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’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).
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.
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.
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.
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.
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
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.
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
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
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.
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
Client posts request and masters respond with their offers.
What you think about that guys?
...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.