Year ago with enough prompting it could write a function for me that did something conceptually simple but tedious.
I tried Windsurf over the last 3 days. It can write complex game mechanics logic in multiple files and refactor code when told to. I can talk to it using language of the domain I just made up and I can talk to it using language that developers use when they talk about the code. It can solve bugs that would take me half an hour to figure out at least. It went from a poor junior to clever junior and half a senior in one year. It's not great at one shot architecting yet, but when told how to architect a thing it can move stuff around to conform with the desired architecture. It creates and modifies SVG's, understands colors and directions. It does all of that blind, without running the code. All it has is compiler (language server really) and linter.
You mention it's an autocomplete. I barely used auto-complete at all. I just told it what to do. I touched the code a little bit only because I wanted to, never because I had to. When I wanted a change to the code, 95% of the time I just told it what change I needed and it made it.
It basically replaces about 1/3 of software team already enabling (and pushing me towards) a role of (tech lead/architect/qa/product owner) and this fraction is rising.
It's not flawless. Sometimes it makes things that don't work on the first try. But you don't always need to revert. It's capable of fixing what it made when told about how the desired behavior differs. And all that without running the program.
I anticipate that further development is going to be letting it run and observe what it built so it's gonna move towards QA role. And the other thing might be slowly learning to differentiate between well architected code and badly architected code but that's probably harder as it requires careful preparation (and creation) of training data. While moving towards QA just requires figuring out how to let it run and inspect stuff. Maybe use the debugger.
I'm doing everything you're doing and probably more with Claude code (I 'let it run'). I'd say I too have been through a few of the different phases of the gartner hype cycle now.
As I said it's a good tool, it's far from being an "auto-pilot".
I don't care about the hype. If you put me in a cave for last 5 years and today gave me Windsurf and told me nothing beyond that a machine does that, I wouldn't believe you. All of it was literally science fiction 5 years ago. This literally replaces a person that I would have to hire to pursue my ideas.
Ok I guess that's where we differ, I don't really find it to be science fiction, the more I use it the system, the more I understand that magic of it, what it's going to do, and I've learned that to get good results, you absolutely have to drive it like almost every other tool.
I guess if anyone went into a cave and came out for 5 years things would be like science fiction. FPV drones on the battlefield would be another example...
As I said earlier, not downplaying it, really just warning you that if you lean to heavy on it without guiding it quite precisely, you will get burned, properly.
FPV drones on the battlefield were absolutely obvious. The only strange thing about it was, and still is, that powerful armies of the biggest players were "nah, I'll pass" on this obvious application. Wedding photographer figured out that you can hit people with drones. Farmers figured out they can carry and release significant payloads. Yet for some reason, generals seen them as toys and preferred to stick with their bombs and artillery shells costing order of magnitude more. We needed a war and soldiers on the contact line to basically start doing this obvious thing out of necessity and demanding supply from commanders.
Computers being able to keep the conversation going about complex topic like development of a computer program were 5 years ago farther away than fusion energy. Perpetually 30 to 100 years away.
Sure, for the best results you need to guide it because without it you'll get some program but not a good program. If you don't help it to control the complexity of the code with sensible architecture it will paint itself into a corner. But a junior developer would also do that if you just piled up requirements and never steered his output.
I agree with you that it's a tool. But more like a combine harvester, that you still need to drive, for now, than a hammer. Already more like a CNC machine than a power drill. And still getting better.
Everyone can have their own opinion on this. But I think next 5 years will change those opinions.
For me it is at least as magical as cell phones. When I was a kid the peek of wireless personal communication was walkie-talkie with a very limited range. The idea that you could cram enough electronics, RF stuff and a battery in a small enough package and route the calls globally with vast networks of two way communication stations seemed absolutely ludicrous. I haven't even seen SMD element before I had my first cellphone and I picked apart a lot of electronics. Then it happened. It was the kind of technology that got me completely surprised, like LLMs. I remember not being that shocked because I was young and was expecting new things. I was reading a lot of SF and there instant wireless communication was everywhere. But after 4 decades of basically no progress in the machines that do anything that resembles thinking this innovation caught me off guard completely. I would be less surprised if we already had fusion reactors providing most of our electricity.
On a more practical note, do you switch a lot the models that you use? I can't really tell them apart because I haven't worked with them all that much but I heard good things about GPT5 and when I used it it was very good (but others, like Claude or swt-1 were great too).
I've done a lot of wood work and there is almost zero doubt in my mind it's a router.It's flexible, brilliant, time saving, repeatable (with templates) and you can do a lot of damage with it in a very short amount of time. I'd never call it a Hammer.
With regards to the models I use. I'm pretty much using Claude Sonnet 4 with Claude Code and Gemini 2.5. In my opinion Sonnet 4 is the most superior model I've used yet, although I've not used GPT-5. Where some of my attitude comes from is even in the way you get used to using differnt models, like you really kind of get a grasp for how to prompt them and kind of know what to expect. That's where I don't see them so much as an intelligence, but an approximation engine. I feel that what many "coders" fail to realize is how extremely common the patterns they're employing to solve a problem are. The approximation engine can see that as numbers, match it and give you a similar eyt slightly modified (to your context) version of what you want.
These models have been trained on probably hundreds of millions or billions of lines of code after all, yet there probably isn't all that much variation in what most code looks like. What you'll notice is that it's horrible at coding in anything it's not trained on. Recently I tried to use it with FastHTML, and it knows almost nothing about it, and the author of that project will even provide you with a context file you can give your "AI" to use so it can "understand" how to code with it.
We also follow like common architectural patterns to get them most out of "AI". We've found that to be very helpful to reduce the amount of random and unhelpful code we get out of them.
I tried Windsurf over the last 3 days. It can write complex game mechanics logic in multiple files and refactor code when told to. I can talk to it using language of the domain I just made up and I can talk to it using language that developers use when they talk about the code. It can solve bugs that would take me half an hour to figure out at least. It went from a poor junior to clever junior and half a senior in one year. It's not great at one shot architecting yet, but when told how to architect a thing it can move stuff around to conform with the desired architecture. It creates and modifies SVG's, understands colors and directions. It does all of that blind, without running the code. All it has is compiler (language server really) and linter.
You mention it's an autocomplete. I barely used auto-complete at all. I just told it what to do. I touched the code a little bit only because I wanted to, never because I had to. When I wanted a change to the code, 95% of the time I just told it what change I needed and it made it.
It basically replaces about 1/3 of software team already enabling (and pushing me towards) a role of (tech lead/architect/qa/product owner) and this fraction is rising.
It's not flawless. Sometimes it makes things that don't work on the first try. But you don't always need to revert. It's capable of fixing what it made when told about how the desired behavior differs. And all that without running the program.
I anticipate that further development is going to be letting it run and observe what it built so it's gonna move towards QA role. And the other thing might be slowly learning to differentiate between well architected code and badly architected code but that's probably harder as it requires careful preparation (and creation) of training data. While moving towards QA just requires figuring out how to let it run and inspect stuff. Maybe use the debugger.