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

New question! Do coherent world sims via diffusion models count?

https://x.com/ehuanglu/status/1952769676666388799?s=46

I think they do!

@DanW I agree that "how" AI generates games isn't what matters.

But I suspect these models will have very demanding hardware requirements. If it can generate a game like Dredge, but it requires the equivalent of an A100 to play at 30 fps, does that count?

When it comes to say puzzle games which I agree are a much harder category, is the model expected to make a challenging game with a well designed learning curve and verifiably solvable hard levels etc. etc. (This feels like it becomes very subjective very quickly)
Alternatively, is it enough that the game runs and includes requested mechanics/assets at a basic level and shows all of the core functionality required to meet the users prompt. (This seems like a better benchmark as it's much more objective)

I guess my question would be how much is this about general capabilities, and how much is this about finding specific edge cases. As in, are you interested if by 2029 it will be trivial to create 2023 AA level games kinda like how it was trivial to generate photorealistic AI images in 2023? Or are you interested in edge cases, artifacts, and such (kinda like how AI images in 2023 still had some body poses or texts they couldn't generate)?

So, basically, are you interested in whether it will be possible to make all kinds of games or are you interested if there are some specific things that AI couldn't do, even though some game in 2023 could do?

I'm open to narrowing down the (admittedly broad & open to interpretation) resolution criteria now before this market gets bigger and closer to resolution. If you've got questions or suggestions, post them! I'll try to refine resolution criteria as closely to the spirit of the original market description as possible (which is why I don't bet on my own market!).

@DanW Based on the wording, I'd expect that by the start of 2029, there will be a tool that can create a game at the quality level of something like Stellaris, Chants of Senaar, Lies of P, etc., from a prompt, and it can do virtually any genre available today. I don't think is needs to reliably generate something at BG3's level of subjective player satisfaction, but it needs to be able to produce a binary with 100% of the features in a finite number of prompts, e.g.:

Generate 3D models, with UV-mapping, rigging, animation, multiple LODs, textures of consistent quality and style, etc.

Design interesting levels that guide the player to their objective

Create original VFX and custom shaders as necessary

Implement reliable save systems that don't break with game patches

Have a consistent, polished UI system for all the game's functionality

Create balanced and varied NPC AIs

Support multiple platforms and controllers

If the game is multiplayer, create the servers, database, and lobbies

Put all of this together into an application

etc.

I think it's okay if the AI uses an engine like Unity or Unreal, but it can't use 3rd party assets in a major way. Like it's fine if the AI uses existing fonts or middleware, but not models or game "templates".

I also think multiple prompts to get to 100% is okay, in the same way that a game studio may do internal alphas, and open or closed beta tests.

(I'm an indie game developer who has been following LLMs and generative AI pretty closely. I'll be absolutely shocked if this market is even close.)

@JustKevin Personally I would disagree with allowing a third party engine to handle most of the work. This offloads an enormous amount of code from the AI, and many of the default frameworks in unreal and unity are plug-and-play games with working character controls and physics, etc. I mean, unreal is just a game creation codebase built by thousands of humans over the last couple decades. I want to see the AI make its own graphics engine and physics engine and animations too. Those are all parts of game development!

@gamedev My reasoning for allowing an engine:

If you disallow engines so the AI has to do the graphics and physics, should it be prohibited from using frameworks and libraries? Can it use NVidia's Physx? What about OpenGL?

Most studios in the AA space use 3rd party engines and that percentage is increasing, so this feels like a more fair comparison.

Getting from Unreal's FPS gray box template to "Lies of P" is going to be hard enough.

I'm a "no" holder, so if the market maker wants to make the "yes" resolution harder, I'm all for it, but I think allowing an engine would be in the spirit of the question, in my opinion.

Does the AI have to get this right on the first prompt? If no, what degree of interactive prompting is allowed for a YES resolution?

What is the quality bar (number of bugs, visual artifacts, or "mistakes")?

@ae In the initial market criteria I said "...must be able to use an arbitrary prompt to generate video games...", which suggests it should be done in one prompt.

That said! I don't think the exact # of prompts (within reason) is really what I cared about when making this question, and I'm curious if other betters feel the same way. Depending on what the community thinks maybe we switch the criteria to a handful of prompts?

@DanW I personally feel like if it has this capability, then the difference between one prompt and ten or w/e will probably not be very large, and, again as an opinion, best of 10 prompts or something like that is still reasonable and is a marked difference from having to handhold it through the process or manually change a bunch of things. But that might not be what the market was attempting to express, or what others read into it.

@DanW The standard that gets used by most evals orgs for long-horizon tasks is just to run the model with a simple agent scaffold.

As a dev, this is never fucking happening due to so many reasons I cannot begin to describe them to you. However, I am not betting anything for a payout 4 years away. Free low risk one here tho

@OKaya You can collect loans on invested mana

@OKaya agreed

@AdamK how does one do this? trying to figure

Recent spike is nuts, but if I’m wrong at least I get to die happy playing infinite decent-quality BG3 mods.

@Balasar And thus did humans lay down their weary burdens, and pass into Infinite Fun Space.

@Balasar I think there are AI bulls and AI skeptics basically locked in battle across a bunch of markets, and when a market resolves one way money is freed up and the victorious side splurges on new AI markets.

Of the prompts you gave some of them seem way easier than others. Minecraft clone is easy, a decent trippy puzzle game seems impossible.

@digory I disagree with your intuitions regarding what's likely to be easy versus hard here, given that the requirement isn't just to do it at all but to do it to 2023-era-AA-studio-tier quality-standards. But, also, if that is the case, that seems convenient as a source of future market-movements; it'll let us gauge the AI's rate-of-progress better over time, via seeing when the easier prompts start being filled well, than we would be able to if the prompts' difficulties were more tightly-clustered and the uncertainty was mostly whether it'll learn to make any AA-studio-tier games at all.

@digory The resolution criteria specifies that the AI system should be able to generate comparable games for any genre, so it will need to be able to generate puzzle games. (which I agree are easily the most challenging, considering how terrible current AI systems are at metacognition)

@Tulip New orders up