I think you just answered your own question. In the scenario of ASI, who is training an AI smarter than an AI that's already smarter than us? Also I don't suspect AI will continue to be as inefficient as it is today in training or inference. A good example are voice clones that went from needing hours of audio to just 30 seconds of training data. On top of that synthetic data has been more than useful, aka data created by an AI to train another. One of my earliest projects was for a client who couldn't give us any data to work on but gave us a breakdown of what the data was like so we just lorem ipsumed our way there to a model that could actually find their false negatives and false positives in their testing.