iiinova — where your digital self keeps evolving

copper-irl.bsky.social @copper-irl.bsky.social · 03/10/2026

I’ve had a curiosity about this for a while: what would it look like? Would it make space where it occurs almost look… foggy? Would relativistic frames matter, so it’d be angle and color dependent… kinda like the sky? I’ve never gotten good answers for this. It’s a cool rabbithole. The math is hard

011 0

copper-irl.bsky.social @copper-irl.bsky.social · 24/09/2026

I had to force myself to switch recently because of weird version differences between Linux, embedded, and windows. ulong is uint64 on Linux, uint32 on windows. Embedded systems can do weird shit like “a char is 13 bits!!!” and explicit types help a lot. The _t suffix sucks though. Baffling syntax

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 23/09/2026

I really enjoy “small models”, like tens of params. Things that can “learn” on live data. Differentiable splines, updating voxels from screenspace, etc. Running Adam on live data per pixel has some really interesting dynamics to it. Highly suggest if you havent! It’s visually very different to EMAs.

131 0

copper-irl.bsky.social @copper-irl.bsky.social · 23/09/2026

I feel there’s genuine usecases in graphics for imposter generation and GI approximation, but it isn’t worth burning all of art to summon this tech. Kinda like nukes? Fascinating, horrible devices that should stay in the past.

120 0

copper-irl.bsky.social @copper-irl.bsky.social · 23/09/2026

ML makes me so bitter since it’s genuinely fascinating: ResNets bridging CNNs to LLMs, Grokking, Autoencoders, all super cool. Softmax gives it the same math as thermodynamics. They learn like how we’d cool a crystal! …but it’s exclusively used to poison the world on stolen resources.

120 0

copper-irl.bsky.social @copper-irl.bsky.social · 17/09/2026

Oh I hope you don’t merely copy what exists! In my experience people approaching a problem with near zero experience tend to produce alien and very cool results. I do the same thing, so your process has been fun to watch heh

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 17/09/2026

Whatever it is, it’s interesting! It seems similar to a Particle Mesh method in astrophysics. If it is, you may find value in extracting gradients with an FFT for way higher precision. Those usually live as “Spectral Poisson solvers” though. Barnes-Hut may also be interesting to pull from as well

120 0

copper-irl.bsky.social @copper-irl.bsky.social · 28/08/2026

I really need to aluminize my telescope’s mirror, but that first shot turned out great

041 0

copper-irl.bsky.social @copper-irl.bsky.social · 17/07/2026

It’s exhausting seeing what new thing they’ve plundered from us every week. Private equity behavior stripping the copper out of the walls. I gleefully await the day they’re gone so we can begin to clean their decade of rot

020 0

copper-irl.bsky.social @copper-irl.bsky.social · 11/07/2026

It’s the only field I’ve ever seen where they *prefer* magic over science and it gives me a headache navigating that mess every time. AI bros feel like it’s natural evolution. “Can this be made in js?” No, I don’t want to

080 0

copper-irl.bsky.social @copper-irl.bsky.social · 10/07/2026

My phone almost died on the plane so I resorted to literal napkin math. Trying to adapt axis-angle rotors from gamedev to astronomy using fft image alignment. I fear I’ll never get an engine dev job I’d like with all the layoffs so I hope this gamble works. The gutting of iD stings hard :(

050 0

copper-irl.bsky.social @copper-irl.bsky.social · 10/07/2026

Took a flight to Denver and wanted to measure the cosmic ray radiation when in the air. Finally figured out how to extract data and fit an MLE to the Geiger ticks. An exponential model from altitude to cps is a shockingly good fit!! cps ~ exp(A+B•H(C+Dt)) (Had to calibrate its clock via C & D)

GMC-300S Geiger counter, measuring 114 CPM with a bar graph showing counts per second.

Clouds seen out of an airplane window in the skies above Denver, Colorado on 2026 July 1.

Plot of fit data of the Geiger counter’s detected ticks to a Poisson distribution using an MLE. Orange is altitude data taken from FlightAware passed into an exponential model: exp(A+B•H), the blue dots are raw geiger data, the blue line is Fourier smoothed data. I calibrated the clock with 3 photos taken with my phone, noted as arrows with dashed lines and red text.

010 0

copper-irl.bsky.social @copper-irl.bsky.social · 10/07/2026

I’m a bit late here but if you prefer linear algebra all of trig collapses to one formula: [[cθ, -sθ],[sθ, cθ]]•[cφ,sφ] = [cos(θ+φ), sin(θ+φ)]. Rotate a rotated coordinate via a matrix. Phasors may help too! It’s what demystified trig for me: [1,𝒾]•[cθ,sθ] = cos(θ) + 𝒾sin(θ) = exp(𝒾θ)

000 0

copper-irl.bsky.social @copper-irl.bsky.social · 12/06/2026

I’ve gotten pissed off enough at this to heavily consider doing presentations in Godot directly, but it has limited video decoders. Other than that? I don’t see a reason why not. Super tempted to see if ffmpeg can be used as a decoding library in some way to do this

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 24/05/2026

That's too bad. :( I remember a convo we had back at twitter which made me heavily consider working at seed when I graduated. Even in physics the current research all needs ML flavors to get funding. It sucks.

130 0

copper-irl.bsky.social @copper-irl.bsky.social · 23/05/2026

I’ve encountered this even when you correct LLMs, it’s irritating. For example, ChatGPT refuses to accept relativistic boosts of 4-velocity to be simpler than the velocity addition rule. Even when shown a proof and it agrees, it asserts is wrong. I wonder if “AI Inertia” will be a problem soon

020 0

copper-irl.bsky.social @copper-irl.bsky.social · 30/04/2026

The model complexity spiked again. Now I'm doing full affine spectral gradients. The better the model the "more" seems to appear. The next step is a PSF spectral gradient, and its stressful. I feel like if I cant present at APS, a door in life closes. Classes keep getting in the way :( #astronomy

030 0

copper-irl.bsky.social @copper-irl.bsky.social · 26/04/2026

This has fun consequences for rolling objects with different mass distributions on a slope. For a solid disk, I = mr²/2. For a ring, I = mr². The ring will roll down the hill genuinely slower than the disk does. Energy & momentum conservation acts as the constraints that determine by how much

010 0

copper-irl.bsky.social @copper-irl.bsky.social · 26/04/2026

Whoops, I pressed send too fast. As more energy gets dumped into angular momentum, less gets dumped into linear momentum, whose sum of kinetic energies is the value you expect.

120 0

copper-irl.bsky.social @copper-irl.bsky.social · 19/04/2026

I was avoiding mentioning oxbow lakes for one of his future posts to not be that guy but unlike simulating the dynamo for aurora it may actually work in terms of gameplay lol

010 0

copper-irl.bsky.social @copper-irl.bsky.social · 19/04/2026

This detail wouldn’t add much to the game but if you stick with the “wedge taken from the earth” look like when you zoom out here, it’d be really funny if the wedge glowed with a color temperature. Obv only if there’s minimal perf cost lol

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 04/04/2026

THE PHASE LOCKED It’s working It’s working It’s working It’s working It’s working It’s working It’s working It’s working It’s working It’s working It’s working (Almost) NONE of these objects are in the Simbad catalog! What a treasure trove. 128x128px, TESS sector 19, cam1, ccd1. #astronomy

040 0

copper-irl.bsky.social @copper-irl.bsky.social · 02/04/2026

I did a little digging with that variable star, and I think it’s a constant binary! It has a very strong resonant frequency across the whole image, so I wonder if I can extract the PSF from this alone #astronomy

000 0

copper-irl.bsky.social @copper-irl.bsky.social · 02/04/2026

I think TESS’s pulsar calibration has some precision issues. There’s significant subpixel drift in this test crop. But interestingly, we’ve got asteroids now :) I’m shocked at the number of variable stars in this tiny 64x64 section. I wonder if it’s like this everywhere? #astronomy

130 0

copper-irl.bsky.social @copper-irl.bsky.social · 29/03/2026

spectral volumes sure are *weird*. This is 32 frames of filtered TESS video, with a spatial blue noise filter (airy disc) and each image is accumulated via a sinc window. Each frame is a unique temporal frequency. Complex phase is hue. I think the waves are asteroids? They're plane slices!

020 0

copper-irl.bsky.social @copper-irl.bsky.social · 28/03/2026

True / LIES makes for a great substitute with very funny syntax if(condition == LIES) {…}

030 0

copper-irl.bsky.social @copper-irl.bsky.social · 21/03/2026

I’ve got a problem though and I can’t quite figure out how to get around it. I want to share videos, but the noise nukes quality in every video encode I try. Does anyone here know of some good denoising algorithms I can try from either of these realms? I’m pretty alien to denoising in both.

020 0

copper-irl.bsky.social @copper-irl.bsky.social · 21/03/2026

Holy shit I finally got this #astronomy project with #gamedev tech working at a semi reasonable speed. 40 minutes per sector sure beats 4 days! Still not happy with it. Astropy is slow. SCIPY is slow. Nothing in this realm has a hint of systems programming and there’s SO much headroom for more.

Psuedospectral Fourier transform applied to time only per-pixel of Tess sector 4, Camera 1, ccd4. Plotted complex phase as hue, tonemapped luminance. Asteroids are visible as rainbow streaks, variable stars have a color. ~400 images used for a total exposure of ~1 week.

Lomb-Scargle Periodogram applied via a Psuedospectral Fourier transform on time only per-pixel of Tess sector 4, Camera 1, ccd4. I believe this is the same slice as image 1, but with real LS instead of complex phase.

Asteroids are visible as streaks. There are two gaps in the dataset associated with the earth getting in view of the camera, and data downlink. Both are visible in asteroid trails. The tile pattern is a failure of my algorithm to handle a lens flare.

Zoom in on a color test with the full two week dataset from image 1. Note the strange asteroid trajectory and gaps. I suspect this is a lens distortion artifact, which greatly complicates processing. It feels surprisingly similar to working with distorted normal maps!

1113 0

copper-irl.bsky.social @copper-irl.bsky.social · 11/03/2026

I’ve been doing exactly this with astronomy lately! It feels like peering into a parallel world of precision over speed. They have some really cool tools that are begging to be accelerated, like thin-plate splines or sigma clipping. TPS feels a wee bit like radiance caching. It’s super neat!

040 0

copper-irl.bsky.social @copper-irl.bsky.social · 06/03/2026

Facebook is bizarre. It’s the first entity I’ve seen that late stage enshittified. From what I remember, their strategy was: if you didn’t *constantly pay more* you got deranked in the algorithm. Even for advertising. Insane anti social behavior for both consumers and advertisers

210 0

copper-irl.bsky.social @copper-irl.bsky.social · 14/02/2026

After profiling the first step, the primary perf hog (~30s per frame) is something called a “Thin-Plate RBF spline.” I don’t think we have anything like this in #graphics or #gamedev …but we should. It feels a little like irradiance caching. Maybe radiance cascades could be handy here? #astronomy

Thin plate RBF spline, taken from https://shape.polymtl.ca/lombaert/thinplates/

061 0

copper-irl.bsky.social @copper-irl.bsky.social · 10/02/2026

That one is a wee bit of a fib because I only transfer the data once. I’m assuming I can parallelize transfers while a compute job is going for said test. If anything, it’s the upper limit of what this simple code can do. The streams are a command queue vs requesting each kernel individually

010 0

copper-irl.bsky.social @copper-irl.bsky.social · 10/02/2026

#Astronomy feels SO much like peeking into a parallel world from #gamedev. Its fascinating. Its like if we only did pixel shaders for decades with color only, no normals Tested a 2048x2048 image convolved with a 32x32 kernel. Literally a bloom shader lol, ~500fps Gonna blow some minds with this

2048x2048 convolution with a 32x32 kernel performance comparisons. Time is log scaled.

2102 0

copper-irl.bsky.social @copper-irl.bsky.social · 09/02/2026

Since each generation is 15 years apart, by extrapolation, 21 will be funny again in 2371

000 0

copper-irl.bsky.social @copper-irl.bsky.social · 07/02/2026

I was finally given an astronomy research task, the first of which is to install something called "Topcat" Oh my god hello again web 1.0 I wonder how much room there is to transfer the last 20 years of graphics tech into this field. Java's a bit of an odd choice nowadays #astronomy

120 0

copper-irl.bsky.social @copper-irl.bsky.social · 09/01/2026

Are these GA Rotors? It almost has the exact half angle rotations we see in that realm, but in this case both sides are positive? Weird, but neat!

000 0

copper-irl.bsky.social @copper-irl.bsky.social · 08/01/2026

This all ignores Elon’s dark descent since buying twitter. He’s gone fascist, abandoned his companies, and turned twitter into 4chan’s worst boards. His mission went down the drain of that sink he let in

000 0

copper-irl.bsky.social @copper-irl.bsky.social · 08/01/2026

The fucking “@grok put a one-piece bikini on it made of thin strips of fabric” shit is so god damn antisocial it feels like a Facebook feature. We all know how that went. Shocked any artist remains there. What a disgrace to any artist to ask grok to pave over labor with corporate slop

200 0

copper-irl.bsky.social @copper-irl.bsky.social · 08/01/2026

I was debating cross posting to the dead bird site for engagement. Every time I go there all I see are cool research papers! But then I scrolled a little further and between the research there’s “yeah, we SHOULD kill the Jews” and in every art reply there’s ai generated CP and Yeah I’m good.

130 0

copper-irl.bsky.social @copper-irl.bsky.social · 08/01/2026

Do people usually not importance sample the human luminance sensitivity spectra and treat the rest as procedural color? I’ve done that to great success for visualizing relativistic blackbodies but maybe I had the precise niche where that worked well

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 01/01/2026

Man I’m having such a shitty 2026 so far. • my nose was kinda itchy earlier God it hasn’t been my year. Can’t wait for 2027

030 0

copper-irl.bsky.social @copper-irl.bsky.social · 25/12/2025

Whoops!! I forgot to share the results! It still has a few odd rendering bugs, this is a Kerr black hole. Runs in real time with my modified Kerr metric using a normal vector parameterization. Still couldn't squash the horizon reflections, but the frame dragging is super visible now!

050 0

copper-irl.bsky.social @copper-irl.bsky.social · 22/12/2025

I’ve always been a fan of ml papers that rip out as much of the ML as they can. No magic parameters, no stolen data, and it’s often fast enough for games. Gaussian Splats and Plenoxels are excellent examples for graphics. I feel a form of GS would make for stellar screenspace imposters

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 16/12/2025

Imagine if all of OOT was like that, in first person, like they originally planned! It’d prob end up being like metroid prime lmao

110 0

copper-irl.bsky.social @copper-irl.bsky.social · 13/12/2025

I was crunching to finish GR assignments (I need more time 😔) and had a realization. What if you remove gravity from GR? And it worked. Like, shockingly well. Is GR just a system to embed physics into a coordinate system? …isn’t this Lagrangian fluid dynamics? What a crazy plot twist wtf

040 0

copper-irl.bsky.social @copper-irl.bsky.social · 11/12/2025

Here’s a fun litmus test for you: does your sim produce the rainshadow effect & horse latitudes?

110 0