The Hardest Games for AI May Be the Smallest
retained text
I’ve been using AI pretty heavily for game development lately, and I’ve started to think the phrase “game development has no barrier anymore” is only half true. AI is getting ridiculously good at building things that already have well-known patterns. Platformers, shooters, RPG systems, 3D environments, dialogue, assets, animations, UI — there are tons of existing examples for AI to learn from. And generative AI is also rapidly cutting the cost of 3D asset production, which used to be a huge barrier for solo developers and small teams. But I’ve had almost the opposite experience with small puzzle games. If the mechanic is genuinely new, AI suddenly has much less to copy. The hard part isn’t drawing the UI or writing Swift/Kotlin/C++ anymore. It’s things like: How do you randomly generate effectively unlimited levels? How do you guarantee every generated state is solvable? How do you avoid “technically valid” levels that are strategically doomed? How do you generate them fast enough that players don’t think the game froze? I ran into this repeatedly while developing my own Playable Image games. With iaReveal, if I didn’t describe the chain-reaction initialization logic precisely enough, AI naturally drifted toward something much closer to a normal puzzle/jigsaw implementation. With iaBrick, random generation could work perfectly from a coding perspective, but without explicit solvability constraints some boards could become effectively unwinnable. With iaStroke, the generated boards could be valid, but a naïve search algorithm could take far too long to initialize a new game. That’s when I realized something: AI is removing the implementation barrier much faster than it is removing the invention barrier. If a mechanic already exists, AI can often reproduce it amazingly well. But if the mechanic is new, somebody still has to define the state space, constraints, invariants, solvability conditions, search strategy, and optimization approach. And that part can be much harder than writing the code. So weirdly, I’m starting to think some tiny original puzzle games may be harder for AI to build well than visually spectacular games based on familiar mechanics. Not because they need more code. Because there’s less existing knowledge to copy. That’s been one of the most interesting lessons for me while building Playable Image / iaReveal with AI. Curious if other game devs using AI have run into the same thing. \---------------------------------------------------------------------------------------------------- Edit: Fair point about the writing style. English isn't my first language, and I did use AI to help turn my notes into clearer English. I probably should have mentioned that from the start. The ideas and examples in the post are from my own development experience though — the mechanics I was trying to build, the solvability problems, the search/initialization issues, and the conclusions I came to while working through them. I use AI heavily for both coding and writing. That's actually part of what I was trying to talk about here: once I know exactly what the rules and constraints should be, AI is often very good at implementing them. Figuring out what those rules and constraints should be has been the harder part. So yes, AI helped me write the English. The experience and the argument are mine.
document fetched 2026-09-19 00:26 · canonical https://reddit.com/r/gamedev/comments/1wjr5h0/the_hardest_games_for_ai_may_be_the_smallest