Let's talk about AI
-
Out of curiosity - I have two questions to people who use AI for coding daily. I'm kinda interested in how that works for you:
- As someone who iterates on code a lot, tests, rewrites, measures, rinse and repeat all day long - I'm kinda curious how do you make that work with the costs of LLMs. Does it impact how you think about designing stuff, e.g. "oh, I'm not gonna try this or that because it's gonna take too many tokens/requests". Do the costs sit in your mind at all when you code or do you just pay up whatever your usage amounts to? Do you use it for small stuff or large design of whole systems? Do you make one or two requests and then take over? From what I see, depending on service and plan it's usually a couple hundred requests a month. Thinking about how I code it would probably last me about a day or so. How do you deal with that? Do you just don't code as much? Do you have private local models that are any good?
- Since all those requests pass your code to 3rd party services do you have any company policies about that? What about paid and patented algorithms that you use. What about "secret files" in your projects, like crypto keys, serial numbers etc? What about closed source license code you work on - are your companies ok with sharing those things with 3rd parties for them to learn on for free and with no legal framework? What's the legal and security side of your AI usage is probably what I'm trying to ask. Or do you just don't think about it until there's actually a problem.
-
@Chris-Kawa:
To make this clear: I don't use AI tools for professional coding.
I use them around it though. Bisecting and pinning a bug to a potential lines of code is something where AI is really good at.
Fixing the bug in an atomic manner, eliminating the exact root cause, without wading neck deep in regressions - that is something where I claim to be way better than the Claude's of this world.
I tried it to make something work very quickly and explore how I feel about it.
I got something done fast, that is really helpful and it would have taken me a lot of time otherwise.
Won't qualify for any beauty contest. But I learned a lot. -
That's also a bit in the nature of the beast, isn't it?
If you break down a desktop application into pieces, just about everything was done many time before. So LLMs chew up whatever has attractive robustness scores.
New features of a framework, or its long standing bugs are unique. Never done before, never fixed before. -
@JoeCFD Are you thinking about the Qt AI Assistant ?
-
@J.Hilk said in Let's talk about AI:
Funny enough though, I already use AI instead of googling things myself – just ask it to find me relevant links rather than wading through pages of SEO nonsense. I'm sure I'm not the only one.
I find it harder every single time I try to use Google for something more specific (not just googling a celebrity or city or movie or song) to get good results. It's been a very long time that you could put in
ANDandORinto your search. And even adding double quotes or+does not work reliably anymore. An LLM is the much better alternative to find something on the internet. BTW, the internet nowadays does forget–in many cases you need to know the original website and dig it up on the wayback machine.@JKSH said in Let's talk about AI:
I've been using https://www.ecosia.org/ for the past year and I've found that its search results are comparable to Google's.
If that is true, then I don't need to try it. Google is bad and this statement says, that ecosia is as bad as Google. I need a "retro" search engine that works in the way they did in the '90s/'00s. Don't try to outsmart me!
@J.Hilk said in Let's talk about AI:
AI makes a lot of things more convenient, but honestly I haven't encountered anything it does that I couldn't live without.
The problem is that things that have worked fine in the past have been replaced by AI. One very noticeable thing is spellcheck in Word: It used to be rule-based/dictionary-based and it worked in many cases and you'd understand why it had occasional problems (I'm speaking of German spellchecking). Most recently the suggested corrections aren't even words (and no suggestion is anything plausible). So, we replaced something that worked reasonably well in all cases with something that fully fails regularly. Why don't people understand that AI is not the solution to every problem?
@JKSH said in Let's talk about AI:
So has a movement pushing for reversal: https://quitgpt.org/ (Well, it's more a push towards alternatives, but hopefully that slows down the enshittification process of all AI at least)
I have recently learned about https://chatjimmy.ai/ . They are putting LLMs into silicon (on their website this is currently quantized LLama). Having the LLM inside a chip makes it really, really fast and uses a lot less power.
@SimonSchroeder said in Let's talk about AI:
@JKSH said in Let's talk about AI:
I've been using https://www.ecosia.org/ for the past year and I've found that its search results are comparable to Google's.If that is true, then I don't need to try it. Google is bad and this statement says, that ecosia is as bad as Google. I need a "retro" search engine that works in the way they did in the '90s/'00s. Don't try to outsmart me!
I certainly wasn't expecting anyone to interpret my words that way 😅
I was trying to say this: Everything that I needed from Google Search, I could get from Ecosia (including double-quotes, ANDs, etc). No sacrifice was required for the switch -- I didn't miss anything after switching -- so it was a no-brainer.
@JonB said in Let's talk about AI:
@SimonSchroeder said in Let's talk about AI:
It's been a very long time that you could put in AND and OR into your search. And even adding double quotes or + does not work reliably anymore.Very true! :( Modern AI ~= Google with ANDs, ORs and quoting ;-)
AI is great for complex research: ANDs, ORs and quoting PLUS organizing info/links on multiple sub-topics in a coherent way. That last part is a big time-saver.
For simple searches though, Ecosia still has the ANDs + ORs + quotes. I'd rather scan through a list of candidates and pick one myself. Not looking for an exposition on the search results.
-
Out of curiosity - I have two questions to people who use AI for coding daily. I'm kinda interested in how that works for you:
- As someone who iterates on code a lot, tests, rewrites, measures, rinse and repeat all day long - I'm kinda curious how do you make that work with the costs of LLMs. Does it impact how you think about designing stuff, e.g. "oh, I'm not gonna try this or that because it's gonna take too many tokens/requests". Do the costs sit in your mind at all when you code or do you just pay up whatever your usage amounts to? Do you use it for small stuff or large design of whole systems? Do you make one or two requests and then take over? From what I see, depending on service and plan it's usually a couple hundred requests a month. Thinking about how I code it would probably last me about a day or so. How do you deal with that? Do you just don't code as much? Do you have private local models that are any good?
- Since all those requests pass your code to 3rd party services do you have any company policies about that? What about paid and patented algorithms that you use. What about "secret files" in your projects, like crypto keys, serial numbers etc? What about closed source license code you work on - are your companies ok with sharing those things with 3rd parties for them to learn on for free and with no legal framework? What's the legal and security side of your AI usage is probably what I'm trying to ask. Or do you just don't think about it until there's actually a problem.
@Chris-Kawa said in Let's talk about AI:
Do the costs sit in your mind at all when you code or do you just pay up whatever your usage amounts to?
From what I've seen so far (I'm not an AI user myself) people pay whatever amount it is. It's usually something like $200 per month. Compared to an employees wage it is basically negligible (especially if it really saves time).
I mostly share your concerns which is why I don't really use AI for this kind of work. One additional concern is the technical dept you are introducing by using AI. You'll end up with a system that can only be programmed any further with AI. But, for this you need to hope that AI evolves fast enough that it'll be able to do this job in the future because it is certainly not up for the task right now. "Best case scenario" everybody learned to rely on AI and doesn't know how to code themselves and we'll get high paying jobs because we are the only ones left to be able to fix it (although I'm gonna think hard how much pain I'd be willing to suffer through when fixing these programs).
-
@SGaist Yes. I am using other tools. I do not know Qt AI assistant. Good to know. Thanks. Will try it out.
@JoeCFD Note that Qt AI Assistant is ...:
Qt AI Assistant is available only for premium commercial Qt developer license holders. For more information on licensing, check Qt pricing on the qt.io web pages
If you are on the open-source Qt Creator you can try out:
The latter is also available in the Qt Creator's Extension's pane.
-
@JoeCFD Note that Qt AI Assistant is ...:
Qt AI Assistant is available only for premium commercial Qt developer license holders. For more information on licensing, check Qt pricing on the qt.io web pages
If you are on the open-source Qt Creator you can try out:
The latter is also available in the Qt Creator's Extension's pane.
@cristian-adam Thanks!
-
@JoeCFD Note that Qt AI Assistant is ...:
Qt AI Assistant is available only for premium commercial Qt developer license holders. For more information on licensing, check Qt pricing on the qt.io web pages
If you are on the open-source Qt Creator you can try out:
The latter is also available in the Qt Creator's Extension's pane.
@cristian-adam isn't the "legacy" copilot plugin also available with Qt Open Source? I believe that's what I'm using for basic auto-completion.
-
@cristian-adam isn't the "legacy" copilot plugin also available with Qt Open Source? I believe that's what I'm using for basic auto-completion.
@GrecKo said in Let's talk about AI:
@cristian-adam isn't the "legacy" copilot plugin also available with Qt Open Source? I believe that's what I'm using for basic auto-completion.
That's still available. You need the copilot access though. And it's just code completion, no chat, no agents.
