Comment by idopmstuff | Hacker News Reader

Comment by idopmstuff | Hacker News Reader

Parent Full thread

I'm a PM who is relatively technical, but I haven't written more than a few lines of code (stuff like minor modifications to my Shopify site) in a decade.

I saw a bunch of people talking about how GPT helped them code stuff on Twitter, so I thought I'd give it a try. Right now I'm building a sort of simple, mock version of the type of software that integrates with my company's APIs. I've successfully managed to create a simple web application that creates a new object, hits my company's API endpoint to create a corresponding object on our software, allows me to upload a document locally and then allows me to upload that document to our software via API as well. It's all a little messy and clearly not production-ready, but it works. It would've taken me probably a few months on nights and weekends do this (mostly refreshing myself on JS and Python). Instead I've done it in <24h (would've been shorter except for GPT-4's message limit).

I'm sometimes able to spot and fix GPT's bugs, but even when I'm not, it walks me through adding more logging and successfully debugs issues. Sometimes it takes a few tries and a little direction as to what I suspect the issue is, but so far it's fixed everything that's come up. I don't think this would be doable for a totally non-technical person, but I do think it'll get there pretty soon.

I'm just absolutely blown away.

One of the reasons I'm not blown away is that everything we've tasked it to do has resulted in rather terrible answers. A lot of them outright didn't work. When they did, the way GPT created our solutions were often in a way that wouldn't work well over time. Unfortunately you wouldn't necessarily know that unless you really know your tools. In many ways, this isn't too different from people performing brute-force programming through their favorite search engine, but at least most people know that "Google programming" is sort of bad. I think we're going to be cleaning up the GPT messes for decades to come because it's very confidently incorrect and much more accessible as you point out.

I think we're going to see a lot of programmers who are going to trust GPT a little too much, and I think that's sort of scary. For the most part that is going to work out just fine. Often the quality of your programming isn't actually going to matter that much, because as long as it solves the business needs okish, then it's frankly great. That's not always the case, however, imagine someone using GPT to get your healthcare software wrong.

I'm still impressed with it in other areas. I think it'll do wonders in the world of office automation because it seems to have the ability to succeed at this much better than any previous "no-code" attempt where the logic would almost always end up requiring people who are basically programmers for it to work. I think GPT will help here, requiring less "superusers" for a department to move their data flows into automation. Especially in areas, where efficiency and stability aren't necessarily that important if the automation-tools mean you don't need three full time employees moving data from one system to another.

Have you tried gpt-4? It's not perfect but it is a clear improvement for code.

Are people actually allowing it to perform actual software development?

State your assumptions.

Read and summarize the pervious documents

generate a data flow diagram.

generate a data model.

Get it to inquire about use cases and requirements

generate tests for these uses cases and requirements.

I've played with getting it to ask me about my requirements. I think that was on ChatGPT3.5? It wasn't great, but it wasn't horrible.

> ...but at least most people know that "Google programming" is sort of bad. I think we're going to be cleaning up the GPT messes for decades to come because it's very confidently incorrect and much more accessible as you point out.

Speaking of which: https://meta.stackoverflow.com/questions/421831/temporary-po...

> Overall, because the average rate of getting correct answers from ChatGPT is too low, the posting of answers created by ChatGPT is substantially harmful to the site and to users who are asking and looking for correct answers.

> The primary problem is that while the answers which ChatGPT produces have a high rate of being incorrect, they typically look like they might be good and the answers are very easy to produce. There are also many people trying out ChatGPT to create answers, without the expertise or willingness to verify that the answer is correct prior to posting. Because such answers are so easy to produce, a large number of people are posting a lot of answers. The volume of these answers (thousands) and the fact that the answers often require a detailed read by someone with at least some subject matter expertise in order to determine that the answer is actually bad has effectively swamped our volunteer-based quality curation infrastructure.

If we're lucky ChatGPT will poison itself by pissing in its well, but it will take a lot of good things with it.

> I think it'll do wonders in the world of office automation because it seems to have the ability to succeed at this much better than any previous "no-code" attempt where the logic would almost always end up requiring people who are basically programmers for it to work. I think GPT will help here, requiring less "superusers" for a department to move their data flows into automation. Especially in areas, where efficiency and stability aren't necessarily that important if the automation-tools mean you don't need three full time employees moving data from one system to another.

Yeah, at this point I think this is a valid use case for GPT-4 in its current form. I would be comfortable using it to build internal process tools or standalone things like a simple browser extension. Nobody in engineering at my company would be dumb enough to let me start monkey around with our actual codebase though.

> everything we've tasked it to do has resulted in rather terrible answers

Maybe this is the chatGPT equivalent of "learning to google search properly". You got bad answers, but maybe someone more competent at chatGPT prompts and workflow would have gotten to a better solution more quickly, and we need to figure out what that means

Is there a primer on how to engineer prompts? I didn't like my results, tried engineering my prompts a bit, but it kept introducing different error. I had the domain knowledge to see them, but it felt like whack-a-mole.

There are a lot of posts of different prompts on some subreddits. I have seen people selling a live course on prompts to make titles for social media or blogs.

There is also some research to show giving examples and answers sets up ChatGPT to give better results in the style of examples.

But overall, I have not seen anything that covers a process that would work most of the time.

I've never seen so many people grinding to make an autocomplete engine produce "better solutions" to randomized output.

I think your parent is on to something: it won't be the programmers trusting GPT a little too much, it will be the PMs, and there won't be a programmer :-)

I see an Idiocracy inspired future for programming.

People claim that AI can write code so they start firing programmers. Universities stop software engineering programs as there is no one taking the courses. People stop writing blogs or stackoverflow. Software engineers either move to other fields or start living offgrid. No new innovation or new line of code written by human.

Meanwhile, software quality get worse by each passing day and there’s no one to fix. AI poisons it’s own well by generating shitty code and now even simple tasks are taking 30 seconds. People say, “In the good old days, we used to get response in under 1 second”. Just like how they talk about cars and their durability in the good’ol days.

It seems like you're assuming that the only way for it to get better is more human-written code. Also, why would software quality get worse over time, as opposed to stagnating at worst? Attempting to use it for more complex applications?

There is already so far beyond enough data available online for an AI to be a super-human senior software engineer/architect/whatever. Because there's enough for the humans, and an AI can do it better in theory. It just needs to be a good enough model.

Barring the most extreme edge cases, it's never going to be worse than it is right now.

Yeah, but we don't have that kind of tech yet. When someone figures out how to build something that learns like a human does--where, even though all the books I ever read sucked, I am still a better programmer than any of those authors ever were and notably better than I was yesterday despite not reading anything new--then the world is going to change in a fundamental and potentially devastating way... but while that day might come literally tomorrow, it also isn't obvious it will ever come, and just because GPT exists doesn't imply it is going to happen. GPT is not that, though, and no matter how much data you throw at it it isn't going to become that without a fundamentally different architecture.

Because AI doesn’t have it’s own mind. It’s gonna put some bugs here and there, sooner or later, and then it’s gonna repeat that same thing everywhere, after enough iteration the bug will be so prominent that it’s gonna be a pain in the neck for user (also the premises of almost every AI bad Sci-fi movie/book). It’s a blackbox. It’s like infection spreading.

And because of the corp greed. Cheap solution will always triumph good solution, so I am pretty confident that they will remove the guardrails way sooner that they should. I mean Facebook literally fanned the flames in Myanmar[1] so my bet is on corp greed this time too.

1. https://www.amnesty.org/en/latest/news/2022/09/myanmar-faceb...

The biggest difference between a beginner and a veteran is systematic problem solving. Software development has already been reduced to bashing libraries together like rocks... It still is very effective, but frustrating to the end user. If software development gets reduced to brute forcing a prompt until it starts to look like it's working, then we might as well call it a day and promote the users to eternal QA.

It's already becoming a strain to review ChatGPT code committed by others, who try to save time by using it instead of thinking for a moment and then coming up with a proper solution to a problem.

>>it will be the PMs

IMO PM's are the clueless bunch of the whole set up. Often they are just making up things as they go and making it right through 'authority'.

PM's remind of me of early days of Sudoku craze when people thought they can put in any numbers in any squares and then later on puzzle would automagically rearrange itself to help them win. Backtracking never clicked for such people.

Once you are beyond simple toy apps, and have to build large engineering artifacts in iterations and realise it can't be from scratch in every iteration. The real fun begins.

> I'm still impressed with it in other areas.

You mean in areas you are less expert in? :) Maybe it's bad at everything, but each person can only tell when it comes to their own area of expertise.

It'll definitely help people learn faster - the "what to try next" problem is HUGE in programming discovery.

Modern languages (and tools like autocomplete) have already helped that a lot compared to assembly code or binary, this looks like the biggest jump in a long time. The path of programming so far has been moving from "describe how to do something" to "describe what to do" which this is certainly in line with.

Honestly, it seems the people most blown away by things like this are people who haven’t much experience in the various fields.

It’s easy to get amazed by something that can halfway do something you can’t do at all automatically. But as others have pointed out, it’s not that great at it and not knowing enough to do it yourself means you don’t know enough to catch and fix bugs.

So this move to using chatgpt and similar in production by people who otherwise wouldn’t be able to do things in production is worrisome, imo.

Reply on news.ycombinator.com