Let's look at an example, assuming we're using mean squared error as our loss. For example, if we make our first split based on whether or not mpg <= M, we'll have some average MSRP for the low-MPG cars (below M), along with residuals if we used that average to predict the MSRP for all the low-MPG cars. Similarly, we have the same information for the high-MPG cars (above M). A good value of M will make the mean squared residuals for the two groups as small as possible. So, at each node, we choose a split that makes the mean squared error on each side as small as possib