You can hardly get online these days without hearing some AI booster talk about how AI coding is going to replace human programmers. AI code is absolutely up to production quality! Also, you’re all…
My most recent usage of AI was making some script that uses WinGet to setup a dev environment.
This is a good example. What I’m saying is that pre-AI, I could look this up on StackOverflow and copy/paste blindly and get a slightly higher success rate than today where I can “AI please solve this”.
But I shouldn’t pick at the details. I think the “AI hater” mentality comes in because we’ve got this thing that boils down to “a bit more convenient than copying the solution off of StackOverflow” when used very carefully and “much worse than copying and pasting random code” when used otherwise. But instead of this honest pitch, it’s mega-hype and it’s only when people demand specific examples that someone starts talking like you do here.
Yeah, I can see that. Search has gotten worse. While AI slop is undoubtedly responsible for this, there are cases when some things are essentially best solved by reading thousands of code examples because the documentation is rather vague. Searching on Stackoverflow still relies on some people having already been presented with a similar situation and shared their solution. Also, you’d assume the solution is the correct one. (I’ve been burned and I’m sure the majority of my stackoverflow answers end up being corrections well after trying something else touted as the correct/popular solution.) That’s really my push back.
That’s really one of the strengths of AI: a large feeding of data until it finds a common pattern. It correlates to simple things like syntax. That means it’s pretty good there. But it also correlates to saying “a lot of people set up scripts like this”. That’s where I’m reminded of working with people who I assign a task to and they come back with stuff they got from SO. It has the gist of it being right, but not all there.
That’s kinda the key, though. I could be okay with an 80% workable state. That’s like asking somebody to compile all the search results and give be back a result as best they could. It doesn’t mean it should be treated as hot pluggable code.
Full disclosure, my main experience is CoPilot and VSCode. It’s… neat. Some of the auto complete is useful when what I’m writing has an obvious pattern. Some is laughably unrelated. There is another AI that has some level of training to it, which I think is Facebook’s. It can be “trained”. I’ve tried those models, but all those offline models don’t have the ability to combine web results. CoPilot lets you link to a spec page and it’ll read it in “realtime” and correct itself. I find that much more valuable than some pretrained model. The saddest part is that’s all proprietary in ChatGPT which was supposed to be Open (OpenAI). You basically have to buy-in to their models at least until something else comes along.
This is a good example. What I’m saying is that pre-AI, I could look this up on StackOverflow and copy/paste blindly and get a slightly higher success rate than today where I can “AI please solve this”.
But I shouldn’t pick at the details. I think the “AI hater” mentality comes in because we’ve got this thing that boils down to “a bit more convenient than copying the solution off of StackOverflow” when used very carefully and “much worse than copying and pasting random code” when used otherwise. But instead of this honest pitch, it’s mega-hype and it’s only when people demand specific examples that someone starts talking like you do here.
Yeah, I can see that. Search has gotten worse. While AI slop is undoubtedly responsible for this, there are cases when some things are essentially best solved by reading thousands of code examples because the documentation is rather vague. Searching on Stackoverflow still relies on some people having already been presented with a similar situation and shared their solution. Also, you’d assume the solution is the correct one. (I’ve been burned and I’m sure the majority of my stackoverflow answers end up being corrections well after trying something else touted as the correct/popular solution.) That’s really my push back.
That’s really one of the strengths of AI: a large feeding of data until it finds a common pattern. It correlates to simple things like syntax. That means it’s pretty good there. But it also correlates to saying “a lot of people set up scripts like this”. That’s where I’m reminded of working with people who I assign a task to and they come back with stuff they got from SO. It has the gist of it being right, but not all there.
That’s kinda the key, though. I could be okay with an 80% workable state. That’s like asking somebody to compile all the search results and give be back a result as best they could. It doesn’t mean it should be treated as hot pluggable code.
Full disclosure, my main experience is CoPilot and VSCode. It’s… neat. Some of the auto complete is useful when what I’m writing has an obvious pattern. Some is laughably unrelated. There is another AI that has some level of training to it, which I think is Facebook’s. It can be “trained”. I’ve tried those models, but all those offline models don’t have the ability to combine web results. CoPilot lets you link to a spec page and it’ll read it in “realtime” and correct itself. I find that much more valuable than some pretrained model. The saddest part is that’s all proprietary in ChatGPT which was supposed to be Open (OpenAI). You basically have to buy-in to their models at least until something else comes along.