Again, this sounds like a very simple problem to solve. But the fact that very smart people have not solved it implies that there's some brick wall of functionality I don't understand. Imagine a technology so useless you can't even run Doom on it... - ThreadSky

tomforsyth.bsky.social • 65 days ago

Again, this sounds like a very simple problem to solve. But the fact that very smart people have not solved it implies that there's some brick wall of functionality I don't understand.

Imagine a technology so useless you can't even run Doom on it...

Comments

bartwr.bsky.social•65 days ago

Here is the thing: it's not that it cannot be done or doesn't make sense. What you suggest makes 100% sense, and I agree.
It's a philosophical rift. "OpenAI camp" (not just them, they represent it) of young ML researchers who never worked on graphics or games claim domain knowledge is useless. 1/2

bartwr.bsky.social•65 days ago

That 3D rendering, geometry, lighting, tracing rays, or even programming gameplay will be obsolete. And you could solve everything with prompts, more data, and by (pre)training on more videos and by predicting next frames. ¯\_(ツ)_/¯ I wish I was making a strawman, but I heard it multiple times. 2/2

bartwr.bsky.social•65 days ago

Btw: I don't dismiss the research and people verifying hypotheses. But what annoys me the most is: a) confidence, arrogance, and dismissal of art (whether of making fun gameplay, or beautiful assets) and effort. b) Pure bullshit. Holtz or Altman promised a "full holodeck" by the end of 2024. Oops.

bartwr.bsky.social•65 days ago

Many of my colleagues, who have been doing graphics for decades and are generally confident about their ideas, started questioning themselves and even got anxious. What if those "kids" are right? What if we are dinosaurs? This is a healthy, open-minded response. But it's sad if it's because of bs.

mostech6502.bsky.social•65 days ago

Dan Luu had a recent article about using some LLM to code up a puzzle solver. There were cases where it made something that had a compiler error, and he had to prompt the thing "this doesn't compile please fix", which made him wonder about why the human needs to do that:

mostech6502.bsky.social•65 days ago

The link to "The Bitter Lesson":
http://www.incompleteideas.net/IncIdeas/BitterLesson.html

basically, it's worked so far to brute force your way through any AI problem, the expectation is that it will continue to work that way, and spending time on anything like "special cases" or "optimization" is a waste of time

tomforsyth.bsky.social•65 days ago

It's called being a manager :-(

thibaulttricard.bsky.social•63 days ago

I wouldn’t be so sure about that, ML based rendering have made a lot of progress by ignoring important properties and leaving them to future works, and when they try to solve go back to graphic method

thibaulttricard.bsky.social•63 days ago

Nerf was slow as hell until they started using gaussian splat, 3D texture, or voronoi cell as rendering primitives.
Turns out having efficient rendering primitive is important 🤯. Who knew ?
Wait until they realise primitives need to support animation …

demofox.bsky.social•65 days ago

Yeah real true. There are valid places for ML in game dev and graphics (and as a tool in the CS tool box), but these folks are so out in left field, they aren't even interested in real progress in these areas. It's ridiculous.

cyphersignal.bsky.social•65 days ago

I still think of this as one of the most thorough, singular, and visible self-portraits of that kind of lunacy in that regard that I’ve seen from that space.
https://www.linkedin.com/posts/drjimfan_if-you-think-openai-sora-is-a-creative-toy-activity-7163976340042448896-1WFQ

bartwr.bsky.social•65 days ago

Yeah, it's like training a word-matching model (not even an LLM, but an autocomplete) on a sentence "I am conscious!" and then when it reproduces it saying "wow, look, it achieved conscience!". There is zero proof of world building or physics understanding. https://arxiv.org/abs/2501.09038

cyphersignal.bsky.social•65 days ago

Oh this should be a neat read.

nicole.ashley.kiwi•61 days ago

All those words to say, "I missed the part where the ships are literally mutating shape before my eyes."

frankhghtwr.bsky.social•64 days ago

...I need some brain bleach after reading that bullshit

sosowski.bsky.social•65 days ago

I like your point very much.

They claim to be making tools for creating art, writing text and making games, yet they know nothing of drawing, writing or game development.

bartwr.bsky.social•65 days ago

This hubris is hilarious. One after another, big tech companies look at billions in the video game market and say, "How hard could it be?" And then, one after another, Daydream, Amazon, Stadia, and Meta fail to make any game or dent in the market.

bartwr.bsky.social•65 days ago

The less you know about sth, the simpler it can seem from the outside. I heard numerous times at Google "we solved THE INTERNET, games are comparatively trivial."
It's like tech bros believe it's easy to solve hunger in Africa with "brilliant" ideas, but super hard to solve gun violence in the US.

skarolewics.bsky.social•65 days ago

I saw this firsthand at one company. Stuff like discussing revenue for a game like Fortnite to suggest sales expectations. There was an unspoken assumption of not only success but success measured in billions.

knarkowicz.bsky.social•64 days ago

It reminds me of FMV games, which were a cool idea, but it also was a dead end due to various limitations of the tech.

sosowski.bsky.social•64 days ago

FMV games were ART, tho.

okonomiyonda.bsky.social•64 days ago

AI: the 3D0 of gaming

saschawillems.bsky.social•64 days ago

That's an insult...to the 3DO ^^

okonomiyonda.bsky.social•63 days ago

Oh man, I’m gonna have to apologise to RJ now…

nothings.bsky.social•65 days ago

Yeah. It's definitely possible to imagine feeding the RAM into a model. Compare, for example, tom7's semi-jokey Learnfun/Playfun effort in 2013 doing ML on an NES's 2048 bytes of RAM. (1/2)

https://www.youtube.com/watch?v=xOCurBYI_gY

sosowski.bsky.social•64 days ago

This is Machine Learning, but it's not a Neural Network. NNs are incapable of operating on binary data by design, and all contemporary AI implementations run on NNs.

That said, that Mario ML attempt is 1000x more impressive than this slop.

nothings.bsky.social•64 days ago

I never said it was an NN. But if your only goal in this thread is to defend your claim that NN on "binary data" is impossible, here's a counterexample. https://bsky.app/profile/nothings.bsky.social/post/3lm6voyiubs2h

sosowski.bsky.social•64 days ago

An LLM could be as well a counter example, but the truth is, there's no way to have a neural network operating on binary data that actually works. It's just not possible.

I am not trying to have an argument here, this is just how it works. NNs operate on real numbers, and need to be able to err.

nothings.bsky.social•65 days ago

The thing is, (a) image-based generative AI has been so successful that they can imagine that's a path, & (b) training on the "RAM" of a program that allocates memory or worse yet moves things around in memory is extremely difficult or impossible. So their incentive is to push on image genAI. (2/2)

nothings.bsky.social•65 days ago

To be clear, Learnfun was a system based on using ONLY the domain knowledge of "number goes up", showing how it was possible to discover which hidden variables mattered, sorta. So if really you wanted to go down a domain-knowledge-free path, there are some possibilities. Doubt they'd go 100% though.

nothings.bsky.social•65 days ago

Oh yeah, I forgot, Deepmind actually did related research from 2015-2020 using the Atari 2600 (128 bytes of RAM + screenshots as input). https://deepmind.google/discover/blog/agent57-outperforming-the-human-atari-benchmark/

feliceenellen.bsky.social•54 days ago

It WOULD be simple to solve.

If you can get AI image-gen to work on 360-degree skyboxes.

Then just render the game into a full skybox and nothing will ever go offscreen. You'd only show the player the bit they can "see", of course.

Still, you can leave a room and find a new one when you go back.

sosowski.bsky.social•65 days ago

I'm gonna have to frame this one.

cidwel.bsky.social•64 days ago

This is so true lol

dulwichquantum.bsky.social•64 days ago

Quantum computers?

tranquil-abattoir.bsky.social•64 days ago

When I look down at the floor and back up.

nolram.bsky.social•64 days ago

And this is why Tom Forsyth is a true icon.

raptornx01.bsky.social•61 days ago

faultierfan.bsky.social•65 days ago

I think the actual solution is to not use generative AI for the entire thing, but only for specific parts of the rendering and/or NPC behavior.
But then you're back to human beings making games, just with some additional helpful tools, and we can't have that. /s

binarymarionette.neocities.org•65 days ago

that's pretty much back to human beings making games period

flootzavut.bsky.social•64 days ago

If you removed all but the first eleven words, I'd completely agree.

faultierfan.bsky.social•64 days ago

That's a bit too simplistic. It's possible to use machine learning (including aspects of generative AI) in useful & positive ways that don't align with the goals & world views of the neo-fascist AI bro culture.

flootzavut.bsky.social•63 days ago

Unless & until it's ethical (ie not based on stolen data) & not an environmental disaster, GenAI is never gonna be okay with me 🤷🏻

flootzavut.bsky.social•63 days ago

(& using it for video games? Like yeah there are uses it might be valuable for where it could be worth the cost in some way, but video games are not it.)

ghostkumo.bsky.social•65 days ago

That's maybe the most damning thing Ive ever heard, considering Doom runs on literal bacteria

spaceprez.bsky.social•65 days ago

The floats to binary isnt' the problem, you can always represent binary values with floats after all.

The problem is the nature of these models is they aren't code. They're a huge database of weight values. Its not something you can read and modify. You can only train it.

davidjayindie.bsky.social•65 days ago

That's what pisses me off the most about this tech. It's all about approximation. But it's framed as this magic entity that's infallible and if it's not, a few tweaks here and there will fix it somehow.

If you need anything that has to be consistent and precise, ML is absolutely not that.

spaceprez.bsky.social•65 days ago

That's the problem with all these models and why google now says you should put glue on pizza. They can't simply tell it "don't say glue on pizza" they can only train on tons and tons of pizza data that doesn't include glue and hope that finds the right weight to change it.

The best they can do

spaceprez.bsky.social•65 days ago

The best they can do is add a traditionally programmed filter that runs after the genAI model and looks at what it generated and checks some basic logic to say "wait this isn't good" and then runs it again or generates some error.

But then that's dependent on traditional programming logic.

spaceprez.bsky.social•65 days ago

And so its entirely possible for the traditional filter to not catch something the genAI did that was "bad"

The whole stack is a mess of smoke and mirrors

raptornx01.bsky.social•61 days ago

funnily enough, this was a basis for characterAI's content filter. you had the bot, but then had a secondary AI model that scanned the output to try to catch "banned content" which then reprompted the bot. the result was the bot became braindead after a filter trigger AND the filter could be tricked

safty.me•64 days ago

That's another problem with all of the LLM slop. To make it even remotely useful you need to start reintroducing programming logic and a lot of overhead in order to control its outputs.

safty.me•64 days ago

Microsoft has been many things I've despised, but under Nadella it has increasingly been selling fantasies and complete bullshit.

The "fantasy" being sold here? The idea you will be able write a paragraph about your "dream video game" and magically it will materialize out of thin air.

readyuser.one•65 days ago

They only know what something probably looks like, not what it does.

claralaflore.bsky.social•64 days ago

Similar shit has been done with Doom and it was fucking awful too, hallucinating the shit out of monsters, maps, hud elements, and comprehensively not understanding things like how taking damage works in the game.

https://www.youtube.com/watch?v=p6RzS_mq-pI

Comments

Posting Rules

Reply