Last night I was nearly tearing my hair out trying to get a crappy inkjet printer installed when I turned to ChatGPT for help; 5 minutes later it was done (driver issues).
I realised that I use it a lot for all sorts of things but haven’t really seen any posts on here about it, and was wondering if other people do the same.
I’ve used it to create bash scripts, write small, niche python programs (eg sort photos into separate folders based on their orientation), and create cron jobs among other things.
In particular, I find it useful when I want to use a command line program that has a complicated set of options / syntax, but I might only ever want to use it once, or very infrequently. So I tell it what I want to accomplish, and ask it to give me the correct command to run with the options I specify in plain language.
Example:
User
Give me the correct rsync command to recursively copy all files in /home/user/Documents that are image files of any type but were created more recently than 1/1/2024. The destination is /home/user/foo
ChatGPT
rsync -avm --include='*/' --include='*.jpg' --include='*.jpeg' --include='*.png' --include='*.gif' --include='*.bmp' --include='*.tiff' --include='*.svg' --exclude='*' --prune-empty-dirs --min-age=$(date -d '2024-01-01' +%s) /home/user/Documents/ /home/user/foo/
User
Break down the command and explain each part of it in dot point format.
ChatGPT
- rsync`: Invokes the rsync program, which is used for efficiently transferring and synchronizing files across systems or within a system.
-avm:-a: Archive mode, which preserves permissions, symlinks, modification times, group, and ownership. Implicitly includes recursion.-v: Verbose mode, which provides detailed output of what rsync is doing.-m: Prune empty directory chains from the file-list.
--include='*/': Include directories themselves, which is necessary for recursion.--include='*.jpg' --include='*.jpeg' --include='*.png' --include='*.gif' --include='*.bmp' --include='*.tiff' --include='*.svg': Include files that match these image file extensions.--exclude='*': Exclude all other files that do not match the previously stated include patterns.--prune-empty-dirs: Exclude empty directories that result from the inclusion/exclusion pattern.--min-age=$(date -d '2024-01-01' +%s): Exclude files that are older than the timestamp for January 1, 2024. This usesdate -dto convert a date string into a timestamp, whichrsyncuses to filter files./home/user/Documents/: Source directory from which files are being synchronized./home/user/foo/: Destination directory to which files are being synchronized.`
It’s also really good for explaining concepts / commands in plain language.
It’s like having a 24 hour on call Linux expert :)
#Things to note:
- Get the subscription. ChatGPT 3.5 is pretty useless. ChatGPT4 is fine, but I’m pretty sure you need the subscription to access it.
- Give it pre-instructions. I have told mine what distro, what shell I’m using and the make and model of my laptop. If you have a subscription you can add these as permanent pre-instructions, so to speak. That way it will be much more likely to give you correct answers.
- It’s not magic In order to get what you want, you have to be able to ask the right questions. It will boost your ability but won’t turn you in to a 1337 haxx0r
-Ask questions As always, don’t run any commands that you don’t understand. Ask it to break down any commands it tells you to run if you don’t understand them.
-Sometimes it goofs For various reasons, sometimes it will ask you to install a package that no longer exists, or will give you a command that doesn’t work. When that happens, I just feed the error message back into ChatGPT and it usually is able to correct itself.
-Ask “Is there a better or easier way to do this?” This is probably the most valuable question I’ve found to ask chatGPT. Sometimes it gets so far in the weeds looking for a solution to a problem that you need to pull back and start fresh.
Someone excitedly demonstrated to me how easy it is to code with copilot. They generated a bunch of code easily. And then proceeded to debug subtle bugs for longer than it would have taken to write it yourself in the first place.
And in the end they were still left with badly structured and maintainable code.
LLMs will do exactly what Stackoverflow has done, but more efficiently: allow profileration of bad/outdated solutions to problems, and application of those with no real understanding.
More garbage code and more work for the few people who continue to actually read manuals and understand what they are doing.
Perhaps, but I’m not really suggesting its use for professional programming in this post.
What it is good for is helping with simple stuff like terminal commands, learning python etc. Stuff that has a low risk profile that you’re not relying on for anything too important.
User: “ChatGPT, write me a script to clean up my hard disk on Linux”
ChatGPT: sudo rm -rf / 😁
Squeaky clean 😅
Chatgpt does not know truth. It does not know if the info it provides is true. It does not know if the code actually works. It just concatenates strings based on probability. You may be lucky or you aren’t. The easier the task, the more likely it’ll succeed. But a low difficulty is no guarantee for success.
It is great for layouts, structure and for the basic concept. “For loop in fish”. But it may struggle to convert a video from x264 to av1 with ffmpeg. It depends on info that’s provided online. If it uses misinformation, then that’s in there as well.
The command you got is just wrong. What about avif, jxl or most other image formats? Use it, but think.
I was wondering how long it would take the gatekeepers to show up. The command works, and is perfectly fine. If I had any uncommon formats, I would tell gpt to include them.
I’m quite sure it won’t be long until some bad practice spreads like this. Giving clueless “Linux pros” top advice on how to enable a back door.
LLMs can be poisoned and as datasets increase and complexity grows it will be harder to contain.
Cgpt works great for some stuff, but all you know is that someone somewhere wrote something similar. They are no better than Google in predicting what is good material and what’s wrong, and training is statistics.
In order to poison a LLM, you’d need access to the training process, which is locked down by openai. Just posting false info on the net isn’t enough. GPT doesn’t simply repeat what’s already been written.
More than that though, you can find plenty of wrong and bad advice posted confidently by legions of Linux gatekeepers on any forum.
Anyone who has ever spent any time on stack overflow will tell you why they’d rather talk to an LLM instead of posting there.
chatgpt only generates text. that’s how it was supposed to work. it doesn’t care if the text it’s generating is true, or if it even makes any sense. so sometimes it will generate untrue statements (with the same confidence as the ‘linux gatekeepers’ you mentioned, except with no comments to correct the response), no matter how well you train it. and if there’s enough wrong information in the dataset, it will start repeating it in the responses, because again, its only real purpose is to pick out the next word in a string based on the training data it got. sometimes it gets things right, sometimes it doesn’t, we can’t just blindly trust it. pointing that out is not gatekeeping.
Gatekeeping ain’t bad. It keeps the trash out.
Then why are you here?
Cute
don’t run any commands that you don’t understand. Ask it to break down any commands it tells you to run if you don’t understand them.
You need to pay extra attention to this, as ML models will spit out commands and parameters that doesn’t exists if there was not enough examples in training dataset for that action. Especially with explain as it could just spit out totally wrong but “sounding good” explanation for parameter etc as it not always will tell the magic keywords like “typically” that indicate that it doesn’t have confidence as it’s “based on other similar command/knowledge”.
In your example it spit out:
-m: Prune empty directory chains from the file-list. --prune-empty-dirs: Exclude empty directories that result from the inclusion/exclusion pattern.which is actually exactly the same parameter with 2 different explanations, you can confirm this with
man rsync--prune-empty-dirs, -m prune empty directory chains from file-listSo the more edge case you have the bigger chance it will spill out bad results, but those new models are shockingly good especially for very common use cases.
I’ll confess that I only tried gpt 3.5 (and the mistral one but it was actually consistantly worse) given that there’s no way in the world I’m actually giving openAI any money.
Having said that I don’t think it fundamently changes the way it works. Basically I think it’s fine as some sort of interactive man/stackoverflow parser. It can reduce frictions of having to read the man yourself, but I do think it could do things a lot better for new user onboarding, as you seem to suggest in the comments that it’s one of the useful aspect.
Basically it should drop the whole “intelligent expert” thing and just tell you straight away where it got the info from (and actually link the bloody man pages. At the end of the day the goal is still for you te be able to maintain your own effing system). I should also learn to tell you when it actually doesn’t know instead of inventing some plausible answer out of nowhere (but I guess that’s a consequence of how those models work, being optimized for plausibility rather than correctness).
As for the quality of the answer, usually it’s kind of good to save you from googling how to do simple one liners. For script it actually shat the bed every single time I tried it. In some instances it gave me 3 ways to do slightly different things all in the same loop. In other straight up conflicting code blocks. Maybe that part is better in GPT 4 I don’t know.
It also gives you outdated answers without specifying the version of the packages it targets. Which can be really problematic.
Basically where I’m going with this is that if you’re coding, or maintaining any server at all, you really should learn how to track the state of your infra (including package versions) and read man pages anyway. If you’re just a user, nowadays you don’t really have to get your hands in the terminal.
At the end of the day, it can be useful as some sort of interactive meta search engine that you have to double check.
I’m really not getting into the whole “automated garbage that’s filling up the internet, including bug reports and pull request” debate. I do think that all things considered, those models are a net negative for the web.
As an example of what’s possible with GPT4. Client wanted DNS auth in Letsencrypt instead of HTTPS, so we can close incoming port 80. They’re using a registar with a proprietary API. With ChatGPT I created a certbot plugin in about 10 minutes, feeding it a pdf with API description.
I know how to do every step of this myself, but it’s a 4-8 hour task to research the registar’s API and how certbot plugins interface. Instead, I took another 15 minutes to review the code, ran it, and it’s done.
I’m not opposed at all to using LLMs for such purposes, however, please consider a solution that aligns with the values of GNU/Linux and the Free Software Movement. If you have sufficient RAM and a somewhat modern CPU, you can do inference on your very own machine, locally, with no connection to any external servers. And at very respectable speed.
Serious question: Can running locally be as good as ChatGPT-4?
It’s worth doing anyway to get a sense of how computationally intensive it is. Then consider how many people ask for the daily fart joke and you get a sense of the environmental impact.
Is this as ad?
You could also use free LLMs, check out FMHY.
Good tips! I do the same, also for rsync.
Ask it to explain everything, you dont need to pay to have a permanent prompt saved (here “if I ask for a command, when writing to a file always do that in one command, dont describe beforehand but explain every argument”)
I think GPT 3.5 is okay, but makes some errors and has hallucinations. Sometimes the script is also wrong or has flaws.
Its basically a syntax translator, which is great!
I agree. I’m learning go right now and it has answered every question with absolute excellence. It probably accelerates my learning time by 10x at least.
It’s much better than searching the web, because you can’t share your exact code with the web and get a detailed walkthrough of how it works.
Gpt4 of course.
