Notes on coding with ChatGPT
- This is real.
- There's a sort of endurance that coding with ChatGPT uniquely requires. You have to maintain the discipline to keep asking ChatGPT where appropriate; it's tempting to plow through a problem unassisted, especially when you've just asked ChatGPT for help a half-dozen times. You have to remember that ChatGPT is not getting tired or annoyed, and that you're not keeping ChatGPT from working on something more important. There's a certain social instinct that must be slightly repressed: entering into conversation of any sort prompts us to spin up various emotional capacities. Some of these are in fact good to maintain even when the interlocutor is ChatGPT, but many stop being worth the energy when it's not a human being on the other side.
- The skill of dealing with code you didn't write, which has always been an important part of software engineering, is perhaps now even more important, given that you won't even have written a bunch of the code that you wrote.
- There's a particular skill of remembering to double-check sources that are right perhaps 95% of the time. As I get better with ChatGPT, it has been harder to catch its mistakes, because its accuracy rate is sometimes so high that it comes to seem extremely convincing. (At Google there was an absolutely magical development tool that hot-swapped fixes into a development environment as you coded. It worked perhaps 99% of the time. I tended to lose hours chasing my tail when a 1% case hit.)
- If you are coding with Copilot, don't be afraid to type "This is confusing because" and wait for the autocomplete.
Home page
Published 2023-04-10.