Improve Your ChatGPT FreeBSD Queries

B

BawdyAnarchist

AI/LLMs have been hugely beneficial to my FreeBSD experience, but you'll notice that responses bias significantly towards Linuxisms. You can overcome this somewhat by specifying obvious opening tags like: "In FreeBSD {command, config, system, /etc}, how/why/do {X,Y, and/or Z}. POSIX preferred"

But if you want to massively improve the response quality and avoid Linuxisms, upload the relevant manpages. Not copy/pasted as text, but as a file. Upload your config file too. I've found improved quality responses with statements like:
  • Take a look at the manpage and let me know if you can find {options, syntax, explanations, etc}
  • Be careful not to make things up. Read the manpage carefully, and let me know if there is any clarity regarding {Y}
  • [Copy/pasting terminal output with diagnostic errors]
  • Are you completely sure about that? Can you double check the manpage because I thought that {Z}, but I'm not totally sure.
  • It's okay if you dont know. If you need the manual for {command} or additional reference material, I can provide that.
Another important note is conversation management. If the thing starts hallucinating early on and making mistakes, scrap the thread and try again, or else it's likely to just keep on faulting. Adjust your opening verbaige to avoid the original errors. Conversely, I've found that threads can get into a sweet spot, where the AI understands the assignment.

Interested in what other tips some of you have found for improving AI/LLM experience. Personally I used Claude.

EDIT: I dont know why so many people interpret this to mean the exclusion of reading the manual and handbook for yourself. At no point did or would I ever suggest such a thing, and anyone who accuses me implying that, isnt operating on good faith.
 
Google's AI is kind of annoying, usually not exactly correct, but I'm not sure I want to spend the time teaching it FreeBSD. I might as well spend some time learning myself how to solve a particular problem in FreeBSD... 😩
1730909108869.png

Not to mention that ChatGPT has a drinking problem and a money problem, too. The datacenters that power it - they take a lot of water to cool the expensive computer hardware, y'know. Friends don't let friends compute drunk!
 
astyle said:
Google's AI is kind of annoying, usually not exactly correct, but I'm not sure I want to spend the time teaching it FreeBSD. I might as well spend some time learning myself how to solve a particular problem in FreeBSD... 😩

Not to mention that ChatGPT has a drinking problem and a money problem, too. The datacenters that power it - they take a lot of water to cool the expensive computer hardware, y'know. Friends don't let friends compute drunk!
Click to expand...
There's alot to unpack here.
  1. You're not "spending time" ..'teaching'.. it anything. You're uploading a manpage and maybe a config or terminal diagnostic. The point is to save time, which these tips do for me regularly.
  2. I guess you never use grep either to find information. You just read the whole manpage from start to finish every time you need to find an obscure option. LLMs are advanced linguistic search tools. Sometimes you're not quite sure what the canonical phraseology is for some question, and an LLM can help you find it.
  3. It sounds like you maybe have some kind of ethical or moral hangup on datacenters and power consumption? Kind of a weird complaint.
 
BawdyAnarchist said:
There's alot to unpack here.
  1. You're not "spending time" ..'teaching'.. it anything. You're uploading a manpage and maybe a config or terminal diagnostic. The point is to save time, which these tips do for me regularly.
  2. I guess you never use grep either to find information. You just read the whole manpage from start to finish every time you need to find an obscure option. LLMs are advanced linguistic search tools. Sometimes you're not quite sure what the canonical phraseology is for some question, and an LLM can help you find it.
  3. It sounds like you maybe have some kind of ethical or moral hangup on datacenters and power consumption? Kind of a weird complaint.
Click to expand...
1. Somehow, it takes zero time to find and upload a manpage, and that does not qualify as 'teaching'. Y'know, when you upload a manpage, it's the same as providing teaching materials to someone. 'Teaching materials' is the very manpage you're uploading. Instead of you RTFMing, you expect ChatGPT to RTFM and learn FreeBSD? :rolleyes:
2. I do use grep to find information in a manpage. Well, I use Konqueror to read them, so it's really Ctrl-F...
3. I don't have an ethical/moral/whatever hangup on datacenters and power consumption. I do have a problem with over-selling of AI at the expense of promoting critical thinking inside people's brains.
 
astyle said:
Google's AI is kind of annoying, usually not exactly correct, but I'm not sure I want to spend the time teaching it FreeBSD.
Click to expand...
Well, I'm quite sure I don't (unless properly paid).

astyle said:
3. I don't have an ethical/moral/whatever hangup on datacenters and power consumption. I do have a problem with over-selling of AI at the expense of promoting critical thinking inside people's brains.
Click to expand...
The critical thinking case seems to be lost already.

I am always wondering, when wikipedia tries to squeeze money out of me, why this is actually some managers behind wikipedia entertaining their greediness, while the actual work of writing wikipedie is done by people who get nothing. So it is actually a scam: money is collected, apparently for the good information in Wikipedia, but which in fact is created by unpaid volunteers.

Apparently the over-selling AI now uses that same scheme: let some internet-users do the teaching work for the product that can then be sold. And it seems to be a general characteristic of the internet nowadays, to exploit the labour of the millions of internet-users in order to then profitably sell the result.
 
PMc said:
I am always wondering, when wikipedia tries to squeeze money out of me, why this is actually some managers behind wikipedia entertaining their greediness, while the actual work of writing wikipedie is done by people who get nothing. So it is actually a scam: money is collected, apparently for the good information in Wikipedia, but which in fact is created by unpaid volunteers.
Click to expand...
In case of Wikipedia, somebody needs to run the physical datacenters and computers where it's hosted, because rank-and-file users cannot exactly do that.

In case of AI - some of that capacity is actually moving to individual devices, rather than data centers. As in - a late-model phone can translate English dictation into Chinese kanji in real time, even if you turn off the wi-fi AND gsm antennae.
 
  • Like
Reactions: drr
Not only do I agree with above but consider Wikipedia legal costs.

You know some rich folks don't like what posted about them and sue.

Somebody has to show up in court even in frivolous cases.
 
  • Like
Reactions: drr
astyle said:
1. Somehow, it takes zero time to find and upload a manpage, and that does not qualify as 'teaching'. Y'know, when you upload a manpage, it's the same as providing teaching materials to someone. 'Teaching materials' is the very manpage you're uploading. Instead of you RTFMing, you expect ChatGPT to RTFM and learn FreeBSD? :rolleyes:
2. I do use grep to find information in a manpage. Well, I use Konqueror to read them, so it's really Ctrl-F...
3. I don't have an ethical/moral/whatever hangup on datacenters and power consumption. I do have a problem with over-selling of AI at the expense of promoting critical thinking inside people's brains.
Click to expand...
1. It takes all of about 3 seconds to man cmd > man_cmd.txt, so okay not infinitely fast, but pretty quick. I'm not drawing on a chalkboard or writing a dissertation.

Moreover, I have looked very hard, but can you please show me where I suggested doing this in place of reading the manual or Handbook for yourself? If you jumped to that conclusion, then maybe that moreso confers something about your mentality rather than my own.

2. Yes ofc. As do we all. LLMs are an advanced linguistic search, and for complex topics. Whether it's coz I'm green to a new subsystem/component and need help with bridge gaps in general Unix understanding, or I've merely forgetten the syntax/options of a command_config I use 1-2x per year, LLMs can be used like an advanced search function that save me time.

3. I'm not overselling anything. You'll notice that my post explicitly acknowledges some of the problems you might experience with LLMs. It's a post dedicated entirely towards helping you turn what might be a lousy and limited LLM experience into a good one.
 
Phishfry said:
Not only do I agree with above but consider Wikipedia legal costs.

You know some rich folks don't like what posted about them and sue.
Click to expand...
Okay, never mind. I'm not rich and I neither do like what is posted on wikipedia.

Now I'm wondering, why should we pay other people for sueing each other about having some different opinion on wtf, while not having enough food for ourselves? Sometimes I think I'm in a madhouse. Don't we have real problems?
 
BawdyAnarchist said:
1. It takes all of about 3 seconds to man cmd > man_cmd.txt, so okay not infinitely fast, but pretty quick. I'm not drawing on a chalkboard or writing a dissertation.

Moreover, I have looked very hard, but can you please show me where I suggested doing this in place of reading the manual or Handbook for yourself? If you jumped to that conclusion, then maybe that moreso confers something about your mentality rather than my own.

2. Yes ofc. As do we all. LLMs are an advanced linguistic search, and for complex topics. Whether it's coz I'm green to a new subsystem/component and need help with bridge gaps in general Unix understanding, or I've merely forgetten the syntax/options of a command_config I use 1-2x per year, LLMs can be used like an advanced search function that save me time.

3. I'm not overselling anything. You'll notice that my post explicitly acknowledges some of the problems you might experience with LLMs. It's a post dedicated entirely towards helping you turn what might be a lousy and limited LLM experience into a good one.
Click to expand...
1. I don't want to do the extra clicking in a browser to upload a manpage or a config file. Pay me to do that. Besides, doesn't ChatGPT charge users for access to that feature?
2. There's no substitute for spending a bit of time to internalize a topic inside your own brain. Yeah, that includes figuring out which terms to use in a search. And you just might discover that the solution to the puzzle is stupid simple, but no, not what ChatGPT is suggesting.
3. I'm not accusing you of over-selling... When I said I have a problem with overselling of AI at expense of promoting critical thinking, I'm talking about how mainstream press presents AI to the general public. The downsides of AI get very little attention in mainstream press, one has to be smart enough to figure out the downsides for themselves - just like you figured out how ChatGPT is skewed towards Linux-flavored answers, and that it takes some time and brain power to tease out an answer that applies to FreeBSD.
 
astyle said:
1. I don't want to do the extra clicking in a browser to upload a manpage or a config file. Pay me to do that. Besides, doesn't ChatGPT charge users for access to that feature?
2. There's no substitute for spending a bit of time to internalize a topic inside your own brain. Yeah, that includes figuring out which terms to use in a search. And you just might discover that the solution to the puzzle is stupid simple, but no, not what ChatGPT is suggesting.
3. I'm not accusing you of over-selling... When I said I have a problem with overselling of AI at expense of promoting critical thinking, I'm talking about how mainstream press presents AI to the general public. The downsides of AI get very little attention in mainstream press, one has to be smart enough to figure out the downsides for themselves - just like you figured out how ChatGPT is skewed towards Linux-flavored answers, and that it takes some time and brain power to tease out an answer that applies to FreeBSD.
Click to expand...
Suit yourself. Maybe you're a superstar Unix pro that basically doesnt need an advanced linguistic search tool. Or maybe for you it's all about "the journey," of the slowness and pain of being an official-docs-only purist.

As for me, I'm happy to ...

save 5-10 minutes reacquainting myself with a command_config I use once per year and always have to look up the syntax/options; which AI typically gets correct in 5-10 seconds,

have it double check my scripts/configs for errors that shellcheck and/or my dyslexia might miss,

shave minutes to hours troubleshooting complex configs/scripts that arent working as intended,

reduce time to proficiency for a component or subsystem,

ask pointed specific questions that I would spend literally hours scrolling forums, man pages, and google search,

get quick feedback to on what error messages might mean, and ideas on what to troubleshoot next ...

avoid bothering the SmartGuys with a help! post borne out of a syntax mistake or slight misunderstanding of an ambiguous sentence in a manpage,

Again, I am not suggesting that AI is a substitute for a thinking brain or deep understanding of the system. This is not an either/or thing where you can ONLY read man pages or ONLY use AI. It's both/and.

Yes I pay for access to an actually capable LLM (Claude sonnet). The older versions and the free versions are marginal for things beyond simple queries. I'm not advertising that you do the same, but personally, I'd probably pay 5x or more simply because of the time it saves me.
 
I spend too much time noting to these AI bots, "What you said sounds more like a commercial promoting said product." To which it replies, "You're right! I'll rewrite that so it has less bias."

Or "I thought we weren't supposed to do it that way" which returns, "You're right! I stand corrected."

This morning it replied, "I didn't really know how that worked."

One cannot trust them any more than one would trust results of search engines which, essentially, is what these are.
 
drhowarddrfine said:
I spend too much time noting to these AI bots, "What you said sounds more like a commercial promoting said product." To which it replies, "You're right! I'll rewrite that so it has less bias."

Or "I thought we weren't supposed to do it that way" which returns, "You're right! I stand corrected."

This morning it replied, "I didn't really know how that worked."

One cannot trust them any more than one would trust results of search engines which, essentially, is what these are.
Click to expand...
For sure these are persistent problems with AI. The idea of the post was to reduce those kinds of problems, so that you might get something faster and more useful. Upload the manpage(s), your config(s), simple diagnostics, and short simple instructions.

You might still have to go through some iterations, depending on how arcane and complex the task is, but used properly and cautiously, can be a big time saver.
 
Well, this is how the AI bots are coming for the jobs that require brainpower. The training datasets and LLMs get so large that eventually, they have to specialize on a topic. Like FreeBSD, medicine, law, whatever. Like a chatbot that can search the PubMed.gov archives... it already costs money to curate those bits of specialized knowledge so that they are machine-redable. Specialized knowledge is a money-maker, so it's more valuable than general-purpose knowledge. Oh, and don't forget that there's university degrees that cost money to obtain, and then it costs money to employ those with specialized knowledge... you get the drift in context of this thread?
 
Wikipedia can be filtered out by appending "from academic sources" to an AI query. Search engine operators (eg site:freebsd.org) are also honored.

In my case, AI acts as an amped-up search engine filter.
 
You must log in or register to reply here.
Back
Top