Hacker News Reader

Comment by whywhywhywhy | Hacker News Reader

This will change within 6 months I promise you, EA/Ubisoft/etc will ALL be shipping AI generated textures in games before the end of the year.

View on HN

That's not a change from this though, as long as those AI generated textures are from systems that were trained on images that they have permission to use (or are copyright free).

No such system exists so they won’t.

Even Adobes system has questionable training data mixed in

Every major game publisher has an enormous archive of in-house artwork, and probably a decent amount of hardware they could use for training. They'd be stuck with open source software like Stable Diffusion - or have to cut a deal with Midjourney or whoever - but there's no barrier to creating such a system.

Stable Diffusion nor Midjourney can be used to create artwork under these Valve guidelines.

And only a matter of time until a major game lets you talk to NPCs by generating dialog with an LLM.

It'll be interesting to see if a game can make good use of that, but it sounds like it would lead to some very frustrating interactions.

The first versions of this won't be using it to create characters with plot significance and great depth, they'll be generating infinitely large piles of background character chatter. No matter how big your budget is, you'll never be able to to make a world full of random civilians who don't sound repetitive if every line has to be recorded in advance.

Any prerecorded line having a slight bit of personality to it ("I used to be an adventurer like you, but then I took an arrow to the knee") becomes very noticeable after the first time you've heard it, so the safe thing to do is have lots of the most inane chatter possible ("Nice weather today!") to make it stand out less.

If the game can write new lines and synthesize voices on the fly, you could have more interesting lines without the repetition.

Have you tried it? As in, prompt the LLM that it's the character you want and then converse?

I've used GPT for DnD type stuff. It's really good if you don't have an existing expectation of the world you're in, as everything should be a hallucination. Within the context window LLMs are pretty internally consistent, so within a chat session it'll be surprisingly coherent and human-like. But inside of a full game world like Skyrim, each hallucination could lead to wasted time finding the magic scroll the LLM told you about but never existed. If you can think of a style of game where hallucinations aren't a problem, or a way to put the LLM on rails so that it won't hallucinate too much then I think players will have a good time.

Of course it will change and Valve will even release their own game that's generated with AI.

Governments and companies everywhere trying to lock out small time people today before they get too much traction with AI generated content. They know indie devs will never be able to prove their model is only trained on their art. Only massive companies with billions of dollars can do that right now.

Every big company is trying to create rules to ban AI but keep a small enough loophole that they can use it when the time is right.

Reply on news.ycombinator.com