✨ Nora Tindall  🪐

Training an AI based on information doesnt mean copying the code / violating licenses. Take a look at how the copilot network works - its not much different than a human looking at your (public) code and learning from it

That's not true - unlike a human, it does not conceptually understand the material. Rather, it is a set of weighted dice which sometimes produce useful results. Also, it clearly does contain copies of at least some of the training set, because it can regurgitate verbatim.

Actually, that is debatable. There is no clear precedent for that. We would need now to define if a license protects the code against that or if it don't.

The whole thing is dubious at best, it is essentially offloading the task of "taking inspiration" from existing source code to AI that really doesn't understand what the code does. You can learn *what* a code snippet is doing and figure out *how* to do the same without copying it

What makes it even worse in this case is that GitHub Copilot is like a black box for the human using it: it magically spits out recommendations. By obscuring the process of sourcing where the code comes from it becomes almost impossible to catch cases leading to verbatim copying

GitHub Copilot is basically the equivalent of taking a very large collection of completed Legos and throwing them in a pool: a lot of them will be broken in small pieces but out of the lot you'll get large enough chunks still intact that will be suggested by the AI almost as-is

I don't know how well you understand deep learning and NLP, but you should research GPT-3. It is not exactly what you're suggesting. Github Copilot is nothing special, it is just an implementation of the GPT-3 transformer by specializing it in coding.