English as input language works in simple scenarios but breaks down very very quickly. I have to get extremely specific and deliberate. At some point I have to write pseudocode to get the machine to get say double checked locking right. Because I have enough experiences where varying the prompting didn't work, I revert to just writing the code when I see the generator struggling.
When I encounter somebody who says they do not write code anymore, I assume that they either:
1. Just don't do anything beyond the simplest tutorial-level stuff
2. or don't consider their post-generation edits as writing code
3. or are just bullshitting
I don't know which it is for each person in question, but I don't trust that their story would work for me. I don't believe they have some secret sauce prompting that works for scenarios where I've tried to make it work but couldn't. Sure I may have missed some ways, but my map of what works and what doesn't may be very blurry at the border, but the surprises tend to be on the "doesn't work" side. And no Claude doesn't change this.
I definitely still write code. But I also prefer to break down problems into chunks which are small enough that an LLM could probably do them natively, if only you can convince it to use the real API instead of inventing a new API each time — concrete example from ChatGPT-3.5, I tried getting it to make and then use a Vector2D class — in one place it had sub(), mul() etc., the other place it had subtract(), multiply() etc.
It can write unit tests, but makes similar mistakes, so I have to rewrite them… but it nevertheless still makes it easier to write those tests.
It writes good first-drafts for documentation, too. I have to change it, delete some stuff that's excess verbiage, but it's better than the default of "nobody has time for documentation".
Exactly! What is this job that you can get where you don't code and just copy-paste from ChatGPT? I want it!
My experience is just as you describe it: I ask a question whose answer is in stackoverflow or fucking geeks4geeks? Then it produces a good answer. Anything more is an exercise in frustration as it tries to sneak nonsense code past me with the same confident spiel with which it produces correct code.
"It's a translator. But they seem to be good/bad/weird/delusional in natural translations. I have a"
(Google translate stopped suddenly, there).
I've tried using ChatGPT to translate two Wikipedia pages from German to English, as it can keep citations and formatting correct when it does so; it was fine for the first 2/3rds, then it made up mostly-plausible statements that were not translated from the original for the rest. (Which I spotted and fixed before saving, because I was expecting some failure).
Don't get me wrong, I find them impressive, but I think the problem here is the Peter Principle: the models are often being promoted beyond their competence. People listen to that promotion and expect them to do far more than they actually can, and are therefore naturally disappointed by the reality.
People like me who remember being thrilled to receive a text adventure casette tape for the Commodore 64 as a birthday or christmas gift when we were kids…
…compared to that, even the Davinci model (that really was autocomplete) was borderline miraculous, and ChatGPT-3.5 was basically the TNG-era Star Trek computer.
But anyone who reads me saying that last part without considering my context, will likely imagine I mean more capabilities than I actually mean.
Enter technical specifications in English as input language, get code as destination language.