@rohrich_steven Hey Steven,
I'm glad to hear you like the time stretch player. Feel free to send me an email with suggestions. I'll gladly consider them for future updates.
Cheers,
Jonas
@JacobJoaquin I think https://t.co/b8WZlQmC8H is where I first saw that. What I find interesting is that it while it does reproduce the code fairly exactly, including comments, it leaves out the copyright notices. Would be interesting to know whether that was intentionally trained in RLHF :)
The sun rising over the alps. Another shot from my over-engineered #rustlang hill shading path tracer. I didn't properly account for the curvature of the earth but I think it looks quite pretty. :)
The DEM used is ASTGTM v003.
@joelprestonsmit@cyb_detective@BenDoBrown@Istihbarat1800L@LHreports Didn't think of that but I guess it could be useful to some extent. I guess in practice things like cloud cover will have a fairly big impact though. Also I've heard the earth isn't really all gray outside. :D
Having a bit of fun over engineering hill shading by path tracing digital elevation maps with an atmospheric simulation. (:
Coded in #rustlang and running on the CPU. Very slowly.
@nyuriks I think as it is it would only make sense to prerender hill shades. Building an interactive webg(pu|l) version with perspective projection could be fun though. It's been a long time since I've played with terrain LOD. 10 years judging from the URL. :) https://t.co/cDmw7MzRdE
@ludikalell In it's current form it wouldn't make much sense. This took about an hour to render on 16 cores - it's far from real time. It also depends on gdal which probably doesn't want to be compiled to wasm without a fight.
But the 1000x speedup to realtime is surely possible so maybe. :)
@kinkeyding You could automate it using any web automation/testing tool. Alternatively I could give you the relevant bits of javascript that do the dithering in case it helps.
There is also an unfinished version two around in case you want to give it a try: https://t.co/iZRV9YD1eH
@philkellr Training models with a more narrow scope also seems to be within reach.
But there is also no denying that not having access to the amounts of data, labor and compute available to big tech companies is limiting.
@philkellr But it's actually something I struggle with as well as I spend more time playing with deep learning. Training costs are high and I can't offer inference for 100'000 people on a little hetzner box. The UX of trying to run it on the client is generally quite bad as well. :(
@philkellr The largest language models are out of reach for most hobbyists such as myself even if there seems to be some progress https://t.co/NR8xMaPnwQ
Fine-tuning smaller models is definitely doable and can yield surprisingly good results.
@philkellr In those rare moments where I have a bit of time to spare at work I play around with tutti and ricardo data. :)
So in short running around mildly whacking things with my new golden hammer. :D
@philkellr At this stage it's just playing around. Tried to model (tube) guitar amps. Playing around with a learned version of smartcrop.js. Want to try to use a GAN for tuning the phase-vocoder in the timestretch player and experiment with "neural paint brushes" on top of stable diffusion.
@philkellr Hey Philipp (:
It's 29a, mostly the time stretch player thing that gets users. Mostly static, mostly cached. Keeps it cheap and somewhat maintainable. :)
@GuitarKidGames @threejs It's a random initial velocity, a random force field based on 3D simplex noise with a little bias towards the center so fewer particles escape. :)
Made a little GPU particle system with @threejs. 16'777'216 particles dancing in a force field. Somehow completely forgot how much I enjoy playing with particles. :)