Something I didn't see mentioned is also that over time this is going to feedback loop into the training set and compound (possibly exponentially). I.e. as mor

Something I didn't see mentioned is also that over time this is going to feedback loop into the training set and compound (possibly exponentially). I.e. as more lower quality code hits github and is used for training, the output of the code will decline, which causes lower quality code to hit github, which causes output to further decline, etc. I am an experienced dev but new to ML so take with a grain of salt, but I really wonder if the future is going to be quality in the training sets rather than quantity. I have heard that the "emergent properties" don't seem really affected by bad data as long as the set is large enough, but at what point does this cease to be true? How do we guard against this?