Machine Learning and TAS Videos for Speedrunning

There's been a couple instances of how the Canon list set up was discovered in Mario 64 where you clip into a wall to get a star. It used to be you would just run into it and get it maybe 20% of the time and it was essentially random. But this guy's sock folder figured out this really elaborate setup that he could do by testing pretty much every single series of inputs possible at this point. So I think we're okay for now with machine learning, but once you get more than like a few seconds of gameplay it turns into millions and millions and then quickly billions and trillions of different inputs you'd have to test. That becomes kind of exponential and maybe hard to replicate