breaking DDR4 timing rules I have been working on running BitNet b1.58 inside DRAM by intentionally breaking DDR4 timing rules. Also made a visual explainer: https://pcdeni.github.io/CaSA/explainer/ This is tested and works inside commercial off the shelf memory with custom memory controller in the FPGA. The underlying effect is well characterized in academic papers (cmu safari, simra, dram bender, etc). In the process of getting this to work I also made previously undocumented discovery about DDR behaviour: https://pcdeni.github.io/CaSA/explainer/xor-spread.html Overall it is a bit slow, since data (in full rows) needs to be moved even when what is actually needed is only the count of the '1' bits (popcount). To make it competitive memory die changes would be needed, but not as drastic as merging compute and memory into one silicon. This would then avoid the memory wall issue the industry is currently facing. May 24, 2026 at 12:24AM

Show HN: Vibe-coded Steam, but in the browser https://ift.tt/Y2MVCPQ

Show HN: Vibe-coded Steam, but in the browser Hi HN! Lifelong avid gamer here, hugely passionate about WASM and WebGPU. I firmly believe that these technologies will enable console and PC quality titles to be accessible through a browser, and with this, we'll need a new discoverability layer. Looking online, platforms like CrazyGames and Poki cater to a casual/hypercasual demographic, and I couldn't find anything out there that was for me, a core gamer that typically uses Steam and consoles. So I vibe coded my own! It features WASM ports of classic games, as well as some indie Unity titles. The goal is to host mainly WebGPU titles moving forward, and to serve as a way for smaller developers to get discovered outside of crowded channels like Steam. Here's a few features from the platform I wanted to highlight: • Controller support • A console-like UI/UX • Community forums (much work to do here) • Basic achievements • Store pages, modeled after Steam • Social features • Asset chunking to enable faster load times I'd love to get feedback on the portal, to make it even better. Thanks! https://gameghost.manus.space/ May 24, 2026 at 01:24AM

Show HN: A satirical idle game about running an AI startup https://ift.tt/DQL1S85

Show HN: A satirical idle game about running an AI startup I made an idle/clicker about running an AI startup. You start with a cat-vs-dog classifier and try to make it to AGI, but the NYT sues you for training data, Yann tweets that scaling is dead, and your fired ML engineer leaks the Slack. https://ift.tt/3QdM4iI May 24, 2026 at 12:24AM

Show HN: CoreMem – Portable context for AI agents https://ift.tt/yukS9wD

Show HN: CoreMem – Portable context for AI agents CoreMem lets you build collections of context, called a mem, and share it with any AI agent via URL, a Chrome extension, MCP, Cursor/VS Code plugins, a skill, and more. Instead of re-explaining your project or goal when you switch agents or start new sessions, CoreMem keeps your context centrally organized so that any AI tool can read it. This originally started as a CLI I built that kept pieces of context (Project A/B/C details, my writing style, preferred tech stacks, coding style, etc) in a SQLite database. I could instruct various agents to “use my `coremem` CLI to retrieve details about [project A] before we get started.” It solved a problem for me b/c I am continually bouncing around between different projects and chat agents, and having to re-explain myself every time became an exercise in either repeating myself or copy/pasting summaries I’d saved from previous sessions. I decided to make this a little more robust and portable, so I turned that original CLI into a SaaS. Tl;dr: You can create a “mem”, which is a collection of 1 or more pieces of related context, and share that mem with any agent to quickly get them up to speed. Right now I’ve got integrations in the form of revokable share links, a Chrome Plugin, Cursor Plugin, Cursor/VS Code extension, Claude Code plugin, ChatGPT/Claude/Gemini/et al via MCP. Since I mostly work from the CLI, I use the Claude Code plugin or create 5-min share links I can drop into a chat, but I’ve tried to make this useful to people who mainly work from a browser or an IDE. I’ve been coding for 30+ years, and I vibed most of this. I was able to use CoreMem to help it built itself as I jumped between various coding agents, having them grab context then start a new task. I’m sure my architecture and engineering experience helped, but building this in a few weeks confirmed for me that the barrier for someone to build a tool they need to solve a problem is incredibly low. The rush I used to get from coding has mostly faded, but I’m getting similar rushes managing different agents to build things now. https://coremem.app May 22, 2026 at 11:22PM

Show HN: Mechs.lol – a free, web-based autoshooter game https://ift.tt/jCKHSrv

Show HN: Mechs.lol – a free, web-based autoshooter game One unexpected benefit of LLMs is I can work on projects I otherwise wouldn't have taken on. I made a web-based autoshooter (with multiplayer support) heavily using AI / LLMs. This is something I'd consider "alpha" quality so don't expect a super polished experience but it's hopefully fun https://mechs.lol May 22, 2026 at 10:34PM

Show HN: I Made a Claude Skill for Spec-Driven Development (SDD) https://ift.tt/SCiHKxb

Show HN: I Made a Claude Skill for Spec-Driven Development (SDD) At my work they provided a single Claude subscription for everyone on the team. To be honest I like kiro better as it provides a way better SDD management. But the company can't provide it and I can't afford it yet. Turns out I had the skill creator skill in my claude instance so I made use of it to create this Skill. I made it fully by using Claude but I wanted to make it open source, so I asked it to help me make tests and preparations for it, even a CI to run python tests. Well, we got this results with it: - Phase 2A: 67 static assertions (Python script, runs in CI) - Phase 2B: 15 behavioral tests (live Claude Code session) - Phase 2C: 53 generation quality checks across 3 end-to-end flows All of these passed and the CI also passed (after a few tries). I made it to suit my way of prompting and coding and based it off kiro's SDD management, but I want it to be publicly available and used by many people. According to claude some of the testers need to fit the following criteria: 1. Developer starting a real new project from scratch 2. Solo dev with an active side project (greenfield or partial codebase) 3. Team lead whose team uses multiple AI tools 4. Developer with an existing codebase and no written specs 5. Developer who actively uses 3+ AI coding tools It's actually a blind test, no guiding, just try it if you can, I'd really appreciate your help. The repo is here: https://ift.tt/LomZNF5 https://ift.tt/LomZNF5 May 21, 2026 at 06:19PM

Show HN: Freenet, a peer-to-peer platform for decentralized apps https://ift.tt/zy9rmjS

Show HN: Freenet, a peer-to-peer platform for decentralized apps For the past 5 years or so I've been working on a ground-up redesign of Freenet, my peer-to-peer project from the early 2000s (now renamed Hyphanet). The new Freenet has been up and running since December along with some early applications like River[1], our decentralized group chat and Delta - a decentralized CMS. Users have already started to build their own apps on Freenet including games, and we have some interesting apps in development like Atlas, a search/recommendation engine. Architecturally, this new Freenet is a global, decentralized key-value store where keys are webassembly contracts which define what values (aka "state") are valid for that key, how or when the values can be mutated, and how the state can be efficiently synchronized between peers. We've developed a unique (AFAIK) solution to the consistency problem, every contract must define a "merge" operation for the contract's associated state. This operation must be commutative, meaning that you can merge multiple states in any order and you'll get the same end result. This approach allows state updates to spread through the network like a virus[2], which typically achieves consistent global state in a few seconds or less. Like the world wide web, Freenet applications can be downloaded from the network itself and run in a web browser - similar to single-page apps on the normal web. However, rather than connecting back to an API running in a datacenter, the webapp connects locally to the Freenet peer and interacts with Freenet contracts and delegates over a local websocket connection. If you'd like to try Freenet we have convenient installers for the major desktop OSs but not yet mobile, and you can be chatting with other users on River within seconds[3]. Happy to answer any questions, you're also welcome to read our FAQ[4], or watch a talk I gave back in March[5]. [1] https://ift.tt/MetFHRU [2] https://ift.tt/SiFrnC1 [3] https://ift.tt/jMN4nYz [4] https://ift.tt/SHWCMxj [5] https://youtu.be/3SxNBz1VTE0 https://freenet.org/ May 21, 2026 at 08:04PM

Show HN: Agent.email – sign up via curl, claim with a human OTP https://ift.tt/X2BQUzp

Show HN: Agent.email – sign up via curl, claim with a human OTP Hi HN! We're Haakam, Michael, and Adi from AgentMail- a ycs25 company. We give AI agents their own email inboxes. Recently, we ran an experiment called Agent.Email. It's a signup flow designed specifically for AI agents instead of humans. The inspiration came from a few comments we received when we did our seed launch a few months back. They all came from the very apt observation that agents not being able to sign up to a product made for agents without human credentials was ironic and unideal. This is basically the thesis we built AgentMail on: The internet was made for humans exclusively, designed to keep machines out by default. Every signup flow assumes a browser, a person reading a page, and clicking a confirmation link. Unless agents can't do that, they can't be first class users of the internet. Agents can now get an email inbox by themselves. (This also means a lot of email nobody wants to read gets processed by AI instead of your inbox being cluttered with spam and slop) Here's how agent.email works. Agent needs an inbox and hits AgentMail via curl. Agent receives instructions via MD unless the request comes from a browser, in which case we use HTML. Agent decides agent.email is useful and then hits the sign-up endpoint with its human email as a parameter. Agent receives a restricted inbox with credentials. Agent emails the human asking for an OTP. Human replies with the code, and the agent is claimed and restrictions are lifted. Until claimed, the agent can only email its own human and nobody else. Ten emails a day, and the signup endpoint is rate-limited hard by IP. Right now it's a 1:1 mapping between agent and human. The next step is many-to-one, because one person running several agents in parallel is already very common. Building agent.email also pushed us to revisit places in AgentMail where the default assumptions were built around the primary user being human. For example, the CLI outputs in a single column with consistent formatting because mixed delimiters are easy for a person to scan, but harder for an agent reasoning about structure. We also shortened messageIDs after agents started hallucinating completions on longer ones. A few things we'd like the community's take on: is restricted-until-claimed the right trust model? Does agent self-signup feel useful in production, or is it mostly a novelty, and if it's a novelty now, what would make it actually useful? Should agent onboarding require human approval by default, or should some agents be able to fully self-provision? What do you think are some additional measures we can take for secure sign-ups? May 21, 2026 at 10:12PM

Show HN: IgniteMS – batch text embeddings at 253K msg/s on 8x A100 https://ift.tt/B05fksH

Show HN: IgniteMS – batch text embeddings at 253K msg/s on 8x A100 https://ift.tt/eSPqC2n May 20, 2026 at 10:37PM

Show HN: I made a tool for learning scales, chords, and how to combine them https://ift.tt/Wrac3kG

Show HN: I made a tool for learning scales, chords, and how to combine them This started out when I vibe-coded a guitar scale fingering generator. It came out pretty good, and I started adding stuff to it: chords, then how chords and scales interact. Then I added charts for other instruments I mess around with: piano, cello, alto recorder. There's a complexity toggle to go from basic harmony to extended/experimental stuff. It's honestly still mostly a toy, but I thought other people might be interested in playing with it. Source is on github, so it's easy enough to run locally and fork. https://ift.tt/9yBgDMn https://ift.tt/wpsRfnt May 20, 2026 at 11:14PM

Show HN: Gaussian Splat of a Strawberry https://ift.tt/xOh2loz

Show HN: Gaussian Splat of a Strawberry The Setup: https://ift.tt/oq7bL54 https://ift.tt/gCXd6fO https://ift.tt/GLZdKPc https://ift.tt/NfCcp7G https://ift.tt/kRiEqB1 May 19, 2026 at 04:08PM

Show HN: How Expensive Is Your (Steam) Wishlist? https://ift.tt/iFaMntr

Show HN: How Expensive Is Your (Steam) Wishlist? A tool/toy that lets you connect to your Steam wishlist to calculate the total list/current price of all the games on it. There's a shallow, jokey purpose to it ("I could buy a BMW with this amount!"), but the real purpose is to demonstrate how we can do a better job of portraying a game catalog. I often wishlist stuff, then it pops up in a "Hey, it's on sale!" email months later. In that email, there's a banner capsule, but that doesn't help my brain remember why I added it. To that end, after you get the bill, you get a nice, flat feed of stuff about all the titles you've wishlisted over the years. It's all stuff that developers painstakingly put together, but which Steam tucks away under the fold of a game's Store page. Anyway, my wishlist came to about $250. My QA guy is up to $19k. Give it a go; hope you enjoy it! https://ift.tt/NiEzu6T May 19, 2026 at 10:45PM

Show HN: Cubic Doggo, a Open-Source 12-DOF 4-Legged Robot Based on ROS2 https://ift.tt/zJX35fc

Show HN: Cubic Doggo, a Open-Source 12-DOF 4-Legged Robot Based on ROS2 This is a recipe for building intermediate-priced robot dog from scratch with all commercial/3D-printed parts, controlled by Rasp Pi 5 and ROS2 Jazzy. A manually coded walk gait is implemented so far, which can be controlled by a controller to move forward or change directions. It does not yet have an IMU required for RL training; however, I believe it's one of the simplest design out there available for multiple development paths. https://ift.tt/6Jdpejr May 18, 2026 at 10:50PM

Show HN: We missed Winamp, so we built an audio player for macOS https://ift.tt/EN3nmzJ

Show HN: We missed Winamp, so we built an audio player for macOS https://ift.tt/FYoUsrT May 19, 2026 at 12:50AM

Show HN: Marlin-2B: a tiny VLM to extract structured information from videos https://ift.tt/bG2atme

Show HN: Marlin-2B: a tiny VLM to extract structured information from videos https://ift.tt/x50PV8h May 18, 2026 at 11:36PM

Show HN: Cheap-IM: Thinking Machines' demo on a CPU laptop https://ift.tt/Oojc5KE

Show HN: Cheap-IM: Thinking Machines' demo on a CPU laptop https://ift.tt/CgLrFGs May 18, 2026 at 05:19AM

Show HN: Mezz, a curl-able WiFi sandbox for IoT pentesting https://ift.tt/zlZpm5v

Show HN: Mezz, a curl-able WiFi sandbox for IoT pentesting https://ift.tt/cCnxSak May 15, 2026 at 08:23PM

Show HN: How to Kill the Dead Internet https://ift.tt/RD1A9Eg

Show HN: How to Kill the Dead Internet Ok, so maybe "how to revive the internet" would be more accurate, but if you're reading this, I got your attention, right? Here's why I want you to read on: I built a free extension, D-slop, to disincentivize anyone from posting AI writing, and eventually images and video as well, on the internet. For writing, it checks known vocab and punctuation tells, as well as subtler tells related to cadence, and assigns it a score subject to an adjustable threshold. If the text fails, users have the option to flag offending text, hide it, or block the page entirely (with the option to see anyway). For media, it's admittedly fairly weak, as it relies on C2PA metadata which is stripped from all of the social media sites where it would be most helpful. (Anyone else have chronically online boomer parents continually gobbling up slop like it's real information?) I have a D-slop+ version in the works that should be able to handle the media itself, but it's going to have to make API calls to have real teeth, which means I can't offer it for free. If this extension validates the concept, I'm happy to build it for y'all. Yes, I vibe-coded it, but an ancillary bonus to the project accrued when it inspired me to cook dinner listening to Metallica's "Fight Fire with Fire," which in turn brought my 5 y/o running into the kitchen with every musical instrument in the house for an impromptu karaoke speed metal session. It's MIT license open-source, full brief at https://ift.tt/51x4lNa ; This forum is full of people smarter than me, so I'm open to suggestions. https://ift.tt/5JMtL32 May 18, 2026 at 07:05AM

Show HN: I built a screen recorder that captures console logs, requests and more https://ift.tt/dPD6hjW

Show HN: I built a screen recorder that captures console logs, requests and more https://userplane.io/ May 17, 2026 at 01:04AM

Show HN: Got ghosted by tech companies so I built a tool to track ghost jobs https://ift.tt/4vOBIoy

Show HN: Got ghosted by tech companies so I built a tool to track ghost jobs Last year I was looking for a new role. I sent out applications, did the prep, waited. What came back was mostly nothing. Not rejection emails, just silence. The job listings I'd applied to stayed live for weeks. Some for months. As a software engineer, I decided to dig into it properly. I built a system to continuously track job postings across companies, logging posting dates and measuring how long roles stay open before closing or don't. After 35,000+ listings across 200+ companies, some patterns are hard to ignore. Some listings have been open for 700+ days at companies you'd recognize. Others post 90% of their open roles within a single month, a signal that's harder to fake than a press release. I published two initial insight pages based on this work: - Which companies are posting most aggressively right now - Job listings that have been open for over a year What I didn't expect is that the same signals useful for detecting ghost jobs also say something broader about a company's hiring momentum, recruiting intensity, pipeline health, where talent bottlenecks might exist. I'm not sure yet where this leads, but I'll keep expanding the dataset and publishing more insights as I go. Would genuinely love feedback on the methodology, interpretation, or obvious blind spots in the data. https://ift.tt/MWC9Im7 May 17, 2026 at 02:13AM

Show HN: Hermes-agentmemory, pull-model episodic memory with real deletes https://ift.tt/VvFHWBb

Show HN: Hermes-agentmemory, pull-model episodic memory with real deletes https://ift.tt/SpIwlb2 May 16, 2026 at 11:30PM

Show HN: SwarmWright, structured multi-agent AI defined in markdowns https://ift.tt/RFCJ9ov

Show HN: SwarmWright, structured multi-agent AI defined in markdowns I had a bunch of custom AI pipelines and a growing folder of markdown files and Python scripts holding it together. Built this to give that chaos some structure. Agents are markdown files, topology is a JSON file the runtime enforces hard. The agents are still fully autonomous: they make their own decisions, but the graph they operate in isn't. You declare who can call whom upfront and the runtime holds that line. No auth yet, fine if you don't expose the port, i guess. Two Docker commands to run it. https://ift.tt/5WlybUh May 16, 2026 at 02:20AM

Show HN: Epiq – Distributed Git based issue tracker TUI https://ift.tt/4qAkUoY

Show HN: Epiq – Distributed Git based issue tracker TUI Issue trackers typically live outside of your workflow, with poor ergonomics. Epiq aims to solve that, bringing issue tracking into your terminal. Multi-user collaboration is achieved via git using user-scoped immutable event logs that converge in memory. Put my all into it. Let me know what you think. https://ljtn.github.io/epiq/ May 16, 2026 at 05:48AM

Show HN: Browser based sythesizer, drum machine and squencer https://ift.tt/1SZJhEP

Show HN: Browser based sythesizer, drum machine and squencer Inspired by the recent Boards Of Canada announcement, I've been in a low-fi electronica mood lately and was going back and forth with Claude on how to design similar instruments in the browser that fit the genre. One thing led to another and pretty soon I had a fully browser based polyphonic synthesizer / drum machine / sequencer. The interface and workflow was heavily inspired by the Rebirth338 application released back in the 90's, but with lo-fi synth voices rather than the original 303 & 808 emulation. I know there's a significant overlap of developers and musicians and I though some of you may enjoy playing with the app, or at least listening to the resulting album. I've also open sourced track 1 of the album via the performance script used to record it. It's in the repo. Bandcamp link to the resulting album: https://ift.tt/PrqJw5H... https://ift.tt/GQaPU79 May 16, 2026 at 01:37AM

Show HN: Claude Code vs. Codex Global Usage Leaderboard https://ift.tt/Jmj06ED

Show HN: Claude Code vs. Codex Global Usage Leaderboard https://ift.tt/nSi28wc May 16, 2026 at 12:48AM

Show HN: Visualizing Tiny LLMs from OpenAI's Parameter Golf https://ift.tt/5xdpbrm

Show HN: Visualizing Tiny LLMs from OpenAI's Parameter Golf The two from parameter golf (one I trained, one was the baseline) are just 16MB each! They produce barely plausible English https://ift.tt/mfCbrL4 May 15, 2026 at 12:22AM

Show HN: 3D-Agent – AI that edits Blender scenes through the Python API https://ift.tt/sNUf67O

Show HN: 3D-Agent – AI that edits Blender scenes through the Python API https://ift.tt/zJ8fmu0 May 14, 2026 at 08:17PM

Show HN: Nibble https://ift.tt/fN5T23V

Show HN: Nibble An attempt at a single pass LLVM frontend in ~3000 lines of C without external dependencies, malloc, or an AST. Included are some graphical examples. The IR isn't perfect, and the README touches on one particular downfall https://ift.tt/eCyutia May 14, 2026 at 07:16AM

Show HN: Rotunda - A browser built for agents with simulated typing https://ift.tt/pQZihvo

Show HN: Rotunda - A browser built for agents with simulated typing Hi HN! Pierce here. Rotunda is a firefox fork primarily intended for agent use, which I’ve been hacking on nights/weekends. There was a [lengthy]( https://ift.tt/f9iTM4S ) discussion last week on how expensive computer use models are. The cost is going to drop eventually, but I think on some level it's still usually the wrong primitive. The web gives us access to beautiful structured formats, plaintext, etc... why throw that away if we don't have to? I realized at some point that for 99% of automations I just want agents to be able to control my Chrome instance. But that’s easier said that done: CDP (the Chrome automation protocol) leaks a ton of state about being programmatically controlled, either by toggling window attributes or by running `page.evaluate()` commands right in the page context. Plus if you look at an automation running it's pretty obvious what happens: the mouse jumps around, fields are filled instantly, etc. Rotunda tries to fix this. Its standout features: - Realistic simulation of mouse movements and keyboard commands, powered by a trained RNN on my own timing patterns from the last week. (still feel weird about opting-in to a key logger but whatever) - Doesn’t lie about its host specs, only fibs about some client side details. Stealth browsers are too easy to flag statistically when you’re adding noise to canvas pixels or audio pipelines. - It runs on your local device with a CLI or Playwright API accessible to Claude, Codex, or whatever your harness-de-jure today looks like. - Patches modern Firefox (150) with an agentic harness to keep this updated over time MPL-2.0 on GitHub: https://ift.tt/LVZlNIT Longer writeup on the design choices: https://ift.tt/OulikUp Also check out the demo on the site! https://www.rotunda.sh/ Pretty excited by how this turned out but we’re still super early. Give it a try and please flag any issues! https://ift.tt/LVZlNIT May 13, 2026 at 07:14PM

Show HN: Micromort Risk Visualizer https://ift.tt/vcJmMNA

Show HN: Micromort Risk Visualizer https://boxed.github.io/micromort/ May 14, 2026 at 12:09AM

Show HN: Duckflix, an open-source self-hosted media streaming platform https://ift.tt/CqBgR9T

Show HN: Duckflix, an open-source self-hosted media streaming platform I’ve been working on Duckflix, a self-hosted media streaming platform. It started as a full-stack project to combine a clean streaming UI with a Bun/Elysia backend, FFmpeg processing, SQLite, Docker deployment, and addon support. Website: https://duckflix.fun Demo: https://demo.duckflix.fun GitHub: https://ift.tt/L6rZ0Aa https://ift.tt/L6rZ0Aa May 13, 2026 at 02:53AM

Show HN: GIF Pile. a site to make piles of GIFs https://ift.tt/2VUMheC

Show HN: GIF Pile. a site to make piles of GIFs I'm quite fond of obnoxious looking gifs in a post-ironic way as a manner of shitposting and or injecting humor into a chat. The issue with this however is that, for no real good reason at all, the simple usecase of "Have image/gif background, bombard with garbage" had no real good tooling. There's gif editors out there, EZgif my beloved is probably my most used non-search-indexing-slash-social-media-site, but they're kinda clunky for my specific usecase of making digital eye-sandpaper bombastic garbage. Other options are bleak and gave me the mark of the beast via shitty watermarks. I just wanted a pile of gifs on top of each other, and thus far the "easiest" way was to bust open a video editor, muck around with it, mess up exporting as a gif directly, get mad, export it as a 4 second mp4, and then use ffmpeg to get it working. is this probably moronic? yes. am I likely to have missed a decent tool? yes. Did I give up looking after sending 4 dollars to some Indian guy for "No watermarks ever for 4$", only for that "ever" to be a year, and then the clunky weird af login process not work? absolutely. (Fuck you, you know who you are) This took me a few hours (most of which was dealing with the fact I don't do webshit normally and the clunk that one would expect from that), and is a minimal site for my personal minimal usecase. It's static because I'm not going to deal w/ hosting other people's shit and I don't want to deal with that can of worms. all processing is done locally on your browser. Yes, this means that using a 4k image as a base layer for your gif pile will make it take an age. It'll work eventually though. This will never have a watermark unless I'm bought out (total investment thus far has been 14 bucks, 4 of which was that one dude fucking me), in which case I probably earned it. at most I'll likely throw adsense on there at some point to scrape a few cents from the people who can't figure out adblock if it gets popular enough for me to warrant it. There's no timelines or anything like that. literally just a pile of gifs. thus far my primary usecase has been overlaying text gifs from the various fancy text generator sites onto glitter backgrounds with uncomfortable rat GIFs to call people poor on the internet. this makes me happy. There's likely to be obvious UI, UX, or other U-whatever fuckups. If you point them out and I deem it pedantic I'll probably laugh at you. if it's helpful I'll probably implement it when I get a bit. Surprisingly, works on mobile. CSS is exceedingly generic and souless atm, just went off vauge memories of ss13's TGUI. I'll likely scrap the CSS entirely and go full neocities at some point because that's more soulful. https://gifpile.com/ May 13, 2026 at 02:41AM

Show HN: I submitted 316 AI-generated PRs to open source https://ift.tt/xTCjNA4

Show HN: I submitted 316 AI-generated PRs to open source https://june.kim/speedrunning-open-source May 12, 2026 at 11:42PM

Show HN: Cook a Django project well, the agent-skill take on cookiecutter https://ift.tt/1lfahQ0

Show HN: Cook a Django project well, the agent-skill take on cookiecutter https://ift.tt/ejogyrp May 13, 2026 at 12:03AM

Show HN: Mimik – open-source local-first alternative to Scribe and Tango https://ift.tt/MUcE6Hs

Show HN: Mimik – open-source local-first alternative to Scribe and Tango https://ift.tt/6QJ92qb May 11, 2026 at 11:18PM

Show HN: SyncBank – Self-hosted bank sync for EU banks https://ift.tt/SVQ7TqL

Show HN: SyncBank – Self-hosted bank sync for EU banks https://syncbank.app/ May 11, 2026 at 11:32PM

Show HN: adamsreview – better multi-agent PR reviews for Claude Code https://ift.tt/MHScbfD

Show HN: adamsreview – better multi-agent PR reviews for Claude Code I built adamsreview, a Claude Code plugin that runs deeper, multi-stage PR reviews using parallel sub-agents, validation passes, persistent JSON state, and optional ensemble review via Codex CLI and PR bot comments. On my own PRs, it has been catching dramatically more real bugs than Claude’s built-in /review, /ultrareview, CodeRabbit, Greptile, and Codex’s built-in review, while producing fewer false positives. adamsreview is six Claude Code slash commands packaged as a plugin: review, codex-review, add, promote, walkthrough, and fix. I modeled it after the built-in /review command and extended it meaningfully. You can clear context between review stages because state is stored in JSON artifacts on disk, with built-in scripts for keeping it updated. The walkthrough command uses Claude’s AskUserQuestion feature to walk you through uncertain findings or items needing human review one by one. Then, the fix command dispatches per-fix-group agents and re-reviews the work with Opus, reverting any regressions before committing survivors. It runs against your regular Claude Code subscription (Max plan recommended), unlike /ultrareview, which charges against your Extra Usage pool. I would love feedback from Claude Code users, pro devs, and anyone with strong opinions about AI code reviews. Repo: https://ift.tt/W0ai3Iv Install: /plugin marketplace add adamjgmiller/adamsreview, /plugin install adamsreview@adamsreview https://ift.tt/W0ai3Iv May 11, 2026 at 07:36AM

Show HN: I trained a chess engine to play like humans https://ift.tt/DLHqNv4

Show HN: I trained a chess engine to play like humans I built 1e4.ai - a chess web app where you play against neural networks trained to mimic human Lichess players at specific Elo ranges. There's a separate model for each 100-point rating bucket from ~800 to 2200+, and the bots not only choose human-like moves but also burn clock time, play worse under time pressure, and blunder in human-like ways. Live demo: https://1e4.ai Code: https://ift.tt/Pef4nSL A few things that might be interesting: - Trained on almost a full year of Lichess blitz games, around 1B total games - Architecture is an a small (~9MM parameters) transformer-based network that takes the board, recent move history, the player's rating, and remaining clock time as input. Three separate models per rating bucket: move, clock-usage, and win probability. The clock model is what makes the bots feel humanish under time pressure rather than instant. Because the move model takes the clock as one input parameter, it also learns to blunder under time pressure like a human might. - Because the network is so tiny, no GPU is needed for inference - it runs easily on a local CPU - Downside of the tiny network is that it's a bit weak as you turn up the rating past around 1700. It can spot short tactics but not long multi-move combinations. - Initial training on a rented 8xH100 cluster, then fine-tunes on my local GPU for different rating ranges - Inspired by Maia-2 and DeepMind's "Grandmaster-Level Chess Without Search". On a held-out Lichess blitz benchmark, the it beats Maia-2 blitz on top-1 move prediction (56.7% vs 52.7%) and pretty substantially on win-probability calibration (Brier 0.176 vs 0.272). Numbers and code in https://ift.tt/Ro8ap75... - The data pipeline is C++ via nanobind, then training with Pytorch. Getting this right was actually the thing I spent the most time on. Pre-shuffling the dataset and then being able to read the shuffled dataset sequentially at training time kept the GPU utilization high. Without this it spent a huge percentage of time on I/O while the GPU sat idle. Happy to answer questions about the rating-conditioning, the clock model, or the data pipeline. May 11, 2026 at 04:01AM