AI - The Death of Human Discovery In Speedrunning?

A few weeks ago, I watched this video from YouTuber Linesight about a new Machine Learning model that he trained to beat Trackmania, an arcade racing game that's been around since 2003, world record player times. In the video, he talks about how this ML model was able to find shortcuts and skips in a matter of hours, when the Trackmania community took up to 20 years to find these same time savers. That point stuck out to me in particular, it put a question into my head that I've been stewing on ever since: Is the oncoming increased use of these machine learning models going to rob us of the joy that comes from human discovery? Let me explain what I mean.

I, am speed.

I love speedrunning, a lot. There is just something about people going as fast as humanly possible by any means necessary that really actives my monkey brain neurons. Watching someone accomplish things that don't look even remotely possible by human hands is downright impressive. Learning the often decades long history of these speedrunning communities is fascinating. These runners dedicate thousands of hours honing their technique and skills to be the best of the best. Not only does it require skills, but it also requires extensive game knowledge. Speedrunners will absolutely tear a game to shreds, sponging up any knowledge there is to know about how these games work to get a leg up.

Personally, I feel one of the more hype moments in speedrunning communities is when something new in the game is discovered. Be that a new route, a new mechanic, a newly found bug, or even a new way of holding the controller to be just that one frame faster, they are large events in these communities that often leave permanent effects to the world record time. These moments can be quick to come in the early stages of a game's life cycle as people find the more obvious or easier mechanics, but as everything there is to discover gets found, new finds can be as sparse as once every 10 years, as it takes someone with either insane luck or insane knowledge and skill to find something like that. It is often a true feat of human skill that leads to these discoveries. This makes it all the more exciting when someone does find something. The community often goes into an absolute frenzy trying to either master the new skill, see how the mechanic could be applied, or see if the route truly is shorter or not. That's all to say that these are large events that can hype-up and shake-up a whole community of often tight-knit people

I fear that AI is at risk of taking that away from us...

Another funny header

Back to that Trackmania video. In the video, he talks about the ML model's progression of improvement. At first, it can't come close to world record track times, but as he improves the training method, it gets better. After several iterations, it is setting its own world records, notably the map named A02, beating a record that stood for 6 years previously. This isn't an instance of the model just having tighter lines and perfect inputs (it did, to be clear), but it found a faster route, on a 10-year-old track that thousands of people have tried perfecting over that time. As far as I am aware, that is unprecedented for these ML models. The rest of the video shows the model crushing the rest of the human world records he tested against. To be clear, it is impressive and hats off to Linesight for the great work making such a sophisticated model. But I'm worried about what happened in A02 specifically.

Are we going to get to a point where instead of dedicating time and energy to grinding these out, we rely on what the "all mighty AI" shows us the most optimal route to take is? I fear that as these tools get better and better, and easier and easier to use, more people will start incorporating these tools into learning how to speed run a game. I worry that this is going to rob communities of new and exciting moments of watching someone discover something never thought of before. Instead of spending the time and energy to thoroughly hunt for new routes, will we default to just throwing an "AI" at the problem until we have the answer?

There is a tangent here about how this applies to creative endeavors more broadly that maybe I'll go down at time point, but not today.

bUT wHaT aBOuT TAS?!?!?

Quick aside: I think TAS (Tool Assisted Speedruns) are different. They still require human input. Yes, they are theoretically input perfect runs that no human could ever hope for, but they are limited by the creativity of the human generating the TAS. A TAS will not discover and optimize on its own.

Conclusion

I do fear for speedrunning. I don't actually know if anything I said is true. Maybe speedrunning will continue to thrive completely unfazed, and I hope it does. I am just worried that yet another cool outlet for human achievement could be potentially stripped from us like we are seeing companies trying to do to the other outlets I enjoy.

Please let me know what you think! This is my first ever "real" post I've written that's not technical. Its probably really bad, so please take everything I say with a grain of salt. I enjoyed trying to flesh out this idea into something worth reading, hopefully you agree :)