Ask HN: What are you working on? (May 2026)

Posted by david927 20 hours ago

What are you working on? Any new ideas that you're thinking about?

207 points | 744 commentspage 4

jballanc 16 hours ago|

I've been working on RVW, my adaptation of the standard transformer model that is capable of online continual learning without catastrophic forgetting. I finally published the first pre-print of my early experiments: https://doi.org/10.5281/zenodo.20064617

Now I'm working on expanding the work into more parameters and improving performance. I just finished an extremely harsh test of a Nemotron-flavored RVW that consisted of stretches of a random assortment of domains interspersed with long runs of single domains. Across all of it the model didn't forget (and actually improved on some of the more challenging domains). PPL on SmolTalk is still in the ~18 range, which I'd like to get lower, but this is all with only 4B params.

Currently, I'm training a Llama 3.2-flavored RVW with only about 2B params to see how that turns out. Depending on results of that, I may take it to Gemma 4 next.

schipperai 11 hours ago||

A better permissions layer for coding agents. The tool works like auto-mode for Claude Code, so you can stay in the flow and only get prompted to allow or deny tool calls when it truly matters, but it is fully deterministic. My benchmarks surfaced that most Bash calls don’t need an LLM to be classified as safe, ambiguous, or dangerous. A deterministic classifier can auto-allow or block 95% of Bash tool calls as safe or dangerous, with only the remaining 5% being truly ambiguous or unknown.

Conclusion is permission reviews with LLMs like Claude’s auto mode or Codex auto review are like using a data center to flip a light switch - overkill.

The main benefit is that your agent’s autonomy can be governed deterministically through policies that can be stored at the user and repo level. The bonus is that you save tokens vs using auto modes.

https://nah.build

jstrebel 6 hours ago||

I am trying to build a simulation that lets a simulated organism come up with its own small language, purely learned from sensory input: https://github.com/JoergStrebel/VirtualZoo/blob/main/compute... I would like to implement the ideas put forward by Stevan Harnad in his symbol grounding problem paper (Harnad, 1990).

krogenx 2 hours ago||

I am in the process of launching my poultry RFID tags, “eFlock Smart Poultry Tags”. They allow farmers to count and identify all of the birds that they have. The project goes along with a mobile app, Manger, a tool that can be used to gather data about livestock (not just poultry). Tags should be available sometime this week, perhaps early next week on Amazon. Here’s a video of scanning: https://youtu.be/wilixJiyPYA Tags on Amazon: https://www.amazon.com/dp/B0GWTB1DR9 Manger app: https://www.manger.app/

jeanlucas 1 hour ago|

You lost the opportunity to call it io-chicken

t_mahmood 3 hours ago||

I am working on a minimalist journal app that is for really quick single line jots. Can be used for idea dumping, project management, quick calculations, unit conversion, task management etc. Have unlimited undo/redo support, assign tags etc. Will be adding scripting support using Python next. The data is stored in text file, in really human friendly manner, but also in way so that the *nix tool users can easy to navigate the file using text processing tools.

Native application, no web UI, built using Rust + iced.rs, minimal dependency. NO AI.

I am putting the best effort to make it performant. Target audience is the users who want's the simplicity of the notepad [non-sloppy one], but still with some bells and whistles to note without worrying about managing the metadata manually.

I think with scripting there will be infinite possibilities to play with linear notes, and I want to make that happen.

Continuous challenges while implementing features are:

  1. It should load instantly
  2. Keeping it extremely simple to use
  3. Keeping the interface minimal
  4. Still have ways to let the user find the features easily.

Will have a demo version ready soon

thisisjedr 2 hours ago||

I'm working on https://docx-editor.dev/, the open-source, mit licensed word editor library.

We are building this because we such library it in our core business, and a lot of other engineers seem to need it too. We have contributors showing up with bug reports and fixes, and real interest from people building apps around .docx docs.

My previous show hn post (https://news.ycombinator.com/item?id=46947229) got a lot of skepticism because we're developing heavily with AI, but with active community feedback and proper ai oversight (mostly me), I'm super proud of what we have now.

maxbond 11 hours ago||

I've been learning to crochet. I'm trying to do more hobbies with my hands, but it's also pretty interesting from a mathematical perspective. The fundamental primitive (the chain stitch) is like a series of slip knots, and each stitch is reversible. So the piece is actually a series of reversible transformations. The yarn is sewn in at the end to secure it.

This has some interesting implications. If you make a mistake, you can always backtrack and try again. If you have a crocheted piece, at least in principle you could find the lose end, free it, and work back stitch by stitch to reverse engineer it. (In practice people don't seem to do a stitch-for-stitch reverse engineering just like you probably wouldn't bother reimplementing something line by line without a compelling reason, you figure out what's going on in the challenging places just by look and feel and improvise from there.)

I'm oversimplifying somewhat and there are some forms of crochet that include irreversible stitches, yarn can be felted together (entangled, like a cotton ball) to create irreversible bonds between adjacent strands, and often several panels/pieces are joined together irreversibly to create a larger piece.

maxbond 7 hours ago|

I guess I should be clear that by "irreversible" I mean a transformation like the following: "to cut the yarn with scissors, to untie a knot that was strongly bound, or to felt together." So a slip knot is "reversible" in the sense that if you tug on it, it easily comes undone, whereas an overhand knot would just get tighter. You can think of felting as being equivalent to tying a lot of overhand knots between adjacent strands, they become permanently attached and could only be torn from each other.

planckscnst 5 hours ago||

I'm working on [Context Bonsai][1] - LLM harness tools that allow the LLM to prune messages out of the context, leaving behind a summary and keywords instead. In addition to a "prune" tool, there is a "retrieve" tool that allows it to recall the messages if needed.

In addition to these tools, I'm also building automation that will port the tools from the reference implementation (OpenCode) to other harnesses (Claude Code, Cline, Pi, Gemini, Kilo, Codex, others to come?). As well as automation that will either cherry-pick or re-implement commits onto the latest head from upstream.

[1]: https://github.com/Vibecodelicious/context-bonsai-agents#con...

[2]: https://blog.vibecodelicio.us/posts/how-i-fixed-context-wind...

letharion 3 hours ago||

Building a GNSS/RTK-based lawnmower.

The existing ones were quite expensive, especially when I started out. A friend had the idea to get a cheap/non-functioning lawnmower second hand, and tear out the circuit board. We're in the process of coding up a new ROS2 based stack that will roam the lawn on GPS with RTK in the charging station. My friend does most of the electronics stuff, and I focus on the software.

I'm at the point where I will start testing a simple bounding box soon and just have it drive around until it "hits the edge" and then randomly pick a new direction.

It's fun so see the software I build "in real life" instead of as a web-site, as is the case for my my daily job.

Jemm 2 hours ago|

Would love to see your progress as I am very early on in the same project. I was thinking of using ArduPilot and some custom waypoint generators.

letharion 1 hour ago||

Feel free drop me an e-mail on [nick] @ [the big G].com and we can discuss. :) Haven't seen ArduPilot before. Since I'm using ROS already, on of my goals is to lean as much as I can on that and avoid custom work where I can. Nav2 appears to be the most common "go from point A to B" framework in ROS-space.

hxtk 14 hours ago|

I made a Python tool to build distroless container images for projects managed by uv. It draws inspiration from Ko from the Go ecosystem and works with/depends on uv from the Python ecosystem, so I smashed them together and called it Kuvo: https://github.com/hxtk/kuvo

It’s a hobby project in a very early state where it technically works but it’s missing several things I think it needs before I’d use it for anything serious. As of right now it isn’t even complete enough to dogfood a minimal container for itself without an intermediate base image because it can’t target a platform compatible with the distroless uv container image.

More comments...