[hutzlibu]

"As far as I know there's a sizeable number of devs who don't intend to ever rely on copilot"

Is that really a thing?

I mean, I also don't want to rely on Microsoft and therefore also not on Copilot, but not using AI tools in general out of principle is probably a very rare minority. I simply would prefer my own local LLM.

But in the thread linked above I read "AI never had and never will have it’s place in art." And this stance would be very weird for me for devs.

[southerntofu]

> "As far as I know there's a sizeable number of devs who don't intend to ever rely on copilot"

I'm one of those. I've been programming for a while now and there's no way i'm gonna trust a neural network with my code. Debugging is painful enough without having to deal with subtle bugs hallucinated by ML.

Some machines are really useful to reduce human suffering and augment our collective capabilities. Some machines are just useless, polluting gadgets. I think ML sits in the middleground: if your job is pissing meaningless code all day that's very repetitive it can probably do it for you... but if you have to actually do R&D to develop new tools i don't think ML will be any use.

So yes AI can reduce work, but arguably work that was never required nor beneficial to humanity to begin with. I would be way more interested in society reflecting on "bullshit jobs" and how to actually share the workload so that we can have 1-day work-weeks planet-wide, just as the scientists from the 19th/20th century envisioned. Instead of continuing to destroy the planet so we can run bullshitting neural nets in the cloud that produce arguably little value.

But sure, ML is fun. Let's just pretend we don't see the whole world burning outside the window.

[hutzlibu]

"So yes AI can reduce work, but arguably work that was never required nor beneficial to humanity to begin with"

Hm, just a suggestion, I would be careful with such statements, if you don't want to insult peoples work you know nothing about.

Because LLMs enable a very broad spectrum of work. I don't use them in my current workflow(nor am I that easily insulted), but the times I did use them, they were useful. My problem with them was mainly ChatGPT4 was out of date, but it did produce very useful results for me for WebGPU and Pixijs, which I had not used before and the solutions it gave me, I could not find on the internet. So for my novel work, they don't help me in general, but they do help me if I need a new custom part, without having to reinvent the wheel.

And then of course there are people who greatly benefit from them, who did not study CS, like a friend who is ecologist and all he wants are some custom python scripts, to modify his GIS tool. I think he is doing useful work and with LLMs he is indeed spending less time on his (freelance) work and has more time for his children. Isn't that, what you are also hoping for?

[Affric]

I think to a large extent this is correct but, being the devils advocate, perhaps if the GIS tool were better then your friend wouldn’t have had quite as big a gain in free time?

Answering my own question somewhat, I think that LLMs are becoming a kind UI layer over many applications/tools for many users. Which is interesting. And in some ways they show signs of fulfilling the promise of AI.

[hutzlibu]

I think even with the best GUI (and I seldom have seen worse GUIs than with GIS tools), scripting is way more powerful, if you have advanced tasks.

[Affric]

Definitely more powerful but well crafted GUIs can enforce constraints on inputs and preview what changes would look like.

Agree that I have never seen a GIS tool where I thought the GUI was top notch.

[elcomet]

It seems like you tried used copilot did you? To me the best thing about it is not the full function generation, which doesn't work very reliably, it's to finish the end of the line, when you already know what you will type, and it just types it for you faster. It feels like magic and checking the code is extrême fast as it's just one line, much faster than writing it.

[sussmannbaka]

So it’s "just" providing the vim advantage to typing speed? As a vim user myself, I’d like to point to the tired argument that typing takes up the least amount of time yadda yadda

[bombela]

I found copilot most useful as an code editing assistant. A smart `sed` of sort.

Moving code around a loop, extracting a series of variables etc.

Like a sibling comment I also find it useful to complete the end of the line.

As for writing code, it mostly produces very convincing looking code at a glance, but full of shit on a second look.

But as a smart completion and local refactoring tool, I really find it useful.

[throwaway4aday]

We don't have 1 day a week workweeks not because it isn't possible but because we don't live late 19th / early 20th century lifestyles anymore. We have modern cities, infrastructure, transportation, manufacturing, power generation, diets, and recreation. I don't think many would want to go back to an agrarian lifestyle where you live in a one or two room brick cottage, walk everywhere, till a field with a very simple, small tractor, eat only what you grow seasonally, have two changes of clothes, and own basically nothing but the bare essentials to clean and feed yourself. If you did that then sure you could share the work between a little commune and maybe get by on a rotation of duties if you had enough up front capital to buy all the labor saving devices and could manage to keep everyone happy enough to share it all equally but I have a feeling it would still wind up being a hard life of poverty. There's also the question of how you keep all of the manufacturing and professional services going with so little demand for their outputs. There are very good reasons why the vast majority of the populace used to be stuck in subsistence farming for life and why that only changed with the advent of mass production and market economies.

[apetresc]

> I don't think many would want to go back to an agrarian lifestyle where you live in a one or two room brick cottage, walk everywhere, till a field with a very simple, small tractor, eat only what you grow seasonally, have two changes of clothes, and own basically nothing but the bare essentials to clean and feed yourself.

You could not possibly have made it sound more attractive and compelling.

[throwaway4aday]

I suppose for some people but if you've ever done any farm work then you'd know the reality is very different from the fantasy.

Also, there's nothing really stopping you from giving up all your worldly goods and doing this now.

[stolsvik]

You should really try it. I’ve been coding professionally since 2000, and some 10+ years before that as a kid with a hobby. The AI takes away quite a bit of tedious stuff. When you think “I’ll just have to code out that annoying little dumb piece”, where it is obvious what you need, but boring to write it, the AI typically knows what you want. Sometimes it is really mind-warping, like wow, that’s pretty heavy context you dragged in there! And sometimes it comes with a little left field that you hadn’t thought of, and even if it didn’t nail it, you got a new idea.

It’s like having an over-eager coworker to pair-program with - and can kinda boss around as you please, it never tiring or needing a break. Not senior-level (outside of knowing “deep” small pieces and snippets), but not fresh out of school either.

And it is great for fleshing out comments (if you’re into that, as I am), picking up your style and notation as you go.

[LorenzoGood]

I dont use them out of principle, but because I am still learning, so it just takes away from that.

[lionkor]

Yes, that's a thing, I'm one of those people.

While impressive, the two issues I have with codepilot and other AI tools are:

1. The code is usually the same code I'd get a few web searches away, except then it would have the appropriate copyright. As a FOSS developer (in my free time), I do not want to risk using code I don't have a license for, and thus dirtying up my entire project and putting it in danger of being taken down.

2. I really don't need it. At very few points in a project do I both think "I want to continue this" and also "I want my code written for me". I like autocomplete, I use autocomplete, and I like Visual Studio's suggestions, too. It's only wrong 50% of the time, around about. I have no interest in a tool that writes my code for me, because I have learned everything I know from solving problems myself.

Edit: Clauses in the AI's ToS like "all code generated is yours" or something is akin to a sign on a bar saying "if you hit someone in here it's not assault" -- it doesn't change the facts whatsoever, and the fact is that it's still a crime to hit somebody, even if the bar's ToS say otherwise.

[Kiro]

> The code is usually the same code I'd get a few web searches away

My impression is that people normally don't use Copilot as a substitute for finding solutions (ChatGPT is much better for that), but as a way to help with otherwise tedious tasks that are really specific to your codebase. Check out 6:05 and 6:25 in this Andreas Kling video for a good example: https://www.youtube.com/watch?v=8mxubNQC5O8

Regarding your second point Copilot helps me when I least expects it. I think the video illustrates what I mean with that as well.

[Nullabillity]

> Check out 6:05 and 6:25 in this Andreas Kling video for a good example: https://www.youtube.com/watch?v=8mxubNQC5O8

Oof. So not only was it a poor replacement for a macro, it didn't even generate the correct code!

[tambourine_man]

It really is amazing and Andreas Kling is awesome. Thanks for the link

[philote]

3. To be truly useful, you have to send your company's proprietary code to a 3rd-party AI, which may or may not use it for training their AI, or which may or may not have security issues and leak your proprietary code. Yes, we do this already with GitHub/GitLab, etc. but those are mature and (AFAIK) haven't had big security issues like OpenAI has had in the past year. 4. For ChatGPT at least, you have to give them your phone number to sign up. For me this is a deal-breaker, but I get others are fine with it.

[lionkor]

GitHub and GitLab are sufficiently useful, AI is not (for me)

[bheadmaster]

> Is that really a thing?

Yes.

I'm a Vim user with 100 WPM typing speed, and I can say with confidence that Copilot isn't that useful to me. Typing boilerplate is not an issue - understanding what I wrote is most of the work. And having an AI spew code that I have to read is more work for me than just writing it myself.

[radlad]

Just as a data point counter to yours, I'm a Vim user with a 150+ wpm typing speed, and I have to say I find Copilot massively useful.

In Go, it's great for dealing with a lot of the repetitive code one finds themselves writing.

When writing Android apps, it's useful for API discovery!

[DonHopkins]

What a tragic waste of fast typing speed. If only you were using Emacs, your typing speed would be multiplied by Emacs's superior capabilities, with multiple shell windows and keyboard macros and many other powerful packages thanks to its deep and flexible extensibility, and you would be so much more productive and powerful! Typing speed isn't everything. ;)

Disclaimer: My cat is named Emacs, so if you say anything bad about Emacs, I will take it personally, because he is such a fine cat, named after such a fine text editor.

PS: https://github.com/zerolfx/copilot.el

[OOPMan]

Are you guys seriously involved in a WPM pissing contest?

[bheadmaster]

Do you know how many hours I've spent grinding gtypist? Of course I'm going to involve myself in WPM pissing contests.

Not doing so would be like going to the gym and not posting your muscle flexing photos on social media. It's just taking away the joy.

[disqard]

Ssshhhh! You just scared away a 200+ wpm person from adding on to this interesting conversation.

[radlad]

The point I was _actually_ trying to illustrate was that wpm and vim are irrelevant, and that Copilot is worth a try even if you're a fast typist that uses (your favorite editor here).

What does this comment add to the conversation?

[bheadmaster]

The initial question was "Is there really a sizeable number of devs who don't intend to ever rely on Copilot?", not "Are all Vim users who type fast anti-Copilot?".

I don't mean to dismiss your data point, just to put things in context. Since Copilot is generally recognized as a useful tool, there will of course be Vim users with great typing speed who still find it useful regardless.

Now, when it comes to Go, I find that there isn't much repetitive code to write (especially since generics landed in Go 1.18). Some people say that error handling is repetitive, but I find that those people just bubble up errors without adding appropriate context, which makes them less useful. But I haven't personally found a scenario in which I explicitly thought "damn, I know exactly what I need to write, but it's so long - I wish someone would write it for me".

[radlad]

And yet the comment I'm replying to didn't say "I won't ever rely on Copilot," it said "I can say with confidence that Copilot isn't that useful to me."

I made my comment because I hope others who are fast typists, and familiar with their tools, do give Copilot a try. I expected to hate it, didn't try it for a long time, and was quite surprised when I did.

> But I haven't personally found a scenario in which I explicitly thought "damn, I know exactly what I need to write, but it's so long - I wish someone would write it for me".

Test case setup comes to mind. Another place it's useful is for writing long function-interface signatures. Or adding a bunch of similar "case" statements to a switch.

Generics still have a ton of limitations.

[makeitdouble]

In an ideal world I'd choose stacks with not enough boilerplate to warrant copilot.

I had my share of auto generation with enterprise Java stacks, and tried as hard as I could to move to stacks where what we write is concise and relevant (rails is the closest I came to this, not perfect but clearly going in the right direction).

I think AI has its place, but I also hope to be lucky enough to not have to use it.

Illustrators might have similar issues, where some of them need to produce boilerplate drawings a lot, but I think they'd also prefer working on project that aren't that.

[bee_rider]

Yeah these Copilot type tools shouldn’t help much if a language is well-designed, or a project is structured in a fashion that doesn’t require a ton of boilerplate. If we’re doing things well, we’ll only have to tell the computer something once, right?

If it is possible to guess what we’re going to write, then we aren’t transmitting much information to the computer.

Copilot seems to be very popular though.

[dartos]

Well think of how AI was trained. GitHub trained copilot on _mostly_ open source data with permissive licenses.

Giving away code and their rights for free is commonplace. Also it’s not like you can use “by Ryan Dhal” to make the output from copilot better.

But these art AI were trained on. CC, CC-BY, and closed license pieces of art. And you can use “by Greg Rotowski” to get art in that artist’s specific style.

I don’t think comparing the use of AI or the general attitude towards AI between artists and devs makes sense. Very apples and oranges.

[hutzlibu]

I see the licence issue, so hypothetically, if a LLM was only trained on code that explicitely allowed LLM training, would you use that?

[bee_rider]

Hmm, is it possible to include a bit of prompt in copilot, something like “follow the formatting guidelines of the Linux kernel,” or something like that?

I dunno if it would help, but people do seem(?) to be improving their ChatGPT responses by telling it to answer as if it is an expert on a topic.

[dartos]

I’m sure it is, but the Linux kernel is freely available already AND it’s much less personal than directly naming someone to use their style

[jwells89]

I’ll use ChatGPT to help surface docs and examples or to get a very rough high level overview if what a task might involve, but I don’t really want an LLM in my IDE.

Not only does it feel like something that might be dangerous to become reliant on (what happens when it’s not working or I don’t have access to it), I have no idea what material it was trained on which makes it ethically gray. I might be more receptive to a local LLM where I can personally vet what it was trained on (primarily, I’m concerned with if the material was obtained fully consensually or not).

My attitude towards image generators is similar. Adobe’s is totally out of the question for example, because though they claim it’s 100% ethically trained because all material came from their stock image service, I know that’s bullshit because I’ve seen stolen art put up for sale there more times than I can count (and worse, they’re unresponsive when theft is reported).

[tambourine_man]

I’m a Vim guy and I’m very impressed with VS Copilot, but:

- Not enough to switch (yet, at least)

- I would have to carefully review the generated code, which is not as fun writing

The license issue is something I expect will be solved in the next few years (a dropdown menu to choose from, maybe).

[happymellon]

> The license issue is something I expect will be solved in the next few years (a dropdown menu to choose from, maybe).

I'm not sure it will, as everyone who uses it don't appear to really care about other people's licences anyway. It's just a method of BSD washing GPL code.

[Filligree]

You can use copilot with vim, no problem.

[freedomben]

Yes, original copilot, but not new copilot with expanded features. I'm a vim person and wold rather give up copilot than move to vs code, so I do hope they aren't going to leave vim behind and focus only on vs code moving forward.

[RunSet]

I see a different future where people continue to write their own code rather than trust Microsoft AutoPlagiarist™. Perhaps I am wrong and this no-code solution will at last relieve us of our onerous cognitive burdens.

[nine_k]

There is a big difference between AI generating a piece of picture which you completely observe, and AI generating a piece of code that may contain a subtle, hard-to-spot bug.

In both cases you're also risking being accused of plagiarism, when the model literally remembers, or reconstructs, a piece it has seen, and finds it perfectly matching your request.

I think "AI" tools in Krita may have their place: object detection and selection / tracing, upsampling, seamless resizing, cutting and pasting, texture generation, light adjustment, stuff like that. An integrated analog of DALL-E or Midjourney would likely be a poor for.