New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Which AI tool has actually saved you the most time as a sysadmin or developer?
ReliableSiteHosting
Member, Patron Provider
in General
Which AI tool has actually saved you the most time as a sysadmin or developer?
Comments
None (since i didn't use any).
He wasn't asking luddites.
Oh, i guess i somehow missed the part where @OP limited participation to curmudgeons. My deepest apologies.
Kind of funny @OP is an actual account
Haha, yeah. If he ever comes back the amount of notifications from all the tags he got over the years is likely to make his browser explode
Yeah, but you're not a very capable dev or sysadmin if you've failed to make so many useful tools available to be productive for you. All the best devs I know are constantly learning new things and using new tools to get a whole lot more shit done than those who have not. You're trying out to be a dinosaur.
I am what i am and i'm content with that
Have we considered that just because a new tool exists doesn't necessarily mean it is better? The whole field of computer science has an embarrassing tendency to not only fail to learn from its mistakes but also to fail to learn from its successes. I find that sticking with boring, known-good technology is the most underappreciated yet successful way to complete most tasks, and when it does fail, I can comprehend its failure modes. AI tools are the opposite: they subtly fail you all the time and have incomprehensible, potentially unknowable failure modes.
To answer OP's question, I find that the only actual use I've found for gen AI is that it's good to run some local model in your CI/CD to look for bugs/style improvements you may have missed. Doesn't cost a lot of computation, no subscription, and it's easy to quickly tell if it's wrong (because you always write good tests to go with your code, right?). Far too unreliable at present to trust it for anything else.
I wouldn't claim to be an "A" grade sysadmin, (perhaps C-) but I do run a small hosting empire. I also absolutely love AI. I am always trying to figure out what it can help me with. Based on my experiments, I think the hierarchy is ChatGPT > Claude > Grok. Now, I freely admit it is pretty easy to stump ChatGPT with hard problems. But, to me, it is amazing how many problems it has solved for me. BTW NEVER ask Grok to help you with Server 2022 problems, ESPECIALLY database problems. It's advice sucks.
It's kinda funny now you say it. I actually had a long comment typed out on how and why i feel AI might not be as great as the hype may suggest (well, at least in relation to programming when you aren't in a position where maximum volume at any cost is the prime concern - as far as administration is concerned i haven't seen much practical use outside the better Google approach @MTUser2012 describes yet) but then i remembered that the poster in question usually tends to be a little confrontational and skipped on the potential hair splitting contest.
The TL;DR was that it's mostly yet another layer of abstraction (even if it's an exceptionally thick and opaque one) and that human language in general isn't very good at describing code (all those human-like-programming-language projects over the years turned out to be total clusterfuck failures - why would the other direction suddenly turn out great?) forcing people to paint in very broad strokes since trying to control details rapidly decreases the actual efficiency until the gain becomes negative at some point.
It's worse than a new abstraction layer. A new abstraction makes certain invariants that are helpful and always true: when I program in Python, I can trust that Python will free up memory for me reliably and safely, which saves development time. When I ask AI to write some code for me, there's NO guarantees, which makes me quite cautious to adopt it. It might always work for small things, but I know that I can't always trust it to maintain any invariant at all. Maybe it'll do what I want it to, maybe it won't, there might be some level of control with careful prompting but it ultimately devolves to a roll of the dice past a certain code/system complexity level. Look at the recent (partial) Bun rewrite in Rust done by AI, that's 700k lines of code no one human will ever be able to reason about and it's constantly failing to work at random...
I agree with you that human language is not precise enough to actually describe code (which is why writing good technical documentation is such a hard skill to master). I've seen many colleagues try to use AI for everything and it really makes me question if all this stuff is just performatively productive since it's probably faster to just write the damn thing yourself rather than coax Claude into doing it for you with a novel-length description of the task.
There are so many good ones. Among the main services there isn’t a single one I can say has never helped me. And as for which one helped me the most, or saved me the most time, the answer would naturally be biased in favor of the oldest service, which is Chatgpt.
Yeah, exactly. That's pretty much what i meant by broad (and as you said likely rather generic) strokes. I'm pretty positive that telling an AI to write a calculator in language X, which kinda looks so-and-so, will produce an acceptable result. Asking it for example to write an efficient linked list implementation in C would likely produce a mixed bag though. I don't doubt that AI knows what a linked list is and how to come up with something that satisfies the efficiency constraint but then there's also at least 1000 little details in regards to something like that, which have no definitive right or wrong, and getting it to take those into account will end up with the novel you mentioned.
If it's even possible to describe them precisely enough for the AI to understand the request and the AI is also capable of satisfying it. AI doesn't really write or understand code after all. All it does is recognizing patterns in the dataset it's trained on and trying to reproduce/combine them when told to. So once the descriptions divert from the obvious there is only a constantly shrinking amount of information to draw from and at some point the descriptions will likely have to almost be copy-pastes of singleton data pieces to be recognized (if those singletons existed in the dataset at all).
I figure the situations where AI is most effective involve either very rigid and mundane code and/or full projects so consistency/interoperability doesn't become much of a factor. In regards to the latter there is the 700k lines gibberish problem you describe though.
I guess if AI use keeps expanding in development it might end up playing out somewhat similar to how being able to write assembly mostly became extinct just in a much, much more drastic way. Programming in high level languages still mostly uses the same concepts as its low level counterpart even if the results might look foreign when coming exclusively from either this or that direction. AI removes the concepts completely. Sure, even if AI doesn't just hit a brick wall at some point a future where programmers for the most part just role play science fiction authors is still a bit off but not unimaginable given the human tendency to avoid tasks perceived as cumbersome.
I'm not really sure that's too much of a positive direction but it's not like companies are going to care all that much as long as it quickly produces something they can sell at an (at least superficially) reduced cost. I wonder how much of that excitement will last once they realize what AI succeeding would practically mean.
Claude, Gemini
Claude (Code)
I will write on behalf of the company, but about personal experience that is not related to us. I use Gemini when I'm stuttering in development and just can't get any further. Although I refer to documentation or forum posts, sometimes this is not enough and I connect Gemini, I can send him a piece of code and describe the idea of what should be next and he edits it to me or just points out the mistakes that were made. In general, using AI in development is correct at some stages and even greatly facilitates software development.
I use Claude for simpler tasks and I have to say that it has done amazingly well and saved me lots of time.
I'm not a software developer but sometimes I need a script or a simple piece of code to do a specific task. Usually it's things like migrate a database from one product to another or doing specific snmp queries to map a network or something like that. I am fully capable of doing these myself but it takes time, usually hours or even days. With Claude I just say "I've got a database that looks like this, I need all data converted and inserted into a database that looks like this and the api is here" and it takes literally seconds to whip out a script that works at first try.
Things that you execute once and never run again does not really impose a security risk so I am totally fine with AI generated code there, but I would be very hesitant to use it in anything permanent or internet facing without a proper and thorough code review.
My multi-AI 2nd opinion verification skills https://ai.georgeliu.com/p/deepseek-v4-in-claude-code-kilo-code + context7 MCP these allow me to use Claude Code and verify all coding/plans/implementations across several AI models for a more accurate understanding.
Antigravity