Diving into Generative AI

looks great, just wonder about the limits.

As someone who’s been personally experimenting with ChatGPT to help create code, I can say this could be a cool idea for people learning how to code. In my experience, I’ve realized that I still need knowledge on coding for this, so I can tell it what to do. Trying to do things on a whim doesn’t go far.

I’d love to know where is the AI getting the data from? For example the scripting-AI from where does it get the code? Same for the material generation.

Is it taking the data from the DevForum or popular games?

One thing I just realized that would make this very useful is if AI can be used within the an experience during runtime. Now I’ve already mentioned generating stories and natural speech in a previous post of mine, however another useful feature would be image interpolation. This would be useful for many things, including animation. It would also be useful for uploading videos to Roblox as a video could have half of its frames removed, and then generated back with AI. I think these more technical uses for AI would be far more useful than typical ChatGPT usage (outside of natural language, as that will be useful regardless).

I think you’re referring to something like Flowframes or DAIN, RIFE, etc. AI used for video interpolation, typically to smoothen frame rates, restore old camera footage or cut off render time in 3D animations since interpolation is sometimes cheaper than rendering highly complex raytraced scenes many times.

That’s exactly what I’m referring to (and I’m totally not writing this because I just saw a video about it and currently downloading it to see what it can do). However, as useful as that is, that’s all done in post. I want to see this done in runtime so that it can have the benefits that AI gives in gaming. AI generating frames is a huge one, I don’t think anyone would complain about an instant 2x in their framerate if their device supports the AI. It would also help with developer-made things, as well as reduce the cost of things like video.

One big reason I’d want this is because of something like my Gif UI element, which would greatly benefit from AI interpolation during run-time, depending on how Roblox would implement it. It would mean that fewer frames would be needed for the same result.

1 Like

Currently I don’t think using AI to double framerates is doable.

Right now it has no awareness of timing or key frames and rendering 3D graphics is actually faster than interpolating a few frames so right now it’s more useful for restoring camera footage and interpolating pre-rendered 3D animation with raytracing involved.

But I’m excited to see what’s next, you never know with how fast tech is going. Maybe one day AI-enhanced framerates will be a thing.

Some TVs already do this when watching movies although sometimes results in slightly messed up timing (especially for cartoons and stop-motion), for gaming it currently just gives too much input lag.

For games it doesn’t really make sense to go one frame back to interpolate when the next frame already exists.

This is gonna be an interesting update.

Could you explain this a bit more? I see AI interpolation working great in games like Cyberpunk 2077 using DLSS. There’s also another solution that doesn’t require AI at all, and I’m not actually sure why all games don’t use it.

1 Like

This is great! I could definitely see myself using this to generate Materials or make simple scripts that would normally take about a minute or two. Very excited to see where this goes.

This sounds like it could be useful in some ways, but right now I feel that it isn’t a good idea.

Right now, AI is in a legal grey zone since the government hasn’t caught up with it yet, but in the near future there will most likely be laws and regulations put in place.

With this in mind, I just ask that you don’t jump into this trend full-force until that happens and tread carefully until the legality and ethics of AI become more clear. This is especially important right now as there have been serious legal, ethical, and moral issues going on related to how AI technologies have:

  1. Collected their data.

  2. Been used to put people out of work.

And multiple lawsuits from various groups of people are coming from those who have been affected by it.

2 Likes

One thing that’ll probably be important for Roblox’s AI to last through all of this is to make sure the datasets are responsibly sourced and documented. I was very happy to see your comment on consent, but there’s a lot of freely available datasets like LAION that are under fire right now for containing millions to billions of copyrighted works. And any future regulations will most likely target those.

If they ever do this they are most likely to use the code they made, so there’s no copyright etc.

This would be the ideal scenario, but from what I know it’s extremely hard, even impossible to make a well functioning AI with the little data they would get from using their own code.

A good example for why this is would be github copilot. github copilot is a code generating AI that is good from time to time, but makes mistakes a lot and generally can’t write code as good as a professional. This AI scraped the entirety of github, which is where tons and tons of developers put their code, and yet it still isn’t perfect.

With this in mind, it would be much better to get data from the users of Roblox who willingly consent to it, but even then it would still be extremely difficult and take years and years of gathering data to even be usable.

What I fear is that because of this, they take the easy route and base the AI off of something that had already stolen its data elsewhere to get a worthwhile result. Kind of like how Corridor Digital(a YouTube channel) recently made an AI animation using their own techniques and software, but the whole thing was based off of stable diffusion(A well known art AI that stole its data).

I could be proven wrong as to how difficult it is to do this, but from what it seems, this is the reality.

Yes, it seems very hard, but what I meant with they will use their own code, is that they will probably use code that is form them directly, which has (hopefully) no errors, and as they made the language, shouldn’t

Also one problem with this is that as @Roblox is always updating their code, the AI will always need to be updated, which makes it actually a bad use for Programming

Are you meaning that they would have pre-written code instead of letting the AI create it?

No, they have the code they have in the Documentation and that they written, and the AI looks at it and learns

Ah ok. Then my point still stands on that end. Unless they’re holding on to millions of pages of code, then I don’t think they would be able to get enough data to create their own AI from that.

I definitely agree that it would be a bad use for programming though, either way.

1 Like

Mostly because they are always updating their code, this the AI has to rewrite its code

But for materials it can be a good idea

Another interesting thing about generative AI too is that it has been shown to repeat the data it has in its dataset. It’s actually called over-fitting.
For example, people have recognized their own artwork being shown to them by art AI like stable diffusion. And other people have noticed their own fanfiction they’ve posted online in the past being shown to them by chatGPT.
So besides being a nightmare for plagiarism and copyright, they could very well leak Roblox’s source code by doing that.
That is, unless they go the chatGPT route and use cheap labor to censor what the AI puts out. Even then though, AI is unpredictable and may as well do it anyway.