[deepGem]

I spent a few weeks trying to build an alternative to self attention that scales memory linearly. I I got surprisingly good results. While in principle this makes a lot of sense, I am struggling to push the test accuracy above 86%.

Some of the alternatives I am about to consider:

1. Diffusion with sparse attention layers. 2. Hierarchical diffusion - next token diffusion combined with higher order chunk diffusion.

Still figuring out the code and I would love any feedback on these approaches.