OpenAI GPT-4 - wow

AI based code generation is moving very fast.

Hopefully soon we can ask OpenAI to fix the Delphi LSP + compiler bugs that haven’t been fixed in 5 years.

1 Like

Make sure that you validate any code that you get from her. (Ie thoroughly unit test it) I had her generate me a small method which didn’t work over some basic combinations. I told her that and she made some adjustments to the original code , and then again some more and then eventually she said that she couldn’t do it.

It’s impressive but like the presenter says in the video (around the 15min mark),

“it’s an augmenting tool, it’s still important you’re in drivers seat knowing what’s going on.”

Will it ever surpass that role? For building complex systems I don’t think so, but perhaps mostly because of the human side of the equation rather than the AI side.

A complex system requires a complex specification. How many of your customers are capable of properly specifying their software specifications without you being there to ask the probing questions of them? That early role we fulfil of being the interrogator rather than just a builder isn’t something I see reflected in these AI systems capabilities at all.

Perhaps programming may just change in nature, rather than writing precise programming code, we may find ourselves having to write precise English specifications for an AI to work from. At that point you could maybe argue that you’re just using a very high level programming language.

If that happens you’re still going to have people who’s job description is some sort of developer, because you know your customer is never going to have the patience to fully spec out their requirements, even if it is in English.

Also right now there seems to be very quickly diminishing returns once you get past the clean slate of a new project. This will keep improving but could an AI ever fully grasp the purpose and inner workings of a 1 million line code base with all its years of associated real world history, coding styles, contradictions, errors and workarounds?

Helping you write some new methods for a large existing codebase, sure. Performing system wide modifications on that codebase, I’m less sure of. I’d love the AI’s help finding the code to be modified, but I don’t think I’d trust it to make those changes itself.

Also something to be aware of when using ChatGPT (and maybe GPT-4). I have read that anything you enter into it will be used to help further train the model. So before you paste your trade secrets into ChatGPT for it to improve, be aware you may be giving them away for future ChatGPT users to discover.