One question I have about the potential widespread use of AI music generation systems like MusicLM is whether AI will have the same capacity for creativity as humans do. Right now, my understanding of current AI music generation systems is that they are trained using existing music datasets and are capable of producing music similar to the data that they are trained with. Humans similarly learn by using information from the present and the past. However, I think humans also have the unique ability to be creative and innovate beyond the information they have been provided. I feel like this is the reason that music is constantly evolving and changing and the reason that music today doesn’t sound like the music from 50 years ago. Could AI systems be creative in the same way and be capable of developing new types of sounds, new instruments, and new genres? Or would music just continue to generally sound the same without the input of humans? 


So far, I think AI systems have shown that they have some capacity for imagination and creativity, but still not to the same extent as a human. My first exposure to this was when I learned about AlphaGo, an AI that DeepMind developed to play the board game Go. I grew up being a huge chess nerd and had a basic understanding of how chess engines worked. Traditional chess engines are able to play chess incredibly well by traversing the tree of all possible moves and heuristically evaluating each resulting position to determine the best possible move. This strategy doesn’t work in Go because the number of possible moves in any given position is much higher, making it impossible to traverse the entire tree of possible positions. It is also hard to algorithmically evaluate a board position because of the complexity and creativity involved in Go strategy. 


AlphaGo was the first AI capable of beating the top human Go players in the world. Unlike chess engines, AlphaGo uses a neural network and is trained on a large number of Go games to learn what kinds of moves would be best in different scenarios, more similarly to the way a human learns. In AlphaGo’s highly publicized matches against famous Go players, commentators stated that some of the moves that AlphaGo played were extremely unusual and that a human would never have thought to play them. Many said that these unusual moves were brilliant. 


In this way, I think that AlphaGo was able to achieve a type of imagination and creativity that extended beyond a human’s capacity for creativity. However, Go is a confined system with defined rules. There is a limit to the inputs, outputs, and potential actions in this system. It is a lot easier to be creative within a confined system than it is to be creative when the possibilities are endless. When creating new music, for example, humans can be inspired by so many things outside of existing music, such as the emotions they feel or events that occur around them. Humans are also capable of pushing the boundaries of what’s possible in the limitless realm of music creation. When the options are infinite, can an AI possibly mimic the creativity of a human? How could an AI understand what it means to feel an emotion and learn how to fuel that into a song? What kinds of data could inspire an AI to invent an entirely new genre of music?