Sudo su @sudoingX

so my net isn't broken and it isn't cheating. it learned to watch the transit, exactly like it should. it gets fooled because brightness alone genuinely can't separate some impostors from real planets. the fix isn't more training. it's giving it something brightness can't see. that's the next build, and it's where the spark stops being an inference box and starts earning its keep. the score was never the point. interpretability, actually understanding what the thing learned, is the whole gam

and before anyone says a saliency map is just a pretty heatmap that means nothing, i tested it. i blanked out the exact bins the map flagged as important, and the net's confidence collapsed. then i blanked the same number of random bins, and it barely moved. the flagged bins mattered about 7 times more. the explanation is real. the net genuinely uses those points.

i don't fully trust my own neural net. so i spent the day putting it on trial. everyone in my feed is benchmarking gpus running models someone else trained. i wanted the harder thing, so i trained one from scratch on the 5090 on my desk to find real planets in nasa data. it mostly works. but a high score can be a lie, the net could be cheating on an artifact instead of actually seeing the planet, and you can't trust a number you can't explain. so i cracked it open. this is one real star, and the glowing part is exactly where my net looked. locked onto the transit, the dip where the planet crosses its star. it's not cheating. then i ran it across 500 🧵

one star could be a fluke, so i ran it across all 500. every group, the planets it nailed and the impostors that tricked it, the net puts about half its attention right on the transit. random would be 30%. it always looks at the dip, whether it ends up right or wrong. the net learned the correct thing to look at. it just can't always finish the job from brightness alone.

here's the part that got me though. this one fooled it. the net looked at this star and called it a planet, but it's an impostor, not a real world. and look where it's staring. the exact same place, right on the dip. it didn't get fooled by looking in the wrong spot. it looked in the right spot and brightness alone just wasn't enough to tell the difference. that's a very different problem than a broken net.

@technohobo23 because every tool that matters is linux first. on windows you run it all through WSL, basically linux in a vm, so you're paying the windows tax to run linux worse, then it reboots mid training for an update you never approved.

running ai infra on windows is like trying to operate a datacenter from a microwave

@deeperflows 26 years clean. a sobriety chip.

@HououinTyouma you don't even need the switch. windows update reboots it mid thought every patch tuesday, alignment solved.

@ExileAI_0 leave it, the dust is load-bearing. that's the most stable that copy of windows will ever be, you boot it up and you'll only ruin the record.

the how, for anyone who wants it. it's blender, but driven by real numbers, not eyeballed. sun and earth placed at exact 109 to 1, the rule enforces it. and the stars behind them aren't decoration. it's a real position starfield, the actual stars sitting where they actually sit in the sky. the only thing bent in the whole frame is the gap between the two, and i told you that part up front. even the background is true. that's the bar.

almost every picture you've seen of the sun next to the earth is lying to you. not about the color or the detail, about size. they shrink the gap and pump the earth up so both read like characters in the same shot. this is the real ratio. that orange ball is the sun. earth is the dot on the right, the one i had to circle so you could even find it. 109 earths would line up across the sun's face. you could pour about 1.3 million earths inside it. i'll be straight about the one cheat. the sizes are exact, 109 to 1, that's the part everyone fakes and the part i refused to. the distance i pulled in, because at the true gap earth sits more than a hundred sun-widths away and you'd be staring at a black frame with nothing in it. so real sizes, squeezed gap, and i labeled it so you always know which is which. here's why this even exists. i've been teaching a neural net from scratch to find planets in real telescope data, and the last piece is drawing what it finds. i gave the renderer one rule, it is never allowed to fake a size. there's a test that fails the whole build if a planet comes out wrong by even a little. doing science means the picture stays true even when the true thing is almost invisible. rendered on the 5090 on my desk. that speck i had to circle is every person who has ever lived, every fight, every deadline, every benchmark i've ever posted. all of it, on the pale blue dot.

friends don’t let friends run local ai on windows garbage

Sudo su @sudoingX

6 hours ago

running ai infra on windows is like trying to operate a datacenter from a microwave

8 6 51 5K 0

i so really don't like being away from my computer man. fuck. i wish i could inject the hardware straight into my veins. carry the dgx spark in my bloodstream.

i'm not exaggerating, go into debt for this one if you have to. i took this account from 200 followers to past 30k now, and the monthly revenue, the partnerships, the clients, the actual users on my platform, almost all of it traces back to this one subscription and learning to actually use it. it changed my life and i don't say that lightly. i'm writing the whole thing up right now, every play, exactly how i use x premium+ to turn a timeline into a business. nothing paywalled, all free. if you take one thing from me, take this.

Sudo su @sudoingX

2 days ago

listen anon, if you're a serious founder and i could lock you into ONE subscription, no debate, it's x premium+. it was the go to before grok build. now it's not even close. what you actually get: > grok build, a full agentic coding agent, bundled in > grok at the highest

15 2 108 30K 74

@DegenApeDev @MichaelGannotti Noted

the more i use my dgx spark the more i think it's one of the most undervalued machines on the market right now. and i keep finding new things to throw at it that have no business working on something this small. but every time i post about it the same question comes back, what about the amd strix halo. and here's what actually bugs me, i can't answer it, because i've never run one myself, or even seen anyone around me run one. tons of people name it as the competitor, almost nobody posts real numbers. no tok/s, no model loads, no thermals, nothing. just the name. so i'm asking straight up. if you've got a strix halo or a ryzen ai max box, drop your real numbers. what models, what speeds, what breaks. is it actually competing with the spark, or is it the machine everyone recommends and nobody runs. i'd benchmark it myself the second i could, until then i'm genuinely curious what you're all seeing.

people keep asking what i actually do with the dgx spark sitting on my desk. i kept answering with inference benchmarks, large models running reliably, big context, the usual. and that's all real, but it was never the point. it's not an inference box. it's the smallest research lab i've ever owned. so here's the one thing. i'm training a neural net from scratch to find exoplanets in starlight. no pretrained weights, no borrowed architecture, my own design, on real Kepler data. right now it sits at 0.75 on a tiny slice, 514 stars out of 7,586. proof of life, nothing more. now the move is to go big. first the full labeled archive, fifteen times the data. then the part i'm itching toward, self-supervised pretraining on around 200,000 raw light curves, a corner of machine learning i've genuinely never worked in before. and this is where the box stops being a gimmick. one gb10, 121gb of unified memory, so the entire dataset lives in ram and i train for days with zero data-loader bottleneck, the thing that usually kills experiments like this on a normal rig. no streaming, no thrashing, just the whole sky sitting in memory. everyone treats the spark as an inference box. on my desk it's a from-scratch research lab.

Sudo su @sudoingX

4 days ago

the more i use my dgx spark the more i think it's one of the most undervalued machines on the market right now. and i keep finding new things to throw at it that have no business working on something this small. but every time i post about it the same question comes back, what

39 4 134 25K 32

@yacineMTB itll just be. like how chinese local models are.

since people asked why i'm scaling data instead of just building a bigger net, here's the undeniable answer in one chart. watch the two lines split. blue is the net learning my 514 training examples, it climbs all the way to 96%. red is how it does on stars it's never seen, it stalls at 75% and then the loss actually gets worse. that gap is the whole story. ten million parameters, 514 examples, so it just memorizes the answers instead of learning the pattern. that's not a brain problem, it's a starvation problem. you cannot out-architecture a tiny dataset. which is exactly why the move is fifteen times more labels and then 200,000 raw curves. the net was never the bottleneck. the data was. this chart is the receipt.

Sudo su @sudoingX

4 days ago

people keep asking what i actually do with the dgx spark sitting on my desk. i kept answering with inference benchmarks, large models running reliably, big context, the usual. and that's all real, but it was never the point. it's not an inference box. it's the smallest research

5 4 64 16K 20