Best AI for software development and sysadmin scripting?

D

Daveis

Inspired
I’m looking for AI to aid me in PowerShell and C# and database work. I have CoPilot at work but I need something better that doesnt suck. My budget is under $300 year. Any recommendations?
 
Anthropic is getting good press from Gizmodo and Axios about the latest release of Claude. Whether it supports PowerShell and C# I dunno. SQL is probably supported.

I use Warp as my terminal and it has a bunch of different AIs (including the latest Claude) it can call upon when it senses a code-crash. I let Warp automatically decide which AI is the best choice and t's interesting to watch whichever one diagnose the problem and see its recommendations. They've gotten much better in the last year. And it's useful when I'm not sure what path I want to take to do something, especially when I remember there's a way to get from here to there but can't remember the actual command at the terminal level. And sometimes I like to f*ck with it and tell it to try again using a different method because I already know how I'd do it and have code that works really well, or I want it to optimize or give me an idea what it thinks are the best practices would be.

I still won't trust it to do something automatically though. I always dig into the code to make sure it won't explode or eat my files.
 
Last edited:
We're mandated at work to use CoPilot... And it's ok.

Except when it's not. Often it's more like a semi-smart search engine "consolidator".

I've heard Grok is pretty good for IT related stuff, but have no experience with it.

My personal experience so far is that you definitely should already be experienced to a certain degree before using these AI tools because otherwise you're going to get sent down rabbit holes chasing incorrect information.
 
Grok for single scripts of almost any kind, SQL, Splunk SPL queries, Python, including Django, I like it better than ChatGPT but this one is also good. Claude Code has an excellent reputation to handle full projects but it may go above your budget.
 
Google feeds you Gemini responses when you google the usual stuff.

If you know what you’re doing this is a massive timesaver of unbelievable orders of magnitude. What used to take a week can take only a couple of hours.

If you don’t, well, this:
1770414439169.gif
 
I've had good experiences with Claude, on my limited experience. For sql and other very targeted questions i asked, i got back code that just worked, faster than i could have typed it if i knew had it fully formed in my head, which i didn't. I frequently build up solutions in layers, making sure the pieces are doing what i mean them to do before adding on and integrating them. Claude just went straight there.

I didn't leave a lot of room for it to fantasize, and i can recognize bs when i see it and test it. I didn't even have to tell it "be accurate" like i did to keep chatgpt from just making shit up, pffft.
 
I have been using Copilot and find it goes on tangents I didn’t ask for. I usually give it 10-20 numbered rules at the very beginning. But over time it reverts to doing whatever it wants. It’s given me a lot of code that is missing closing braces.
 
If someone tells you they are vibe coding what does that really mean?

The warp terminal only uses their builtin AI?
 
unix-guy said:
Often it's more like a semi-smart search engine "consolidator".
Click to expand...
^Zactly^

SOOO much of the time they're regurgitating some hodgepodge of various search results using who knows what algorithm to determine their suitability.

I've seen some interesting code reviews come from them but haven't had a chance to vet those. :-/
 
Daveis said:
The warp terminal only uses their builtin AI?
Click to expand...
No, it has a good-sized list of AIs and we can choose or it can make a determination based on their assessments of the various AI's abilities.

I think it's somewhat of a crapshoot. But I spent some time messing with the latest version of Claude this morning having it go through some Python code I'd written and its assessments made sense. I haven't had a chance to run the code to see if they will make the code go out in flames yet, but at first and second glance the suggestions were good.
 
Dave Merrill said:
I frequently build up solutions in layers, making sure the pieces are doing what i mean them to do before adding on and integrating them. Claude just went straight there.
Click to expand...
I think that's the difference between how we humans code vs. an AI.

We're designing, inventing, on the fly a lot of the time. The AI has the benefit of looking through a lot of code we, and others, have written and can integrate and combine it.

WE have a pretty good idea why we're writing what we do. I really doubt that AI does, at least currently, but it can find what is used most often and, because that algorithm has the highest hit rate it "must be best". Statistics basically. ¯\(ツ)/¯

And, of course, next year a better algorithm might show up that blows the doors off what is the accepted best path to solve a particular problem, and AI won't recommend it until enough "statistics" show that the new way is better than the old. That is the next frontier, for people to explain to the AI that the goal is to solve problem X and let it chew on that. There are lots of problem X out there to keep it busy for a long time, but once it gets past regurgitating old code and actually creates new algorithms then we'll be seeing something.
 
M365 Copilot just removed 200 out of 1050 lines of code from a PowerShell function. it then argued with me that it hadn’t. It lies and hallucinates. Unfortunately its the only AI my work allows me to use and its junk. Microsoft product cant tell me how to easily load MS SQL drivers using Powershell 7 on Windows 2019 Server. It took two hours for it to make a usable suggestion. That or Net8 Core stuff is garbage.

Maybe I should switch to Python or Java?

Also earlier today couldnt get it to write AD Group query for a single domain. Idea was to list all local administrators. I switched from Powershell AD module to “net group /domain” commands which worked well.
 
Absolutely none. Don't use it to write code for you. Don't use it to do anything. If you see somebody using it, make passive aggressive remarks about their ethics and question their intelligence. Be especially hostile to forum users who try to copy-paste slop at you. Scroll as fast as possible away from generated "content", don't engage with it in any way at all other than possibly reporting it. Make using "AI" as socially unacceptable as public indecency.
 
Daveis said:
M365 Copilot just removed 200 out of 1050 lines of code from a PowerShell function. it then argued with me that it hadn’t. It lies and hallucinates. Unfortunately its the only AI my work allows me to use and its junk. Microsoft product cant tell me how to easily load MS SQL drivers using Powershell 7 on Windows 2019 Server. It took two hours for it to make a usable suggestion. That or Net8 Core stuff is garbage.

Maybe I should switch to Python or Java?

Also earlier today couldnt get it to write AD Group query for a single domain. Idea was to list all local administrators. I switched from Powershell AD module to “net group /domain” commands which worked well.
Click to expand...
Maybe try github copilot (if you're confined to MSFT), M365 copilot is focused on Modern Work products like Teams and Outlook, not writing code.
 
KevinP said:
Maybe try github copilot (if you're confined to MSFT), M365 copilot is focused on Modern Work products like Teams and Outlook, not writing code.
Click to expand...
Took me a single code reviews to identify GitHub copilot vibe coding.

It’s not just plain bad for production. It’s not even good for parsing text files in development.
 
google.com free AI is frequently better at coding than M365. I need to try Claude. I have not heard if anyone has admitted to how they obtained training data for their models.
 
You must log in or register to reply here.
Back
Top Bottom