Aside from boilerplate code, when a system of any significant complexity is required, the prompt would effectively be the detailed specifications, in which case it would be more efficient to code it from scratch, instead of incurring the additional overhead of humans having to validate the output. Change my mind.
Honest attempt to changing your mind :)
Complexity rarely requires a lot of code but a lot of iterations. You don’t know the specifics yet in detail.
That’s actually one of the good parts of LLMs to quickly iterate on design and architecture in human language without having to write the code. After all we only want to decide left it right in the complex stage of development - doesn’t have to necessarily be a smooth ride to figure that out.
Only once the specifics are clear you have to get the good stuff.
That said: it’s still not yet there to really save time to the most of the seniors I’ve talked to -but it’s the lack of reliability and not prompt complexity that I’ve heard as criteria.
Fair line of thought, and it also seems quite a few of the ais we hear about as ‘successes’ (eg that train to drive a car virtually around a track, or play a game) seem to follow the approach you described.
However, by coming to a working solution by throwing billions of iterations (in the case of playing games) and having a human score the results for the next 10000 attempts, it seems to be a very laborious process, and hardly efficient. It ‘saves time’ by using the speed of the computer, yet consumes way more resources than simply having a skilled human. I’d even argue that paying a skilled human to do the task is orders of magnitude cheaper than assembling an array of GPU racks to attempt the task. It seems it can hardly be called ‘intelligent’, nor ‘learning’. It’s like the monkeys and the typewriters. Even if one did do a Shakespearean work, could the monkey be called a playwright?
Oh I agree with that completely. The intelligence party is a marketing agent that works really well.
I only see (potential) use cases -I’ve yet to see them being efficient in the development context. :)
Man, this section just kills me:
As the market gets flooded with AI-generated MVPs, the products that will stand out are those built by developers who:
- Take pride in their craft
- Care about the little details
- Focus on the full user experience
- Build for the edge cases
- Create truly self-serve experiences
The irony? AI tools might actually enable this renaissance. By handling the routine coding tasks, they free up developers to focus on what matters most - creating software that truly serves and delights users.
I’m agreeing with everything they’re saying, but then they bring in “routine coding tasks”.
These don’t exist for developers who take pride in their craft. That’s like going up to an artist and telling them they could color in their blue sky with a paint roller. Yeah, they could, but no one would want to look at it, much like no one wants to look at code that’s not very intentional about how it’s written.Sure, you typically have some boilerplate for converting between different data formats, but my head is racing even while writing that, how I could eliminate more of this boilerplate. Because no one wants to look at boilerplate, even if you’re the fastest at producing it.
I took a relatively simple go app, that has some Javascript front end and asqlite dB, and tried to use chatgpt and Claude to change one specification that should not be that complicated, but it still could not do it. It kept thinking the file name of the. Js file was paste.js when that was never the name. It couldn’t load the whole project for context and I could only really provide it once file at a time… It’s been pretty ok at some very simple coding tasks but nothing more than that so far, unless I’m using the wrong tools?