✨ Nora Tindall  🪐

Also, the majority of the data used to train Copilot (which came from GPT-3) is not code. The biggest part of it was the GPT-3 part that was trained with public internet content (which could also be licensed). So, you can't also state that "everything" (+)

that Github used to train Copilot was potentially licensed since the majority of the training was done in the transformer (GPT-3). So, the "code" part is just the tip of the iceberg. Just to specialize the GPT-3 algorithm to write code (+)

Anyway, we need to do a LOT of debate on this subject. Not only for copilot but for ANY licensed content. For example: - If we train an AI to write songs, what should we do? Since we used licensed music to train it (+)

- If we train an AI to write movie scripts, what should we do? Since we used licensed movie scripts to train it - If we train an AI to take pictures, what should we do? Since we used professional pictures to train it (+)

This is a WHOLE new area. We can't simplify it as "OMG! COPILOT IS WRITING THE SAME LINE AS WE CAN SEE HERE IN THIS SOFTWARE" Common, it is way more complicated than that!

Actually, we can, and we should simplify it, because no matter how complex the AI system is, it boils down to the same end result. It is designed to provide useful recommendations, not to avoid copyright infringement, that's the problem.

You're mixing concepts, maybe because (as you've said) you're not an expert in machine learning. Recommendations are different from predictions. What Copilot (and also GPT-3) does is predicting your expectations based on a previously stated context. (+)

As I've said, if we go to this road, we can cancel any machine learning initiative, since every ML initiative uses publicly available data, not considering if they are licensed or not.