MuZero, the new AI developed by DeepMind — which was bought by Google and is now part of Alphabet — is capable of doing anything without being designed for it.
The AI agent is developed to learn any rule by itself and without the help of human beings, and in future, MuZero could improve YouTube by reducing the consumption of data from videos.
With the new DeepMind AI model, MuZero is able to learn the rules of the game on its own, something that makes it even more like human beings and their ability to adapt. Thanks to a combination of a binary search tree — a data structure that allows finding information belonging to a class or ‘sub-tree’, and a learning model, the two most common techniques in AI development.
The big problem with depending on a learning model is that the results are better the more precise that model is, and therefore the worse the more complex it is (for example, if you have to learn shapes from an image). To fix this, MuZero only models the parts that it considers important to the decision process.
For example, if you are facing a Go board, MuZero creates a representation that updates with each iteration depending on the actions it performs; the model predicts the action (such as movement), the value of the function (whether to win), and the immediate reward (to win a point). In this way, MuZero invents its own rules, which at the end of all iterations coincide with the real ones.
This allows you to get similar performance in chess, Go and shogi, and play classic Atari video games successfully. But its creators don’t just want to ‘crush’ the players; explain that one use of this AI would be to create a new video compression algorithm, which would be more efficient.
By not having the limitations that humans have, AI has turned out to be quite good at compressing videos without loss of quality. And although DeepMind has not confirmed it, it is to be expected that Google and its services, such as YouTube, will benefit from it.