Posted by david927 1 day ago
Ask HN: What Are You Working On? (March 2026)
My favorite output so far is that I asked it what life was and in a random stroke of genius, it answered plainly: "It is.".
It's able to answer simple questions where the answer is in the question with up to 75% accuracy. Example success: 'The car was red. Q: What was red? ' |> 'the car' - Example failure: 'The stars twinkled at night. Q: What twinkled at night? ' |> 'the night'.
So nothing crazy, but I'm learning and having fun. My current corpus is ~17mb of stories, generated encyclopedia content, json examples, etc. JSON content is new from this weekend and the model is pretty bad at it so far, but I'm curious to see if I can get it somewhere interesting in the next few weeks.
A site for anti patterns in online discourse.
Example: https://odap.knrdd.com/patterns/strawman-disclaimer
Need to gather more patterns then create tooling around making it easier to use.
The goal is to raise the quality of comments/posts in forums where the intent is productive discussion or persuasion.
2. boulderinglist.com - a catalog of climbing gyms across the world
3. livedsupport.com - AI driven support for colorectal issues
4. radiusing.uk - find where people are willing to do something to improve the social fabric of where they live
5. llmstxt.studio - AI-SEO via sitemaps, llms.txt, and AI search
6. probe.bike - tell stories with your cycling data
need to find a way to get more sleep
Bringing back the hobbyist self-made PCB workflow since it can be a headache to get designs back from PCBWay / JLC between customers, tariffs, shipping..get more quick to close the prototyping loop.
Here's a YouTube playlist of how to make PCBs at home for anyone who wants a deep dive. https://www.youtube.com/watch?v=Hm_JrACrmVs&list=PLWDQgxl-jH...
I have been using online courses and youtube forever and they all have converged on a similar format. Basically, adapt a textbook to slides and add voiceover. Sometimes they'll be animated. Not for everyone but I like it for passive learning.
Being a web developer I always thought video was a strange way to deliver this information - you can't even copy the text! Videos are also hard to make and heavy on bandwidth. So after iterating on different approaches to this over the last few years, I finally started on a new iteration called useful.
Currently on rev1, so early days: https://github.com/ohmstone/useful/tree/rev-1
There are a few more things I want to add to it but I want to get back to what I was doing (the bass guitar stuff). So I will make few of these website-as-video courses based on my projects to try and prove the concept.
Some of the nerdier features useful has:
- Uses state of the art CPU-based TTS with voice cloning, realistic enough to not be distracting
- Very simple markup language to create the visuals
- Extensible slide content with simple plugin system
- Full website export with complete SEO/social metadata
- Export is a PWA, so it caches nicely and can work offline
- Self-hostable
- Export is optimized for low bandwidth, so it loads way faster than a video and uses <1/10th of the data when served with brotli
- Minimal dependencies
Beyond my own use-case I figured it might be useful for others creating courses. One stretch-goal would be for people to turn what they are learning via LLMs into low-bandwidth courses like this so we don't have people burning energy asking the same questions and watching the same 4K videos.
Applications on the public cloud raise strong concerns about data protection. As an architect, I spend a meaningful part of my time ensuring the security of customers’ data in the cloud.
Bao introduces an innovative approach where data remains on local devices while the cloud provides encrypted storage for synchronization and peer exchange. Because cloud providers cannot access the data, the need for due diligence is reduced.
Any feedback is welcome.
With nao we are exploring the fact that agents are good to work with filesystems, so we help people getting the context into the filesystem and then you can plug our agent loop on top (on anything else).
The most exciting part is working with sandboxes and emulating filesystems. So at the moment the filesystem is local, but it could be a bucket or a database that we emulate as real filesystem for the agentic loop, all of this can also be mounted to sandboxes, and in the end you can do data transformations all in isolation with control on the context.
The original Python icloudpd is looking for a new maintainer. I’ve been building a ground-up Rust replacement with parallel downloads, SQLite state tracking, and resumable transfers. 5x faster downloads in benchmarks, single binary, Docker and Homebrew ready.