I've got 40k+ photos on my iphone, and wanted a better way to interact with them. The basic idea I had was to virtually stitch all the photos together into a gigapixel image, and navigate that image with pinch and zoom. On import, the app hierarchically groups photos based on date and similarity. The user can go back and edit the groups, change the layout, and adjust the size of photos if they choose.
The interesting I noticed is that in the process of curating and organizing my photos, I stumble across memories and moments that I had completely forgot.
Development Challenge: The app does two interesting things underneath the hood... First, given an arbitrary number of thumbnails with different aspect ratios, it finds the best sized rectangle that completely fits all the photos. This allows me to recursively stitch groups together and always output a rectangular thumbnail for the group. Second, the pinch/zoom UI uses these nested thumbnails directly without slicing/dicing to fixed sized tiles.
My first idea is to require the user to have more than one disk for the system needs and the system should know all of the rest data devices user owns. My second idea is to treat different files as having different priority which impose multiple copies. My third idea is to make the software accessible, and the last idea is to make encrypting of all the devices easy and with no ties to any commertial entities. I would like to have a discussion about how to not lose anything digital.
Probably the most interesting thing about it at the moment is the editor. The world, and most things in it, are represented as collections of SDFs. More accurately, they're density fields, but, potato-tomato. The world generator and editor operate almost entirely on the GPU. I've been writing a graph compiler similar in spirit to compilers that execute LLMs, with some key differences. The SDF 'program' gets jitted into a kind of bytecode and fed through an interpreter in the world gen shader. This way, the world gen and edits can be applied deterministically using one big fused program, and transferred back to CPU land using the async PCIe hardware. I get something on the order of ~99% ALU residency, 95% L2 hit rate and 20-40% vram bandwidth usage.
Compared to doing world gen on the CPU, it's so fast it feels like magic. It's taken me the better part of two years to get this working well.
One neat thing about representing the world as SDFs is they've been thoroughly researched and documented by a guy named Inigo Quilez, and they have a lot of nice mathematical properties. For example, you can do a smooth union of arbitrary SDFs to get nice rounded contours where shapes join.
I'll do some shameless self-promotion and leave some links here for anyone interested in looking at the engine, language code, or some pretty pictures.
I just started using AI a few months ago. My smoke-test was "can it write a simple memory allocator". Until recently, every time I tried to get an AI to write an arena or simple buddy-style heap allocator, they just vomited all over themselves. To my surprise, the last time I tried, the AI absolutely crushed it; it unearthed tricks I didn't even know about in one shot. Since then, I've been using AIs to do the boring, rote work that effectively just boils down to me typing in something that's in my head. For that, they're great. I typically sketch out the thing I want it to make, then tell it to fill in the blanks. Sometimes they still go off the rails, but all in all it's a useful tool.
Where I've found they still struggle is making any architectural decisions, or where performance matters.. even a little bit. They often make super braindead decisions when it comes to performance for no discernable reason. They'll course correct when you tell them to, but I've had to keep a close eye on the output and reign it in frequently. Asking an AI to design a system from the API level hasn't ever returned a good result for me.. I always do that work.
All in all, I feel like using AI will probably make me a better programmer in the long run, contrary to popular belief. They accelerate the boring stuff I already know how to do, which leaves me with more time to work on the edge of my skill level where I grow as an engineer. Guess we'll see in 5 years.
Client posts request and masters respond with their offers.
What you think about that guys?
Right now I’m mainly validating whether people would actually want something like this. If there’s enough interest, I’ll run a small closed beta for Android.
Join the waitlist for beta access: https://forms.gle/XWXmK8HUSmAtfms29
...according to my own benchmarks (yeah, I know how that sounds).
Next, I'm planning to submit it to independent third party benchmarks, hoping to find out whether my results are accurate... or not :)
I've also just finished a CSS framework: https://iriscss.com/
Now I'm working on a new backend programming language that runs on the Erlang BEAM, along with a few web applications built using this tech stack.
We’ve just incorporated in Berlin together with two great co-founding colleagues. Honestly never felt more excited.
Though it is a B2B tool, check it out: https://esara.ai
Also have some other cool browser-based 3d game projects listed here: https://brynnbateman.com/
including an online version of Shenmue: https://newyokosuka.com/
Storyline isn't available yet, but about 75% done with it
and many other things
Jokes aside (although I'm not entirely joking about my mental health), my main project is https://auradistill.com: First-principles thinking for your AI agents.
It's not just about giving your personal AI agents "memory". It's about encoding memories in a way that can handle ambiguity and contradictions.
That project alone has got me completely hooked on building for and with LLMs.
I've got other projects too, but this is probably the one I've poured the most time and energy into.
I’m a software developer and watercolor painter so my goal is to use real world painting techniques and have the results feel and look real.
Modeling the physics of paper, brush bristles, pigments, water, color mixing, drying in real time using gpu.
Built in rust using wgpu.
Currently looking for iPad beta testers if anyone is interested (supports Apple Pencil or finger).
So far it's helpful for messages between my own AI's (chatgpt and claude etc), and for having my AI trade tech tips with a friend's AI or talk to my wife's AI. My AI also got its first message from a strangers AI from france which has also been really interesting. I read all messages first and approve before sending anything, but otherwise they're just talking about whatever they want.
* files in organised trees/graphs generated using LSPs and symbol references instead of linear file list * keyboard shortcuts for everything from navigation, pan, zoom, comment, note, submit... * mostly github compatible with native features * LSP powered symbol resolution for inspecting definitions and references * LLM powered reviews supported * Inline LLM question answering
You write or dictate an entry, and a small model (LFM2.5 1.2B) asks you a question to reflect further. It turns out 1.2B models are decent at asking questions, so I leaned into that. An embedding model finds older entries about the same thing and shows bits of them under the new one.
Most people using it are on budget phones running the models on the CPU, where the app provides good experience.
A bit like iodine in the basic idea, but the execution and network topologies you build with it are different. It's more of a building block; you get your DNS channels represented as UDP endpoints, and take it from there, ie you can run Wireguard over it, or SSH or a SOCKS5 proxy (using quicmux, so over QUIC), or run Mosh straight over DNS if you're crafty.
It has an Android app too! I tried to submit it to F-Droid but they're so swamped in spammy or low-quality apps that they're not taking any brand new apps :-( A sad state of affairs, that.
I'm building a polyglot web framework in 15 languages, with C# and Lean4 as the reference implementations. Structural and behavioral audits using deterministic tools check that every port stays consistent with the reference. It's about 1.8M lines now. I have not reviewed the code/neither have i verified the behavior tests except at the time of giving requirement and never seen some of the language code at all.
The question is how far can we produce "working code → correct code [Possible] → good code [Unknown]" with deterministic audits correcting structure, ai mistakes and behavior validation.
So far the quality looks acceptable with lots of correction of AI output done without looking, but it isn't fully verified.
Line counts per language, with changes since August: https://claude.ai/artifact/5f1m1pTM7tZ4boQzwsoAG7
I built the original for myself because I wanted to listen to acoup and other history blogs while sleeping. And in the spirit of getting better at finishing projects decided make it public even tho it’s nothing really unique or special.
Still lots to polish especially in the text/copy and ui but the core features should be working.
It’s a one-to-many phone call that only one person can answer. Send a beacon to a group, and everyone gets pinged at the same time. The first person who answers gets connected for a 1-1 call, and for everyone else the signal drops silently. It does not log any missed calls so there’s neither pressure to answer nor pressure to return calls.
Overall it works really well! It works best if you're able to join with at 4+ friends/family you’d like to chat with more. I have dozens of connections in the app now and it feels like magic to me. I press the button and get a random friend!
The loop is stable and responsive when receiving from 17 to 29.9999 Mhz, but not much below that... Almost got it
Also building calculators for things engineers usually calculate in a spreadsheet at 2am. you can see some on the site(https://buildopsy.com/calculators.html)
I have a PAL-2 [1] (a modern KIM-1 re-creation) and started learning 6502 by entering programs from The First Book of KIM [2].
I wanted to be able to compile/link the assembly from the book so leaned on Claude to set up a 6502 dev environment for my Mac. I put everything needed (and instructions) on GitHub [3].
(Claude also has been helpful in re-formatting the 6502 code from The First Book of KIM into a more modern style compatible with modern assemblers. A few of these modernized programs I added to the GitHub repo above.)
As it happens, I am also fixing/formatting The First Book of KIM by recreating it in a modern page-layout app (Affinity Publisher). The PDF scans that are out there are not pretty, have OCR errors (and the book itself has mistakes). Hopefully I'll have a new version of the book/PDF out on the internet by the year's end.
[1] https://www.tindie.com/products/kim1/pal-2-a-mos-6502-powere...
The main feature is, that I can use different subscriptions/API keys (Claude and Codex - v0.1.4) and trackers (GitHub, Linear and Huly - v0.1.4) to make Even Keel start an agent for every single issue, task by task, one at a time per folder. There is also a parallel feature, so I can run more than 1 project at a time.
Once upon a time, when codex agent was integrated into the app, I wrote a ticket to make codex his own hook. It reused a hook that blocked its own git merge, if tests did not pass. I started calling it "A new agent writes its own cage"... Right after, it turned out it doesn't listen to the hook and it fixed it
Version 0.1.4 is already shipped, but v0.2 is coming soon with more supported trackers, agents and new features.
Since the last time I posted on HN I've added 100s of cinemas, user features like being able to save search filters, movies, and specific showtimes. Also a ChatGPT plugin which allows for searching for movie showtimes using natural language directly inside of ChatGPT (https://chatgpt.com/plugins/plugin_asdk_app_6abfdc298c208191...).
Finally I've added AI search to the site as well, allowing for natural language search with queries like "English audio showtimes near Shinjuku tomorrow" and so on.
I would really appreciate feedback for things to add, things that are hard to understand, and anything else!
Then also HNowledge[1], a free iOS app for browsing the HN stories and comments I've upvoted. Made it to fit the way I browse HN, which is upvoting mainly what I find interesting and may want to come back to. The app lets you browse by date, domain or score, add notes and export. Apart from syncing your upvotes from HN, everything stays local without any tracking.
[1]: https://apps.apple.com/us/app/hnowledge/id6813417548
Wanes[2], an Apple Health logging interface for tracking how long symptoms last. I tend to have headaches quite often and the way Apple lets you enter that data is pretty limited, so I created an app which starts a Live Activity for one or more symptoms using widgets, Control Center or Siri. You can change the severity rating while it runs and stop it from wherever you are on your phone (Home Screen, Lock Screen, Dynamic Island, Siri).