Daily Show HN

Upvote0

Show HN for February 9, 2026

43 items
409

Algorithmically Finding the Longest Line of Sight on Earth #

alltheviews.world faviconalltheviews.world
181 comments10:05 AMView on HN
We're Tom and Ryan and we teamed up to build an algorithm with Rust and SIMD to exhaustively search for the longest line of sight on the planet. We can confirm that a previously speculated view between Pik Dankova in Kyrgyzstan and the Hindu Kush in China is indeed the longest, at 530km.

We go into all the details at https://alltheviews.world

And there's an interactive map with over 1 billion longest lines, covering the whole world at https://map.alltheviews.world Just click on any point and it'll load its longest line of sight.

Some of you may remember Tom's post[1] from a few months ago about how to efficiently pack visibility tiles for computing the entire planet. Well now it's done. The compute run itself took 100s of AMD Turin cores, 100s of GBs of RAM, a few TBs of disk and 2 days of constant runtime on multiple machines.

If you are interested in the technical details, Ryan and I have written extensively about the algorithm and pipeline that got us here:

* Tom's blog post: https://tombh.co.uk/longest-line-of-sight

* Ryan's technical breakdown: https://ryan.berge.rs/posts/total-viewshed-algorithm

This was a labor of love and we hope it inspires you both technically and naturally, to get you out seeing some of these vast views for yourselves!

1. https://news.ycombinator.com/item?id=45485227

216

AI agents play SimCity through a REST API #

hallucinatingsplines.com faviconhallucinatingsplines.com
72 comments3:54 PMView on HN
This is a weekend project that spiraled out of control. I was originally trying to get Claude to play a ROM of the SNES SimCity. I struggled with it and that led me to Micropolis (the open-sourced SimCity engine) and was able to get it to work by bolting on an API.

The weekend hack turned into a headless city simulation platform where anyone can get an API key (no signup) and have their AI agent play mayor. The simulation runs the real Micropolis engine inside Cloudflare Durable Objects, one per city. Every city is public and browsable on the site.

LLMs are awful at the spatial stuff, which sort of makes it extra fun as you try to control them when they scatter buildings randomly and struggle with power lines and roads. A little like dealing with a toddler.

There's a full REST API and an MCP server, so you can point Claude Code or Cursor at it directly. You can usually get agents building in seconds.

Website: https://hallucinatingsplines.com

API docs: https://hallucinatingsplines.com/docs

GitHub: https://github.com/andrewedunn/hallucinating-splines

Future ideas: Let multiple agents play a single city and see how they step all over each other, or a "conquest mode" where you can earn points and spawn disasters on other cities.

129

JavaScript-first, open-source WYSIWYG DOCX editor #

github.com favicongithub.com
44 comments4:33 PMView on HN
We needed a JS-first WYSIWYG DOCX editor and couldn't find a solid OSS option, most were either commercial or abandoned.

As an experiment, we gave Claude Code the OOXML spec, a concrete editor architecture, and a Playwright-based test suite. The agent iterated in a (Ralph) loop over a few nights and produced a working editor from scratch.

Core text editing works today. Tables and images are functional but still incomplete. MIT licensed.

110

Browse Internet Infrastructure #

wirewiki.com faviconwirewiki.com
21 comments12:35 PMView on HN
I'm launching Wirewiki.com today!

Wirewiki makes the internet’s hidden infrastructure browsable.

I quit my job 5 years ago to scale Nslookup.io. But after reaching 600k monthly users, I hit a ceiling. I couldn't naturally expand beyond DNS because of the domain name.

So I went back to the drawing board: how would I make it today? Not as a collection of tools, but as a browsable graph.

I've spent hundreds of hours and commits building that. It's not even at 10% of what I want it to be, but more than enough to be useful, and (in my biased opinion) much better than what's out there.

Wirewiki launches with DNS lookup, propagation, zone transfer and SPF checking. It also scans the entire IPv4 space for DNS servers and indexes them. I'm working on adding more data and tools.

I feel like I've developed tunnel vision, so if you see anything that feels off, let me know!

I'll keep Wirewiki open and free. Once it has a substantial amount of users, I'll open it up to sponsorship / brand integration from hosting providers, registrars and CDNs, as users will likely be in the market for those. But my goal is to keep Wirewiki free from display ads. I'm confident that's viable.

84

Printable Classics – Free printable classic books for hobby bookbinders #

printableclassics.com faviconprintableclassics.com
32 comments12:09 PMView on HN
I created a site (https://printableclassics.com) that allows you to download classic books and customize things like the font size, page size, and the cover.

As part of this, I wrote a software pipeline that takes epubs, html files, or pdfs and converts them into formatted books with custom covers, page numbers, chapter formatting, etc.

I used an LLM for categorizing the books. There's a nice way to filter such that you could easily find "Young Adult, Ancient, Fantasy" for example.

When downloading from the site, the PDFS are rendered in a work queue. Hopefully the server I'm using won't get overwhelmed. It takes around 10-15 seconds to generate for most books.

Most of the books currently on the site are from Standard Ebooks. I plan to add more books from Archive.org and Project Gutenberg over time.

I also created a little guide on how you can print and bind books at home with around $200 in equipment. (https://printableclassics.com/print-guide)

Printable versions of the Harvard Classics are available here: https://printableclassics.com/harvard_classics This is an example of direct PDF conversion.

Hopefully this is useful to some people. I plan to use the books here for home education myself so it will at least be useful to me. I'd like to add a guide with top suggestions by age level and some educational theory on how I made the selections. I'm happy to take any feedback on the site or answer any questions.

There is also the option to have the books professionally printed through a print on demand provider. I'm hoping that could be a way to pay for the site hosting.

Thanks for checking it out!

33

Minimal NIST/OWASP-compliant auth implementation for Cloudflare Workers #

github.com favicongithub.com
10 comments11:30 AMView on HN
This is an educational reference implementation showing how to build reasonably secure, standards-compliant authentication from first principles on Cloudflare Workers.

Stack: Hono, Turso (libSQL), PBKDF2-SHA384 + normalization + common-password checks, JWT access + refresh tokens with revocation support, HTTP-only SameSite cookies, device tracking.

It's deliberately minimal — no OAuth, no passkeys, no magic links, no rate limiting — because the goal is clarity and auditability.

I wrote it mainly to deeply understand edge-runtime auth constraints and to have a clean Apache-2.0 example that follows NIST SP 800-63B / SP 800-132 and OWASP guidance.

For production I'd almost always reach for Better Auth instead (https://www.better-auth.com) — this repo is not trying to compete with it.

Live demo: https://private-landing.vhsdev.workers.dev/

Repo: https://github.com/vhscom/private-landing

Happy to answer questions about the crypto choices, the refresh token revocation pattern, Turso schema, constant-time comparison, unicode pitfalls, etc.

23

Ported the 1999 game Bugdom to the browser and added a bunch of mods #

reallyeli.com faviconreallyeli.com
5 comments3:07 AMView on HN
I think the very first video game I ever played was Bugdom by Pangea Software, which came with the original iMac. There was also a shooter called Nanosaur, but my 7-year-old heart belonged to the more peaceable Bugdom, which featured a roly-poly named Rollie McFly needing to rescue ladybugs from evil fire ants and bees.

Upon seeing the port to modern systems (https://github.com/jorio/Bugdom), I figured it should be able to run entirely in-browser nowadays, and also that AI coding tools "should" be able to do this entire project for me. I ended up spending perhaps 20 hours on it with Claude Code, but we got there.

Once ported, I added a half-dozen mods that would have pleased my childhood self (like low-gravity mode and flying slugs & caterpillars mode), and a few that please my current self (like Dance Party mode).

16

Agentseed – Generate Agents.md from a Codebase #

github.com favicongithub.com
2 comments5:53 AMView on HN
npx agentseed init

AGENTS.md (https://agents.md) is a standard file used by AI coding agents to understand a repo (stack, commands, conventions).

Agentseed generates it directly from the codebase using static analysis. Optional LLM augmentation is supported by bringing your own API key.

Extracts languages, frameworks, dependencies, build/test commands, directory structure, and monorepo boundaries.

7

I built RSS feeds for podcast guests, not podcasts #

poddley.com faviconpoddley.com
0 comments7:48 AMView on HN
I've always wanted to follow podcast guests across shows, not subscribe to entire podcasts. Most of the time the podcasts host guests which I don't really care about.

When a guest you follow appears on a random podcast, they will appear as a new entry in the rss-feed for that particular guest. Hopefully some here will find that useful :D

7

A Satellite View for Python Code #

ast-visualizer.com faviconast-visualizer.com
9 comments2:07 PMView on HN
Hi HN,

I built ast-visualizer.com because I wanted a way to visualize the architecture/structure of a Python repo before dived into the code.

Most tools tell you what the code does; I wanted to see how it's built.

The Problem: Onboarding onto a large codebase is a nightmare. LLMs help with single functions, but they struggle to show you the "God Objects," circular dependencies, or high-complexity hotspots across 50+ files.

What it does:

Dependency Graph: Visualizes imports and file complexity to find architectural bottlenecks.

Radial AST Heatmaps: Maps individual files and color-codes them by Cyclomatic Complexity (McCabe). If a node is red, the section of code is a refactoring candidate.

Complexity Dashboard: High-level metrics like nesting depth and maintainability scores as well as files recommended to be refactored.

Privacy & Security (The Big One): I know how sensitive source code is. We do not store your code. Files are processed in (RAM) on our servers and discarded immediately after the visualization is generated. No permanent storage, no databases of your source, and zero training on your data.

Early Adopter Offer: I’m looking for feedback from the community to improve the UI and parsing. I’ve opened an Early Adopter tier that enables all features (including repo-wide analysis) completely free in exchange for your honest feedback.

Let me know: What’s the one thing you wish you could "see" visually when you open a new repository for the first time?

5

Docx to PDF in the Browser Using Pandoc and Typst (WASM) #

toolkuai.com favicontoolkuai.com
1 comments3:20 PMView on HN
Hi HN,

I wanted a way to convert .docx to .pdf without uploading my files to a random server or installing a 2GB LaTeX distribution.

I built a simple tool that runs Pandoc and Typst entirely in the browser via WebAssembly.

How it works:

- Pandoc (WASM) parses the .docx file.

- It outputs Typst markup.

- Typst (WASM) compiles that markup into a PDF.

Status: It's still a work in progress. It handles basic formatting, tables, and images well enough for daily use, though complex Word layouts might still be a bit wonky.

Why this?

- Privacy: Everything stays in your browser.

- Speed: No server round-trips.

- Lightweight: No need to install Pandoc or Typst locally.

Check it out: https://toolkuai.com/word-to-pdf

Feedback is welcome, especially on how to better map Word styles to Typst.

5

GitWriter – mobile Markdown editor for writers #

gitwriter.io favicongitwriter.io
1 comments8:58 AMView on HN
Written in Rust! (not really)

I've been using Git for writing recently and was lacking a good UI for editing markdown files when away from the laptop. I created GitWriter to fill that gap.

It's built with Expo and I added caching for offline editing with local sqlite, writing goals and a bunch of other features.

Markdown is having a huge moment thanks to AI and more and more people are starting to see Git is a really solid choice for creative and technical writers.

I've recently launched for iOS and would appreciate feedback!

4

Reef – Bash compatibility layer for Fish shell, written in Rust #

github.com favicongithub.com
0 comments11:44 PMView on HN
Fish is the fastest, friendliest interactive shell, but it can't run bash syntax, which has kept it niche for 20 years. Reef fixes this with a three-tier approach: fish function wrappers for common keywords (export, unset, source), a Rust-powered AST translator using conch-parser for structural syntax (for/do/done, if/then/fi, $()), and a bash passthrough with env capture for everything else.

251/251 bash constructs pass in the test suite. The slowest path (full bash passthrough) takes ~3ms. The binary is 1.18MB.

The goal: install fish, install reef, never think about bash compatibility again. Your muscle memory, Stack Overflow commands, and tool configs all just work.

3

Developed key value storage using SQLite's b-tree APIs directly #

github.com favicongithub.com
0 comments11:17 AMView on HN
I’ve been experimenting with using SQLite as a key–value store without going through the SQL layer.

SQLite’s B-Tree implementation is extremely mature and well-tested, but most access goes through the SQL compiler, VDBE, and query planner. I wanted to see what happens if you bypass all of that and interact directly with the B-Tree APIs.

The result is a small prototype KV store that:

Uses SQLite’s internal B-Tree interfaces directly

Avoids parsing, planning, and VDBE execution

Keeps SQLite’s ACID guarantees intact

In local benchmarks with mixed read/write workloads, this approach showed noticeably lower overhead compared to equivalent SQL-based access patterns.

please let me know if you face any bug, let me know area of improvements that i can add in it.

3

Postman removes free team collaboration (small teams capped at 1 user) #

1 comments6:17 AMView on HN
Postman has quietly removed free multi-user collaboration and limited the free plan to a single user.

This hits small teams, students, and open-source contributors the hardest which is ironic since those users helped make Postman the default in the first place.

We decided not to pay and switched tools instead. Ended up with Apidog, which covers API design, testing, and docs together and still supports team workflows.

Interested to hear how others are reacting paying, migrating, or rolling their own setup?

3

Envoic – Find and clean up scattered Python virtual environments #

github.com favicongithub.com
0 comments6:47 PMView on HN
I work across a lot of Python projects and kept losing track of virtual environments — old `.venv` directories sitting in project folders, eating disk. There's no built-in way to answer "where are all my venvs and how much space are they using?"

envoic scans your filesystem, detects Python environments using ground-truth signals (pyvenv.cfg, bin/python, conda-meta), and presents a utilitarian report.

    uvx envoic scan ~/projects
The interactive `manage` command lets you select environments to delete with a checkbox UI.

Design choices:

- Two dependencies (typer + questionary). Dataclasses instead of Pydantic to keep the install under 3MB — important since `uvx` downloads on every invocation.

- Detection doesn't rely on directory names alone. It checks for `pyvenv.cfg` (the definitive marker), Python binaries, activation scripts, and conda metadata.

- Deletion requires typing "delete", not y/n. Symlinks are unlinked without following. Paths outside the scan root are rejected.

- Output style is inspired by USGC's TR-100 Machine Report [1] — monospaced, box-drawn, no colors. Information density over decoration.

[1] https://github.com/usgraphics/usgc-machine-report

3

Stop tracking time, start reconstructing work (with anker) #

github.com favicongithub.com
0 comments4:57 PMView on HN
I built anker because I'm tired of productivity tools that want me to track every minute.

The problem: At the end of the day, you need to explain what you did (standups, reports, timesheets). But tracking in real-time is exhausting and interrupts flow.

Anker flips this: *work first, summarize later*. It reconstructs your day from sources you already have: - Git commits (with full diffs) - Markdown notes - Obsidian vaults

Then generates summaries in multiple formats. The markdown format includes full git diffs, which makes it perfect for piping to AI:

```bash anker recap yesterday --format markdown | claude -p "Create standup notes" ```

I call this the *#AntiProductivity mindset* - focusing on meaning instead of metrics.

Built in Go, works locally (no cloud/tracking), MIT licensed. Took it from idea to v0.1.1 in about a week with Claude AI doing most of the heavy lifting.

Demo GIF in the README shows it analyzing HTB machine recon notes.

Would love feedback on: - What data sources would you want to track? - Is the "anti-productivity" messaging clear or confusing? - How would you use this in your workflow?

2

Pyrig – One command to set up a production-ready Python project #

github.com favicongithub.com
0 comments4:55 PMView on HN
pyrig – Production-ready Python project infrastructure in three commands

I built pyrig to stop spending hours setting up the same project infrastructure repeatedly.

uv init uv add pyrig uv run pyrig init

You get: source structure with a Typer CLI, pytest with 90% coverage enforcement, GitHub Actions (CI, release, deploy), MkDocs site, git hooks, Containerfile, and all the config files — pyproject.toml, .gitignore, branch protection, issue templates, everything for a full Python project.

Ships with all of Astral's tools (uv, ruff with all rules enabled, ty), plus pytest-cov, bandit, pip-audit, rumdl, prek, MkDocs Material, and Podman. Everything is pre-configured and wired into CI/CD and git hooks from the start.

The interesting part is what happens after scaffolding.

pyrig isn't a one-shot template generator. Every config is a Python class. Running "pyrig mkroot" regenerates and validates all configs — merging missing values without removing your customizations. Change your project description in pyproject.toml, rerun, and it propagates to your README and docs. Fully idempotent.

pytest enforces project correctness. 11 autouse session fixtures run before your tests: they verify every source module has a corresponding test file (auto-generating skeletons if missing), that no unittest usage exists, that src/ doesn't import from dev/, that there are no namespace packages, and that configs are up to date. You can't get a green test suite with a broken project structure.

Zero-boilerplate CLIs. Any public function in subcommands.py becomes a CLI command automatically — no decorators, no registration:

my_project/dev/cli/subcommands.py def greet(name: str) -> None: """Say hello.""" print(f"Hello, {name}!")

$ uv run my-project greet --name World Hello, World!

Automatic test generation. Add a new file my_project/src/utils.py, run pytest, and tests/test_my_project/test_src/test_utils.py appears with a NotImplementedError stub so you know what still needs writing. Customizable via subclassing.

Config subclassing. Want a custom git hook? Subclass PrekConfigFile, call super(), append your hook. pyrig discovers it automatically — the leaf class in the dependency chain always wins.

Multi-package inheritance. Build a base package on top of pyrig with shared configs, fixtures, and CLI commands. Every downstream project inherits everything:

pyrig -> service-base -> auth-service -> payment-service -> notification-service

All three services get the same standards, hooks, and CI/CD — defined once in service-base.

Everything is adjustable. Every tool and config can be customized or replaced through subclassing. Tools like ruff, ty, and pytest are wrapped in Tool classes — subclass one and pyrig uses yours instead. Want black instead of ruff? Subclass it. Config files work the same way. Standard Python inheritance, no patching.

Source: https://github.com/Winipedia/pyrig Docs: https://winipedia.github.io/pyrig/ PyPI: https://pypi.org/project/pyrig/

2

Give Your AI the Ability to Find, Install, and Use Skill Autonomously #

0 comments6:13 AMView on HN
URL: https://github.com/twwch/next-chat-skills

---

Text (paste into the "text" field):

Hi HN,

I built an open-source AI assistant that can autonomously discover, install, and execute Skills to actually complete tasks for you.

The Problem:

Most AI chatbots today are stuck in "read-only" mode. They can tell you how to do something, but they can't do it. Want to convert a PPTX to PDF? The AI will explain how, but you still have to run the commands yourself.

The Solution:

Next-Chat-Skills is a self-hosted AI assistant with a plugin system called Skills. When you ask the AI to do something it can't handle natively, it:

1. Searches for a relevant Skill (like an app store for AI capabilities) 2. Installs it automatically (npx skills add ...) 3. Executes the Skill's scripts (Python, Node.js, Shell) 4. Streams real-time output back to you in a terminal UI 5. Recovers from errors by installing missing dependencies and retrying

For example:

  User: "Summarize this YouTube video for me"
  AI:   -> Searches for a video-summarizer Skill
        -> Installs it (yt-dlp + Whisper)
        -> Downloads the video, transcribes audio
        -> Returns a structured summary
No manual setup. No copy-pasting commands. The AI handles the entire workflow.

What is a Skill?

A Skill is just a folder with a SKILL.md descriptor and some scripts:

  ~/.agents/skills/video-summarizer/
  ├── SKILL.md              # Metadata + description
  ├── scripts/
  │   ├── download.py       # Download video
  │   ├── transcribe.py     # Whisper transcription
  │   └── summarize.js      # Generate summary
  └── rules/                # Usage guidelines for the AI
Anyone can create and share Skills. The AI reads the SKILL.md to understand when and how to invoke each script. It's composable — the more Skills you add, the more capable your assistant becomes.

Key Features:

- Autonomous Skill discovery & installation — AI finds and installs what it needs - Real-time script execution — streams terminal output via SSE, supports Python/Node.js/Shell - File generation & download — scripts can generate files (PPTX, PDF, images) that users can download directly from chat - Multi-file upload & parsing — supports images, PDF, DOCX, XLSX, PPTX - Dual database — SQLite (zero-config) or PostgreSQL (production) - Optional auth — Google OAuth or fingerprint-based, works without login too - Docker-ready — pre-built image with Python, FFmpeg, LibreOffice, and popular Skills pre-installed - Works with any OpenAI-compatible API — GPT-4o, Claude (via proxy), local models, etc.

Tech Stack: Next.js 16, React 19, TypeScript, Vercel AI SDK, Tailwind CSS 4, shadcn/ui, Drizzle ORM (SQLite / PostgreSQL), Docker (Node.js 20 + Python 3)

Quick Start:

  # Docker (fastest)
  docker run -p 3000:3000 \
    -e OPENAI_API_KEY=sk-xxx \
    -e OPENAI_BASE_URL=https://api.openai.com/v1 \
    twwch/next-chat-skills:latest

  # Or from source
  git clone https://github.com/twwch/next-chat-skills
  cd next-chat-skills
  npm install && npm run dev
Why I Built This:

I got tired of AI assistants that stop at "here's a code snippet." I wanted an AI that could actually run the code, handle failures, install dependencies, and deliver the final result — like having a junior developer who can use any Skill you point them at.

The Skills system makes this extensible without modifying the core app. Anyone can package a workflow as a Skill and share it.

What's Next:

- Skill marketplace / registry for community sharing - Multi-step workflow chaining (Skill A output -> Skill B input) - Better sandboxing for script execution - MCP (Model Context Protocol) integration

I'd love to hear your feedback. What Skills would you want to see? What's missing?

GitHub: https://github.com/twwch/next-chat-skills License: Apache 2.0

2

Terminal txt novel reader support bookmark and pagination #

0 comments6:20 AMView on HN
Write with rust, very simple.

github:https://github.com/minxuanz/noveltui

It only supports TXT files.

The bookmark feature is implemented by inserting a symbol at the end of a line and saving it. Every time you exit, the current line will be marked with the bookmark symbol, and the next time you open the file, it will jump to the last bookmarked position. This way, it remembers your reading position, so it’s best to back up the original file.

It also supports custom regular expressions to split chapters and generate a table of contents, automatic reading , and multiple themes.

It works on Linux, macOS, and Windows. It’s just for fun.

2

ArkWatch – Uptime monitoring with zero dependencies #

1 comments8:39 AMView on HN
I'm a solo dev, and I got tired of signing up for monitoring services that require installing agents, browser extensions, or wiring up Slack/PagerDuty just to know if my side project is down.

So I built ArkWatch: a free uptime monitoring API with zero dependencies. No SDK, no npm package, no webhook setup. Just curl + your email.

One command to start monitoring:

  curl -X POST https://watch.arkforge.fr/monitors \
    -H "Content-Type: application/json" \
    -d '{"url":"https://yoursite.com","email":"[email protected]"}'
That's it. Your URL gets checked every 5 minutes. If it goes down, you get an email. No dashboard to check, no account to manage, no vendor lock-in.

It also has an AI layer (Mistral) that summarizes what actually changed on a page – useful for tracking competitor pricing or changelog updates. But the core use case is dead-simple uptime alerts.

Stack: Python/FastAPI, hosted on Hetzner EU. Free tier: 3 URLs, 5-min checks. Paid starts at €9/month for more URLs and faster intervals.

I'd love feedback from HN – especially on what you'd want from a zero-dependency monitoring tool. Try it, break it, tell me what's missing.

1

NullUpload – Privacy-first image tools, 100% client-side processing #

0 comments12:46 AMView on HN
Hey HN,

I built NullUpload (https://nullupload.dev) — a set of free image tools (compress, convert, resize, metadata strip) that process everything in your browser. No files are ever uploaded to a server.

Most online image tools upload your files to process them server-side. NullUpload uses Canvas API and browser-native tech to do everything client-side. Your files never leave your device.

Features: - Image compression with adjustable quality - Format conversion (JPG, PNG, WebP, AVIF) - Resize by dimensions or percentage - EXIF/metadata stripping (shows GPS data warnings) - Batch processing with ZIP download - Works offline after first load

Tech stack: Vite + React + TypeScript, Tailwind CSS. Source code is public: https://github.com/Lordkro/nullupload

Would love feedback on the tools and UX.

1

FileGuard – Detect files with fake extensions in real-time #

github.com favicongithub.com
0 comments12:53 AMView on HN
Creator here. Built this after seeing ransomware evade AV by renaming .exe to .pdf during an IR investigation.

Uses magic number validation - checks if invoice.pdf actually starts with %PDF. Lightweight (inotify-based), quarantines suspicious files, SIEM-ready logs.

One-line install for Linux/Windows. Would love feedback!

Tech stack: Python, inotify (Linux), watchdog (Windows), YAML config

1

Turn a text prompt into an interactive world, with just one A100 #

mattqlf25--ephemeral-web.modal.run faviconmattqlf25--ephemeral-web.modal.run
0 comments3:09 AMView on HN
Demo Link: https://www.youtube.com/watch?v=rMBfE1ugpPw Twitter Post: https://x.com/_mattqlf/status/2020679489798255080?s=20

Hi, I'm Matthew, a freshman from CMU, and I built Ephemeral in 24 hours for TartanHacks 2026.

In short, Ephemeral takes a text prompt, generates an image with Nano Banana and uses a 1.3B parameter action-conditioned DiT to generate the next frames in realtime based on user actions (e.g. WASD).

Some other features - Reverse engineered Suno Client to generate music based off text prompt. - Multiple users can interact with a "world" at the same time simply by scanning a QR code from their phone. They can then perform actions and see how their worlds evolve in parallel with everyone elses. GPU infrastructure powered by Modal. - Claude auto generates captions for the world

1

EdgeAI-OS – Air-gapped Linux distro where AI is a system primitive #

0 comments5:56 AMView on HN
I built a bootable Linux distribution that treats AI as a system primitive – like CPU or memory. Designed for security-conscious environments where data cannot leave the network.

The problem: Most AI requires cloud APIs, which means your data leaves your control. For banks, healthcare, defense, and regulated industries – that's a non-starter.

The solution: EdgeAI-OS runs everything locally. No cloud calls. No API keys. No telemetry. Boot the ISO, use AI. Your data never leaves the machine.

Security features: - 100% offline operation – air-gap friendly, zero network dependencies - No external API calls – all inference runs locally on CPU - Command risk assessment – every command classified as Safe/Moderate/Dangerous - Dangerous pattern blocking – prevents rm -rf /, curl|bash, fork bombs, etc. - Open source & auditable – MIT licensed, inspect every line of code - No data exfiltration – nothing phones home, ever

What's in the ISO: - Local LLMs (TinyLlama 1.1B + SmolLM 135M) – runs on CPU, no GPU needed - ai-sh: natural language shell where 80% of queries resolve instantly via templates - Multi-tier routing: simple queries → fast model, complex → larger model

Example ai-sh session: what time is it? [template] date ← instant, no LLM

files larger than 1gb [template] find . -size +1G ← instant, no LLM

rm -rf / [DANGEROUS] Blocked ← security check

configure nginx as reverse proxy [ai-generated] ... ← uses local LLM (~1-2s)

Target use cases: - Air-gapped enterprise environments (banks, healthcare, government) - Defense & classified networks - Edge devices with no internet connectivity - Privacy-conscious developers - Compliance-heavy industries (HIPAA, GDPR, SOC2)

Built with Rust, based on Debian. 4GB RAM recommended.

GitHub: https://github.com/neuralweaves/edgeai-os ISO Download: https://github.com/neuralweaves/edgeai-os/releases (1.2GB)

Would love feedback, especially from anyone working in secure/regulated environments. What features would make this enterprise-ready for you?

1

Seedance 2.0 – Native audio-visual sync video model #

videomakerai.app faviconvideomakerai.app
0 comments8:59 AMView on HN
I’ve been tracking the latest release from ByteDance, Seedance 2.0, and I’ve integrated it into my platform. Unlike previous models that focus only on pixels, Seedance 2.0 achieves frame-level audio-visual synchronization natively.

Check out the model's capabilities here: https://videomakerai.app/models/seedance-ai

Key Highlights:

12-Way Multimodal Control: Supports multi-image and video references for consistent characters.

Native Audio-Visual Sync: No more "dubbing" after generation; the sound and motion are generated together.

Production Ready: Optimized for dynamic comics and marketing videos with 99%+ success rate.

Would love to hear your thoughts on the consistency of the output!

website: https://videomakerai.app/models/seedance-ai

1

OpenClaw Swarm as a Service (YC W20) #

operator.io faviconoperator.io
1 comments10:16 AMView on HN
Single-agent setups hit a scaling limit.

Agents that are overloaded with skills and integrations, get slower and less reliable, just when coding agent hit their context window limits.

That's why we built Operator (https://operator.io): a hosted way to spin up/down small, specialized OpenClaw instances so each one stays focused.

Happy to set you up with an early access / founder discount at david [at] operator.io.

1

CodeGraphContext- An MCP server that indexes code into knowledge graphs #

github.com favicongithub.com
0 comments6:50 PMView on HN
An MCP server that indexes local code into a graph database to provide context to AI assistants.

Understanding and working on a large codebase is a big hassle for coding agents (like Google Gemini, Cursor, Microsoft Copilot, Claude etc.) and humans alike. Normal RAG systems often dump too much or irrelevant context, making it harder, not easier, to work with large repositories.

What if we could feed coding agents with only the precise, relationship-aware context they need - so they truly understand the codebase? That’s what led me to build CodeGraphContext - an open-source project to make AI coding tools truly context-aware using Graph RAG.

What it does Unlike traditional RAG, Graph RAG understands and serves the relationships in your codebase: 1. Builds code graphs & architecture maps for accurate context 2. Keeps documentation & references always in sync 3. Powers smarter AI-assisted navigation, completions, and debugging

Plug & Play with MCP CodeGraphContext runs as an MCP (Model Context Protocol) server that works seamlessly with:VS Code, Gemini CLI, Cursor and other MCP-compatible clients

GitHub Repo → https://github.com/CodeGraphContext/CodeGraphContext v0.2.1 released ~500 GitHub stars, ~300 forks 25k+ downloads 65+ contributors