It’s already been a few years, before the LLM boom, but a neural networked figured out how to play Super Mario Bros by simply looking at the RAM and it mastered the game.
If conditions are the same every run, it will eventually find the mathematically fastest way. I saw a video about a guy doing it, and after thousands of runs the bot noticed a glitch if the car wasn’t on four wheels allowing it to move insanely fast. Like grinding a rail in Tony Hawk, the bot would immediately do it, and run the entire course on the glitch.
That’s not human intelligence tho, that’s the same as when a slime mold can design a transportation network as effectively as we can.
It’s already been a few years, before the LLM boom, but a neural networked figured out how to play Super Mario Bros by simply looking at the RAM and it mastered the game.
Racing games too, especially single player races.
If conditions are the same every run, it will eventually find the mathematically fastest way. I saw a video about a guy doing it, and after thousands of runs the bot noticed a glitch if the car wasn’t on four wheels allowing it to move insanely fast. Like grinding a rail in Tony Hawk, the bot would immediately do it, and run the entire course on the glitch.
That’s not human intelligence tho, that’s the same as when a slime mold can design a transportation network as effectively as we can.
But that’s cheating no? The challenge is playing the game using user input with all the delays and extra steps that can happen along the way.
Also, optimizing a neural network for just one narrow use case is not what’s being discussed here.
ML is very, very good at narrow use cases, same with other tools of automation going back a century, they are talking about generic all purpose ML