I love this, there are so many interesting questions and challenges in the area of how humans will interact with AI. To me the correct model is to think of inte

I love this, there are so many interesting questions and challenges in the area of how humans will interact with AI. To me the correct model is to think of intelligence as a resource like compute as opposed to a personified computer, which opens up the opportunity for so many more applications. An example is the recent text-inversion paper where they found a way to extract the embedding of a visual concept from several images. Or how Midjourney is able to somehow integrate their model with StableDiffusion to get the realism and control of SD but the aesthetics of Midjourney. Basically, it seems like deep learning networks are a lot more composable and controllable than I would have expected, while still demonstrating really interesting features. There could be entirely novel neural net architectures with no analogue in the human brain that still do really interesting and useful things. It’s exciting to see what people are coming up with!