[shon]

I disagree with the idea that handwritten software falls into the same category has handmade woodworking. Unlike software, there is no machine that can do the type of woodworking you describe.

At least for now, we still have the unsolved 'robotics problem' where the data to train humanoid robots to do woodworking better than a human doesn't exist in sufficient quantity to make it work using current AI training techniques. This too may soon change via synthetic training data or some modified technique.

Conversely, my own experience shows that even with context window limitations, ChatGPT is already better than 90% of working software engineers for generating small practical programs. Here's why it's better:

1. It's better because it not only knows how to code but is equally knowledgeable about product management, UI design, business practices, etc which makes it more like interacting with an 'architect level' engineer 2. It knows a good if not 'the best' way to do most things

[JohnFen]

> ChatGPT is already better than 90% of working software engineers for generating small practical programs.

Not sure how you measure this. However, if by "small practical programs" you mean "trivial programs", then I see no reason to doubt your assertion.

That said, at least where I work, ChatGPT hasn't been a real benefit for real work. By the time you manage to craft the right prompts, then check and correct the code produced, then incorporate it into the larger project, it all takes a bit more time and effort than if you just wrote it yourself in the first place.

[shon]

100% John, I do mean trivial programs.

It’s the stuff that you should not bother an engineer to do, but it can’t be done (in a reasonable timeframe) by a business person or even a technical IT person who isn’t proficient at code.

There’s a space there, a practical need for things that are boring or obnoxious for an actual SWE but that fill some business need or use case. ChatGPT fills these in nicely.

[S201]

> Unlike software, there is no machine that can do the type of woodworking you describe.

A CNC?

> Conversely, my own experience shows that even with context window limitations, ChatGPT is already better than 90% of working software engineers for generating small practical programs.

Well if we're doing anecdotal experiences, I find ChatGPT (both with 3.5 and 4) useful for replacing quick search queries to look up documentation but for anything more complex than that and it quickly starts making up blatantly incorrect examples even for small programs making it more than useless most of the time for me at least. But you're welcome to disagree; I don't have the energy anymore to get into my 1000th internet discussion full of anecdotal evidence trying to validate the future merits of generative AI for software engineering purposes.

[shon]

> Making a coffee table out of a solid slab of walnut, for example, will last a lifetime

Not CNC, right?

Yeah I think it's fun to discuss. I think we're saying the same thing. I'm not saying ChatGPT could replace a professional software engineer for most things at this stage. That clearly isn't practical yet.

As you said, it does make mistakes, though it can often correct them if the context window is there. The differences in how we use it for programming tasks and what we find useful is probably related to our SWE skill levels. Yours is likely much higher than mine, as I've never been a professional SWE.

That said, for the many little automation tasks that business people often bug a real engineer for, like iterating through a CSV and hitting an API to do something, then writing a result, it's ability to write simple and useful python scripts has been amazing for me.

I believe it's getting close to the place where I could use to to prototype simple MVPs but it failed during my last attempt.