3 ms·
Hmm, I was thinking about building a 2d game solver with a LLM and I was wondering the other day if there is a better way to represent 2D data instead of just r
by DesiLurker 4y ago
Hmm, I was thinking about building a 2d game solver with a LLM and I was wondering the other day if there is a better way to represent 2D data instead of just raster scanning it to a vector. probably a hadamard transform might work out better but the hilbert pattern looks interesting too.
- gus_massa 4y agoI'm not sure I understood your idea, but somewhat related (of the inverse representation of your idea?) https://xkcd.com/195/ https://xkcd.com/195/
- DesiLurker 4y agoits basically trying to figure out the best way to represent 2D or 3D data to be vectorized and fed to a LLM for a puzzle game. I am rather new to the field so may not be using the right terminology. The idea is to create a bunch of these random game boards and generate various intermediate player states leading to a solution so the LLM can learn from it & then ask it to come up with steps to solve for a new board config. Its not a serious product, I am just trying to get my feet wet and understand how to use these transformers.