Gloomy

Silicon-whisperer, wannabe AI researcher

[Gloomy @GloomyCr]

@girthybologna @H7pernerd I'm sure there are some issues that will be tricky or design choices that would invalidate performance gains but nothing technologically impossible. One turing machine is another turing machine.

On more specific point, what does RTX in Nvidia RTX stand for?

[Gloomy @GloomyCr]

What is the reason why ppl don't compute server tick on gpu to get 100x player count?

Flatten player state into a vector, batch it spatially into a tensor and let CUDA rip.

Yeah there are issues like branching or lookups but what is really your logic complexity vs gpu speed.

[Gabriel Dechichi @gdechichi]

just finished my multiplayer system in C and turns out we can support 50k players in a single server at 60hz simulation tick

of course not much happening on this demo but pretty sure we could support Fortnite level gameplay at 10k players.

[Gloomy @GloomyCr]

@girthybologna @H7pernerd Out of all cases, physics sim is exactly what you would put on gpu and parallelize for an approximate good enough result instead of computing exact one on cpu.

My "handwaving" is hearing objection and looking for that good enough approximate parallel.

[Gloomy @GloomyCr]

Yeah don't use tensor, that was me being NN brained.

But since you wouldn't actually care about matmuls (tensor cores) here, only about threads, would you even use the expensive "new" cards?

[Jebrim @AgileJebrim]

Don’t put it into a tensor lol, but absolutely yes run it on the GPU.

General purpose client-server applications where both run entirely on the GPU is the future of computing and is where I’m making big bets on right now with my current project. If the AI bubble pops, there will be a lot of affordable GPU hosting available and plenty of demand for GPU-based server software to repurpose that hardware.

[Gloomy @GloomyCr]

@H7pernerd Almost all games only respect action order between different server ticks anyway.
Anything that happens on same tick is a diceroll on how was data ordered in memory.

Just allow side effects like players killing each other on same tick.
Leave cpu for transactions, not sequences.

[Gloomy @GloomyCr]

@gdechichi But now you made me think about somehow transforming the state tensor into an image and streaming the ticks from card to players as video stream using gpu codec acceleration.

But that is some voodoo idea.

[Gloomy @GloomyCr]

@gdechichi Although I did ignore the "writing snapshots, bandwith" second part here, fair enough.

[Gloomy @GloomyCr]

So you're upping your P(doom) is what I'm understanding 🤔

[Ben Goldhaber @BenGoldhaber]

the opus 5.5 animations convinced me ai super persuasion is a real and present threat model

[Gloomy @GloomyCr]

This is the equivalent of counting LoC after compiler expands macros, templates and inlines every function.

Your "code" is your inputs to Claude.

You're coding in a new abstraction layer and like any good abstraction, your true LoC went down, not up.

[Anthropic @AnthropicAI]

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-2025.

[Gloomy @GloomyCr]

@model_mechanic I thought fast iteration always applied to work on the product, not tech research.
Unless "didn't work" is considered a valid conclusion of single iteration, but at that point your ideas neither come from any understanding or advance it.

[Gloomy @GloomyCr]

@Vitispi3 It's data scientists in USA fighting data scientists in Russia.
Details become convoluted to the point of incomprehensibility.

They will literally simulate finger gestures on the emulator.
Including realistic chance to missclick the button.

[Gloomy @GloomyCr]

@Vitispi3 Maybe less likes on comment means less comeptition for attention.
Maybe it's more likely for users to check comment likes.
Maybe they know that comment like will push notification.
Maybe the bot rolled a random dice to avoid repeating same action.

I could keep guessing but IDK

[Gloomy @GloomyCr]

@Vitispi3 Mass fishing with comments or DMs would insta-nuke them.
Liking a post won't.

So, they target small profiles and niche communities.
There, just a like will likely get someone's eyes on them.

That's also why they try to pass as models.

[Gloomy @GloomyCr]

@Vitispi3 They are intentionally fishing for one-off interactions with real users to boost their trust score on bot detectors.

One like or even just profile click from real user is enough to count as a score bump.
Comment or DM is a jackpot.

[Gloomy @GloomyCr]

Oh.
So doctor will just watch you slowly die over months if you don't act like a beaten dog during visit.

[Gloomy @GloomyCr]

@txnftxnf @gcolbourn What was the solar power wattage over Sahara alone again?

Double or triple digit multiple of global power production?

[Gloomy @GloomyCr]

This is just a proxy for autism

[Find me on bsky @colin-fraser.net @colin_fraser]

one more excellent finding from the AI poetry paper dataset presented without comment

[Gloomy @GloomyCr]

I got attacked by antichrist in dream tonight.

He jumped me at the bottom of minecraft cave, probably thinking I have noodle fists like most people while dreaming.
I hit him with nuclear punch so hard that explosion lagged out the dream and crashed me back awake with java error.

[Gloomy @GloomyCr]

@Aella_Girl Also, CSV is just a bad file format for a lot of data.
Still, at that size you shouldn't be running into memory problems so you're probably unknowingly holding intermediate partially transformed chunks somewhere in code.

[Gloomy @GloomyCr]

@Aella_Girl The issues sound like improper separation of code.
Chunking should be transparent to the data transform logic and whole process executed in passes.
I guess AI helper made you stack more and more nested logic rather then do a composable transformation pipeline and chunk executor.