I agree, the thought of hoards of developers sending their code to Microsoft (or AWS via CodeGuru) to train their proprietary semantic model is disturbing, because of the potential network effects. They suck right now and have a lot of problems, and I hope they either stay that way, or that something open like this supersedes them.

Hold up. Is everyone talking about this thing as if they're going to use it on proprietary code?

I only ever pictured using it on open source stuff if it turned up decent

If folks are looking at this thing like "that's going right on my work PC" I would highly recommend reviewing your IT usage policy first. Yikes.

I was thinking the opposite, for the audience you're describing.

"Can't get into trouble if it's never published publicly"...which is wrong, obviously.

How useful do you think has it been, for your own personal use cases?

How do you expect to ensure a "transparent pipeline" when GPT-3 itself is not fully transparent?

Unlike GPT-3, GPT-j from EleutherAI is a free dataset: https://6b.eleuther.ai/. This is supported according to the README.

Edit: it may not be supported right now, but it is mentioned in the Goals section.

I gave 6b.eleuther.ai a try on the Copilot fetch-tweets example, several days ago, and it spat out a similar completion: https://twitter.com/abecedarius/status/1408862135267037186

It took several seconds and added some logic I didn't ask for, but this is already interesting.