CoPilot

[Agent Crimson]

Linus and Luke talked about CoPilot on Wan show today. Having access and experimented with OpenAi's Beta I am not sure if I agree about how well CoPilot will be. CoPilot seems to be more intended to provide an outline of code or a general code and not a full to the spec code. 

 

[Eigenvektor]

It is called CoPilot rather than Pilot, so I don't think the intent is to replace developers altogether but rather to provide a generic template (based on comments + method name) that an actual developer can then refine.

 

Their examples feel a bit like doing a Google/Stackoverflow search and accepting the first result as your solution. I'm not sure having to proof-read the code produced/procured by an AI is actually more productive than simply writing the code yourself.

 

Though I highly doubt it can produce domain specific code and will mostly provide templates for things that are commonly used everywhere (e.g. "maximum of two numbers"). Which are most likely provided by standard libraries already, so…?

[pythonmegapixel]

The name "OpenAI" really irks me... a completely proprietary piece of software which a huge company has the exclusive license to the source for and where a registration procedure requesting an awful lot of information (and which looks like it's literally going to prevent you from using it based on certain conditions) is required to even use the thing is about as "closed" as it gets.

 

I don't like everything the open source initiative does, but they have the right idea with their definition of open - and this isn't it.

[Agent Crimson]

12 hours ago, Eigenvektor said:

It is called CoPilot rather than Pilot, so I don't think the intent is to replace developers altogether but rather to provide a generic template (based on comments + method name) that an actual developer can then refine.

 

Their examples feel a bit like doing a Google/Stackoverflow search and accepting the first result as your solution. I'm not sure having to proof-read the code produced/procured by an AI is actually more productive than simply writing the code yourself.

 

Though I highly doubt it can produce domain specific code and will mostly provide templates for things that are commonly used everywhere (e.g. "maximum of two numbers"). Which are most likely provided by standard libraries already, so…?

I absolutely agree and thats exactly my point/argument against what was said on wan show. 

[wasab]

To me it looks like a step above IDE's intellisense and auto complete. Pretty sure it can be a good built in feature of any IDE. 

[igormp]

I've been using copilot for the past week, it's pretty nice.

Most of the times it just gives you a useful outline, but sometimes it can actually give you something that you wanted. For example, it could easily complete a match case in scala for me using different regexes based on some string parameters that I threw at it.

 

It's also capable of reproducing entire codebases, that's why there's a whole copyright discussion going around it: 

 

[Wictorian]

Tools like this have been existing for some time. Examples are Kite and TabNine

[igormp]

9 hours ago, Wictorian said:

Tools like this have been existing for some time. Examples are Kite and TabNine

None of those have been trained on the entirety of github's projects though.

[Pyxlwuff]

I've used copilot for the past few weeks and I've found it to be a massive help when working with libraries I don't know much about. Managed to use it with wordpress extremely well and like others have said in the thread would give outlines or 100% correct snippets for what I asked about. Of course there were a few hitches and confusions with some of the prompts but it did a good job overall.

 

That being said I wouldn't reccommend using this on a mass-scale just yet especially if there's copyright issues at vein still. It's fine the way it is right now.