A randomizer is basically a program that alters the assets of an existing video game to randomly rearrange items, redraw levels, or otherwise create an original experience within that game.

Generally, randomizer logic needs to ensure that the resulting experience can actually be completed by players. This means, for example, that it shouldn't include unreachable rooms, or that it shouldn't place an item behind an obstacle that can't be overcome without already having that item.

As far as I know, there are currently no randomizers incorporating machine learning techniques; their logic is all written by hand by humans with knowledge of the game and its mechanics. Will that change by the beginning of 2025? Resolves YES if there is at least one reasonably popular randomizer by that time that uses machine learning trained on the original game's data to produce some or all of the randomized content.

This may change, but here are my initial tentative attempts to clearly define the criteria. Stakeholders should feel free to suggest revisions here.

• "Machine learning" means any algorithm which gathers statistics about the original game's data set and uses them to generate statistically-probable new data, without direct human assistance. Definitely meant to include deep neural networks, but not necessarily limited to them.

• I don't intend for this to be fulfilled by an academic experiment or proof-of-concept; I want it to be a randomizer that people actually play regularly and enjoy playing. Hence the "reasonably popular" criteria.
  
  As a first guess, let's say that "reasonably popular" means that there are at least three documented races using this randomizer program, where each race has at least four participants. That's roughly the size of the regular races that all the most popular randomizers on racetime.gg are currently getting.

• It's fine if the output of the ML-trained algorithm is then validated or post-processed by some kind of human-coded logic, as long as an ML-trained algorithm is involved somewhere in the process.

• It's fine if the randomizer is integrated into the official game release (as in Axiom Verge) as opposed to a separate community-developed program. But, it should be clearly separate from a non-randomized core game mode. A game that's fundamentally a roguelike wouldn't qualify. (Although a similar market about roguelikes would probably be a good idea for someone.)

• It's fine if the algorithm uses training data other than the original game data. It's also fine if information about the game is included only in the prompt (e.g. "Hey Bing, create a new world map for Final Fantasy IV") rather than in the real training pass.

It's entirely possible there already is a randomizer that fulfills this criteria that I didn't find in the bit of Google research I did, in which case I'll happily resolve YES immediately if someone links to it.

I will not be betting in this market.