AIXI -- Efficient AGI -- Transformers

AGI already exists, it's called AIXI and it's provably optimal. It works by iterating over all programs up to a given length and choosing the shortest one that reproduces the observations. Transformers perform no iteration at all. In particular, they perform the same amount of computation for each input. This means they will never be able to reliably do reasoning or even reliably multiply two numbers together.