Definitely a game changer for speed and productivity for people who know how to use it, just as you showcased. I fear a lot of people misuse it and think it can write everything on its own instead of being an actual 'copilot'

@sebastianjost

Жыл бұрын

I'd say it's a copilot that's still in training. Very good for some things, but doesn't quite see the bigger picture yet. Hopefully Copilot X will be better in that regard.

It's great to see a developer interact with his fanbase. I've been watching all your videos. The one where you explain how to code terrain was awesome.

It's super useful, my job changed to be more of a code reviewer of the stuff copilot generated. Often, I'm still thinking of how I want to write the code, then the copilot suggestion comes in, and it looks generally good for the most part. It even remembers constants from other methods and how you have called a method before.

This is incredible

okay that’s actually cool

Incredible

I don't know the context of this code, but I would guess that maybe a third could better be trivialised by not requiring such verbose code (whether that's a weakness of Java or of some library or of the code base). That would also help with reading code, which is often what developers spend more time on anyway. But the rest seems really cool! For helping to read, maybe there could be AI-created warnings when the comments seem to contradict what the code says, then devs could often just read the comments and skip reading the actual code. And of course it would have to be paired with writing good comments.

So, should I also trust AI to write unit tests for the code it generated? Or will it be biased to wrap the test to succeed around the method? Or should I first let it generate some unit tests for a hypothetical function with certain requirements. And then let it generate that function?

@sebastianjost

Жыл бұрын

Depends on the function. TLDR: three tiers of capabilities below help decide whether to generate code, then always read it yourself. 1. Copilot is very good at doing very common tasks, writing boilerplate code and filling in documentation for function parameters or adding a single line of code. 2. More complex tasks like writing a full function, function description or very task-specific lines are quite unreliable. In those cases filling out some comments for the rough structure often helps. 3. Letting Copilot write entire classes or any project with more than 100 lines at once without extra help is pretty much impossible. In tier 1 and 2, letting Copilot generate code is usually helpful. If it's not right the first time, you can try helping a little to figure out a solution together. Especially with tier 2 problems, you always need to check what it generated. There are often mistakes. Sometimes they're easy to fix, sometimes it's going in a complete wrong direction. In tier 3, letting it generate code is mostly a waste of time. Either write it yourself, find code online or ask GPT-4.

@henrikkniberg

Жыл бұрын

I often use GPT 4 to write unit tests or assist with more complex code. It usually does a great job, but even when it fails and the code doesn't work it serves as good scaffolding and sometimes as educational material (like for example I learned how to do mocking in typescript). Copilot works for simpler things, is right at your fingertips, no copy-paste, but can only add code, not change/remove code. GPT4 + Copilot are great complement to each other.

will ai help in the coding of minecraft updates? its not perfect sure but it can make the process alot easier

Holy mother

Holy cow

Ayo you're back. Do some minecraft

@RohanSlazar

11 ай бұрын

He's no longer working at Mojang, sadly :'(