Comment by diggan
2 months ago
> Creativity is fun. AIs automate that away.
I've been developing with LLMs on my side for months/about a year now, and feels like it's allowing me to be more creative, not less. But I'm not doing any "vibe-coding", maybe that's why?
The creative parts (for me) is coming up with the actual design of the software, and how it all fits together, what it should do and how, and I get to do that more than ever now.
I'm still faster than the cheap bots.
The creative part for me includes both the implementation and the design, because the implementation also matters. The bots get in the way.
Maybe I would be faster if I paid for Claude Code. It's too expensive to evaluate.
If you like your expensive AI autocomplete, fine. But I have not seen any demonstrable and maintainable productivity gains from it, and I find understanding my whole implementation faster, more fun, and that it produces better software.
Maybe that will change, but people told me three years ago that we would be at the point today where I could not outdo the bot;
with all due respect, I am John Henry and I am still swinging my hammer. The steam pile driving machine is still too unpredictable!
> The creative part for me includes both the implementation and the design
The implementations LLMs end up writing are predicable, because my design locks down what it needs to do. I basically know exactly what they'll end up doing, and how, but it types faster than I do, that's why I hand it off while I go on to think about the next design iteration.
I currently send every single prompt to Claude, Codex, Qwen and Gemini (looks something like this: https://i.imgur.com/YewIjGu.png), and while the all most of the time succeed, doing it like this makes it clear that they're following what I imagined they'd do during the design phase, as they all end up with more or less the same solutions.
> If you like your expensive AI autocomplete
I don't know if you mean that in jest, but what I'm doing isn't "expensive AI autocomplete". I come up with what has to be done, the design for achieving so, then hand off the work. I don't actually write much code at all, just small adjustments when needed.
> and I find understanding my whole implementation faster
Yeah, I guess that's the difference between "vibe-coding" and what I (and others) are doing, as we're not giving up any understanding or control of the architecture and design, but instead focus mostly on those two things while handing off other work.
I agree, and my flow is similar
I've made great use of AI by keeping my boundaries clear and my requirements tight, and by rigorously ensuring I understand _every_ line of code I commit
I believe software development will transition to a role closer to director/reviewer/editor, where knowledge of programming paradigms are just as important as now, but also where _communication_ skills separate the good devs from the _great_ devs
The difference between a 1x dev and a 10x dev in future will be the latter knows how to clearly and concisely describe a problem or a requirement to laymen, peers, and LLMs alike. Something I've seen many devs struggle with today (myself included)
2 replies →
Claude code is too expensive to evaluate?
It's 20 bucks a month
[flagged]
Same. I think there are two types of devs. Those that love designing the individual building blocks and those that wanna stack the blocks together to make something new.
At this point AI is best at the first thing and less good at the second. I like stacking blocks together. If I build a beautiful UI I don't enjoy writing the individual css code for every button but rather composing the big picture.
Not saying either is better or worse. But I can imagine that the people that loves to build the individual blocks like AI less because it takes away something they enjoy. For me it just takes away a step I had to do to get to the composing of the big picture.
The thing is, i love doing both. But there’s an actual rush of enjoyment when I finally figure one of the tenets of a system. It’s like solving a puzzle for me.
After that, it’s all became routine work as easy as drinking water. You explain the problem and I can quicly find the solution. Using AI at this point would be like herding cats. I already know what code to write, having a handful being suggested is distracting. Like feeling a a tune, and someone playing another melody other than the one you know.
Yea, I guess some people enjoy both.
> For me it just takes away a step I had to do to get to the composing of the big picture.
You can't successfully build the big picture on the sort of rotten foundation that AI produces though
I don't care how much you enjoy assembling building blocks over building the low level stuff, if you offload part of the building onto AI you're building garbage
Exactly. I loved doing novel implementations or abstractions… and the AI excels at the part where it modifies it slightly for different contexts… aka the boring stuff.
But this is how you learn, how you find better ways, by grinding.
Getting wild ideas badly implemented on a silver plate is a slot machine, it leads nowhere but in circles.
When I say ideas i'm talking in the context of programming... I'm not talking about "I got a great idea for a new social network" and the AI just wrote some spaghetti code for it. When I have the AI write low level code it's stuff like filling out a function implementation of which I already defined high level type classes for... I can focus on high level abstractions, whereas the AI can iterate in the most statistically sensible way to fill in the easy blanks.
By grinding what though? I don't wanna grind "Entering characters with my fingers", I wanna grind "Does this design work for getting X to work as I want", which is exactly the sort of things LLMs help me move faster on.
And yes, if you're just using it as a slot machine, I understand it doesn't feel useful. But I don't think that's how most people use it, at least that's not how I use it.
1 reply →
i've done 20 years of grinding thanks. im happy to work at a higher level of abstraction now.