Show HN: KillSwitch – a programming language designed to be difficult for LLMs https://ift.tt/dSmC7EP

Show HN: KillSwitch – a programming language designed to be difficult for LLMs https://ift.tt/1DO9iT5 September 20, 2026 at 02:40AM

Show HN: CUA-S1 – A System One Model for Computer Use https://ift.tt/eP14pvK

Show HN: CUA-S1 – A System One Model for Computer Use Hello HN! We're Dillon and Francesco from Cua. We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their decisions and steps. Some tasks require thinking about a plan, exploring different paths, recovering from failure. Other tasks are a question of making local decisions, like this value should go in this box, or should I check this box, or this element should be ignored. We wondered how far we could go with a small model trained to only make these kinds of decisions. Our inspiration was Typesafe's Jev and its System One Model framing. This is a nod to the dichotomy between thinking quickly, automatically, and intuitively (system 1) vs. thinking slowly, analytically (system 2), as described by Daniel Kahneman. The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice? This kind of model does not generate output token by token like most LLMs do, but rather scores the options you give it, which you can check, trust, and use to drive your app's behavior. CUA-S1 is our answer for narrow, specialized decision models for computer use. Our first release is CUA-S1-FORMS. We built this from ideas and code in jevlike, and then trained a second model just to handle form interactions. It has 706k parameters, and the original checkpoint is 2.8 MB. The first training iteration took less than 30 minutes on synthetic data. Given a set of structured elements and values extracted from a document, it predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields, and does not consider screenshots. Element decisions are scored together, and your code can order the actions, and Cua Driver will execute them one at a time. A first evaluation of this specialist vs. hosted Jev on our form task: - For the whole decision set: 99.7% correct vs 83.6%. - For the subset of steps that require an action: 100% correct vs 96%. - For the subset of steps that are just leaving already-filled fields alone: 100% correct vs 74%. The specialist was trained specifically for this task and convention (just press skip for already filled boxes), while hosted Jev has not been fine-tuned for it, so this is an experiment in scoped specialization. We measured 7-9 ms to score a form locally vs. 260-280 ms per call to hosted Jev including network latency, though those samples measure different things and are not end-to-end form completion times. Our interest here is in the space between a brittle script and a general agent loop. The content and layout of form fields vary enough that scripts get unwieldy, but the set of available decisions can remain narrow and well scoped. We want to explore the possibility of a general agent encountering something novel, and passing well understood decisions over to specialists like this. That is a direction we are looking into. The current release is for forms only. We're open sourced the synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license. Comments welcome! Especially if you are building computer-use agents and have run into a recurring decision that is too variable to script but is too narrow to call another LLM for. https://ift.tt/SdDUK5P September 19, 2026 at 10:52PM

Show HN: BreachScanner – Check if your email has been in data breaches https://ift.tt/0pXyeiJ

Show HN: BreachScanner – Check if your email has been in data breaches https://ift.tt/2JKb9rP September 20, 2026 at 12:30AM

Show HN: Learn divergently/chaotically with tree-learn https://ift.tt/rQm0tUn

Show HN: Learn divergently/chaotically with tree-learn Chat interface is linear, I never quite liked that for when I would have too many questions, I would end up asking questions in a numbered list and not really have independent branches of conversation. On cli, the pi agent gives you this same functionality. On web, I made Fable make this. BYOK Openrouter - all client side - you can check the network console! https://ift.tt/rTy7dm6 https://tree-learn.srijanshukla18.workers.dev September 18, 2026 at 11:21PM

Show HN: Jev helps you to not run malicous code https://ift.tt/qby5It2

Show HN: Jev helps you to not run malicous code https://ift.tt/hTRxrdK September 18, 2026 at 11:44PM

Show HN: App Launch: GrabThat (NLP Launcher for Windows) https://ift.tt/50NSBvc

Show HN: App Launch: GrabThat (NLP Launcher for Windows) https://grabthat.ai/ September 18, 2026 at 11:15PM

Show HN: Dishlist – my favorite things on the menu https://ift.tt/voWbaQw

Show HN: Dishlist – my favorite things on the menu Hi HN! My friend and I wanted to rank our favorite breakfast burritos. Yelp and Beli allow users to rate or rank restaurants, but not individual dishes. So I built dishlist, an iPhone app around the things on restaurant menus. I schlepped around Dogpatch, Potrero Hill, and Mission Bay taking photos of restaurant menus. Then, I used Claude with my menu schema to turn the photos into structured data. Garry Tan likes to say, "The rocks can talk." Also, "The rocks can see." I represent the menu hierarchy in the database. LLMs can determine what type of thing each menu item is. This lets me index things like breakfast burritos across restaurants rather than just indexing the restaurants themselves. The app suggests possible tags for a menu item. You decide which ones you want to use on your profile. You can build a list of your favorite breakfast burritos, pizzas, burgers, or whatever else you care about. I value my privacy, so you can use the app without logging in at all. If you want to save menu items, you can sign in with Apple. I don't collect names, emails, or phone numbers. If you want to share with friends, you can add a handle. You have to follow another user to see their dishlist, and they have to follow you to see yours. Follow requests are approved or denied by the user. The app is currently available through TestFlight while the App Store submission is being reviewed. https://ift.tt/HzJgIRE I'd love feedback on the idea, the search experience, and especially whether organizing things around individual menu items feels useful. https://ift.tt/HzJgIRE September 17, 2026 at 11:53PM

Show HN: Content-aware PII redaction with Jev in Postgres https://ift.tt/2JiUTG1

Show HN: Content-aware PII redaction with Jev in Postgres https://pg-redact.vercel.app September 17, 2026 at 11:52PM

Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS https://ift.tt/5xh8t60

Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS Simple discovery of the provider's default schemas is possible without any credentials (since they are built into the binary). If you want it to fetch customizations of your specific instance, give it credentials to your org. The repo has more details. https://ift.tt/itPushb https://www.aclif.ai/ September 17, 2026 at 11:48PM

Show HN: Die With Me – Claude and Codex rate limits as AIM away messages https://ift.tt/qSz9sKa

Show HN: Die With Me – Claude and Codex rate limits as AIM away messages made this AIM buddy list but for our token usage because i wanted to bring a little bit of the old internet back! i remember those days where i'd casually check if friends are online and find excuses to say hi and poke them give it a try! i'd greatly appreciate feedback https://ift.tt/Z3Xgja6 September 17, 2026 at 11:25PM

Show HN: Restarted – a 2026 remake of the classic 2015 startup generator https://ift.tt/x5QKkvS

Show HN: Restarted – a 2026 remake of the classic 2015 startup generator The original startup website generator by Tiff Zhang and Mike Bradley landed on Hacker News in April 2015 ( https://ift.tt/7SlXmrD ) and has been one of my favorite little novelties of that era ever since. It perfectly captures the saturated colors, cliché hero shots, gimmicky names, buzzword-heavy slogans, and proudly hirsute team photos of the time. A lot has changed since then, so I thought it would be fun to make a contemporary remake: https://restarted.io/ By default you get the minimalist aesthetic and clean-cut faces of 2026. The classic 2015 look is still available — just click the link at the bottom of the page or change the “z” parameter in the URL to the more familiar “s”. The universe of partner sites and competing startups is just as expansive as it ever was. The original site is entirely client-side and requires downloading all of the data tables locally. It leans on a mix of jQuery 1.11.2, Bootstrap 3.3.2, and Font Awesome 4.3.0, and if you view the source, it instantly gives away all of its secrets. For restarted.io I replaced all of that with a server-side renderer written in Go, so this time view-source tells you nothing. There are many Easter eggs in there — see how many you can find before I write them up. My original goal was to stay faithful to the 2015 appearance, and that turned out to be a technical adventure. The original's sine-based random number generator is... the worst, and different implementations of sine give different results. The eventual solution was to extract the exact sine function from Chrome’s V8 engine, as vendored C behind cgo and as a line-by-line Go port that keeps cgo optional, so the seeds and results line up the way they used to. Both are checked against V8’s own test cases. Then I discovered a bug in the original code that made half of its vocabulary unreachable — the first half of the verb table and the second half of the noun table, exactly complementary, so nothing about the output ever looked truncated. My goal then shifted from remaking the generator as it was in 2015 to remaking the site as the authors intended it to be in 2015. Over the years several people asked for their photos to be removed, so the remake instead draws from a broad pool of era-appropriate AI-generated profiles. A perceptual hash helps keep everyone looking distinct, and there’s a bit of extra care to make sure the Wang Fangs of the world don’t appear as Irish lasses. The hero image pool is much larger now, and all the old Rio de Janeiro shots have been retired, though you’ll still recognize plenty of the 2015 photos. Have fun poking around! https://restarted.io/ September 15, 2026 at 06:34PM

Show HN: Pixel Agents – A pixel-art mission control for your Claude Code agents https://ift.tt/vuTdE6k

Show HN: Pixel Agents – A pixel-art mission control for your Claude Code agents https://mateovalle.github.io/pixel-agents/ September 17, 2026 at 12:12AM

Show HN: YOLO FPS-per-dollar numbers for edge AI boards ($75-$215) https://ift.tt/JZLlxf9

Show HN: YOLO FPS-per-dollar numbers for edge AI boards ($75-$215) https://ift.tt/JaGBv5k September 16, 2026 at 11:20PM

Show HN: TrailVid – Cinematic Travel Animations https://ift.tt/fKPsbjO

Show HN: TrailVid – Cinematic Travel Animations I originally started with this project because my partner wanted a way to make visually clean travel videos for her social media. The ones that existed felt a bit too expensive for the quality they delivered, and many of them leaned heavily on a cartoonish visual style. So I decided I'll build an app for her that fits her needs, and over the last few months it has become bigger and better. I've even created my own map tiles because licensing existing ones was not cost effective. Now it's finally open for public beta and I'd love to get some feedback before I bring this to the App Store. https://trailvid.com/ September 16, 2026 at 12:01AM

Show HN: The bottom 50% of U.S. households are short after essentials (BLS data) https://ift.tt/bQGy6Hj

Show HN: The bottom 50% of U.S. households are short after essentials (BLS data) https://whats-left-over.pages.dev/#s=50_40_10&c=1&u=person&t=0&tm=cbo&b=0&ts=100&k=0&cc=12201&im=nominal&by=2000&ir=2.5&ao=0&ar=20&at=200000&ip=10&rm=sp500&rr=8&pv=portfolio&p=1&iv=left&hp=1&e=food.housing.transport.health.insurance September 15, 2026 at 10:24PM

Show HN: Neobrutalism.dev – just added Base UI support and added new color theme https://ift.tt/Qus81gi

Show HN: Neobrutalism.dev – just added Base UI support and added new color theme https://ift.tt/9VBFutT September 14, 2026 at 11:02PM

Show HN: I made an automated day-by-day itinerary organizer for my trip to Spain https://ift.tt/FbAipye

Show HN: I made an automated day-by-day itinerary organizer for my trip to Spain For my solo food pilgrimage to the Basque Country, I pre-arranged flights in/out of Bilbao, Guggenheim Museum tickets, Michelin Guide restaurant reservations, bus tickets to San Sebastian and Saint-Jean-de-Luz, walking food tours and hotel reservations in each city, and a long list of pintxo spots to hit with notes on each one's food specialty. Most of this was poorly organized (manually) in my phone's Notes app, with references to specific emails for more information. I arrived on the first day in Bilbao, and the hotel didn't have me in their system by name, so I spent a good 10 minutes frantically scouring my emails for the confirmation number, cursing myself for holding up the people behind me. Over the course of the next few days, while I was relaxing at my hotel (after finally getting in) in between pre-planned stops, I started building a personal tool that would organize all of my bookings automatically and keep all relevant information at my fingertips. Thus, Petrel was born and named after one of my favorite birds. It started as a personal pet project but I shared it with some friends that were looking for something just like it so I ended up turning it into an app and website for all to benefit. It's completely free to try without even signing up. On my next group trip, I built out logical collaboration features to allow groups to vote, split expenses, share notes, and add pictures all within the same ecosystem. I've slowly improved the parsing and itinerary logic and built in convenient features such as weather, estimated travel times, an encryption-protected area to store copies of personal information, and a comprehensive travel log with fun stats. I also made it so that the parsing tool can accept almost any file type. However, the tool doesn't do anything to assist on the planning side and doesn't provide true flight alert pushes, despite having live flight status syncing. I would love for more people to test it out, challenge the parser, and let me know their thoughts and experience. https://ift.tt/12Gm0xy September 14, 2026 at 04:47AM

Show HN: VibeWorld – a shared terminal world for developers and scientists https://ift.tt/Z0hxksF

Show HN: VibeWorld – a shared terminal world for developers and scientists Vibeworld is a persistent cyberpunk shared world that runs in your terminal, where you can meet other developers and scientists to have company at 2 AM while screaming at your LLM :) you can connect with other scientist, chat, and includes also use vocal chat if you want to use it. There is a world and a moon, and they are completely 3D rendered and animated. In the moon there is the "complaint crater": a place where you can scream an insult against your claude or codex session and everybody from the space can see it. You can create your pixel avatar! There is also the "stargazing ledge": a place where you can see the stars and whatever happens in the sky. You can chat in the meantime if someone else is there too, or just watch our Earth. Sometimes you can see galactic wars, and there is a plant that grows but dies if nobody waters it. You can leave it on the second monitor while working (it does not use a lot of RAM): watch the sky when you want, and sometimes you may chat with some other scientist in your same situation. In the world, continents are scientific macroareas, and cities are specializations: you can create your corners where you can speak and help others (but for now it is quite empty). You can install in one command and by just typing "vibeworld" in your terminal, you are inside the world. There are some games inside and a sort of Doom where enemies says to you "Retracted" or "dedline tomorrow" when shooting you :). There is also an awkward reason why I am posting the work here. I have just discovered that another product Vibeworld.build uses the same name and the same concept (shared multiplayer world for developers). I don't know how independently it was developed. It is not a TUI, but it is a full, paid application developed by a company after my project was released. To be honest luckly I find my project more funny. I do care about preserving the chronology, my VibeWorld was released before, and there are verifiable timestamps in the GitHub. I just don't want the original public record to disappear if a much larger implementation of a very similar concept gets more attention later. I would like to hear what people think about it! :)) https://ift.tt/uW6Zh8D September 14, 2026 at 03:00AM

Show HN: Kibble – a reader for AI news, tutorials, code and models https://ift.tt/CyYu7Ic

Show HN: Kibble – a reader for AI news, tutorials, code and models https://ift.tt/IXLPi76 September 14, 2026 at 02:39AM

Show HN: Simpleboot and Easyboot https://ift.tt/TlUR4G0

Show HN: Simpleboot and Easyboot These are FLOSS boot loaders and disk utilities which follow the suckless philosophy ( https://ift.tt/Jgfikas ). Despite being totally dependency-free, they come with a fully featured single command to create bootable disk images out of a directory (utility is available on Windows, Linux and other POSIX systems). Usage is as simple as it gets. There are two repositories, depending on the desired use case: Simpleboot is the little brother, an embeddable boot loader: https://ift.tt/NRFmngY - multiplatform (BIOS, UEFI, RPi, coreboot) - very small (less than 64k), single binary - also flashable into ROM (option ROM, expansion ROM, coreboot payload, etc.) - easy to configure (single config file, checked for validity by the disk creator utility) - one backup boot option - capable to boot Linux and simplified Multiboot2 kernels (see below) Easyboot is the big brother, an interactive boot manager: https://ift.tt/MV0vFib - multiplatform (BIOS, UEFI, RPi) - interactive boot menu for dual booting multiple OSes - easy to configure (similar syntax as GRUB's, but much simpler) - up to 9 boot options - transparent payload decompression - on-the-fly, transparent DSDT and DTB patching - plugin system to support multiple file systems and boot protocols - supports FAT12/16/32, exFAT, NTFS (v3, v3.1), ext2/3/4, XFS (SGI), UFS (v2, v4.4), mfs (Minix3), BeFS (Haiku), plus any root file system (zfs, btrfs, etc.) provided the kernel is on the ESP - supports booting Linux, Windows, OpenBSD, FreeBSD, FreeDOS, ReactOS, MenuetOS, KolibriOS, SerenityOS, Haiku, Multiboot1, Multiboot2 (in PE/COFF, ELF, a.out, etc. formats) - comes with a (very minimal but useful) kernel debugger that works on real machines too Full disclosure: forking Simpleboot into Easyboot was sponsored by The Free Software Foundation Hungary ( https://fsf.hu/nevjegy ), created in appliance of FSF.hu's Free Software Tender 2023. Their default "simplified Multiboot2 protocol" is an easier to use superset of the GRUB's Multiboot2 protocol. This protocol handles native 64 bit kernels (no 32-bit trampoline code), programming language and ABI agnostic (works with native C, C++, Rust, Go, Ada, Pascal, etc. kernels, no Assembly preambule needed), supports transparent higher-half mapping, not tied to x86, and has SMP support too (running the kernel on all CPU cores), passes EDID and kernel segmentation information (not in GRUB), etc. etc. etc. An example kernel is provided in the repo: https://ift.tt/CXP94fc If needed, there's an Easyboot plugin to support the old-fashioned, x86-only, GRUB-compatible Multiboot2 protocol (embedded tags, 32-bit entry point, no higher-half, GRUB ABI). Documentation and detailed specification is available in German, English, Spanish, French, Hungarian, Italian, Japanese, Dutch, Polish, Portuguese, Russian and Chinese (some machine translated, sorry). I would really love to hear your opinion and constructive criticism. Simpleboot is feature complete (only bugfix allowed), but Easyboot is expandable with new features via plugins. https://ift.tt/MV0vFib September 13, 2026 at 10:48PM