Not Using Copilot
macwright.com
macwright.com
Humans write too much code at times too. The difference is when I tell a developer they need to be more concise it’s a learning process and takes days per feature to complete and months for the developer to improve significantly. With copilot you can literally just tell it to be more concise and it will do it, straight away, no fussy behavior, no denial, no blaming coworkers, no need to adjust, it just does it.
The author would know this if they tried using the tools. But this is just a grump being grumpy like people still claiming that IDE’s cause brain rot and that assembly is the only “true way to program”. Copilot is just a tool, but it’s an incredibly performance boosting tool, and it doesn’t really have any significant learning code. People should just adopt it and move on with their lives.
Where do you think copilot learned all that behavior, exactly? Like it's funny you sit at the tail end of it and claim it's not a learning process, but literally the core of any LLM is learning from other human behavior. It cannot exist without that learning process.
> The author would know this if they tried using the tools. But this is just a grump being grumpy like people still claiming that IDE’s cause brain rot and that assembly is the only “true way to program”. Copilot is just a tool, but it’s an incredibly performance boosting tool, and it doesn’t really have any significant learning code. People should just adopt it and move on with their lives.
No it's not. I've used the various LLM tools, I've had coworkers use the various LLM tools. When you start to tally up the time spent micromanaging, sifting through jank code and tuning it to get what you want the end result is in my opinion more effort for less. The idea that people should just 'adopt it and move on' seems like you can't actually convince people of its value proposition.
If you want to just shit out some code that barely works for a basic SPA then you can maybe work faster. But you're paying the cost in a larger amount of technical debt that will hit you later [1], with no one actually knowing what the codebase is doing.
[1] https://asim.bearblog.dev/how-a-single-chatgpt-mistake-cost-...
I just can’t believe that to be the case form anyone who has actually spent significant time with real code in real projects. I get that this is someone would say to dismiss it if they hadn’t used it. Like weavers claiming the mechanical loom will introduce too many errors to be with it or typesetters complaining about the kerning of automated processes. Even if you hate all the functions, classes and procedures generated by copilot, it’s a huge productivity gain for the context aware auto completion alone.
> If you want to just shit out some code that barely works for a basic SPA then you can maybe work faster. But you're paying the cost in a larger amount of technical debt that will hit you later [1]
I never once complained that my typechecker couldn’t write a SPA on its own, I never once complained that my linter couldn’t write and SPA on its own. Why should I complain that copilot isn’t capable of doing literally everything unsupervised. It’s cutting away at least 40% of the time I used to spend coding, that’s amazing, why would I not use it just because some set in their ways programmers stubbornly refuse to see the value without even using it?
> The idea that people should just 'adopt it and move on' seems like you can't actually convince people of its value proposition.
For unit testing I needed to explain the value propitiation. For automated build pipelines I needed to explain the value proposition. For copilot I don’t have to. It’s literally just a case of “turn it on, work as usual and you’ll see what all the fuzz is about and it’ll save you tons of time” I have yet to have a single developer who actually started using it need to have the value explained to them, it’s obvious, it’s right there, it’s not a matter of pro’s vs cons.
The only people who want to be sold on it are the ones who won’t turn it on to begin with and instead want some meaningless argument about “the value of real inteligenge” no one ever wanted that argument for unit tests or for ci/cd. I refuse to have those discussions now, it’s just a tool, flip it on, continue to work. If you’re worried AGI will replace you, then not wanting to use copilot makes absolutely no difference either ways. The only thing anyone accomplishes by stubbornly refusing to pick up the major advances we’re getting these days is that they get fired because a fresh out of uni dev is outperforming their 20y or experience and management is never going to buy the “using AI is cheating” mentality they these neckbeards are pushing.
But I see your worry, tech debt is a very real phenomenon which didn’t exist before ChatGPT and therefore people need to decide do they want the old way of always flawless software or the new reality of bugs, tech debt and bad developers which didn’t exist before. /s Of cause you’ll see bugs in code written with copilot just like you’ll see it in code without copilot. But if you have competent people the difference will only be how quickly you get through it not if it’s perfect or not.
The sidebar is a useful tool sometimes, akin to pair programming with a junior programmer who's terminally online and brimming with random ideas.
Where I think they've really hit a sweet spot is with the new Inline Chat feature, I can select a chunk of code, hit CTRL + I, and tell Copilot I want to refactor this code to do XYZ. Suppose you need to extend a function with an extra parameter, the logic to handle that, and some appropriate error handling. You can select your function, type that in one sentence, and Copilot's inline chat will mostly just get it right and propose an amended function for you to accept. This has been a big time saver for me lately.
Notably Copilot is NOT doing much, if any, thinking for me. You really don't have to push it far into unknown territory to turn it into a gibbering lunatic. So in the sidebar it's essentially doing a (mostly) better version of a web search than Google. In CTRL + I it's kind of a macro on steroids. But these are both great tools and very easily justify the subscription fee.
The actual link is https://macwright.com/2024/11/20/not-using-copilot
I have heard more comprehensive “leave the thinking to us” tools like cursor give better results.
Personally I think if your tools or projects are so dull that you require an AI to use them, perhaps you need sharper tools or more interesting projects.
But git use is rather rudimentary anyway :)