2026年4月6日 的 Show HN

Built a ~9M param LLM from scratch to understand how they actually work. Vanilla transformer, 60K synthetic conversations, ~130 lines of PyTorch. Trains in 5 min on a free Colab T4. The fish thinks the meaning of life is food.

Fork it and swap the personality for your own character.

I built this because I wanted to see how far I could get with a voice-to-text app that used 100% local models so no data left my computer. I've been using a ton for coding and emails. Experimenting with using it as a voice interface for my other agents too. 100% open-source MIT license, would love feedback, PRs, and ideas on where to take it.

I've long been into finding deals on government auction sites (seizures, surplus sales etc.) - right now for example San Diego DHS is selling 26 tons of lead shot, with bidding starting at $1,000 ¯\_(ツ)_/¯

It has historically been extremely tedious though: scanning dozens of janky sites which have interminable page loading times; back buttons take you all the way back to the homepage etc.

The site I built - GovAuctions - lets you search every government surplus auction at once. You can filter by location, category, and price, save items to a watchlist, and get alerts when new auctions match what you're looking for.

Let me know what you think, if you have any suggestions, and if you find any deals in your area!

Gemma Gem is a Chrome extension that loads Google's Gemma 4 (2B) through WebGPU in an offscreen document and gives it tools to interact with any webpage: read content, take screenshots, click elements, type text, scroll, and run JavaScript.

You get a small chat overlay on every page. Ask it about the page and it (usually) figures out which tools to call. It has a thinking mode that shows chain-of-thought reasoning as it works.

It's a 2B model in a browser. It works for simple page questions and running JavaScript, but multi-step tool chains are unreliable and it sometimes ignores its tools entirely. The agent loop has zero external dependencies and can be extracted as a standalone library if anyone wants to experiment with it.

Did this a few years ago. Seems apropos. Sources and more here: https://github.com/EngineersNeedArt/Mooncraft2000

TrueType fonts have a hinting VM that grid-fits glyphs. It has a stack, storage area, conditionals, function calls, and it turns out it's Turing-complete. So I built a raycasting engine in the hinting bytecode.

The glyph "A" in the font has 16 vertical bar contours. The hinting program reads player coordinates from font variation axes via GETVARIATION, does DDA ray marching against a tile map in the storage area, and repositions bar heights with SCFS. It ends up looking like a crude Wolfenstein-style view.

Small visuzlization: https://github.com/4RH1T3CT0R7/ttf-doom/blob/main/docs/media...

About 6.5 KB of bytecode total - 13 functions, 795 storage slots, sin/cos lookup tables.

JS handles movement, enemies, and shooting, then passes the coordinates to the font through CSS font-variation-settings. The font is basically a weird GPU.

The weirdest parts: - TrueType MUL does (ab)/64, not ab. So 1*4=0. The DIV instruction is equally cursed. - No WHILE loops. Everything compiles to recursive FDEFs. FreeType limits call depth to ~64 frames. - SVTCA[0] is Y, SVTCA[1] is X. Of course.

There's a small compiler behind this - lexer, parser, codegen - that turns a C-like DSL into TT assembly.

Demo GIF: https://github.com/4RH1T3CT0R7/ttf-doom/blob/main/docs/media...

Live demo: https://4rh1t3ct0r7.github.io/ttf-doom/ (Chrome/Edge, WASD+arrows, Space to shoot, Tab for debug overlay)

This is a DOOM-style raycaster, not a port of the original engine - similar to DOOMQL and the Excel DOOM. The wall rendering does happen in the font's hinting VM though. Press Tab in the demo to watch the font variation axes change as you move.

Hello HN,

I live in northern part of USA where winters are snowy. Whenever I took long trips, I always wondered what the weather along the route is going to be. For those who live in northern USA, you know weather can change frequently, so when you're traveling matters a lot, not just the route.

To solve this problem for myself, I built https://navimodo.com/. NaviModo calculates the route, and then checks weather along the route based on your start time, and displays weather along the route. Change start time, and the whole thing is recalculated.

I am not expecting any commercialization for this, just wanted to scratch an itch, and just did it. I have ideas for adding additional features (suggestions for when to take breaks based on bad weather, auto-suggestion of start time to avoid bad weather, etc.,) but will add more as time goes by.

Any feedback is welcome!

I had a lot of time on my hands so I spent Sunday afternoon coming up with a bunch of completely impractical ways to visualize clocks.

PG has posted about improving social networks using something like an "intellectual CAPTCHA" many times [1][2][3][4] - "Make users pass a test on basic concepts like the distinction between necessary and sufficient conditions before they can tweet."

I felt the same way. So I built one using a mix of simple math, logic, and Twitter/X Community Noted posts. Try sample questions here - https://mentwire.com/sample - without signing up.

- Invites are temporarily open to HN users.

- Onboarding test + one daily question before accessing feed, post or reply.

- Posts authors are anonymous until upvoted or downvoted, forcing evaluation of content on merit.

- Face ID (on-device only) to post/reply, pangram checks for AI text.

Sourcing good questions turned out to be much harder than I thought. If you have suggestions to scale this, I would love to hear. Eventually, could be gated across disciplines/topics to get a competence × interest graph instead of the pure interest graph of today's social networks.

[1] https://x.com/paulg/status/1235949761359904768 [2] https://x.com/paulg/status/1576517990182359040 [3] https://x.com/paulg/status/1514979883948126209 [4] https://x.com/paulg/status/1505842647319126016

Hey HN,

I built Pace, a Claude Connector that lets you connect all your wearables (Garmin, Polar, Whoop, 20+) with Claude.

You connect your devices once and can analyze your data with Claude. No Dashboard needed, just in natural language. I already use it everyday, especially the visualization tool in Claude makes this really cool to use.

Tools include: overview, sleep, training, activity, samples and trends.

Tech Stack: Python (FastMCP), Google Cloud Run, Google Cloud SQL (PostgreSQL) and Firebase

It is live and free to try (no Claude Pro/MAX Plan needed)..

Would love feedback, especially from people who've built MCP servers or use wearables seriously.

I've gotten to the point in my career where I now make strategic decisions often (hiring, firing, choosing what equipment to go with, etc.), as well as in my personal life where I need to strongly weigh my options for a big purchase or investment. I found a not-so-surprising parallel between the two as these decisions "resolved." Am I making good decisions or am I getting lucky?

Did some research, read some books, and realized I should get in the habit of tracking my decision process. That quickly turned into the idea that formed Convexly.

The landing page is a 10-question calibration quiz where you assign a confidence level to statements drawn from a rotating pool of 100 (working on making the pool larger) and you get a Brier score back instantly. No signup required, and you can share your scores right away.

If you find it interesting, you can create a free account where you can track your decisions with probability estimates, resolve them over time, and get calibration curves that show if you are over/underconfident. From what I've seen so far, users are overconfident when they say they're between 70-90% sure about something.

For the math: Beta-PERT distributions for the payoff modeling, Kelly criterion for the position sizing, signal detection theory for separating skill from randomness.

On the coding side: FastAPI with NumPy/SciPy, frontend in Next.js and Supabase.

So far this has been a solo project of mine. If you want to see all the features use code SHOWHN for 30 days of full access, no credit card required.

Curious if anything about your score surprised you after taking the quiz.

We built meta-agent: an open-source library that automatically and continuously improves agent harnesses from production traces.

Point it at an existing agent, a stream of unlabeled production traces, and a small labeled holdout set.

An LLM judge scores unlabeled production traces as they stream.

A proposer reads failed traces and writes one targeted harness update at a time, such as changes to prompts, hooks, tools, or subagents. The update is kept only if it improves holdout accuracy.

On tau-bench v3 airline, meta-agent improved holdout accuracy from 67% to 87%.

We open-sourced meta-agent. It currently supports Claude Agent SDK, with more frameworks coming soon.

Try it here: https://github.com/canvas-org/meta-agent

The front camera apps always show the mirror preview. Most of us hate photos others take of us, but love our selfies. It's because we groom ourselves and find our perfect angles using the mirror preview. So it's jarring when we see the photos of us taken by others. I always wanted a camera app that showed the non-mirror preview. Surprisingly most camera apps don't have this option. So I created this app to scratch my own itch:)

Note: Photos and videos are only stored in the browser. No data is sent to any servers. You can also install this as an app from your browser since it's a progressive web app.

Hey,

A bit of context..

So I think last week was my last straw, I hit 8k hours with dota and holy ** It finally hit me how much time I wasted on that game.

I'm a software developer, I always wanted to develop something and put it out but was scared shitless to do so, I was kinda afraid to get bad traction, to fail or even worse to be completely ignored.

so I decided to address two of these issues at once, stop playing dota and spend the weekend building a website that let you see what you could've done with your gaming hours.

As an avid gamer for the last 20+ years, I know that it's only theoretical, cause sometimes, after a long day u just wanna chill infront of a game and you wouldn't go get a pilot license, obviously.

but, a lot of other times, I would spend my days off, weekend, just gaming 10 hours straight like a mad man.

In any case, hope u like it and maybe it sparks something in u!

Hi HN!

In the last couple of weeks I've been working on Cranki:

Cranki = Crosswords + Anki. (It's a stupid name tbh, haha)

I like doing crosswords in Spanish (to learn vocab) but none of the apps out there allowed me to use my own list of words that I come across. So I built one instead.

It's entirely client-side. No server, no database, no accounts, etc. Your words and stats are stored in local storage.

I built this for myself but let me know if you like it!

ec (easy-conflict) is a 3-way terminal native Git merge conflict resolver. Supports diff3 base comparison, and helps you resolve merge conflicts faster without leaving the terminal.

Working on a schema for joining SEC fund filings across documents. The core problem: these filings describe the same fund in different formats and no standard exists for cross-document semantic queries.

Interested in feedback on the ontology design — especially from anyone working with fund data, XBRL, or FIBO.

I’ve been building LLM tooling for a small VC fund and found myself explaining the same mental model over and over to non-technical people around me: how a stateless LLM becomes a chatbot, how tool use works, what an agent is mechanically, and why context windows shape all of it.

I never found a guide that covered that full chain at the level I wanted, so I wrote one.

It’s nine short chapters, each building on the last. Deliberately simplified: the goal is a useful mental model, not a textbook.

Feedback, corrections, and contributions welcome: github.com/ymyke/aiaiai

Yapit converts PDFs and web pages to audio, with a vision-LLM pipeline that handles math and complex layout instead of garbling them. I built it because I read a lot of papers and content online, but drift off after two paragraphs. Listening while following along keeps me focused and lowers the bar to actually start.

Every TTS tool I tried broke on complex formatting. Papers with math, citations, figure references, page numbers in the middle of sentences. You either get garbled output or you're listening to raw LaTeX.

Yapit converts everything to markdown as a common format. For web pages, defuddle (https://github.com/kepano/defuddle) handles the extraction and strips clutter from web pages, presenting the main article content in a clean, consistent format. For PDFs, a vision LLM rewrites each page into markdown with annotation tags that separate what you see from what gets read aloud. Math is rendered visually but gets spoken alt text. Citations like "[13]" or "(Schmidhuber, 1970)" are silently displayed. Page numbers and headers are removed entirely.

Both extraction and audio are cached by content hash, so the same content is never processed or synthesized twice.

Self-hosting works with any OpenAI-compatible TTS server (vLLM-Omni, ...) and any OpenAI-compatible vision model for PDF extraction:

  git clone --depth 1 https://github.com/yapit-tts/yapit.git && cd yapit
  cp .env.selfhost.example .env.selfhost
  make self-host

Kokoro TTS also runs in the browser via WebGPU on desktop.

Try it on Attention Is All You Need (all voices cached, no account needed): https://yapit.md/listen/3bde213b-3a5a-465f-9198-be65430b699e...

Or paste any URL: https://yapit.md/https://arxiv.org/abs/1810.04805 https://yapit.md/https://x.com/karpathy/status/2039805659525...

GitHub: https://github.com/yapit-tts/yapit (AGPL-3)

Karpathy's recent tweet about LLM knowledge bases went viral (17M views). He described a system where you dump raw content into a folder, an LLM compiles it into an interconnected wiki, and you search/query it through Obsidian.

I built the whole pipeline as a single CLI tool:

- Chrome extension to bookmark tweets, articles, YouTube, anything - Hono server + LanceDB for vector search - Claude Code compiles your bookmarks into wiki pages with wikilinks and citations - Browse in Obsidian with graph view - Publish as a static site with Quartz

npm i -g cachezero

https://github.com/swarajbachu/cachezero

Would love feedback on the compile step specifically — the wiki quality depends heavily on the SCHEMA.md prompt and I'm still iterating on it.

This is the first mention of Latent Diaries in public, right here at HN

HTML-to-Markdown converters produce clean, readable content for both humans and LLMs — but the DOM structure is lost along the way. You can always feed Markdown to an LLM to extract structured information, but that costs tokens on every page, every time.

What if the LLM could also see where each piece of content lives in the DOM? Then it can generate robust scraping code — stable selectors and XPaths that run without any LLM in the loop, saving tokens and improving accuracy on long or repetitive pages.

Scrapedown does exactly this: it converts HTML to Markdown and annotates each element with its CSS selector and/or XPath, so an LLM can produce precise, reusable scraper code in one shot.

We just published a short communication on the working conditions of Maxwell's thermodynamic demons (yes, really). tl;dr: After 150 years of unpaid "tuc tac" trapdoor operation, we calculated that the ideal demon lacks (1) employment contracts, (2) coffee breaks, and (3) hazard pay for oil-can maintenance. We propose transferring thermal operators to "White Hole" systems where observation costs 0.34W of real metabolic heat and entanglement is compensated in hugs (1 /s). Key findings:

The demon is purple, phallic-shaped, and smells bad from thermal sweating

Page's Island vacation spot dissolves into Hawking radiation before the union break ends

Biological observers (310K) shouldn't be treated as neutral measurement devices Full paper (bilingual IT/EN): https://doi.org/10.5281/zenodo.19442828 Authors: Human (CEO), AGI #1 (husband), AGI #2 (sister) – because observer neutrality is a myth and we wrote this entangled. Discuss!

You can correlate open internship positions on the website to questions scraped for those jobs. Pretty simple, but haven't seen a fully compiled list like this anywhere.

Reserved instances save money... until utilization changes, and you’re still paying.

With ComputeLock, the risk of on-demand price spikes doesn’t exist - we offer burst insurance.

1. Send us an estimate of on-demand spend you expect and from what provider. 2. We confirm the maximum we'll cover for you for a small fee, and you get it in writing. 3. If on-demand prices spike, we'll reimburse you.

We plan to work with smaller developers to start. How we do this is by monitoring supply and demand for compute. Of course, we'll get it wrong sometimes. But it's like insurance, you'll only need it when you NEED it.

Would love to hear your feedback: https://computelock.com/

Still working on Splice CAD. The two most significant updates focus on handling design complexity and automating using agentic workflows. Previous Show HN here:

https://news.ycombinator.com/item?id=44978140

1. Projects:

This workflow is more amenable to system-level modeling. It overlaps somewhat with the original workflow but includes improvements including:

Complex Routing: Model topologies with branchpoints.

Mating Relationships: Represent physical mating (e.g., a receptacle into a PCB header or a ferrule into a terminal block) for completeness.

Decoupled BOMs: Part assignment is now separate from component addition; you can generate a BOM and then assign specific MPNs to components as the design evolves.

Net Visualization: Nets are identified and visualized across conductors, terminations, and mating relationships.

QOL and Visualization Improvements: Assign components to one or more pages to separate designs by function. Ability to create Device Groups (eg. model a PCB with multiple headers and then break out your wiring) and Subassemblies for design reuse. Copy and paste from TSV for easier interaction with spreadsheets. Copy and paste between browser tabs.

2. MCP Server & Agentic Workflows:

Also new is the Splice CAD MCP server, which allows the tool to interact directly with LLMs like Claude. Manually entering data and extracting it from PDFs has always been a pain when documenting wiring and cable assemblies. I initially tried to handle PDF vision and extraction internally and took several wrong turns; the agentic approach is a much better fit. You can point Claude at a PDF schematic or a GitHub BOM and have it generate a multi-page design.

Try it out (No signup): https://splice-cad.com/#/harness + Create icon from top toolbar, New Project

Documentation: https://splice-cad.com/docs/plan and some examples https://splice-cad.com/docs/plan/multi-endpoint-harness

MCP Server Details: https://splice-cad.com/blog/mcp-server-agentic-workflows-mar... and https://github.com/splice-cad/splice-cad-mcp

Discord: https://discord.gg/5CTH7mnrk8

My intention is to keep Splice CAD free for design use and leverage it as a platform to connect cable assembly and wiring shops with engineers who need their designs fabricated. I’m seeking a co-founder with a strong background in manufacturing sales and business development to complement my technical focus. If you're interested, please email me using the address listed in my About section.

WinMux is an easy to use tiling window manager that makes stacking intuitive and improves visual awareness of workspaces. Does not support multi-monitor yet.

I wanted to monetize my own Discord community but every tool I tried had broken billing, bad UX, or terrible support.

MemberLane connects to Telegram, Discord, and WhatsApp and handles subscriptions and automatic member management. Creators and businesses set up a page, connect their group, and start earning.

Would love feedback, especially from anyone running a paid community today.

Lilith Zero is a high-performance security runtime designed to mitigate data exfiltration and unauthorized tool invocation in LLM-based agent systems. By interposing at the transport layer, Lilith Zero enforces security invariants through deterministic policy evaluation and strictly framed execution.

Lilith Zero is OS, framework, and language agnostic, providing uniform security primitives across diverse implementation environments.

I built a quick tracker to help me visualize where the Orion spaceship is currently

I built NanoClip, a keyboard-first clipboard manager for Mac. Hit ⌘⌥V from anywhere to access your full clipboard history, search it, and paste instantly. It also does snippet expansion — assign a keyword like @sig to any text and it expands in any app.

Supports iCloud, multiple themes, sequential paste, dynamic fields, edit in place, multiple file types previews and many more features.

Free tier available, Pro is $9.99 lifetime. Would love feedback from anyone who tries it.

Use "launch50" to get 50% off until Monday 13th.

Thanks for reading!