Exploring NLAs

This project is setup as a modal notebook and is some exploration that I attempted with Anthropic's new paper on Natural Language Autoencoders. I wanted to build up on some of the discussions from Less Wrong (credit ryan_greenblatt) and wanted to explore multi-modal capabilities of the model which were not tried by Anthropic. I think NLAs are a really powerful method and my own investigations show the next steps to be generalising one NLA for all models and also attempting to convert the activations to an image instead of text