2025년 10월 10일의 Show HN

I invented Discrete Distribution Networks, a novel generative model with simple principles and unique properties, and the paper has been accepted to ICLR2025!

Modeling data distribution is challenging; DDN adopts a simple yet fundamentally different approach compared to mainstream generative models (Diffusion, GAN, VAE, autoregressive model):

1. The model generates multiple outputs simultaneously in a single forward pass, rather than just one output. 2. It uses these multiple outputs to approximate the target distribution of the training data. 3. These outputs together represent a discrete distribution. This is why we named it "Discrete Distribution Networks".

Every generative model has its unique properties, and DDN is no exception. Here, we highlight three characteristics of DDN:

- Zero-Shot Conditional Generation (ZSCG). - One-dimensional discrete latent representation organized in a tree structure. - Fully end-to-end differentiable.

Reviews from ICLR:

> I find the method novel and elegant. The novelty is very strong, and this should not be overlooked. This is a whole new method, very different from any of the existing generative models. > This is a very good paper that can open a door to new directions in generative modeling.

"Lights Out" is a mathematical puzzle that lives on an grid where each cell of the grid is one of two colors: either red or white. The goal is to eventually get all the cells in the grid to be red.

What's the catch? Clicking a cell will not only flip its color, but also that of all cells sharing the same row or column as it.

To me, this game feels like playing with a Rubik's cube --- every time you think you are fixing one cell, you mess up its neighbors!

There are many ways to arrive at a solution... some mathematical (linear algebra, or combinatorics), others more logical,... and yet others which are brute force.

Try it out and let me know how you do!

Tim Hartnell was one of the most prolific authors during the early days of the home computing boom, writing many popular books covering genres of games on different platforms and, in this case, artificial intelligence.

I've extracted the BASIC program listings from Hartnell's 1986 book 'Exploring Artificial Intelligence on Your IBM PC' and organized them along with a PC-BASIC runtime environment and instructions so you can try these programs out yourself.

Even though the AI landscape has changed enormously since Hartnell first wrote this book, I hope one or two of you will get some value out of these program listings if you're interested in exploring the fundamentals of AI on home-computing platforms as they were in the 1980's.

Tim Hartnell unfortunately passed away in 1991 at the young age of 40, and without his writing I imagine more than a few of us would not have found the start in computing we did. Thanks Tim.

I started this trying to understand two things: why my Asian friends turn red after drinking, and why several friends all seemed to have migraine clusters.

I was reading medical papers and textbooks, but kept getting lost jumping between topics. I thought: what if I could just Cmd+Click through this like code? What if "ALDH2 gene" was actually clickable, and took me to the variant, the phenotype, the population frequencies?

So I started modeling human biology in Rust with my Ralph agent (Claude in a loop, ty ghuntley). Turns out the type system is perfect for this. Every biological entity is strongly-typed with relationships enforced at compile time.

After 1 day of agent coding: - 277 Rust files, ~95k lines of code - 1,561 tests passing - 13 complete organ systems - Genetics with ancestry-specific variants - Clinical pathology models

Try it:

git clone https://github.com/lantos1618/open_human_ontology cd open_human_ontology cargo run --example ide_navigation_demo

Then open `examples/ide_navigation_demo.rs` and Cmd+Click through:

Understanding Asian flush:

AsianGeneticVariantsCatalog::get_metabolic_variants()

// Click through to:

// → ALDH2 gene on chromosome 12q24.12

// → rs671 variant (Glu504Lys)

// → 40% frequency in Japanese population

// → Alcohol flush reaction

// → 10x esophageal cancer risk with alcohol

// → Acetaldehyde metabolism pathway

Understanding migraines: Migraine { subtype: WithAura, triggers: [Stress, LackOfSleep, HormonalChanges], genetic_variants: ["rs2075968", "rs1835740"], ... }

// Click through to:

// → 17 migraine trigger types

// → 12 aura symptom types

// → Genetic risk factors

// → Why clusters happen (HormonalChanges → Menstruation)

Now I can actually navigate the connections instead of flipping through PDFs. Heart → CoronaryArtery → Plaque. VisualCortex → 200M neurons → NeuralConnection pathways. It's like Wikipedia but type-checked and with jump-to-definition.

This isn't production medical software - it's a learning tool. But it's way more useful than textbooks for understanding how biological systems connect.

The agent keeps expanding it. Sometimes it OOMs but that's part of the fun.

Tech: Rust, nalgebra, serde, rayon, proptest

I am not a dr or medical professional this is for my education you can commit to it if you want to or review and open some PR's if you find wrong information or want to add references.

I built a browser extension called Gitcasso which:

- Adds markdown syntax highlighting to GitHub textareas - Lists every open PR/issue tab and any drafts - (Optional, unimplemented) autosaves your comment drafts so you don’t lose work

I made it because I was impressed by https://overtype.dev/ (a markdown textarea syntax highlighter) which went big on here on HN a few weeks ago, and it seemed like a perfect fit for a GitHub browser extension. Keeping up with changes on upstream GitHub would normally be a pain, but with with Playwright and Claude Code it seemed possible for it to be nearly automatic, which has turned out to be mostly true!

This was the first time where I built a tool, gave the tool to AI, and then AI used the tool to make the thing I hoped it would be able to make. I'm pretty sold on the general technique...

GitHub repo (Apache2-licensed, open source): https://github.com/diffplug/gitcasso

Video walkthrough (2 mins of the tool, 12 mins of its development tooling): https://www.youtube.com/watch?v=wm7fVg4DWqk

And a text writeup with timestamps to the video walkthrough https://nedshed.dev/p/meet-gitcasso

Hey all, just wanted to share a little game experiment. It's a rooms & keys kind of adventure with a lot of random deaths. It plays a Veo3-generated video in response to clicks, with Gemini used for coding.

Prompting the videos was fun, but trying to vibe code everything was not. In the future I'll go back to using LLMs more sparingly for isolated functions, or at least try not to have it create anything that requires seeing the output to debug.

Rustacean AI, a weekly newsletter exploring how Rust is shaping the future of AI and Machine Learning. Issue #1: From Models to Data — Rust’s Expanding Role in AI Covers the growing Rust ecosystem around AI frameworks (Burn), data systems (Polars, Qdrant, Daft), and emerging tools that power safe, high-performance pipelines. The goal is to curate news, releases, and experiments from across the Rust + AI community — helping developers follow how the language is influencing the next generation of infrastructure. Read first issue here → https://rustacean.ai/p/issue-1-from-models-to-data-rust-s-ex...

I rebuilt Biblos, my semantic Bible search app, to run entirely in your browser. No more server costs. The main challenge was fitting an accurate text embedding model into browsers. Last year's version cost $20 monthly to host. The new version runs free on Vercel and searches 31k Bible verses without sending data to any server.

I pre-compute embeddings for all 31,000 Bible verses offline using BGE-large-en-v1.5. Each verse becomes a 1024-dimensional vector stored as JSON, compressed into ZIP files by book. When you visit the site, your browser downloads Transformers.js and the BGE model.

The model generates an embedding for your query, then finds similar verses using cosine similarity against the cached embeddings. Everything happens in your browser.

The app is designed to search Bible verses by semantic meaning, not just keywords. There's a separate tab to read daily passages with church fathers commentary from 15+ sources, based on a read the bible in a year plan.

Free to use, no sign-in features or anything like that. Try it out biblos.app. I'd appreciate feedback on the search experience, especially on mobile devices! Happy to discuss the technical details or help anyone building similar browser-based ML applications.

Hello HN, Lucas and Soheil here from Praxos (https://mypraxos.com/)! We’ve been working on an AI personal assistant for a while now, and today we're sharing about our new webhooks feature with you.

You can now add webhooks and triggers by asking Praxos over text or voice to set them up for you. Webhooks listen for conditions that, when they happen, trigger another action. Webhooks can execute one time or be indefinite. They can execute any action currently supported by Praxos.

They are implemented for email and calendar (Gmail, Outlook), Notion, Slack, Discord, Trello, Dropbox, Drive, iMessage, Whatsapp and Telegram.

Examples include:

–"When a new task is added to my Trello ‘Urgent’ list, create a two-hour block on my Google Calendar within my next work window, and send me a reminder as soon as it happens. " [this ties nicely to the next one]

–"If my calendar says I’m in a focus block, auto-reply on Slack saying I’m working on the latest Trello Urgent task."

–“When a meeting transcript is ready from Fireflies and lands on my email, extract decisions and next steps, then post a 5-bullet summary in #team-updates on Slack” [ties to next point].

–“When a meeting transcript from work finishes processing, summarize it, and post key decisions to Slack. But delay the notification until after my kid’s bedtime. If the summary includes tasks due tomorrow, block off my calendar in the morning and text me a reminder after I wake up.”

–"My goal is to read 12 books this year. Every month, send me a list of 10 books and their links to Goodreads based on what I like. Every Saturday morning, ask me how far along I am with my reading. If you find I purchased a book (i.e.: from checking my email) then add it to my reading list, and ask me if I've started reading it."

–"Every time I get a receipt from Uber Eats or Doordash, extract the date, bill, and meal and add those to my personal finances spreadsheet on Google Sheets."

–“When a new transaction appears in my bank email or statement, extract the merchant, category, and amount, then log it in my ‘Spending Tracker’ Google Sheet. If total monthly spend crosses my budget limit, post a summary to my private WhatsApp chat with the top three categories that caused it.”

–"Remind me to pay my credit card and bills each time a new email comes in. Ask me if I have done so one date later if I don't tell you I already have done this."

–“At month-end, compile all financial logs from Sheets, receipts, and transcripts of my finance calls, then generate a single ‘Monthly Snapshot’ PDF. If my savings rate improved, add a green badge, and congratulate me. If it worsened, send me a summary with trends.”

–"Every Sunday, check my latest additions to Google Photos and send a curation to my mom and grandma over Whatsapp."

–"Review user feedback on the Praxos Discord channel and add it to our User Feedback page on Notion. Keep a counter for repeat requests."

–"If I receive an email from Lucas, and I'm coding, respond to him and tell him I'm busy. Also remind him to check what I'm up to on Trello."

Curious? Try it out for free for 7 days at https://app.mypraxos.com/sign-up!

Hey everyone,

I built Indiebooks.io . A free bookkeeping app that automatically organizes your finances and fills out CRA/IRS tax forms for you. It’s built for freelancers, indie devs, and small business owners in Canada and the US who want to stay compliant without spending weekends in spreadsheets.

What it does - Tracks income & expenses automatically - Handles GST/HST, PST, QST (yes, even Quebec math) - Supports US taxes and state sales tax - Auto-fills CRA/IRS tax forms so you can just review and file - Lets you ask AI things like “What’s my profit this quarter?” - 100% free — no limits, no credit card (might have to limit actual receipts image uploads if storage becomes too expansive)

Why I made it:

Bookkeeping tools are either too expensive or too complicated, especially for small businesses. I wanted something that just works, stays compliant, and doesn’t make you learn accounting to use it.

Tech bits - Built with NextJS, AWS, a bit of supabase, and a bunch of automations for tax logic - Handles messy edge cases like multi-province taxes and mixed personal/business expenses - AI is constrained to actual numbers in your books (no hallucinated profit lol)

Would love feedback on: - Anything confusing or missing in the UI - Tax edge cases you think it should handle - Ideas for integrations (bank connections, Stripe, etc.) - What would make you actually switch to it?

It’s still early, but you can try it here: https://indiebooks.io/

Would love to hear what you think, especially if you’re running a small business in Canada or the US.

Thanks!

-Max

I believe claude code plugin marketplace will be a huge impact, so I am building a website to collect all these awesome things, I just start to building, ping me if you have already got asesome stuff

Would appreciate any feedback, I've really been frustrated with modern social media recently, and made this platform to push back.

It promotes genuine connection and deep understanding, topics that we've seemingly lost with the proliferation of short-form and current social media.

Hey HN! I'm Jonathan, one of the co-founders of <>> Compyle.

We're building a coding agent that isn't afraid to stop and ask questions! It checks in with you before it makes key product and engineering decisions, so you can trust it with more open-ended tasks than you can with other coding agents.

Before Compyle, we were building SmartAppetite (AI agents for insurance). Using tools like Claude Code / Cursor, we had a demo in days - then spent weeks cleaning up a codebase we no longer fully understood.

We had to pause feature development just to refactor. That pain is why we built Compyle.

We built Compyle by essentially orchestrating a bunch of Claude Codes. We have multiple "overwatcher" agents checking for various things, as one main coding agent is working. We check for things like bad patterns (based on your project rules) and mismatches between what the coding agent is doing and your plan / decisions.

We also generate artifacts, like research and planning documents.

We built it on top of the Claude Agent sdk (with Sonnet 4.5 on all the agents) and Fly machines to be able to spin up sandboxes quickly.

Compyle keeps you in control through a question-driven approach: - You describe the task - Compyle asks questions until it fully understands - It produces research and planning artifacts before writing code - During implementation, it validates changes against your decisions - If something doesn't match up, it stops and asks

Our vision is to keep engineers in the driver's seat the entire time. We feel that other agents, like Replit, Devin, and even CC are pushing hard to go in the opposite direction (having agents run for 6+ hours autonomously); As engineers, we want more control, not less!

We'd love for you to try our open beta! It's totally free right now -

I would love to hear what you think if you get a chance to test it!

Git Stars helps you discover and track the most trending GitHub repositories. Explore programming languages, topics, and connect with top developers in the open-source ecosystem. We only list projects with over 500 stars and recent activity since 2024-01-01.

Sharing Crock, my Electron/Vite/React desktop wrapper for the croc CLI. It adds a GUI for quick send/receive, transfer history, settings, and uses electron-builder for packaging. Still early—some settings are rough—but basic transfers work. For now only the Windows x64 build is published; macOS and Linux builds will follow. Bug reports and contributions welcome!

A little more than a year ago, I threatened to build RAD tool for egui if no one else got around to it first.

Well, here it is, only a day old, warts and all. GUIs designed with it should build and run, but it still has a lot of rough edges. Much more to do to expose the full range of egui functionality.

Please feel free to kick the tires and leave a bug report or file a PR! Thanks for being a wonderful community!

Hey everyone, I just launched Brinode, a tool that helps teams build automations and AI agents on top of n8n just by describing them in plain text.

Instead of manually connecting nodes and debugging errors for hours, you can simply type what you want -- for example:

“Scrape my website, save new data to Google Sheets, summarize it with ChatGPT, and send a custom email to my contacts.”

Brinode turns that plain text into a production-ready n8n workflow in minutes.

No setup, no coding, no headache.

Would love to get feedback from the HN community -- especially from those who use n8n or build automations regularly.

This is my first project utilizing intrinsics. It proved impossible to make it without having an understanding of Galois Field (GF(2)) math as well as the specifics of CRC implementations.

I heavily documented my work with comments, though I don't know if I did a good job explaining the process, and I tend to notice a lot of grammatical errors when rereading my English writing.

This basically generalizes the algorithm described in Intel's paper to work for any CRC parameters. The paper does vaguely explain how to do handle the various cases that arise when using different parameters, but most implementations tend to be targeted towards a specific type of CRC.

The algorithm uses some clever, complicated math to reduce the data buffer to a much smaller one, and then computes the CRC using the new smaller buffer.

The Paper's algorithm goes this way: Fold-by-4 -> Fold-by-1 -> Fold to 64-bits -> Barret Reduction.

Things get confusing by the end when you need to actually compute the CRC out of the smaller buffer, and the complexity doesn't really help improve the performance.

My simplification of the algorithm: Fold-by-4 -> Lookup Table

I figured that I could do this after reading a section of the paper that stated that the new small buffer is "congruent" to the original. This doesn't affect performance much since the longer section (Fold-by-4) is still utilizing intrinsics. The software table algorithm is only used to reduce the remaining data (fewer than 200 bytes).

I've just finished it so I don't know if it's mature yet and it will definitely require further testing.

What do you think?

Sora Watermark Remover is an online AI-powered tool designed to automatically detect and remove watermarks from videos generated by Sora AI. It preserves video quality while efficiently removing text, logos, timestamps, and other overlays. The platform supports multiple video formats, processes files quickly, and ensures privacy by deleting uploaded videos after processing. Ideal for content creators, marketers, and video editors seeking a fast, professional watermark removal solution.

Everyone in crypto gambling says they’re “provably fair.” But if you look closely, most of them only prove the final number after the spin or roll. They never prove how that number was chosen, or if your bet was even part of the process.

That’s what I wanted to fix.

I built something called the ProbablyFair Verifiability Layer (PF-VL). It’s an open standard that makes real fairness provable for any RNG-based game — slots, dice, plinko, whatever.

It doesn’t use a blockchain. It doesn’t use tokens. It’s just math.

Every bet carries a few guarantees:

1. The exact RNG code and parameters are published and hashed. 2. The game logic (payout tables, reels, etc.) is locked and verifiable. 3. Every bet is included in an append-only ledger with a proof. 4. Any player can replay and verify the result themselves. 5. Independent watchdogs can recompute everything continuously.

ProbablyFair doesn’t replace the casino’s system. It sits beside it. Operators don’t have to rebuild anything — just add a few calls to the SDK.

That’s it. Three API calls and one file hash.

If there’s a free, open, drop-in standard for provable fairness, then there’s no reason not to use it. Any casino that refuses is showing you something: they prefer a system where players can’t see what’s really going on.

If it’s fair, you can prove it. If you can’t prove it, it isn’t fair.

Technical bits: Core: Rust (verifier) SDK: Go (operator libraries) WASM: TinyGo (browser verifier) License: Apache-2.0 / CC-BY-4.0

Spec: https://probablyfair.org/specification | https://github.com/ProbablyFair/pf-specs/blob/main/PF-VL-1.0...

Repos: https://github.com/probablyfair/pf-core https://github.com/probablyfair/pf-sdk https://github.com/probablyfair/pf-wasm https://github.com/probablyfair/pf-specs

ProbablyFair Verifiability Layer (PF-VL-1.0) Every bet: committed, included, reproduced.

https://probablyfair.org

hey guys, I created this tool to make it easier to keep track of the current status of services to help me debug easier without typing in a bunch of commands manually. Feel free to give it a try, any feedback is welcome :) Thanks!

Hi HN! While working on my side projects, I realized something - I have all these apps, but there is no easy way to cross-advertise them. Putting my own ads in my own apps is way easier and cheaper than paying Google or Facebook to do that. Moreover, there is a chance for higher returns if my ads are in the same niche as the app.

At first, I made a basic implementation for myself, but I wanted something more - the number of views, clicks, platform-specific redirects (e.g. Google Play for Android and App Store for iOS), changing links dynamically, etc.

That’s why I made CrossPromote

Picture this: you have a travel tracker, so why not put an affiliate link to a service that offers global eSIM services? Or a multi-currency account? Travel insurance? Your other app that is a travel guide? The possibilities are endless, just like returns.

CrossPromote is a simple tool for app and site owners to drop in custom ads, affiliate links, or cross-promos. Perfect for anyone who has an app or a website in some niche.

The first version is done and ready to use. It’s a 100% solo project and I'm trying to find a product-market fit while gathering feedback that can help me grow it into a better service.

Let me know what you think!

The biggest friction point in building a business is the wasted time on setup. Stop manually creating Notion pages, databases, and SOPs (Standard Operating Procedures). FlowTask merges the structure of Notion with the intelligence of ChatGPT. Simply use a single prompt (e.g., "startup roadmap") and FlowTask delivers a fully populated, tailored workspace in seconds. Our context-aware FlowBot proactively suggests actionable tasks, maximizing your execution velocity. It's time to focus on the long game—solving problems, not setup.