Resources to teach an old dog new tricks (ai)
Posted by Andrew64467@reddit | ExperiencedDevs | View on Reddit | 18 comments
I’ve been building software since 2000 so I’m probably not too bad at designing and building software, although I’ve never been lucky enough to work at FAANG etc.
Someone posted a study here that claimed a 10% improvement in productivity when using ai for coding. My personal experience using GitHub copilot for autocomplete was that it contributed almost nothing to my productivity. It basically helps with the very easy things which don’t take too long anyway and that’s about it. I long ago found ways of doing repetitive tasks quickly and the codebases I’m working on are designed to avoid loads of repetitive boilerplate.
Now it would be nice if this means that ai is all hype for development and I can safely ignore it, but I’ve always had a motto “be careful believing a fact that you want to be true”.
So can anyone point me at some serious resources or tutorials I can use to try and improve my ai usage for development? I want to try as hard as I can to disprove the hype theory for myself.
stevefuzz@reddit
I've been developing since like 1998. The boilerplate autocomplete definitely adds up. I'm a fast typist, and hitting tab is faster. As for all the hype, you're not crazy. Stuff that is obvious to you is like an alien language to a beginner. So, AI coding is like magic. It's not nearly as impressive in a professional environment when you constantly have to review and fix a bunch of AI slop code. Also, it's good as a Google search upgrade or method name brainstormer.
Laicbeias@reddit
im a dev for 26 years. i dislike copilot. usually i just copy together all pieces needed and tell the AI what to do. trick is to define its behaviour and to use claude.
if you know what it can do and cant. you can finish tasks that would take you 3 hours, in 25 minutes (mostly boilerplate or simple services, some sql queries). small scope stuff.
if your task is to navigate a huge ass code base and you have to jump around a lot to find a / b / c. it is useless.
with AI you can basically speed up a lot of things. for example yesterday i started writing a hot reload system for c. used tcc as a compiler. made a cross platform file watcher in that system. implemented error handling with line numbers, even for these reloaded dlls. and now can write my own super set of c with hot reload (i got interested in writing my own language as a side project). 3 hours without any prior c knowledge. so yeah if you know how, AI can speed you up a 10 fold.
So ai isnt hype it is the real deal in iteration and search
(and i now feel old since i .. program 1 year longer than you)
DeterminedQuokka@reddit
So I can get a lot out of copilot but I do have to use it specifically. I don’t think you get a ton out of asking it to do things. But I think you can get a lot out of autocomplete. Like I write a test then I write the name of the next test and copilot auto completes it.
Also really simple functions like flatten_list.
Or the value of a variable. It saves me time typing basically.
fschwiet@reddit
I'd download Cursor and Windsurf to give those a try. I don't use them as a primary IDE, I'll switch to them when I want to run a coordinated change then review what they're doing with source control. I've mainly used Cursor, and found it great for working with stuff I'm not familiar with (some examples: generating Github workflows, converting a powershell script to bash, converting a page from using one css library to another).
Significant-Leg1070@reddit
It’s incredible for anything front-end related. I can explain in plain English what I want to accomplish and it will create 90% of the idiosyncratic angular, components, template widgets, services, etc. with all the (insert specific angular version) syntax and wire it all up. Then I go in and surgically manipulate the last 10%
It’s amazing for getting from blank page to something working seconds later.
I also use it to summarize a class or some insane nested if statement. I ask for trade offs/pros and cons for particular design decisions.
I almost never use it to make sweeping blanket changes on existing legacy code. At most I will have it refactor a complex method into separate more maintainable methods. If I need changes to existing code I always generate a suggestion and then pick and choose what code I bring into the code base.
Andrew64467@reddit (OP)
I must be honest, most fronted comes under trivial stuff. Generating html is pretty simple. I think most full time frontenders just make stuff more complicated cos they are bored
Significant-Leg1070@reddit
Agreed, it’s not anything ground breaking but allow me to clarify my point:
I can type up an angular component and switch back and forth between the docs, etc. but that will take some time.
I can also just pop a plain English request into Claude and have the very same component generated for me in SECONDS bug free with all the idiosyncratic Angular-way to do things.
That’s true time savings and true productivity gains.
I can now spend my time focusing on the harder problems.
Not sure if you are agreeing or disagreeing
Andrew64467@reddit (OP)
I will try this next time I do some react. I’ve not done much frontend of late.
I think though, unless I have an incredibly complex frontend requirement I’ll be avoiding the whole angular and react thing. I think we’re reaching a tipping point where server side html generation in strongly typed languages is going to basically replace all the JavaScript madness as it can generate ui that is just as good in a fraction of the time
ScriptingInJava@reddit
Personally I found both Copilot's autocomplete and chat/suggestions mostly garbage. The autocompletes were typically just a distraction and the chat functionality is massively outperformed by the free tier of ChatGPT.
Copilot was very good for generating tests for me that I would just neaten up afterwards, try this out (assuming you're using an IDE Copilot plugin):
PragmaticBoredom@reddit
With a paid Copilot subscription (or trial) you can choose different models, including OpenAI ChatGPT models.
The ChatGPT free tier will definitely not outperform Copilot on a paid/trial subscription. You could pick the same or better OpenAI model in Copilot, even.
ScriptingInJava@reddit
I’ve picked the exact same model and asked the same question to both systems at the same time, with a paid Copilot license.
They gave different answers, Copilot was relatively unhelpful and ChatGPT was mostly right. It’s entirely anecdotal but still.
The default model on ChatGPT is tweaked though no? It’s not the public LLM model you can run through Ollama I’m fairly sure
PragmaticBoredom@reddit
The default model on ChatGPT is tweaked though no? It’s not the public LLM model you can run through Ollama I’m fairly sure
Ollama is for running local models. OpenAI doesn’t release their models for local use, so ChatGPT doesn’t have anything to do with Ollama.
Your comment about said “free tier of ChatGPT”. The free tier of ChatGPT does not have access to their best models, which are miles ahead of the free tier for coding.
Andrew64467@reddit (OP)
Interesting. I will definitely try this
ScriptingInJava@reddit
It’s worked well in a variety of scenarios, even with integration tests where I’ve created a load of helper extension methods on custom Fixtures. It’s picked them up and used them correctly, I’d assume because I’ve got a lot of tests manually written doing so as well
tetryds@reddit
I have the exact same experience. Yes it gets the code written faster but it does not speed up the time it takes to get the solution shipped and may even delay me sometimes. I'm striving to see if I can find a workflow with net benefits but nothing so far.
nshkaruba@reddit
I tried and have the same experience as you, I'm ignoring it for now. I still use chatgpt as a search engine that basically generalizes info from multiple google search pages into my answers
De_Wouter@reddit
I'm not a big fan of generating too much code with it and just using it as is. I've seen it make small, hard to spot errors way too often.
But what it is really useful for is, like a "more advanced Google search" where instead of asking a simple question, you can ask it "combined" questions. "I use X and Y library together, how do I do Z combining the 2?" (I know kinda stupid example, but you get the point)
Also handy for learning purposes for non-beginners. For beginners it might be bad, because they can't spot the lies and hallucinations that easily. But for more experienced developers it can be pretty handy. If I need to learn a new tool, language, framework or whatever, tutorial are often so tyring because they are mostly aimed at beginners. LLMs can give you information a bit more tailored on your level.
Wang_Fister@reddit
I'm also a huge fan of using it this way, especially the copilot chat extension in VS code. For instance at work I'm learning Vue (and front end) as well as getting acquainted with a pre-existing codebase, it's super handy to be able to provide context using the existing files and ask something like, "I want to do X, what's the best way to go about it? " Or "I'm getting this error/this stupid fucking div has too much margin whyyyy" Then parse the output and ask further questions.