To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/DreamsofCode . You’ll also get 20% off an annual premium subscription.
There's a bias in it anyway even if it is not intentional (it probably is intentional). If you survey people who use copilot, you're going to get more responses from people that have stuck with it than people who have stopped using it. Most of the people I know that have used it, used it for a week, said this is not really worth using yet, and stopped using it, but the people who still use it are the people that liked using it.
Someone told me really early on to disable auto complete and bind the copilot suggest to a key command, it helps a lot with only using it when really needed.
This is what I do. It is then useful for deliberate use cases (like big chunks of repetitive boiler plate that I have already solved, but now need to repeat multiple times in other parts of the codebase - like DTOs, API Patterns etc.) - but I don't use the auto complete as it is too prone to suggesting bad ideas that I can't then unsee putting me in the wrong solution space for the problem I am solving.
That's great advice. I'm a data scientist. I do a lot of custom ML work where Copilot is rubbish. But I love it when I type ggplot( and Copilot creates a decent visualization of my data without having to think about it. The autocomplete just gets in the way, the hotkey is the better option.
having copilot with Golang is so nice to get around writing the if err != nil... every single time you call a function. Another thing I had recently was a huge Json schema from an aws api request which I wanted converted to a Go struct with serialisation and for that kind of stuff it can save you 20 minutes in like 10 seconds
I have been running copilot since very early access, and I used to do the same thing and wait. At some point, I started to ignore it and only used it for boilerplate code, for which I did not have a template. But lately, I have noticed that, at least in VSCode, where the copilot implementation is the most integrated, copilot will present its reply even if I continue typing as long as it is the same as what I typed. This means that there is no longer a delay for me, and it will usually present code that is in the direction I was heading anyway. I also no longer accept full replies and instead use the ctrl + right arrow to accept words, and this is the same shortcut as my "accept lsp prompt", as they are all marked as inline suggestions. I am a naturally slow/deliberate programmer so for me this way of working with copilot did speed up my work and made it drastically more enjoyable
I used AI tools like copilot and tabnine before and felt that once you start having to do more than actual boilerplate the ai just gets in the way and leads you down the wrong paths. I stopped and don't miss it!
Same here! The worst for me is when it makes a mistake so subtle that looks good when you read it, but gives a weird error that takes longer to debug than if I'd written it in the first place. I cancelled copilot, but still use Codeium for boilerplate
I've not found an AI that is actually good. It regularly will give me wrong information and when I explicitly tell it to not do something, it still does. I asked it, "Why do you answer so confidently when given a question with so much ambiguity, wouldn't it make sense to ask for clarification?" to which it responded, "I'm sorry let me clarify..."
i used cody... wich prevented me finishing my mini project 2 days faster... literally debugging whats wrong with the app, since LSP didnt gibe any errors eventually realized the ai was hallucinating and kept rotating and changing my .env variables .. and when use autocomplete it suggested wrong auto-completes... not the same names ive defined... took me a while to realize that lol.. not going to use any AI, if they are not reliable as auto-completers even... isnt there like claim that even the most efficient code ai assistant... gets it right only 17% time? seems like its just holding people back, not helping anyone... only positive use ive found for them is to select a block and auto-describe/comment what the block does... took me 2 days of using AI to realize, it will create a religious-type of blind trust into ai-complete... extra bad was that its personal project, so i didnt use typescript... the ai kept thinking im using mongodb for orm as well... when i was using actually some other db wich also has API like db.method() never used copilot, but is it better than cody (sourcegraph) ? ai pretty decent tho at "describe this block" , whats happening , if u check some foreign codebase and auto-commenting block of code , generally describes it pretty well
Yeah, I think Copilot is being pushed as this tool that will KILL programmers! Like no, it's good at the fundamentals and for learning concepts, but it can and will hamper your developmental progress.
I've always turned off the auto complete prompts. What really gives me value is Copilot chat. 90% of the time I use it to understand what the legacy code actually tries to accomplish and it does a decent job in most languages. The other 10% of the time I'm asking questions about how to structure a project in a language that I'm not particularly familiar with, or asking it to generate boilerplate code, where it does a decent enough job.
@@jamess.2491 my codebase has thousands of lines of code. There's no way outside of using the gpt4 api to give it that context and even with that, I think copilot does a better job
I like Copilot a lot when it comes to pattern completion. If I write functions to accomplish some task and then have a related task that isn’t quite the same it will generally outline new code in the same style as the previous code I wrote. If I write CSS for a red gradient of a specific shade for instance I’ve been able to autocomplete code for a blue gradient that matches the same saturation and lightness as the red one, and continue the pattern for green, yellow, etc. On solo projects I feel like Copilot gives me the power of two people. I just need to be prepared to guide it and review the code and not rely on it for anything too complicated.
it's just a fancy autocomplete often it would give a completely wrong answer but you'd still have to spend time reviewing the code only to find out it was incorrect
Except that it's much more useful. You can do an initial code review, check for errors and explore solutions. It helps you immerse yourself in a new code base and documents what's not... Copilot is therefore an excellent tool to complete your toolbox
@@MarthinusBosman i think this is a rather reductive argument. the point about Copilot being trained on older versions of code and using newer versions of libraries incorrectly still stands, whereas with traditional code completion it's usually based on analysis of the existing API. traditional code completion systems are also fundamentally different regardless; if they weren't, why would people use copilot in the first place? a code completion system that doesn't go further than the language server, static analysis, and the documentation doesn't try to solve problems for you, and it's much harder to write code that you don't understand with it.
I've taken it out as well, missed it in the beginning, but I'm happier without it now. I think it's a bit to much of a crutch, you are better of developing the "muscle memory" in the long run.
I'm an old fella.. I've seen many fads come and go. although AI have some value, right now it's over hyped and overused. and when it comes to software, I like minimal fast setups that help me DO THE JOB. I'm amazed by how some devs are willing to avoid actual learning and work using overkill solutions for the simplest of tasks. I knew from day one that AI is not for me but I think it can be very helpful in many situations.
Ai has been quite useful for me when I have to write code or modify something in a language I don't really know, like the other day I needed to convert a yml file to toml and I had pretty much no clue of the file structure of either but chatgpt could just convert the entire thing for me. Saved a lot of time and got a working answer. If I write in a language I know I tend to write my code myself and use chatgpt to refactor it or make it nicer and more readable + add comments
@@ZÏ̇̃ I do find it amazingly helpful in personal projects and filling minor gaps. but sadly some people are committing code they don't completely understand which is very dangerous and irresponsible. if you use it as a jumpstart and actually learn what it did. and preferably why and what other solutions it might have missed then power to you. but I hope you understand that the fastest answer is not always the best answer
I am also, I hve been writing code for over 25 years. That was then this is now. You will either use it and the productivity bost it provides or you will fall by the side of the road. I am a old dog but I am learning the new tricks.
The best experience I've had recently, was writing C on my job, where I'm not allowed to install anything. It was just a basic Vim (not neovim) setup and the C compiler. Showed me how little you actually need to have fun programming.
For me that’s not really fun. I mean, it is enough for sure depending on what you are doing. But ever since I first used the first IDE (Turbo Pascal in 1992) I don’t like working in a non-ide setting. Of course I think back then it could not do more than what vim could now. But I never used vim for development, I’m an Emacs person…
I remember when I turned off the autocomplete and suggestions on my phone's keyboard. I was suddenly a bumbling fool making tons of mistakes. I stuck with it though, and I'm glad I did. I have no technical reasons for why I stopped, I just felt like I should be in complete control of my thought patterns, and that receiving suggestions would likely influence where my thoughts went so I wouldn't have to type so much.
Totally agree! Using AI is great for quick prototypes or working around a new language (if you don't care for the privacy issues), however over using it can suck away the creative process of writing code, and writing good code. Awesome video as always
My sentiments exactly. I hate having to tweak the code that it produces that I find I may have well just have written the damn thing myself to begin with. It's good for getting up to spead with frameworks and using it as a source of quickly finding documentation tho.
Reminds of a junior programmer I had to work with. I'd written some code that regexed the entire code base to make it Y2K correct and consistent. No further edits required or desired. He edited it, which meant I ended up manually touching 70K lines of code. I'd rather write my own co-pilot.
Part of working through "silly issues" as a junior engineer is learning how to dig into code, understand the nuts and bolts, and build a good solution, a skill you will 100% *need* later on in your career. Without that experience it becomes much harder to describe your problem to other engineers, search the internet for answers, and most importantly and obviously work through problems when neither of those courses of action return results.
I think the biggest issue is that when you tab complete your way through a project you stop thinking about the problem you're trying to solve. The editor will write the code off of an assumption and when it's wrong you have no idea why. You end up creating more bugs than if you just typed the code yourself. I still use the chat feature because it's like a pair programmer or rubber ducky that gives you feed back while solving the problem.
Generating code with llms is a distraction from what they are actually good for. I'm a junior dev and instead i use it as brainstorming partner who can tell me how to solve particular problems. I then add more constraints, such as the language and paradigm i want to solve this problem in. There's some caution required here, because it's quite easy to introduce your own bias and stir the llm away from telling you a more appropriate solution. The biggest gain from this approach is you learn a lot of what's possible in software, so you can make an opinionated decision in which direction to dig further for a solution. Most of what i recently learnt comes from very good lectures on UA-cam, but I would have never found them, if chatgpt or claude didn't tell me about the existence of those topics/discussions in the field. Note this is using it for education rather than "productivity" (imagine how much skepticism i would've gotten from you, if lead my comment with that, huh) TLDR: maybe what people really need, is someone who constantly tells them an alternative perspective on the problem they are solving, rather than some autocomplete that's inherently driven by confirmation bias.
If you have the discipline, a good middle ground it to have a keybinding to quickly toggle suggestions on and off. I find that when I'm learning something or working on something niche, I turn copilot off so I'm not distracted by the output it gives. But if I'm working on something repetitive like unit testing, it's really nice to be able to toggle copilot back on to quickly add test cases.
i have had quite a similar experience especially the frustration of not having copilot afterwards (despite doing just fine before ai really took over) i do miss it when writing tedious bits of code but i much rather keeping and enhancing my skills over saving time
Exactly my thoughts, specifically when it comes to losing the ability to solve problems because you more and more rely on such tools. Also, I hated it when I wanted to think about what to write next just to see copilot already suggesting „something“… can’t count how many times I immediately pressed the escape key to make sure I am in charge of solving the next piece of the puzzle 😊
I've never actually tried copilot for basically all the reasons you've outlined here. I write code (mainly) for fun. I don't have any paid jobs that involve programming (only volunteer work) so when I write I want to have fun. For me that is deeply linked with learning and problem solving which are exactly the things copilot tries to "make efficient" but like you say that feels like it ends up being copilot writing and me reviewing and that's not what I'm after. Then there's boilerplate which certainly is neither about problem solving or learning so copilot could shine very brightly for me here. But the closed source, sends snippets to microsoft nature of the beast really puts me off using it for that alone. Especially when things like templates and snippets exist. Boilerplate is not that bad not to the point i want to give up my privacy even more than I'm already forced to.
Would be better to have an integrated chatgpt style UI which has your editing context + RAG w/ latest docs on libraries being used. No more auto inline complete or out of date suggestions, but can easily pump out boilerplate or consult when stuck
as others have expressed, I like it for writing things with awkward syntax, most commonly flattening a 2D ArrayList in java. anything else and I typically type it faster than I'm able to get a result and process that it gave (or anything big enough that it can beat me - I wouldn't trust enough to take anyway)
I am glad. I used Copilot during the research preview (am researcher) and it made me so... lazy, switched it off when they asked me to pay money for it, and my coding improved tremendously.
I'm a pretty junior software dev working in a big code base, and i find copilot does literally nothing for me, as 80% of the code i write is dependent on in-house software, that copilot doesn't have access to. The company pays for copilot, but the genuinely useless autocompletes it comes with, has made me disable it. It's more an annoyance than a help. It is more or less the same conclusion i have even when I'm writing small projects that just use std libs from whatever language I'm using. It's just annoying to use.
To me copilot is a solution to a problem I've never had. Typing is not the bottleneck in programming. If I find myself pressing way too many keys on the keyboard I use shorter names. I don't write a lot of boilerplate, I use existing libraries for very common tasks and if they don't exist I write my own flexible library and then never have to do that again. And if that doesn't work then there's code snippets. I comment as I go (try writing your prompts as comments instead?) Autocomplete exists and is great and very intelligent. It could already do the creating loops stuff before LLMs and is free. When I don't know how to do something there are tools to make documentation available in the IDE. After asking various chatbots questions about technical things, I no longer trust their answers enough for them to be more useful than stackoverflow. They don't know anything about new technologies. Why would I ever pay for this? And I'd be paying them for the privilege of letting them use my IP to train their model. Plus MS is losing money on it, it's only going to get more expensive.
In my humble opinion, using AI tools for generating new code is not the right way to use AI tools. I'm not expecting to give such tool a complex problem and wait for it to spew a perfectly working solution, the fact that a lot of people use AI that way just sounds so unrealistic to me. In my personal experience Copilot and Codeium are extremely good at generating code that otherwise is a chore. Boilerplate code, unit testing, documentation, repo sensitive code completion and tasks with already defined processes are some of the cases where such tools have had real impact in my productivity.
Copilot makes me think of calculators vs mental arithmetic. It's easy to reach for a calculator but if you reach for it too much, you start to use it for stuff you could do just as fast or faster in your head. It feels like it could be a good resource if you can moderate how/when you use it
A thing that annoys me is that 90% of the time is wrong. Imagine a coworker that is wrong that ratio and you have to constantly correct him, will you say that he boost your productivity?
I've had the same experience using copilot. Initially, I thought it was very cool and was fully enjoying the increased productivity that copilot brought. However, I also realised some time after that I had a copilot pause. My brain would stop working and instead wait for copilot to suggest something. I didn't like that feeling, so I went back to not using copilot and started to enjoy coding more again.
Whe the program I writing becomes deeper and deeper, copilot suggestions becomes worse, and from there it is just distraction. It still helps when I'm writing simple scripts. But it will not helpful if you're writing something serious.
It's funny, I had a similar experience myself. I never used copilot really because I just didn't like it, instead I used tabnine. The original iteration of tabnine was pretty decent because it basically just acted like a smarter language server, giving suggestions for the next two or three words rather than entire lines of code. But once it started suggesting full lines, I found myself ignoring all of the suggestions because they were fundamentally wrong or worse than what I was planning to write. Even in the case of boilerplate, it just wasn't good at predicting what needed to be done. For a while I limited it to only suggesting one or two words again but like you said it fell out of date pretty quickly. I went back to using normal language servers and I don't see myself going back to AI anytime soon especially since AI isn't really AI, if you know what I mean.
After using Copilot most of the time for a couple months I went through the same after stopping. The "Copilot pause" is horrible and ended up not making me more productive but rather taking away fun.
I commented out copilot in my config a while ago and forgot to add it back in, then decided to not put it back in because I noticed with copilot it was much harder for me to get into any kind of flow state because I had to constantly check it's suggestions and I was also pausing and waiting for them and sort of shutting my brain off in the meantime. I'm now copilot-free again and have sort of returned to normal. Now using chatgpt a lot more when I actually want to use ai (I have the chatgpt plugin), but with this plugin, using ai is a deliberate choice which does not break my mental flow as much
I have turned off the autocomplete and may just get rid of it. I still use the chat feature as a rubber duck or asking a question for a language I am not used to or for a common enough algorithm implementation, but that is it. More off than not, it has been suggesting me bad code that has nothing to do with what I am writing. I end up deleting all but a word or two out of a full block it generated. Its especially bad at suggesting symbols and function names that are just close enough to ones that I would use but that are not actually real. So I get confused for a second as to why my code is wrong. LSPs are just better. Event better is when CoPilot adds ``` to the end of something it generates. Which makes me wonder if its just giving me someones random mark down snippet.
If you're writing the snippets code snippets day in, day out, having an elgato stream deck is a good way to save boiler plate buttons when you want them. I use mine for a handful of annoying to write code snippets, but I still write almost everything.
Yeah, I'm a privacy and FOSS centered dev too. I didn't sign up for any of these remote AI services except initially for chatgpt (the free one) but even then I never used it with my real code. Usually if I want it to check something I give it an alternative code and explanation and ask in abstract terms. I'm experimenting with copilot-like addons that run using an ollama server and those are working well though 9/10 times it generates additional code I didn't want, or it actually takes over a helpful language server auto-complete and I accidentally pick up the dumb generated code. So in practice I have autocomplete disabled and instead I trigger it when I want to for boilerplate or configs. I find that is the best way for me to integrate it with my workflow at least. I need it to automate the boring stuff that I wrote 5 000 000 times already, not the new stuff that I'm not even yet sure how to write them myself... That's just counterproductive
In my view it is just another tool, which has some limitations like all tools. I am using it since one or two days and struggle way more then before with the escape key to avoid introducing the code suggestions. Also it felt to me like a fight, I am a Visual Studio user and Intellisense gives you also suggestions... and Copilot as well. 😅 As you said in the video it might be outdated or in my case often wrong. But where it really helps in my view is when you're sitting in a front of a empty document. Or to give you a draft for a function documentation. Or to write boilerplate code. Looks like the good old days with a plain text editor might come back soon.
Removing copilot was the best decision I've made when I was starting out to take programming as a hobby. It's filling out deep and hollow gaps leading to more confusion. It's rewarding not relying on it and do the coding by self instead. Doing it my self gave more context on how to approach something.
I use an offline version of copilot (tabbyml) with a shortcut to activate it, and I only ever use it for boilerplate, testcases and general copypasta. The pauses and results end up being slightly worse on my machine than online solutions, but it doesn't really matter for the level of tasks i use it for, which are dumb tasks in and of itself. But analyzing code written by an AI definitely takes you away from any flow you previously had. It feels like constantly relying on someone to finish off your sentences, which works great half of the time, but half the time they say something completely random that throws you off
The phrase "use it or lose it" is really something I also experienced when writing code. Although I haven't used Copilot or any other comparable product at all yet, I'm nevertheless pretty reliant on the auto completion of my IDE. Writing code in vim feels sluggish and frustrating to me as I always make typos, have to look up symbols and so on. Things that the IDE is gratefully helping me with.
Beta Cody was the best thing I used - they had a pre-step which would tokenise selected private/public repos and then your context was based on that. Holy hecka, once it had all the context it was great, but then they nerfed it to have the same context as copilot. Definitely get the “copilot pause” and have been thinking of ditching it myself. Though has sometimes kept me focused while I talk off topic during a webinar and I’ve added a code comment of what to do next and it autocompletes the thing 😂
I'm thinking the same 'bout Copilot. But i really don't want to miss it with writing boilerplate. I wish there was an ai helper of some sort that's really good at generating boilerplate, but doesn't get in the way of more complex code. If there is please let me know, i'll happily make the switch.
I think for boilerplate it's really good still. The issue for me is still privacy however. Im hoping there's a good open source model out there that I'll be able to use for boilerplate!
Before moving to Neovim full-time I used to use Copilot in VSC. I was still learning programming in baby steps back then, and felt that it giving me (sometimes wrong) code was going to deeply hinder my learning and problem solving in the long run. LSP+cmp is all I need and more, I have to understand what I'm going to do and cmp saves me a few seconds, proper win-win!
Like you said. Enjoy is the reason for me to use Copilot. I don't like to write what I already know in my head. So, I never wait for him until it long part of code that he will do it faster. Like simple long endpoints and data classes For privacy, I disable him in some of my projects. In other words, copilot is not the first action for me when I am codeining. And, Thanks
In my case, it greatly reduced my time for googling stuff. For example, I started using prisma orm for the first time, and without copilot, i would have to go back and forth just to write a simple query but copilot would give me the query suggestion and i would understand that query and implement it. It's just nice being in the code editor, rather than going out again and again.
Co Pilot is a blessing as a freelancer where I do repetitive ui and common features implementation. I would definitely not use it on personal projects or when learning. But I use it my full time and freelance work.
I never really liked having Copilot, but I mostly found that Copilot Chat was the only useful thing for me as a last resort, by using it more as a search engine rather than something to write code for me. And it's mostly because I would sometimes fall into situations where there is very little documentation on a library that I'm using and I exhausted my options of trying to figure it out myself the traditional ways
I use it only to autocomplete repetitive blocks of code, like it helps me map big DTOs quickly. Almost never used it via comments to ask for something or waited for a suggestion. I think we just use it differently and using it, like any tool, only makes you work better.
I put Copilot on a shortcut so it's not constantly bothering me and getting in the way and use it only to generate repetitive code or code comments. But even for code comments it's wrong 50% of the time. Something like Codium or a chat LLM will help me more than in-IDE autocomplete. And Copilot is also resource hungry. Overall, a few of us are already falling in the trough of disillusionment, following a standard hype cycle.
Which Framework laptop do you have (mine is intel 13th gen 1 TB of SSD 64 GB RAM and is running arch linux.) The price was insane because my dad made me use 64 GB of RAM over 16 GB
I agree fully. I setup codieum AI and used it for about 3 days. No enough to get used to it, but it is not the workflow for me. Maybe just as a boiler plate generator. I also noticed the scope of its understanding of a project is meh
I tried copilot, gave it a simple task to sort an existing hardcoded list, it failed 5 times. Tried so it can make maybe documentation but all was wrong anyway and the context way too less. Removed it after 1 week.
Same for me, I love coding and I don’t want a tool to do it for me. Yeah, there are some tasks that are tedious but I find it more fun to come up with some macro recording or meta programming thing. Obviously that’s slower and hard to do when you have a deadline, but on my personal projects I really enjoy that creative aspect of finding ways to make the tedious/boilerplate stuff more interesting and an opportunity to learn something new. It’s like being a painter, you have an idea then AI or a robot painting it for you. That would really suck the fun of it.
I use Copilot mainly through the chat, which I find useful to ask some questions about a Vim command or some general direction about what I want to do. As for the IDE integration, I use a Neovim plugin that adds the suggestions in my completion menu, no more virtual text. That's crazy how it's way less distracting and I rarely wait for it anymore, I sometimes even forget about Copilot. In the end, I'm not super impressed by it, especially the chat that cannot access GH repos. I'm more and more thinking about ditching it and just get a GPT-4 subscription instead.
I use AI mainly for auto completion. For example: You have form on the front end and you want to assign all the inputs in js. You type out 1 element and the next elements are getting auto completed by AI. It helps me a lot to enter my flow state, because all that simple but tedious typing is just skipped.
I don’t use copilot, but I have used codeium. It’s pretty much copilot without the privacy issues. So far it’s been good, but I do notice I wait for the autocomplete when I could just type it myself. I’m going to keep using it, I just know I need to ignore it when I’m writing something a bit more complex/intensive.
It's all about efficiency for the corporate world. If someone or something can write code x10 your speed, then you will just be too expensive to maintain (well if you are a regular coder). Not talking about you specifically, but generally
All your examples were literally me 🤣 Used to work out, not anymore. Used to play the piano, not anymore. Used to cycle a lot, not anymore. What a good wakeup call to start doing those things again!
I agree with your points. I want to understand the library I'm using to solve its issues. Because Copilot can only use solution for already solved problems - not for unsolved ones. If I don't understand the programming, I cannot solve it with Copilot or what ever. If I would write massive boilerplate code - maybe it is useful? But this happens once a year - and to learn a tool (and pay for it) for once a year or so? Nah.
What could go wrong: Prompt an AI for a desired outcome in hopes of producing an artifact, where single character errors can fundamentally alter the behavior of the artifact - An artifact which the prompter no longer understands, from an oracle that the prompter does not control - An artifact with unknown, unknown behaviors - An oracle whose conversion from prompt to artifact is fundamentally non-deterministic. Extrapolated, but not by much.
in my experience; be they code, writing, or even art; at some point the ai output will essentially became extra fluff that would be harder to deal with than just making it yourself in the first place to get what you want. it's also just much more fun and satisfying to do it yourself y'know
I have not used any of the AI coding assistants. I like using a color highlighted editor, but not so much anything that does any kind of auto complete.
I usually just turn copilot of depending on what I'm working on... if it's some more basic stuff like some data transformations in a python notebook it's handy to have around since I'm not going to reuse most of this code or even look at it again. Then typing out a comment describing what should be done, waiting for the autocomplete that uses the correct library function to do it really does boost productivity... if the only thing that matters is the dataset written in the end that's ok. I do turn it off when working on more sophisticated stuff however. Especially in rust, copilot can lead you down paths of strange and bad design choices. Also it stops me from thinking through something completely. When the suggestion appears my train of thought gets derailed and I start reading instead. So in short if the task is easy and I already know conceptually what to do and how to do it copilot is nice because everything goes faster. As a tool to help me do stuff that I don't already know how to do from end to end: not really.
Do you publish your dotfiles? I just like how some elements in your environment are set up and I don't want to spend too much time on replicating some of them :)
As with all your other videos, very clearly explained. I often find myself in the loop of accepting copilot’s suggestions and re-writing what doesn’t work for me. I wonder if I’d be better off writing my stuff from scratch but in my particular case, I feel like I benefit from “one man’s pair programming”. Where copilot is the first one to ride, and then I go and work on that idea.
I use copilot primarily as an input device, where I know what to input but let copilot do the work so I can focus my mind on what to do next. Thankfully copilot is pretty good at copying my style so that makes my coding workflow very efficient!
I use GPT for some tasks, but I have not yet used a Copilot-driven IDE. I believe that we should not overly rely on GPTs. They should serve as a valuable tool, but not as the sole author or co-author.
I'd be really interested in your self hosting of a coding llm. I'm currently in the process of figuring out my homelab server hardware etc, and coding llms are also a consideration for me on that front
now that ollama is cross platform with windows support using that with an openai API compatible plugin should be a consistent experience across windows linux and mac, plus you get the benefit of being able to run inference on another PC or some kind of VPS. originally I disliked ollama's docker style model management but for LAN scale use behind a firewall its actually really nice and I would recommend you try it. A lot of people suggest deepseek coder models and I have had good success there, especially since the 7b quantised models can be ran on a 1080/2080/3080 in vram entirely.
I have never used copilot because I have not found any reason for it. This artificial intelligence is not able to write something that I can't write myself. In the short run you are able to write something faster, but in the long run? I'm sure that in the long run you are an undeveloped version of yourself without using a copilot. If you don't care about basic arithmetic and chatGPT does everything for you, will you be able to take off with something more advanced? I don't think so. Copilot works similarly, it's another tool to make you a bigger no-brainer.
The 'copilot pause' happened to me some days ago when I was working on a local project and for some reason I didn't have internet connection, so GH Copilot didn't work. I imediately felt the change in behaviour that Copilot has given me.
I've noticed that when the code you're writing runs in a specific context (like a specific Javascript runtime), copilot becomes very counter productive. Even though the types are available, it keeps making mistakes and wasting my time. It does excels in writing boilerplate though.
I agree with you far more than I thought I would when going in and I'm considering the same. Although I have found having a chat bot is helpful. It can be used as a rubby ducky and if I have a question it's quick to give an answer. If I need more I go to the docs. The chat bot kinda takes away the middle step of googling the question.
For me Copilot is actually really useful to learn new things. When coding in new languages, Copilot often suggests me ways to do things that I was not aware of, that I would maybe never know before a long time and it's actually really useful.
I am finding codiumate extension for vs code way better than copilot. 1. Codiumate doesn't provide any auto complete and that has not made me depend on AI less and think for myself now, and there's only chat mode which is again better than copilot chat 2. The response is also more focused to the context of my code and even when it gives generic suggestions, they are helpful in pointing me in the right direction.
To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/DreamsofCode . You’ll also get 20% off an annual premium subscription.
Looking forward to a new nvchad 2.5 video!
2:55 I basically refuse to believe any polls or research they show us on their own product. Conflict of interest 101.
There's a bias in it anyway even if it is not intentional (it probably is intentional). If you survey people who use copilot, you're going to get more responses from people that have stuck with it than people who have stopped using it. Most of the people I know that have used it, used it for a week, said this is not really worth using yet, and stopped using it, but the people who still use it are the people that liked using it.
Ahahah like trust me bro
Agreed, same thing with ads in my book
Especially when they don't show their data, metrics and method of measurement. Pure hype for the investors.
Someone told me really early on to disable auto complete and bind the copilot suggest to a key command, it helps a lot with only using it when really needed.
This is a good idea!
This is what I do. It is then useful for deliberate use cases (like big chunks of repetitive boiler plate that I have already solved, but now need to repeat multiple times in other parts of the codebase - like DTOs, API Patterns etc.) - but I don't use the auto complete as it is too prone to suggesting bad ideas that I can't then unsee putting me in the wrong solution space for the problem I am solving.
shockingly obvious when I read your comment - now I feel rather stupid for not thinking of that myself! :)
That's great advice. I'm a data scientist. I do a lot of custom ML work where Copilot is rubbish. But I love it when I type ggplot( and Copilot creates a decent visualization of my data without having to think about it. The autocomplete just gets in the way, the hotkey is the better option.
Where copilot really shines for me is writing boilerplate code
For me is 90% on type definitions and 10% me going "make a loop that does x"
THIS. `if (!canJoinTeam)` and it'll complete the if with an appropriate error response
having copilot with Golang is so nice to get around writing the if err != nil... every single time you call a function. Another thing I had recently was a huge Json schema from an aws api request which I wanted converted to a Go struct with serialisation and for that kind of stuff it can save you 20 minutes in like 10 seconds
One alias away or keybinding away
This!!! That's the shortest way to describe it buddy
I have been running copilot since very early access, and I used to do the same thing and wait.
At some point, I started to ignore it and only used it for boilerplate code, for which I did not have a template.
But lately, I have noticed that, at least in VSCode, where the copilot implementation is the most integrated, copilot will present its reply even if I continue typing as long as it is the same as what I typed. This means that there is no longer a delay for me, and it will usually present code that is in the direction I was heading anyway.
I also no longer accept full replies and instead use the ctrl + right arrow to accept words, and this is the same shortcut as my "accept lsp prompt", as they are all marked as inline suggestions.
I am a naturally slow/deliberate programmer so for me this way of working with copilot did speed up my work and made it drastically more enjoyable
Thanks for the ctrl+right arrow tip. I usually end up hitting tab then deleting 90% of the suggestion
Cmd + Right if you are on mac
I used AI tools like copilot and tabnine before and felt that once you start having to do more than actual boilerplate the ai just gets in the way and leads you down the wrong paths. I stopped and don't miss it!
Yup, experienced the same here.
They help a lot where boilerplate is present.
But when just typing away things in my mind they just get in-between.
I'm quite new and I also expirece the same thing.
Same here! The worst for me is when it makes a mistake so subtle that looks good when you read it, but gives a weird error that takes longer to debug than if I'd written it in the first place.
I cancelled copilot, but still use Codeium for boilerplate
I've not found an AI that is actually good. It regularly will give me wrong information and when I explicitly tell it to not do something, it still does. I asked it, "Why do you answer so confidently when given a question with so much ambiguity, wouldn't it make sense to ask for clarification?" to which it responded, "I'm sorry let me clarify..."
I found them often wrong and very slow. I’ve tried but they just slow me down even with looking up docs.
copilot in generic tasks that are present in its dataset - bing bong boom done,
anything that requires >2% imagination - i halluce
i used cody... wich prevented me finishing my mini project 2 days faster... literally debugging whats wrong with the app, since LSP didnt gibe any errors
eventually realized the ai was hallucinating and kept rotating and changing my .env variables .. and when use autocomplete it suggested wrong auto-completes... not the same names ive defined...
took me a while to realize that lol..
not going to use any AI, if they are not reliable as auto-completers even... isnt there like claim that even the most efficient code ai assistant... gets it right only 17% time?
seems like its just holding people back, not helping anyone...
only positive use ive found for them is to select a block and auto-describe/comment what the block does...
took me 2 days of using AI to realize, it will create a religious-type of blind trust into ai-complete... extra bad was that its personal project, so i didnt use typescript...
the ai kept thinking im using mongodb for orm as well... when i was using actually some other db wich also has API like db.method()
never used copilot, but is it better than cody (sourcegraph) ?
ai pretty decent tho at "describe this block" , whats happening , if u check some foreign codebase and auto-commenting block of code , generally describes it pretty well
Yeah, I think Copilot is being pushed as this tool that will KILL programmers!
Like no, it's good at the fundamentals and for learning concepts, but it can and will hamper your developmental progress.
I've always turned off the auto complete prompts. What really gives me value is Copilot chat. 90% of the time I use it to understand what the legacy code actually tries to accomplish and it does a decent job in most languages. The other 10% of the time I'm asking questions about how to structure a project in a language that I'm not particularly familiar with, or asking it to generate boilerplate code, where it does a decent enough job.
Really? I feel like copilot chat kinda sucks ass compared to just using GPT-4
@@jamess.2491 yea literally only good for autocompleting boilerplate
Yeah copilot chat often just ignores what I tell it or forgets what I told it in the message before
@@jamess.2491 my codebase has thousands of lines of code. There's no way outside of using the gpt4 api to give it that context and even with that, I think copilot does a better job
I like Copilot a lot when it comes to pattern completion. If I write functions to accomplish some task and then have a related task that isn’t quite the same it will generally outline new code in the same style as the previous code I wrote. If I write CSS for a red gradient of a specific shade for instance I’ve been able to autocomplete code for a blue gradient that matches the same saturation and lightness as the red one, and continue the pattern for green, yellow, etc. On solo projects I feel like Copilot gives me the power of two people. I just need to be prepared to guide it and review the code and not rely on it for anything too complicated.
it's just a fancy autocomplete
often it would give a completely wrong answer
but you'd still have to spend time reviewing the code only to find out it was incorrect
Exactly, replace copilot with intellisense and every argument stays the same. They're bad arguments
Except that it's much more useful.
You can do an initial code review, check for errors and explore solutions.
It helps you immerse yourself in a new code base and documents what's not...
Copilot is therefore an excellent tool to complete your toolbox
its also trained on old data
iirc it does not know about react18 yet
@@MarthinusBosman i think this is a rather reductive argument. the point about Copilot being trained on older versions of code and using newer versions of libraries incorrectly still stands, whereas with traditional code completion it's usually based on analysis of the existing API. traditional code completion systems are also fundamentally different regardless; if they weren't, why would people use copilot in the first place? a code completion system that doesn't go further than the language server, static analysis, and the documentation doesn't try to solve problems for you, and it's much harder to write code that you don't understand with it.
Copilot is just intellisense+ and that's amazing when I hate boilerplate.
I've taken it out as well, missed it in the beginning, but I'm happier without it now. I think it's a bit to much of a crutch, you are better of developing the "muscle memory" in the long run.
Agreed!
I'm an old fella.. I've seen many fads come and go. although AI have some value, right now it's over hyped and overused.
and when it comes to software, I like minimal fast setups that help me DO THE JOB.
I'm amazed by how some devs are willing to avoid actual learning and work using overkill solutions for the simplest of tasks.
I knew from day one that AI is not for me but I think it can be very helpful in many situations.
Ai has been quite useful for me when I have to write code or modify something in a language I don't really know, like the other day I needed to convert a yml file to toml and I had pretty much no clue of the file structure of either but chatgpt could just convert the entire thing for me. Saved a lot of time and got a working answer. If I write in a language I know I tend to write my code myself and use chatgpt to refactor it or make it nicer and more readable + add comments
@@ZÏ̇̃ I do find it amazingly helpful in personal projects and filling minor gaps. but sadly some people are committing code they don't completely understand which is very dangerous and irresponsible.
if you use it as a jumpstart and actually learn what it did. and preferably why and what other solutions it might have missed then power to you.
but I hope you understand that the fastest answer is not always the best answer
@@ZÏ̇̃TOML is so nice!
I am also, I hve been writing code for over 25 years. That was then this is now. You will either use it and the productivity bost it provides or you will fall by the side of the road. I am a old dog but I am learning the new tricks.
The best experience I've had recently, was writing C on my job, where I'm not allowed to install anything.
It was just a basic Vim (not neovim) setup and the C compiler.
Showed me how little you actually need to have fun programming.
Also, man pages of c can replace google search.
For me that’s not really fun. I mean, it is enough for sure depending on what you are doing. But ever since I first used the first IDE (Turbo Pascal in 1992) I don’t like working in a non-ide setting. Of course I think back then it could not do more than what vim could now. But I never used vim for development, I’m an Emacs person…
nano + tmux master race
@@MrAlanCristhianno way man nvim nowadays is the meta
I'd hate that. I can't live without my neovim setup anymore.
I remember when I turned off the autocomplete and suggestions on my phone's keyboard. I was suddenly a bumbling fool making tons of mistakes. I stuck with it though, and I'm glad I did. I have no technical reasons for why I stopped, I just felt like I should be in complete control of my thought patterns, and that receiving suggestions would likely influence where my thoughts went so I wouldn't have to type so much.
Totally agree! Using AI is great for quick prototypes or working around a new language (if you don't care for the privacy issues), however over using it can suck away the creative process of writing code, and writing good code. Awesome video as always
My sentiments exactly. I hate having to tweak the code that it produces that I find I may have well just have written the damn thing myself to begin with. It's good for getting up to spead with frameworks and using it as a source of quickly finding documentation tho.
Reminds of a junior programmer I had to work with. I'd written some code that regexed the entire code base to make it Y2K correct and consistent. No further edits required or desired. He edited it, which meant I ended up manually touching 70K lines of code. I'd rather write my own co-pilot.
@@Galahad54 I'm genuinely curious, was there no version control in place? or does your company don't do commits for each input?
Part of working through "silly issues" as a junior engineer is learning how to dig into code, understand the nuts and bolts, and build a good solution, a skill you will 100% *need* later on in your career. Without that experience it becomes much harder to describe your problem to other engineers, search the internet for answers, and most importantly and obviously work through problems when neither of those courses of action return results.
100% agree
I think the biggest issue is that when you tab complete your way through a project you stop thinking about the problem you're trying to solve. The editor will write the code off of an assumption and when it's wrong you have no idea why. You end up creating more bugs than if you just typed the code yourself. I still use the chat feature because it's like a pair programmer or rubber ducky that gives you feed back while solving the problem.
Generating code with llms is a distraction from what they are actually good for. I'm a junior dev and instead i use it as brainstorming partner who can tell me how to solve particular problems. I then add more constraints, such as the language and paradigm i want to solve this problem in. There's some caution required here, because it's quite easy to introduce your own bias and stir the llm away from telling you a more appropriate solution.
The biggest gain from this approach is you learn a lot of what's possible in software, so you can make an opinionated decision in which direction to dig further for a solution. Most of what i recently learnt comes from very good lectures on UA-cam, but I would have never found them, if chatgpt or claude didn't tell me about the existence of those topics/discussions in the field. Note this is using it for education rather than "productivity" (imagine how much skepticism i would've gotten from you, if lead my comment with that, huh)
TLDR: maybe what people really need, is someone who constantly tells them an alternative perspective on the problem they are solving, rather than some autocomplete that's inherently driven by confirmation bias.
If you have the discipline, a good middle ground it to have a keybinding to quickly toggle suggestions on and off. I find that when I'm learning something or working on something niche, I turn copilot off so I'm not distracted by the output it gives. But if I'm working on something repetitive like unit testing, it's really nice to be able to toggle copilot back on to quickly add test cases.
This is a really good idea. Saves the whole "copy and paste" feel of copilot.
Programmers like us will be getting the higher salary jobs. Our approach is superior.
i have had quite a similar experience especially the frustration of not having copilot afterwards (despite doing just fine before ai really took over) i do miss it when writing tedious bits of code but i much rather keeping and enhancing my skills over saving time
Exactly my thoughts, specifically when it comes to losing the ability to solve problems because you more and more rely on such tools.
Also, I hated it when I wanted to think about what to write next just to see copilot already suggesting „something“… can’t count how many times I immediately pressed the escape key to make sure I am in charge of solving the next piece of the puzzle 😊
Hey, what keyboard are u using ?
I would appreciate knowing. Thanks in advance
I'd be interested in videos about how to self-host LLMs
Second
Third
Ollama is a good tool for self-hosting LLMs
check out Continue
i tried Continue. worked well
I've never actually tried copilot for basically all the reasons you've outlined here. I write code (mainly) for fun. I don't have any paid jobs that involve programming (only volunteer work) so when I write I want to have fun. For me that is deeply linked with learning and problem solving which are exactly the things copilot tries to "make efficient" but like you say that feels like it ends up being copilot writing and me reviewing and that's not what I'm after. Then there's boilerplate which certainly is neither about problem solving or learning so copilot could shine very brightly for me here. But the closed source, sends snippets to microsoft nature of the beast really puts me off using it for that alone. Especially when things like templates and snippets exist. Boilerplate is not that bad not to the point i want to give up my privacy even more than I'm already forced to.
Would be better to have an integrated chatgpt style UI which has your editing context + RAG w/ latest docs on libraries being used. No more auto inline complete or out of date suggestions, but can easily pump out boilerplate or consult when stuck
as others have expressed, I like it for writing things with awkward syntax, most commonly flattening a 2D ArrayList in java. anything else and I typically type it faster than I'm able to get a result and process that it gave (or anything big enough that it can beat me - I wouldn't trust enough to take anyway)
I am glad. I used Copilot during the research preview (am researcher) and it made me so... lazy, switched it off when they asked me to pay money for it, and my coding improved tremendously.
I'm a pretty junior software dev working in a big code base, and i find copilot does literally nothing for me, as 80% of the code i write is dependent on in-house software, that copilot doesn't have access to. The company pays for copilot, but the genuinely useless autocompletes it comes with, has made me disable it. It's more an annoyance than a help. It is more or less the same conclusion i have even when I'm writing small projects that just use std libs from whatever language I'm using. It's just annoying to use.
To me copilot is a solution to a problem I've never had.
Typing is not the bottleneck in programming. If I find myself pressing way too many keys on the keyboard I use shorter names. I don't write a lot of boilerplate, I use existing libraries for very common tasks and if they don't exist I write my own flexible library and then never have to do that again. And if that doesn't work then there's code snippets. I comment as I go (try writing your prompts as comments instead?) Autocomplete exists and is great and very intelligent. It could already do the creating loops stuff before LLMs and is free. When I don't know how to do something there are tools to make documentation available in the IDE. After asking various chatbots questions about technical things, I no longer trust their answers enough for them to be more useful than stackoverflow. They don't know anything about new technologies.
Why would I ever pay for this? And I'd be paying them for the privilege of letting them use my IP to train their model. Plus MS is losing money on it, it's only going to get more expensive.
In my humble opinion, using AI tools for generating new code is not the right way to use AI tools. I'm not expecting to give such tool a complex problem and wait for it to spew a perfectly working solution, the fact that a lot of people use AI that way just sounds so unrealistic to me. In my personal experience Copilot and Codeium are extremely good at generating code that otherwise is a chore. Boilerplate code, unit testing, documentation, repo sensitive code completion and tasks with already defined processes are some of the cases where such tools have had real impact in my productivity.
Just disable auto-suggestions, instead use it for code generation, or copilot chat.
Copilot makes me think of calculators vs mental arithmetic. It's easy to reach for a calculator but if you reach for it too much, you start to use it for stuff you could do just as fast or faster in your head. It feels like it could be a good resource if you can moderate how/when you use it
A thing that annoys me is that 90% of the time is wrong. Imagine a coworker that is wrong that ratio and you have to constantly correct him, will you say that he boost your productivity?
I've had the same experience using copilot. Initially, I thought it was very cool and was fully enjoying the increased productivity that copilot brought. However, I also realised some time after that I had a copilot pause. My brain would stop working and instead wait for copilot to suggest something. I didn't like that feeling, so I went back to not using copilot and started to enjoy coding more again.
I need the new neovim setup please
fr
Hey why don't check out mine
Coming soon!
What is the name of keyboard and keys you are using?(the one split in two parts)
came to comments to see that too.
zsa voyager ;)
Whe the program I writing becomes deeper and deeper, copilot suggestions becomes worse, and from there it is just distraction. It still helps when I'm writing simple scripts. But it will not helpful if you're writing something serious.
It's funny, I had a similar experience myself. I never used copilot really because I just didn't like it, instead I used tabnine. The original iteration of tabnine was pretty decent because it basically just acted like a smarter language server, giving suggestions for the next two or three words rather than entire lines of code. But once it started suggesting full lines, I found myself ignoring all of the suggestions because they were fundamentally wrong or worse than what I was planning to write. Even in the case of boilerplate, it just wasn't good at predicting what needed to be done. For a while I limited it to only suggesting one or two words again but like you said it fell out of date pretty quickly. I went back to using normal language servers and I don't see myself going back to AI anytime soon especially since AI isn't really AI, if you know what I mean.
After using Copilot most of the time for a couple months I went through the same after stopping. The "Copilot pause" is horrible and ended up not making me more productive but rather taking away fun.
I commented out copilot in my config a while ago and forgot to add it back in, then decided to not put it back in because I noticed with copilot it was much harder for me to get into any kind of flow state because I had to constantly check it's suggestions and I was also pausing and waiting for them and sort of shutting my brain off in the meantime. I'm now copilot-free again and have sort of returned to normal. Now using chatgpt a lot more when I actually want to use ai (I have the chatgpt plugin), but with this plugin, using ai is a deliberate choice which does not break my mental flow as much
I have turned off the autocomplete and may just get rid of it. I still use the chat feature as a rubber duck or asking a question for a language I am not used to or for a common enough algorithm implementation, but that is it. More off than not, it has been suggesting me bad code that has nothing to do with what I am writing. I end up deleting all but a word or two out of a full block it generated. Its especially bad at suggesting symbols and function names that are just close enough to ones that I would use but that are not actually real. So I get confused for a second as to why my code is wrong. LSPs are just better. Event better is when CoPilot adds ``` to the end of something it generates. Which makes me wonder if its just giving me someones random mark down snippet.
If you're writing the snippets code snippets day in, day out, having an elgato stream deck is a good way to save boiler plate buttons when you want them. I use mine for a handful of annoying to write code snippets, but I still write almost everything.
Yeah, I'm a privacy and FOSS centered dev too. I didn't sign up for any of these remote AI services except initially for chatgpt (the free one) but even then I never used it with my real code. Usually if I want it to check something I give it an alternative code and explanation and ask in abstract terms. I'm experimenting with copilot-like addons that run using an ollama server and those are working well though 9/10 times it generates additional code I didn't want, or it actually takes over a helpful language server auto-complete and I accidentally pick up the dumb generated code. So in practice I have autocomplete disabled and instead I trigger it when I want to for boilerplate or configs. I find that is the best way for me to integrate it with my workflow at least. I need it to automate the boring stuff that I wrote 5 000 000 times already, not the new stuff that I'm not even yet sure how to write them myself... That's just counterproductive
In my view it is just another tool, which has some limitations like all tools.
I am using it since one or two days and struggle way more then before with the escape key to avoid introducing the code suggestions.
Also it felt to me like a fight, I am a Visual Studio user and Intellisense gives you also suggestions... and Copilot as well. 😅
As you said in the video it might be outdated or in my case often wrong.
But where it really helps in my view is when you're sitting in a front of a empty document. Or to give you a draft for a function documentation.
Or to write boilerplate code.
Looks like the good old days with a plain text editor might come back soon.
Removing copilot was the best decision I've made when I was starting out to take programming as a hobby. It's filling out deep and hollow gaps leading to more confusion. It's rewarding not relying on it and do the coding by self instead. Doing it my self gave more context on how to approach something.
0:58 that part about riding a bike, I spent 15 years without riding a bike, it took me 6 hours to learn how to do it again after all that time.
I use an offline version of copilot (tabbyml) with a shortcut to activate it, and I only ever use it for boilerplate, testcases and general copypasta.
The pauses and results end up being slightly worse on my machine than online solutions, but it doesn't really matter for the level of tasks i use it for, which are dumb tasks in and of itself.
But analyzing code written by an AI definitely takes you away from any flow you previously had. It feels like constantly relying on someone to finish off your sentences, which works great half of the time, but half the time they say something completely random that throws you off
I really like your videos and animations. Could you share your general creation process of these? What tools do you use for editing and animation?
Absolutely! I mostly use Davinci Resolve for like 95% of my work but will use After Effects to do some things faster (like the particle smash effect).
@@dreamsofcodeAwesome, thank you!
The phrase "use it or lose it" is really something I also experienced when writing code. Although I haven't used Copilot or any other comparable product at all yet, I'm nevertheless pretty reliant on the auto completion of my IDE. Writing code in vim feels sluggish and frustrating to me as I always make typos, have to look up symbols and so on. Things that the IDE is gratefully helping me with.
Beta Cody was the best thing I used - they had a pre-step which would tokenise selected private/public repos and then your context was based on that.
Holy hecka, once it had all the context it was great, but then they nerfed it to have the same context as copilot.
Definitely get the “copilot pause” and have been thinking of ditching it myself. Though has sometimes kept me focused while I talk off topic during a webinar and I’ve added a code comment of what to do next and it autocompletes the thing 😂
I'm thinking the same 'bout Copilot. But i really don't want to miss it with writing boilerplate. I wish there was an ai helper of some sort that's really good at generating boilerplate, but doesn't get in the way of more complex code. If there is please let me know, i'll happily make the switch.
What's your take on using copilot or copilot-like tools to write tests and/or docblocks?
I think for boilerplate it's really good still. The issue for me is still privacy however.
Im hoping there's a good open source model out there that I'll be able to use for boilerplate!
Before moving to Neovim full-time I used to use Copilot in VSC. I was still learning programming in baby steps back then, and felt that it giving me (sometimes wrong) code was going to deeply hinder my learning and problem solving in the long run. LSP+cmp is all I need and more, I have to understand what I'm going to do and cmp saves me a few seconds, proper win-win!
Copilot free coding is so much stimulating and actually saves you a lot of headaches
What keyboard you got mate?
Like you said. Enjoy is the reason for me to use Copilot. I don't like to write what I already know in my head.
So, I never wait for him until it long part of code that he will do it faster. Like simple long endpoints and data classes
For privacy, I disable him in some of my projects.
In other words, copilot is not the first action for me when I am codeining.
And, Thanks
In my case, it greatly reduced my time for googling stuff. For example, I started using prisma orm for the first time, and without copilot, i would have to go back and forth just to write a simple query but copilot would give me the query suggestion and i would understand that query and implement it. It's just nice being in the code editor, rather than going out again and again.
Co Pilot is a blessing as a freelancer where I do repetitive ui and common features implementation. I would definitely not use it on personal projects or when learning. But I use it my full time and freelance work.
I never really liked having Copilot, but I mostly found that Copilot Chat was the only useful thing for me as a last resort, by using it more as a search engine rather than something to write code for me. And it's mostly because I would sometimes fall into situations where there is very little documentation on a library that I'm using and I exhausted my options of trying to figure it out myself the traditional ways
I use it only to autocomplete repetitive blocks of code, like it helps me map big DTOs quickly. Almost never used it via comments to ask for something or waited for a suggestion. I think we just use it differently and using it, like any tool, only makes you work better.
I put Copilot on a shortcut so it's not constantly bothering me and getting in the way and use it only to generate repetitive code or code comments. But even for code comments it's wrong 50% of the time.
Something like Codium or a chat LLM will help me more than in-IDE autocomplete. And Copilot is also resource hungry.
Overall, a few of us are already falling in the trough of disillusionment, following a standard hype cycle.
Which Framework laptop do you have (mine is intel 13th gen 1 TB of SSD 64 GB RAM and is running arch linux.) The price was insane because my dad made me use 64 GB of RAM over 16 GB
I agree fully. I setup codieum AI and used it for about 3 days. No enough to get used to it, but it is not the workflow for me. Maybe just as a boiler plate generator. I also noticed the scope of its understanding of a project is meh
I tried copilot, gave it a simple task to sort an existing hardcoded list, it failed 5 times. Tried so it can make maybe documentation but all was wrong anyway and the context way too less. Removed it after 1 week.
I recently started using it and my productivity is much higher. I write Go, I don't like go.
I basically use it for writing Boilerplate
Same for me, I love coding and I don’t want a tool to do it for me. Yeah, there are some tasks that are tedious but I find it more fun to come up with some macro recording or meta programming thing. Obviously that’s slower and hard to do when you have a deadline, but on my personal projects I really enjoy that creative aspect of finding ways to make the tedious/boilerplate stuff more interesting and an opportunity to learn something new.
It’s like being a painter, you have an idea then AI or a robot painting it for you. That would really suck the fun of it.
could you tell me what keyboard u r using? thaks
I use Copilot mainly through the chat, which I find useful to ask some questions about a Vim command or some general direction about what I want to do.
As for the IDE integration, I use a Neovim plugin that adds the suggestions in my completion menu, no more virtual text. That's crazy how it's way less distracting and I rarely wait for it anymore, I sometimes even forget about Copilot.
In the end, I'm not super impressed by it, especially the chat that cannot access GH repos.
I'm more and more thinking about ditching it and just get a GPT-4 subscription instead.
I use AI mainly for auto completion.
For example: You have form on the front end and you want to assign all the inputs in js. You type out 1 element and the next elements are getting auto completed by AI. It helps me a lot to enter my flow state, because all that simple but tedious typing is just skipped.
I don’t use copilot, but I have used codeium. It’s pretty much copilot without the privacy issues. So far it’s been good, but I do notice I wait for the autocomplete when I could just type it myself.
I’m going to keep using it, I just know I need to ignore it when I’m writing something a bit more complex/intensive.
It's all about efficiency for the corporate world.
If someone or something can write code x10 your speed, then you will just be too expensive to maintain (well if you are a regular coder).
Not talking about you specifically, but generally
All your examples were literally me 🤣 Used to work out, not anymore. Used to play the piano, not anymore. Used to cycle a lot, not anymore. What a good wakeup call to start doing those things again!
I agree with your points. I want to understand the library I'm using to solve its issues. Because Copilot can only use solution for already solved problems - not for unsolved ones. If I don't understand the programming, I cannot solve it with Copilot or what ever. If I would write massive boilerplate code - maybe it is useful? But this happens once a year - and to learn a tool (and pay for it) for once a year or so? Nah.
I recently added a keyboard shortcut that turns copilot off. It’s nice to be able to turn it on for boilerplate and unit tests
What could go wrong:
Prompt an AI for a desired outcome in hopes of producing an artifact, where single character errors can fundamentally alter the behavior of the artifact - An artifact which the prompter no longer understands, from an oracle that the prompter does not control - An artifact with unknown, unknown behaviors - An oracle whose conversion from prompt to artifact is fundamentally non-deterministic. Extrapolated, but not by much.
I have set up copilot to only be triggered manually. I can trigger it for boilerplate code and write my own code most of the time.
in my experience; be they code, writing, or even art; at some point the ai output will essentially became extra fluff that would be harder to deal with than just making it yourself in the first place to get what you want.
it's also just much more fun and satisfying to do it yourself y'know
I have not used any of the AI coding assistants. I like using a color highlighted editor, but not so much anything that does any kind of auto complete.
100% this. Great video. I agree on everything. I’d rather write my own code than having to wait and then correct someone else’s code
I'm sorry for going offtopic, but what is that sound panel next to your laptop at 0:55?
It's the Rodecaster Duo. A fantastic piece of hardware!
I usually just turn copilot of depending on what I'm working on... if it's some more basic stuff like some data transformations in a python notebook it's handy to have around since I'm not going to reuse most of this code or even look at it again. Then typing out a comment describing what should be done, waiting for the autocomplete that uses the correct library function to do it really does boost productivity... if the only thing that matters is the dataset written in the end that's ok. I do turn it off when working on more sophisticated stuff however. Especially in rust, copilot can lead you down paths of strange and bad design choices. Also it stops me from thinking through something completely. When the suggestion appears my train of thought gets derailed and I start reading instead. So in short if the task is easy and I already know conceptually what to do and how to do it copilot is nice because everything goes faster. As a tool to help me do stuff that I don't already know how to do from end to end: not really.
Do you publish your dotfiles? I just like how some elements in your environment are set up and I don't want to spend too much time on replicating some of them :)
They are available on my personal GH! If you check out my dreamsofcode-io org you should find them that way
Me who doesn't even use an LSP: "Look what they need to mimic a fraction of our power"
Real programmers use cat.
@@yondaime500 No, real programmers use ed.
@@yondaime500 real programmers don't use higher level abstractions like programming languages and just write straight binary
@@gusryan Cowards, I manipulate the electrons on the metal directly.
@@gusryan yeah, the real ones don't even use assembly, they write ones and zeroes one by one, too god-like to write lowly abstractions.
As with all your other videos, very clearly explained. I often find myself in the loop of accepting copilot’s suggestions and re-writing what doesn’t work for me. I wonder if I’d be better off writing my stuff from scratch but in my particular case, I feel like I benefit from “one man’s pair programming”. Where copilot is the first one to ride, and then I go and work on that idea.
I use copilot primarily as an input device, where I know what to input but let copilot do the work so I can focus my mind on what to do next. Thankfully copilot is pretty good at copying my style so that makes my coding workflow very efficient!
I use GPT for some tasks, but I have not yet used a Copilot-driven IDE. I believe that we should not overly rely on GPTs. They should serve as a valuable tool, but not as the sole author or co-author.
I'd be really interested in your self hosting of a coding llm.
I'm currently in the process of figuring out my homelab server hardware etc, and coding llms are also a consideration for me on that front
Hi, would you share with me your Neovim configuration?
now that ollama is cross platform with windows support using that with an openai API compatible plugin should be a consistent experience across windows linux and mac, plus you get the benefit of being able to run inference on another PC or some kind of VPS. originally I disliked ollama's docker style model management but for LAN scale use behind a firewall its actually really nice and I would recommend you try it. A lot of people suggest deepseek coder models and I have had good success there, especially since the 7b quantised models can be ran on a 1080/2080/3080 in vram entirely.
I have never used copilot because I have not found any reason for it. This artificial intelligence is not able to write something that I can't write myself. In the short run you are able to write something faster, but in the long run? I'm sure that in the long run you are an undeveloped version of yourself without using a copilot. If you don't care about basic arithmetic and chatGPT does everything for you, will you be able to take off with something more advanced? I don't think so. Copilot works similarly, it's another tool to make you a bigger no-brainer.
The 'copilot pause' happened to me some days ago when I was working on a local project and for some reason I didn't have internet connection, so GH Copilot didn't work. I imediately felt the change in behaviour that Copilot has given me.
I've noticed that when the code you're writing runs in a specific context (like a specific Javascript runtime), copilot becomes very counter productive. Even though the types are available, it keeps making mistakes and wasting my time. It does excels in writing boilerplate though.
I agree with you far more than I thought I would when going in and I'm considering the same. Although I have found having a chat bot is helpful. It can be used as a rubby ducky and if I have a question it's quick to give an answer. If I need more I go to the docs. The chat bot kinda takes away the middle step of googling the question.
For me Copilot is actually really useful to learn new things. When coding in new languages, Copilot often suggests me ways to do things that I was not aware of, that I would maybe never know before a long time and it's actually really useful.
I recently re-did my nvim config and lost co-pilot also. I haven't missed it really.
What did you do with the extra money?
@@rawallon my employer pays for an enterprise copilot license so no change there
have you tried ollama as offline solution?
Copilot is amazing for generating doc strings and other boilerplate, and for that alone, it's worth it.
Link to those keyboards ?
I am finding codiumate extension for vs code way better than copilot.
1. Codiumate doesn't provide any auto complete and that has not made me depend on AI less and think for myself now, and there's only chat mode which is again better than copilot chat
2. The response is also more focused to the context of my code and even when it gives generic suggestions, they are helpful in pointing me in the right direction.