@anirudhbv_ce@MankyDankyBanky@modal Thanks! Seen in the videos, our rocket league models learned to stop contact farming and actually learned some pretty cool ball techniques along the way!
Over the weekend @mankydankybanky and I trained AIs to play Rocket League, Street Fighter, and Tetris, and won @modal’s challenge at SF’s AI Research Hackathon for it.
We built a recursively self improving RL framework where AI agents supervise training, spot failures, and run experiments to fix them with the end goal of training stronger models with less time and compute.
All we need is an arbitrary environment and objective, and our framework handles the rest.
Train your own AI on our framework using the repo in the thread.
this project caught my eye the moment they showed their trained LLM for Street Fighter III! https://t.co/VBsVEV281T
amongst the auto-research submissions at the hackathon, i was particularly impressed by the generality and branching nature of their framework. more to come from these folks 👀
The winner of our “best use of Modal” challenge at the latest @HackIterate’s AI Research Hackathon!
A general framework for recursively self-improving RL to prevent reward hacking and produce stronger policies.
Over the weekend @mankydankybanky and I trained AIs to play Rocket League, Street Fighter, and Tetris, and won @modal’s challenge at SF’s AI Research Hackathon for it.
We built a recursively self improving RL framework where AI agents supervise training, spot failures, and run experiments to fix them with the end goal of training stronger models with less time and compute.
All we need is an arbitrary environment and objective, and our framework handles the rest.
Train your own AI on our framework using the repo in the thread.
Our policy in Tetris can make up to 258 placements a second and reaches ~75% of the maximum theoretical efficiency of play, making it virtually undefeatable in a battle.
Went to #InferCon today, was awesome getting to listen to and chat with the teams from @modal, @nvidia, and @gimletlabs!
Much more exciting stuff coming soon, stay tuned 👀
you can tell it was an intense panel because i spent most of the time with my head down, staring intently into space and contemplating system design
silver lining: that meant i was hardcore repping my new @sgl_project hat