Quoted By:
I really want Neuro to do streams where she "learns" to play hard games using AI, not with the LLM, but through actual reinforcement learning like in Osu. Like imagine her playing the popular flash game "The World's Hardest Game", where she struggles at first, dies in dumb ways, slowly gets better, and finally beats it in one stream. Then, in another stream, she comes back and completes it faster with fewer deaths, and eventually attempts a no-death speedrun. It could be done with a lot of games QWOP, Geometry Dash, Celeste, Super Meat Boy, Jump King, Getting Over It, and a myriad of retro games.
The fun part isn’t just her beating the game, it’s the illusion that she’s learning and improving live. Vedal could pretrain several RL agents of varying skill (generations), and swap between them to simulate “progress” if it can't be done live fast enough. Add her reacting to failures and success, plus responding to Twitch chat, and it’d be absolute gold.
Feels like something that should be core to what Neuro is.
