People always analogize this, but I've found "pair programming" with LLMs to be extremely underwhelming. I've tried it with Aider+Claude, Copilot Chat, 4o and o1, and the experience is always the same: I spend most of the time correcting misunderstandings before eventually giving up and doing it myself.

To this day I haven't found an LLM application that works better than regular autocomplete Copilot. It's sufficient to get over the blank canvas problem in a lot of cases without being overly ambitious and biting off more than the LLM can chew.

I've yet to try Cursor, but that's because I don't have a lot of hope: I first heard the same kinds of great things about Aider and Claude and o1, and all of those have disappointed. It's hard to want to switch editors to see if this time is for real.

I used Copilot and dropped it when it stopped being free. But Cursor is a different beast - the autocomplete is far better and intuitive. The chat functions like a personal StackOverflow (or just paste an error message for debugging).

For me as a senior eng, Cursor is where AI turned the corner from "maybe helpful for fringe / basic things" to "actually amplifies my productivity". Took about 30 min to flip the switch for me, so I suggest you give it a try.

The trouble is that I heard all of this—personal stack overflow, error messages—already with Claude, and I'm skeptical that anything Cursor can do on top of that will be worth losing the productivity of the JetBrains IDEs.

I imagine it's easier to switch for someone who's already making do with VS Code, because at that point the main hurdle is just being willing to pay for an editor. I already pay for an editor (a bundle of them), but "VS Code with better AI integration" just... doesn't appeal, especially when the word-of-mouth recommendations say all the same things that I've already heard about other tools that didn't work for me.

Same.

I had the perfect use-case for LLM-assisted coding a few days ago: I had a well-defined function to implement, for which dynamic programming was a great fit. I haven't used DP in years but it's a well-trodden path, why not throw the problem at the LLM and let it implement it?

Well despite careful prompting, it kept getting it wrong. At some point it generated some credible code, but I spent the day finding problems, nudging it to fix it, asking for explanations for stuff that looked (and was) incorrrect... Resulting code was bug riddled and horrible (LLMs tend to fix problems by adding more code rather than rearchitecting the existing code to eliminate edge-cases). I ended up spending a whole lot more time, I had to spend ages carefully nudging the LLM to no avail, understand LLM-generated garbage code _and_ still solve the problem myself.