[agentultra]

It can generate example data for tests but why would I want that when, if I use a good property testing framework, I already have that; and it can shrink examples sensibly so that I get the minimal failing example for my test?

Generating code seems worse too. I agree that you don’t need “perfect,” software if you’re building a blog for your local soccer team. Maybe you want more precaution if you’re initiating trades in securities. And definitely a lot more scrutiny if you’re building real-time control systems for helicopters. Verifying that code is correct with regards to its specification first requires a precise, formal specification; and second, is still a lot of work.

Eyeballing every line of code to catch errors ChatGPT introduces into your code isn’t going to work. You’re not a compiler or theorem prover and you’re bad at finding errors in code. Really bad. It’s not your fault; code becomes incredibly complex quickly.

At best, right now, it’s a curious toy and maybe useful for some folks as a “smart” autocomplete. But beyond that ChatGPT seems to struggle. And I struggle to find a use for it myself.