I’ve been using it since the execs at my job have been pushing engineering hard to pick it up. Like any tool, it’s useful sometimes but not all the time.

The line-by-line recommendations are generally helpful and it’s great at small tasks such as writing shell scripts or generating boilerplate for a new project. It almost never gets anything right the first time, and you have to keep telling it that it’s wrong for it to fix all of the problems.

For fun, I did ask it to fix something with my codebase that genuinely had me stumped. It came up with a very convincing solution, but it didn’t work at all. So yeah, it tends to fall flat for hard problems.

For all the AI hold-outs here, please be aware that Copilot Enterprise provides very granular details of Copilot usage per user to your leaders. i wouldn’t put it past any company to factor that into the next layoff decision when it comes time.

I don’t use copilot itself, but I do use an IDE that has AI assistance built in. I have tried all the parts of it, and the only thing I keep enabled is the single line prediction. Even that is right only ~65% of the time, but it’s nice to just hit tab and have it fill in what you were about to type. I also use it to get a first draft on comments and commit messages. Basically trying to use it to remove tedium, it sucks at anything remotely complicated.

Yes. Used it to comment the hell out if an old nearly dead code base. Used it to create markdown wikis for it based on all the comments. Used it to convert an old asp.net site to blazor. My son’s using it to learn how to make minecraft mods. Im using it to make a tv channel plugin for kodi so i dont constantly have to change shows. I dont have enough time to go through all the ideas.

I am not a developer. I have a tenuous grasp of my own language. I don’t understand how Grok can be in Copilot.

That said, my worst AI experience has been copilot. Chatgpt is a great sounding board, a bit wordy but mostly good for organising projects and day to day suff.

I avoid google , Grok is funny to fuck with Elon but copilot will fuck up an excel sheet everytime I ask it to do something.

It is amazing how it gets formulas wrong that chatgpt gets right despite Microsoft having the inside track.

I have never come away from copilot and been satisfied it was worthwhile. And now my 365 subscribed company has blacklisted chatgpt.

Yes, I used it to vibe code a JavaScript single-page application using the Scyfall API to get Magic: The Gathering card set data. The app makes printable inserts for card binder sleeves so you know what cards go in the holes. I made it for my friend who owns a game store.

It knew everything about the Scryfall API. I never had to read the docs. I’m also not very good with JavaScript, CSS, or HTML5. It used elements I had never even heard of in the HTML. Data like the set list and individual card lists are cached and DOM storage to avoid repeat hits to the API. I had to know to ask for such things, but when I did, it gave me all the code I needed

It even generated an SVG logo based on my description, although that took a lot of back and forth, and I recently ended up installing inkscape and fixing it myself to make it look better, but it really wasn’t a bad start.

The page has advanced auto-complete stuff that I never would have programmed myself (out of lazziness). You can start typing the name of a set and it will automatically pick the set from a list box if it matches the set code or the set’s name.

I had to do all the card grid layout math myself, though, for the printable inserts, because the math was pretty specific. I wanted to make sure the printed “cards” made a perfect grid, even if the printer’s margins were not equal on all four sides. I mean, it definitely helped, but it kept getting it wrong, and I had to explain it in detail how to do it, and then it would do it ok. I really wanted to see if I could write none of the code, but I ended up having to write a tiny amount.

Was it perfect? Hell no. I had to hold it’s hand a lot and tell it to stop being dumb. I also had to make a custom agent in VS Code to give the AI a back story. I had to tell it things like, “you don’t just put new code at the top of the file willy-nilly, you put the code in the section where it belongs in the file.”

If I weren’t a professional programmer, I think it would have been a bad time, but as it was, it was great. I definitely would have been too lazy to make this website if I didn’t have the AI to do it for me, since I wasn’t being paid for it. Or I would have made it, but it would have been a really meager, featureless Python script that just spits out basic HTML for printing.

Thing is, I kept adding features and it kept doing it well. Like towards the end, I wanted it to add URL hashes to keep track of where you were in the application so you could bookmark specific pages or email them to people. (Plus it would enable the back button to leave “print mode” that hides the whole UI.) It’s successfully added that feature the first try.

It’s open source and the website is hosted on GitHub pages. You can take a look and tell me how bad the code is. 😂 It’s definitely not optimized for mobile, but it’s “usable”. Try it on desktop. It’s specifically made for printing, anyways, and I don’t know many people who print from mobile phones.

https://github.com/CamelCityCalamity/magic-binder-placeholders

(A link to the page is in the readme.)

The friend I made it for loves it and has used it several times, and will likely keep doing so in the future.

I could make it work with the default/base model (GPT 4.1) and it’s pretty decent. It’s even better with Claude 4 Sonnet but that will use up your credits and the $10 plan gives you (I think) 350 questions a month with these so-called premium models, while with the base you have literally unlimited queries, which is something you don’t really see much with these coding agents (believe me, I’ve looked!).

So I’d say the base model is at around 80-90% of the premium, and it’s free. Most of the time it works just as well. It’s most obvious when you push it or start piling up the code and it gets messy. I initially built a feature in a web app with the base (GPT) which took around 2000 lines total, and after realizing that there was a lot of repeated code and such, I asked the premium (Sonnet) to analyze the code and tell me how it would refactor it… And it did. In one go. Cut it down to 1000 lines. And it actually worked! I know luck factored into it, but that right there proved why Sonnet is superior to GPT. And I used both through Github Copilot.

Basically what I’m saying is that Github Copilot is essentially Cursor if you use the Sonnet model. But the problem is that you can’t get as many Sonnet queries with Github as you get with the Cursor $20 plan (which is around 2000 questions a month). In my case, since my company pays for both, I switched to Cursor. For free Claude access with an obscenely high daily token limit, I recommend Rovo Dev from Atlassian (the JIRA company).

the more vocal support he gets the more likely it is he can sue for wrongful dismissal, and ultimately we all need to accept that faceless corporations are the enemy and we need to hold the people in charge accountable

Fecursion.

SHIT. I thought those flake88 errors were a typo! Time to anti fash the git bash.