[ArchitectAnon]

Here's my perspective on this as an Architect: Most construction details have been done before, they could be easily reproduced by an AI surely? There's usually just a few things that are different from the last time I have drawn it. A few 3D interactions with other components that need to be reasoned about. They are not that complicated individually.

But yet I see this problem as well just using old fashioned automation let along AI to save time. I find that if you haven't drawn the 2D section through all the different edge cases of a particular thing you are trying to design, you haven't done the analysis and you don't really understand what's happening. I've made mistakes where I've been working in 3D on something complicated and I've had to hide some element to be able to view what I'm working on, only to find later that when I turn everything on again I've created a clash or something impossible to build. That's why we still do 2D drawings because they are an analysis tool that we've developed for solving these problems and we need to do the analysis, which is to draw section cuts through things, as well as building 3D models. After all, if models were such a good way to describe buildings, then why weren't we just building physical scale models and giving them to the builders 100 years ago; it's because you can't see the build-up of the layers and you can't reason about them.

Reading this article I get the same sense about software engineering, if you haven't solved the problem, you don't really understand the code the AI is generating and so you don't really know if it is going to do what you've tried to describe in your prompt. You still have to read the code it's generated and understand what it is doing to be able to tell if it is going to do what you expect.

[IsaacL]

> Reading this article I get the same sense about software engineering, if you haven't solved the problem, you don't really understand the code the AI is generating and so you don't really know if it is going to do what you've tried to describe in your prompt. You still have to read the code it's generated and understand what it is doing to be able to tell if it is going to do what you expect.

Yes, this is pretty much exactly the way I've been using GPT and it works tremendously well. (GPT4 works especially well for this style of programming.) My prompts include things like:

- "read the function below and explain in detail what each section does" -- this prompts GPT to explain the code in its own terms, which then fills in its context with relevant "understanding" of the problem. I then use the vocabulary GPT uses in its explanation when I ask it to make further changes. This makes it much more likely to give me what I want.

- "I see this error message, what is the cause? It appears to be caused by $cause" -- if I'm able to diagnose the problem myself, I often include this in my prompt, so that its diagnosis is guided in the right direction.

- "this function is too complex, break it up into smaller functions, each with a clear purpose", or "this function has too many arguments, can you suggest ways the code could be refactored to reduce the number of arguments?" -- if you go through several rounds of changes with GPT, you can get quite convoluted code, but it's able to do some refactoring if prompted. (It turned out to be easier to do large-scale refactoring myself.)

- "write unit tests for these functions" -- this worked phenomenally well, GPT4 was able to come up with some genuinely useful unit tests. It also helped walk me through setting up mocks and stubs in Ruby's minitest library, which I wasn't experienced with.

In brief, if you expect to just give GPT a prompt and have it build the whole app for you, you either get lame results or derivative results. If you're willing to put the effort in, really think about the code you're writing, really think about the code GPT writes, guide GPT in the right direction, make sure you stay on top of code quality, etc, etc, GPT really is an outstanding tool.

In certain areas it easily made me 2x, 10x, or even 100x more productive (the 100x is in areas where I'd spend hours struggling with Google or Stack Overflow to solve some obscure issue). It's hard to say how much it globally increases my productivity, since it depends entirely on what I'm working on, but applied skilfully to the right problems it's an incredible tool. Its like a flexible, adaptive, powered exoskeleton that lets me scramble up rocky slops, climb up walls, leap over chasms, and otherwise do things far more smoothly and effectively.

The key is you have to know what you're doing, you have to know how to prompt GPT intelligently, and you have to be willing to put in maximal effort to solve problems. If you do, GPT is an insane force multiplier. I sound like I work in OpenAI's marketing department, but I love this tool so much :)