They setup an LLM to play a game and it sucked at it. What were they expecting? That’s not what LLMs are made to do.
Yea but getting something to do something that it wasn’t meant to do is part of the fun lol
Sir that’s called sexual assault.
They were doing this on youtube years ago. I remember watching this during quarentine. I’d watch for 2-3 minutes, and mario would die in the same place every time.
But it would try something new each time. You’d watch it run into the same goomba each time for 2-3 minutes, but with very slight variation. Then you’d see a few days later it got to halfway through the stage.
Then by the end of the month, it was to the 3rd world. I think it took 6 months to beat the game. Then they’d save that file, start a NEW file, do it again, and then they’d combine the two files. Supposedly each generation of combining AIs would find the most efficient way to win. Meaning each generation is smarter than the last.
I believe that was a genetic algorithm. Something that would be a lot more successful than an LLM in this context.
Probably reinforcement learning? LLMs are a bad architecture for something like real time video games
Yes Timmy, we spent months finding the best Marios and then made them have offspring. The wonders of AI!
I feel like a good test of any supposed AGI would be to hook it up to a feed of a classic 3D platformer like Mario 64, give it input control, and see how long it takes to progress through the game. We’re seeing sparks of this with Claude Plays Pokemon, but any self-respecting superintelligence (or even human-equivalent) should be more than capable of learning the control scheme, navigating the 3D environment, solving puzzles, and generally playing through as competently as any 10-year-old seeing the game for the first time.
