It has value only if you know what it is talking about, thus it can only assist.
If you ask it about AWS - you need to already know AWS, same with business etc.
And if you can't make your own elevator pitch - the odds that you will be able to run a successful company are little to none.
Website builders and ready-to-use templates have been around for decades and they only increased the demand for devs.
Even if GPT will be able to produce complete system from start to end - those systems will need support, fixes, etc. Sure you will need less engineers per system but the amount of systems will increase.
To add to this - GPT is good at synthesizing existing information, but it's bad at combining it in truly novel ways.
OP's startup idea is in a well-understood domain. There are lots of existing competitors offering basically that exact service. There are clear, obvious paths to building it. GPT can draw from blog posts and tutorials.
On the other hand, it does poorly with significantly novel ideas. I tried it with one I've been working on that combines 3D printing and crafting in new ways. Its answers just regurgitated blog posts about existing combinations. It even ignored the grammar of my questions, which switched the usual direct object with the object of the preposition. Which makes sense, because it isn't doing critical thinking - it's pattern matching.
This is my experience too. I've tried to do simple task like building small JS scripts and while it's doable, is like having a kid who needs oversight and can't clearly connect ideas.
And coding is supposedly an area where it shines.
GPT is waaay better than us humans at finding and summarizing information. Way way faster. But I've tried to write a blog post about some technical stuff and I finally gave up after many tries.
It's good. It's frightening already. But it's still not "there". It's like having a extremely fast but bit dumb virtual assistant.
Knowing about AWS was just a shortcut where steps were skipped because of prior knowledge. The user could have been more naive and asked more leading questions about “how do I write a program for the internet?” Or “what is cloud software?” Or “tell me about popular ways programmers solve problems for internet users” or some such. Become the 5 year old asking infinite “why?” And “how?” This is a conversational chat, not a one-and-done Wikipedia entry.
It really can't and it doesn't really work the way you think it does which is ironic that you'd think people on this site would be informed but I guess ycombinator has a financial conflict of interest in doing so. I wouldn't be surprised if they make posts on this topic with a very heavy bias like they do for a lot of their own investments.
I remember being shown this software almost a decade ago in dealing with chat in online video gaming and it ended becoming a serious conversation at the ACM SIG CHI about it's use and abuse for spreading certain things. At the end of the day, we realized it was an arms race and the only way to win, was to simply not play. But of course these ideas were rejected for it went against the financial interests of the parties involved.
The point is not what we (techies) think about the usefulness of ChatGPT. The point is that managers will definitely think ChatGPT is worth doing... Just like they think Jira/Scrum/Agile/etc. is worth our time. I definitely see management paying for ChatGPT.
It’s also impossible to imagine correctly. Back in 2009, I was completely convinced that we’d be able to go into a car dealership and buy a brand new vehicle which had no steering wheel because the self-driving AI was just that good, within 10 years. Seemed reasonable at the time on the basis of the DARPA Grand Challenge results, but even 13 years later, it didn’t happen.
I think this is the crucial point, just as in all AI applications the way it deals with corner cases will decide its impact on the job market. And corner cases are usually where AI has consistently been performing badly.
Just as a reminder, after we had spectacular results on ImageNet, highly respected AI researchers were predicting the end of the radiologist occupation. Turns out that even when some state-of-the-art CV classification algorithm is used on any kind of scan, you still need a radiologist to look at the image basically in the same way as before.
If you write large-scale applications with the help of a system like ChatGPT you will still need to create accurate test coverage and an understanding of the problem that is essentially equivalent to that of the people writing the code themselves. Whether all of this would in the end actually lead to large enough productivity increases depends on how error-prone the AI generated code will be and given that it takes a lot more time to dig into unfamiliar codebases than those you've written yourself, I think it's anything but obvious that this will have a huge overall impact on the industry. But obviously I might be biased here, since I have a stake in the game.
There's also mathematically excellent reasons why that happened.
Self-driving cars are an impossibly complex problem.
Statistics are statistics.
Predicting the minority class correctly 99% of the time isn't good enough for autonomous driving. A car has to break for little Suzie 100% of the time.
However, generating 1,000 lines of code for a CRUD app? That's 99% bug free?
That's a helluva lot better than I can do.
As with all things. The solution is watch what the domain experts do.
The equivalent is closer to a CRUD app that serves 99% of requests correctly. Which is nowhere near good enough to use.
But even if we do go with 99% bug free for the sake of argument, the usefulness depends on the type of bug. How harmful is it? How easy is it to detect?
I had my wife (a physician) ask ChatGPT medical questions and it was almost always subtly but dangerously and confidently wrong. It looked fine to me but it took an expert to spot the flaws. And frequently it required specialist knowledge that a physician outside of my wife’s specialty wouldn’t even know to find the problems.
If you need a senior engineer to read and understand every line of code this thing spits out I don’t see it as providing more than advanced autocomplete in real world use (which to be fair could be quite helpful).
It frequently takes more time to read and really comprehend a junior engineers PR than it would have to just do it myself. The only reason I’m not is mentoring.
Just because your prediction was wrong doesn’t mean that we aren’t leaps and bounds ahead of where we were. Seems like that is the crux of people’s argument. Because it’s not perfect yet it’s not impressive.
Hm, well that's not the impression I want to create. I certainly think any human intelligence task can be equaled by an AI at some point, I just feel uncertain about any specific timescale.
And GPT-3 et al has a lot of knowledge, even if it messes up certain expert level details. Rather than comparing against domain experts, my anchor point here is the sort of mistakes that novelists, script writers, and journalists make when writing about any given topic.
Whatever you saw a decade ago, it definitely wasn’t this.
I do recommend you play with it. But if you don’t feel like signing up with their free account, here’s a screen recording of me asking it some random general knowledge questions and instructing it to use a different language in the response each time: https://youtu.be/XX2rbcrXblk
I had to look up MegaHal, apparently that was based at least in part on a hidden Markov Model.
Using "it" in this way to refer to both that and GPT-family LLMs, or similarly saying "I remember being shown this software almost a decade ago" like the other commenter, is like saying "I remember seeing Kitty Hawk fly, and there's no way that can get someone across the Atlantic" when being presented with a chance to a free seat on the first flight of the Concord. (Actual human level intelligence in this analogy is a spaceship we still have not built, that's not where I'm going with this).
It’s not clear to me MegaHAL language models are less powerful than transformer models. The difference is the “largeness”, but that’s a hardware/dataset/budget detail.
While it's never going to replace the man behind the machine.
It still seems highly likely that "stitching libraries together" development workflows in 10-15 years will involve large amounts of copy-editing the output of large language models.
The trajectory of improvements, from GitHub Copilot to ChatGPT, is too steep.
Web development workflows honestly are often already at the stage of stitching together the output of large language models (Stack Overflow being the most well known such language model). I'm still surprised it pulls the salaries it does.
In my opinion, the only way it takes 10 years to get there is if all progress stops within the next 30 days.
Because it can literally almost do that stitching libraries together task now, if you give it a compiler and runtime environment and have it iterate on errors. Open AI has said they will release a big update before Christmas. This could include an API. And if we assume a text-only environment. But we already have the first text-to-video models, so we should assume that ChatGPT like systems will be built with multimodal models such that they would include information about UI interactions etc. in the near future. No reason to suppose that those advances would take ten years. We are seeing major improvements every 6-12 months.
Maybe we have another decade until it can handle its own fingers...