Yeah, it is harder than other things, but if we can train a model to explain collections of pixels in human language then we might be able to do similar with co

Yeah, it is harder than other things, but if we can train a model to explain collections of pixels in human language then we might be able to do similar with collections of activations. I don't know if that is the direction, but just an example that comes to mind easily. If someone figures out how to do this, I think their models will be far more capable and reliable.