Comment by jah242

Google actually recently went some steps further and combined the PaLM LLM (bigger than GPT-3.5) with a 22 billion parameter Vision Transformer to do this -

https://palm-e.github.io