GitHub CEO: manual coding remains key despite AI boom

AI is good for boilerplate, suggestions, nothing more.

For you, perhaps.

For me, too. On the poorly-documented stuff that I get into, where I really need help, it flails.

I wanted some jq script and it made an incomprehensible thing that usually worked. And every fix it made broke something else. It just kept digging a deeper hole.

After a while of that, I bit the bullet and learned jq script. I wrote a solution that was more capable, easier to read, and, importantly, always worked.

The thing is, jq is completely documented. Everything I needed was online. But there just aren't many examples to for LLMs, so they choke.

Start asking it questions about complexity classes and you can get it to contradict itself in no time.

Most of the code in the world right now is very formulaic, and it's great at that. Sorry, WordPress devs.

It's a powerful tool, but it's not that powerful.

Have you tried feeding it the documentation of jq? And which LLM(s) have you tried?

I had great success with niche programming languages by feeding it the documentation and a couple of projects.