AI Code Copilot Innovation: GitHub Introduces Feature to Cite AI-Generated Code

0:00GitHub has just announced a big new product that I believe is going to solve an issue a lot of people have had specifically with GitHub AI copilot. So in the past, a lot of people complained when this thing first launched because essentially GitHub's co-pilot was trained on all of the repositories on GitHub, all of the code on GitHub. But the issue was a lot of private repositories were used for training. And people were finding that when they were getting co-pilot to recommend code, it was recommending code from repositories that were supposed to be private, like things coming out of Apple. And of course, this is a big no-no if people are able to get code straight out of Apple's private IP, their private repositories.

0:44So GitHub had to do some quick shuffling back then. I believe they pulled all the private repositories out of their dataset and retrained it. They were able to get rid of that. But still, people to this day are complaining about essentially having code that is from other people's repositories. They did try to put another solution in place. In any case, this has been a problem that they've had for quite a while. and today they may have come up with a solution. So on the podcast, we are going to be diving into what exactly happened and what they did here. So in 2022, GitHub put in motion a feature that essentially permitted users to automatically block suggestions of matching public code, right?

1:23This makes a lot of sense. This is public code. They don't want the suggestions. If it looks like they just straight up plagiarized and copied it somewhere, you know, they could be liable. So while this did address one issue it didn't answer all of the concerns according to github and a spokesperson there um the system would come into play less than one percent of the time so you know it's kind of an interesting idea maybe gave some people peace of mind but in reality it wasn't actually used very much so um developers might be intrigued to inspect these code fragments either to incorporate them or perhaps to kind of explore the entire library from which the snippet originated right of that one percent of the time so in an effort to kind of bridge this divide between both desires, GitHub has announced the launch of a private beta of a code referencing feature for GitHub Copilot.

2:07So rather than an outright blockage of matching code, developers are now empowered with a choice. When code referencing is activated, Copilot showcases the matching code in a sidebar, and it gives the developers the freedom to decide essentially their fate, right? What they want to do, if they want to take this code that's somewhere else or not. Now, the thing that I actually see being valuable here, you know, they say, oh, this is less than 1 % yada yada and I know they try to downplay it but if you if you're trying to code something you can't figure it out it's really complicated you see someone else has done it it's part of that you know 1 % sometimes the 1 % of code is still the valuable percentage that a lot of people would like to use or it's really valuable or it's the trickiest part right so let's say it's there I think developers would like the option of being able to see that code and you know some people are like oh then they can copy it in reality I think they would like to see it and maybe they would just like to see how the how it was coded they can write it their own way but i think that it's actually a really useful even if not for anything but educational purposes i think it's a really useful tool so this feature is planned to expand to co-pilot chat eventually so it was previewed last november but the feature required some fine-tuning before its release and github ceo thomas domk expressed that major players including microsoft github and most quote pilot enterprise customers utilizing the original block were you know using the original blocking feature so however he also acknowledged its limitations stating quote it gives you a little control to decide for yourself whether you actually want to take that code and attribute it back to an open source license it doesn't actually let you discover that there might be a library that you could use instead of synthesizing code it prevents you from exploring these libraries and submitting pull requests you might be reproducing everything that already exists in some open source repo so dom highlights that this was particularly relevant to commonplace computer algorithms like sorting the newly introduced feature will allow developers to reject the code use it directly in the if the library allows or have copilot reformulate the code so the current version doesn't actually allow for results filtering by specific licenses but the team is working, they're getting feedback to kind of gauge the necessity of adding that.

4:25They said, quote, we're letting people understand the match and then go on and explore or go and make the right decision. So Domk explained, adding that he believes it fills the gap in the original solution. So the code referencing feature pretty much tends to be more active in scenarios where Copilot, it kind of lacks context. So when there's tons of context and that's all available, you know, from existing code, Copilot is a lot less likely to suggest matching public code. However, at the initial stages, I think the likelihood of generating matching code actually increases. So central to this whole kind of initiative is a fast search engine, which is aiming to essentially keep latency within 10 to 20 milliseconds.

5:16I think this is really big. No one wants to wait for their code to generate. No one wants to wait a long time for anything. And I think increasing the speed is a big, uh, is a big bonus on this. So it can rapidly essentially identify matching code and its licenses. Um, you're not waiting around for that. And currently the matching snippets are listed, um, in the order found, but in line with GitHub's original announcement future functionality to is going to be to sort by repository license, commit date, and, you know, etc. And that can be expected in the near future. So in essence, GitHub's new feature manifests a nuanced approach to code referencing, catering to the multifaceted need of developers, right?

5:57And I think that the ever changing landscape of software development continues to be shaped. We're seeing a lot of, you know, rapid developments here. And I think that GitHub's latest move is a reflection of the you know delicate balance between innovation and also responsible coding practices so i think this is going to be an interesting story to follow in the future seeing how they roll this out and how this is actually adopted by enterprise