Definition of "AI-rendered video game"
To qualify, a video game must have its graphics rendered by AI in real time. An AI should be rendering graphics at least 50% of the time in an average playthrough. When the AI is rendering graphics, it should generate at least one image per second. The images should generally depend on the very recent actions of the player - for instance, the player character immediately starts moving left when you press the left button.
It's okay if certain parts of the game are not AI-rendered, like UI elements. However, it doesn't count if the AI is just starting with e.g. a 3D mesh rendering and postprocessing it to make it prettier or fill in frames. It has to generate the video more-or-less from scratch. Specifically, I'd look for a system that generates visual elements like characters and scenery during gameplay and simulates their motion, doing this entirely through ML rather than programmatically, similar to Oasis. It's okay if the AI starts with a premade textual or visual prompt, but has to be able to generate and animate new visual elements.
Definition of "1 million paid users"
This market resolves based on 1 million people having paid a subscription or one-time payment for an AI-rendered video game. The number of paying users at any one time is not important - rather, I will be using the number of distinct users who have ever paid for a game. If it's unclear how many paid users a game has, I will lean towards waiting to resolve the market until more information is available. I might try to contact the company selling the game to ask them whether the market should be resolved. It may eventually be obvious that at least 1 million people have paid for an AI-rendered game, even if there are no official numbers from the company; in this case, I will resolve the market.
There may be multiple AI-rendered video games. If this is the case, I will use the number of users who have paid for any of these games, while trying not to double-count any users who have paid for multiple games. This may require some common-sense judgement of how many duplicate users there are, but I will only resolve the market if the answer is pretty obvious.
Other details
A year resolves to YES if at any time during that year, or any time before that year, there are 1 million total paid users of AI-rendered video games. If the year ends before this happens, it resolves to NO.
I will not bet in this market.
1,000People are also trading
https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/
Not available to the public, probably very expensive to run, and not good enough to be worth paying for unless you're just curious about the technology. But given the pace of progress, I would be shocked if this market didn't resolve by 2030 (assuming money still exists at that point)
@CDBiddulph To cover a future corner case you might encounter, would you consider a game that implements Gaussian Splats to have AI-rendered graphics?
12 minute related video:
@Quroe It looks like this process takes a real-world location and turns it into a 3D rendering, rather than generating novel objects on the fly, so this straightforwardly wouldn't count.
One situation that I would consider an edge case is if an AI model generates an original Gaussian splat scene (or NeRF field, or 3D mesh) and then generates the rest of the scene (e.g. enemies running around in the scene) using renderings from the scene it generated as a reference. After some consideration, I think this wouldn't count either, based on the following piece of the market description:
it doesn't count if the AI is just starting with e.g. a 3D mesh rendering and postprocessing it to make it prettier or fill in frames
Even if the scene itself was generated by an AI, the fact that it's constrained to a certain 3D scene hamper its flexibility to do certain things (e.g. terrain destruction, teleporting to another location)
@ProjectVictory IMO this seems unnecessarily wasteful from a technical perspective. What's the benefit over having AI generate the 3D object on a as-needed basis and coordinate them but have everything rendered by a regular engine?
@Enlil This doesn't generalize very well. As just one example, suppose you generate a 3D asset of a building. If the player tries to destroy the building with a rocket launcher, you'll near-instantly have to load in new 3D assets for a destroyed version of that building, as well as the shrapnel flying through the air, the explosion itself, etc. With a full-AI solution, the AI can just render all that stuff, with no more technical complication than if the player took a few steps to the left to view the building from a different angle.
Also, any assumptions you build into the code will be there forever - if you code up an Animal Crossing clone where the town layout and each new villager is AI-generated, the gameplay will still only be as complex as the code you wrote. But if AI creates the whole game on-the-fly, the player can decide that they want to do a cooking minigame, and the AI will just make something up (as long as the model was trained on enough content from games like that).
You could try to come up with a system for the AI to write the game's code on-the-fly as well, but that system will still be limited by your assumptions, the player will have to wait for code to be written every time the AI adds a new mechanic, and in all likelihood the result will be bad-looking and glitchy absent thorough playtesting and adjustments. Granted, pure-AI video games will also have all kinds of glitches in the form of hallucinations, but AI can at least attempt pretty much anything the player could ask for.
This is basically just the Bitter Lesson: http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Probably a better response to this objection is that AI-rendered games already exist - there was a previous version of this market that already resolved because of http://oasis.decart.ai. The only difference is that Oasis has 1 million free users instead of paid users
@DavidBolin I mean, that is basically why I made this market, because Oasis was kind of an edge case. The point is, people are working on this and are already getting some results. GPT-2 was also pretty useless at the time
@CDBiddulph I'll be honest I worry about edge cases here too. Like OpenAI launching a similar experiment for payed users and everyone paying for subscription for OpenAI counting as a payed game user.
@ProjectVictory That's a good point! That is a little tricky, but I'm going to say that if something like that happens, it will count as long as 1 million paid users actually try out the game.
@CDBiddulph In this case I believe your criteria for this market as well as the other one are way too broad.
@ProjectVictory I don't think that particular edge case is all that likely, but feel free to bet accordingly. In any case, I feel like if OpenAI makes a video game model and puts it under their paid plan, odds are it will be fairly polished and 1 million people might start paying OpenAI primarily because of the game rather than their ChatGPT-related offerings