The two models were trained jointly, without labeled data, in a self-supervised way, over “millions of code snippets”, the researchers explained. Even though the idea was to create a program that can identify arbitrarily complex bugs, these are still “outside the reach of modern AI methods...