[manchmalscott]

On one hand I can see how this would be incredibly useful (I can never remember what ffmpeg flags I need) but I also just,,,,, I have this immediate uncomfortable reaction to the thought of just piping AI output into the terminal with the express purpose of not needing to know what the flags do. If the AI makes a mistake and the only way to catch it is to know what the flags do anyways, I think this becomes less useful.

[ajuc]

This reminds me how people in lesswrong community were arguing whether it's inevitable that when general AI is created - it will escape its sandbox. There were elaborate schemas how it will persuade or blackmail the scientists to let it out.

In practice it won't need to, somebody will make a plugin that uses it to download funny images or highlight code in console :)

[visarga]

"escape its sandbox" ... it's more useful outside the sandbox, so... we just let it loose

[eurasiantiger]

Someone somewhere is training an AI to speak TCP/IP directly.

[jamesgreenleaf]

I'd be worried about the same thing, and wouldn't execute blindly. It seems like a useful way to quickly find out which flags you should be checking the man pages for.

[bryanrasmussen]

this makes me think of a really great developer I worked with one time who had one major failing, he was too trusting.

So one time he got some code from the senior backend Java developer in Argentina and went to past it into the place he was told to paste it into for sending alert notifications to customers, some months later we are going through a major crisis because alerts have not been working for several months and all the customers who are paying thousands per month for the service are getting pissed.

He worked home a day he could really focus on debugging and finding what could possibly have caused our systems to fail!

I was done with some little task I was on and decided to look through code changes and it really stood out the code he had pasted because if it worked that would mean Java was some magic language with incredibly terse code.

So later on we were having a major emergency call to figure out what could possibly be causing this extremely serious major bug and I said well I was looking through the code and you could see him manfully controlling his exasperation because hey, Bryan is not a Java guy and there is zero chance I could have a meaningful contribution to the current problem.

But yes it turned out he had trustingly pasted in code that we all agreed would never have worked (without caring to basically even read it), which despite the fact that in many ways this guy was better than me is something I would never, ever have done because I am a really hyper-paranoid guy about trusting code.

Anyway tldr, I would never use this tool but maybe some really skillful people would because they don't have the paranoid mindset.

[abhagsain]

Yes, you're right. I have been using it for only a week now and I search for common commands like how to reverse tunnel, find a pid etc.

Unless the action that I'm doing is destructive I just run the command (YOLO)

[scraft]

What if the nondestructive thing you are querying GPT3 for returns a bad result which is destructive?

[abhagsain]

then I'm screwed but I don't think the model be that off from the actual question. If I ask "copy a file from local to AWS S3" it won't return the delete command. But yeah you gotta be careful.

I show a warning after every command you run "Please don't run commands that you don't understand "especially destrictive ones")

[vocords]

Common sense :)