text
stringlengths
0
2.35k
**Alexey Palazhchenko:** Yeah, yeah. So yeah, it works great for Go, but also it works great for many other places, for example, forgive me, Sometimes Copilot writes better even than I do, even with augmentation of something like Grammarly, for example.
**Natalie Pistunovich:** Yeah. I guess it is based on a natural language engine, right? So that's an interesting point that you pointed out, actually. I didn't to write any readme with it, but that makes a lot of sense.
**Alexey Palazhchenko:** Just recently I was writing a Docker Compose file... It also made a lot of sense.
**Natalie Pistunovich:** Yeah.
**Alexey Palazhchenko:** \[12:09\] So essentially, it can write the whole project for you. There is a very fun video on YouTube where the guys writes a game in six minutes using Copilot. It's basically a game where you pick a random number, and whoever chooses a number closer to the random number wins. And it just work...
**Natalie Pistunovich:** It does work well, from my experience. So what would you say the current state of AI-driven development these days -- do you know of people who are already using it at their job, or for fun? I guess for fun we know some, but for personal projects that are more than tab-tab-tab?
**Alexey Palazhchenko:** Yeah, so I actually use it for work now, and it's kind of scary sometimes. Again, the project that I'm working on right now - we have this problem of MongoDB being quite aggressive with their license... So we should be very careful not to touch the SSPL license code. When I write some handler f...
**Natalie Pistunovich:** Interesting.
**Alexey Palazhchenko:** In the same time, in that particular case I actually went and tried to search this code, and I did not find it. So maybe it actually this code knowing that I am working with MongoDB, and pulling in some other code, maybe from some other languages. I don't know.
So that kind of thing kind of scares me... If you work in open source, you have to be very, very careful about licensing. That could be a problem for you. But in most other cases, that's just a miracle.
**Natalie Pistunovich:** Yeah, that is an interesting point that you're bringing. I'd never thought of it until now... When you use AI to help you write code, you actually have to add one more test to everything that you are allowed to use the code that it generated... Like, when you search, that this code does not exi...
**Alexey Palazhchenko:** Yeah. So actually, after that happened, I did some research what other people think. Yeah, a lot of people think that GitHub Copilot is basically a code laundering machine, because they publicly say that they teach the code on all the public code, not caring about the license. So I can assume t...
So I would say that by the time we have some precedents and we actually know whether it was legal or not, our judges will be AI tools. So that would be an AI judging an AI.
**Natalie Pistunovich:** Maybe they will be the robots.
**Alexey Palazhchenko:** Yeah...
**Natalie Pistunovich:** \[15:59\] One of the interesting arguments in the conversation about AI writing code is that it's trained on open source code, and not all open source code, especially the one that is -- I mean, maybe it's even safe to say that most of the open source code is fun projects of people; it's not ac...
On the level of "Did it learn from a good example or from a bad example?", because GitHub code is not labeled with "good repo", "bad repo". And also, even if the language is written correctly. Does it have good safety practices? Is the code that is out there secure or not? So you have to look out for so many things whe...
**Alexey Palazhchenko:** Yeah, but at the same time, how do you know if the code you want to look on GitHub as you, as a human or whatever, is good or not? Let's say you want a new go developer you join a community and you want to say "Okay, how you organize my Go code." You google it and find this famous Go standard l...
**Natalie Pistunovich:** I guess if it has some people thumbing it up, or even commenting... If it's active enough, it means enough people don't think it's bad.
**Alexey Palazhchenko:** Yeah, maybe. But maybe all those people's AI is too -- maybe it just found it, and started it, and then more people are coming and will start it... But the very first people did not know that it's maybe not good.
**Natalie Pistunovich:** Just skipped it for later...
**Alexey Palazhchenko:** Yeah. So I would say it's always for the human to decide, and I wonder how this would affect the coding styles of a human for example. It may be very different from what language designers expected it to be.
**Natalie Pistunovich:** Yeah. This brings me to the question if Go would be a good choice for AI-driven development, for the style of the language. What do you think?
**Alexey Palazhchenko:** Yes, all this power of go fmt, and you have the same standard for all the Go code - that seems to be very easy for Copilot to understand Go code. And I would say maybe Go is the single best language for that. I don't know another language that has this simple code structure, this regular struct...
**Natalie Pistunovich:** Yeah, that's actually another interesting point, that is AI code trained on up to date code, that is made for our machines, and not for the past machines.
**Alexey Palazhchenko:** Yeah. Can you imagine Go with generics generated by Copilot?
**Natalie Pistunovich:** Maybe we will have to... \[laughter\] Eventually we'll have to imagine how that works, yeah.
**Break:** \[19:32\]
**Natalie Pistunovich:** I've been trying to do the Advent of Code on different programming languages with using Codex, which is the underlying engine of Copilot. So basically, Copilot is kind of you take whatever you have in your IDE, and then you add the prompt to that, so kind of like the secret sauce, and then you ...
For C, for example, you'll probably start with an import, but maybe not... Not all languages have such a signature, if you will, to also help the AI generate, or even understand what language you wanted to use. I guess that's another big benefit of Go, in addition to having this standard formatting, that it always has ...
Last week was GopherCon, and there were some conversations around that, and one of the points that I made in a talk about that was that Go will not land you in the uncanny valley of code. That graph that makes you feel awkward about robots, and the one that can make you feel awkward about code. Because if it looks like...
**Alexey Palazhchenko:** Well, I would say that generated Go code sometimes looks like uncanny valley. So it kind of makes sense, but it also kind of doesn't, and it looks very weird... Look at it careful and it doesn't even compile. It doesn't work. But it looks like it almost does what you need it to do.
This video has a great example when the guy was making a game where you should guess the number, and the one who chooses the closest number wins. But Copilot always steered him to the game of "You have to choose the correct number exactly." So it basically tried to redefine the rules of the game. That was quite strange...
**Natalie Pistunovich:** Interesting. So was it doing that instructions via comments, like the documentation of the function, or...?
**Alexey Palazhchenko:** Yeah, yeah. He basically wrote the first line of comment, and then Copilot suggested even comments for that, and then code.
**Natalie Pistunovich:** And what is your experience of using Copilot with other languages that are not Go?
**Alexey Palazhchenko:** So the second-best language Copilot works for me is actually English... As I mentioned, readme generation is just great. And comments -- yeah, sometimes it writes better comments for my code than I would do.
And you know, they have this new feature in beta when they allow you to describe the code as English. For example, you have a snippet of code, and -- yeah, most importantly, it works with regular expressions. It translates regular expressions to plain English. That's just great for people who don't know regular express...
**Natalie Pistunovich:** Would you say it works great in both ways? Regex to English and English to regex?
**Alexey Palazhchenko:** Yeah, I did not try that. It would be interesting. Actually, the problem there - it's not an algorithm; it's just machine learning. And maybe it's not even correct. That would be also interesting, to check if that's correct. Yeah, I tried that with Python. I mean, it still works, but in my opin...
**Natalie Pistunovich:** \[24:17\] I tried writing some Bash scripts with it and that was also successful in doing what it was supposed to do. I guess Bash is similar to Go in that sense, that it's a kind of one way of doing things.
**Alexey Palazhchenko:** Yeah, maybe just my way of writing Python is outdated. The last instruction I was using was using Python 2 when I switched to Go. So maybe Python 3 is very different.
**Natalie Pistunovich:** I imagine Python has multiple ways of doing things, so you cannot just set a one-one. And what languages did you get to try that were not really a good choice?
**Alexey Palazhchenko:** Actually, I don't think I tried a lot of languages. At my current job I use Yaml. Yaml works great; English works great. I used configuration file, JSON works great. Everything was pretty much great for me.
**Natalie Pistunovich:** Well...
**Alexey Palazhchenko:** I imagine something like very esoteric, I don’t know, like Malbolge programming language would not work very well.
**Natalie Pistunovich:** LaTeX, LaTeX might be interesting to see how that goes. English to LaTeX.
**Alexey Palazhchenko:** Imagine you could write a Ph.D. with Copilot.
**Natalie Pistunovich:** "Here is my thesis in one line... Please write my expanded thesis, with correct formatting." This will be it. So AI can be used to augment your life as a developer, in helping you write code and helping you write readmes... But I can think of all other types of help that it can do. For example,...
**Alexey Palazhchenko:** Yeah, absolutely.