I’ve been playing around with Copilot for the past several months, including some minor production use cases. I can give the following assessment. At first, it took more “risks” and thus produced better predictions for what I was about to type (none of them were perfect, however, though some were hilarious). Later, after about 10 or so updates, it became much more conservative. As a result, it can now rarely offer anything better than whatever the autocomplete hint of the IDE itself offers as part of code analysis.
As with DALL-E and the other imagen-y AIs, I am personally more fascinated/terrified by the prospect of what happens if/when their dataset inputs begin to include significant amounts of their previous output.
DALL-E’s present output is far from perfect. So all this could lead to is to further degradation of output quality.
This “pollution” effect is a very interesting question that all these systems will have to face. We can expect that a primary use of these systems will be to generate enormous quantities of mediocre work and spammy nonsense. As this material appears on the public internet and becomes part of future rounds of training data, the model will produce progressively worse results, and magnify the influence of garbage.
So, after the trial period ended, I unsubscribed from copilot. As it became worse, if I would OK its suggestions, I would actually introduce bugs into the codebase.
It couldn’t even pick up on the variable names in the source (which it could do previously). So it all boiled down to the only helpful feature that was comment completion (kind of auto-doc).
Seems like a hard pass, at least for now.
I’m now part of a legal team that last week filed a lawsuit challenging GitHub Copilot.