> This is what’s behind all the “vibe coding” you’re hearing about. AI agents — basically, a program that keeps applying AI over and over until a task is complete — are now taking over much of software engineering.
No, agents and LLMs help you with coding specifically, but they aren't good at software engineering yet, which is pretty obvious once you start playing around with the various vibe-coded projects out there and especially if you start reading the code that all these agents produce without strong hand-holding.
Once you can fire off "Build a hacker news clone" and you actually end up with a maintainable and properly built thing, then I'd agree they're suitable for software engineering. But until then, they're more "code monkeys" than anything, and require a human software engineer to prod them into the right direction, otherwise they'll just run with it, not unlike a junior developer who haven't yet figured out software engineering but they know how to code.
> But a submarine can go faster than any fish
What? Guess I should have finished reading before commenting, this is obviously not true and I don't think a human would make such a mistake... I fell for vibeslop yet again.
Although I agree they're better developers than they are engineers for similar reasons to you, and would also make a distinction on the headline* between "do you mean 'smart' as in 'learned a lot' or as in 'learns from few examples'?" (because the answer is different):
> Once you can fire off "Build a hacker news clone" and you actually end up with a maintainable and properly built thing, then I'd agree they're suitable for software engineering.
I've not seen the HN source code, but superficially it seems quite bare-bones (which I appreciate!). Are you sure you wouldn't prefer a harder target?
* The body shows that Noah Smith himself is well aware of, if not this specific argument, then at least the general point
> I've not seen the HN source code, but superficially it seems quite bare-bones (which I appreciate!). Are you sure you wouldn't prefer a harder target?
Yeah, I think it'd suffice. It's less about if it's possible or not (as apparently building something resembling of a browser "from scratch" isn't that hard), and more about exactly how it ends up being built, as that's the "software engineering" part that I'm arguing is still the hard part, and LLMs aren't that great at it (yet?).
A smaller test case would also be faster to execute, rather than spending 3 days of an engineers time, but I guess ultimately that's less important, just a happy side effect.
> This is what’s behind all the “vibe coding” you’re hearing about. AI agents — basically, a program that keeps applying AI over and over until a task is complete — are now taking over much of software engineering.
No, agents and LLMs help you with coding specifically, but they aren't good at software engineering yet, which is pretty obvious once you start playing around with the various vibe-coded projects out there and especially if you start reading the code that all these agents produce without strong hand-holding.
Once you can fire off "Build a hacker news clone" and you actually end up with a maintainable and properly built thing, then I'd agree they're suitable for software engineering. But until then, they're more "code monkeys" than anything, and require a human software engineer to prod them into the right direction, otherwise they'll just run with it, not unlike a junior developer who haven't yet figured out software engineering but they know how to code.
> But a submarine can go faster than any fish
What? Guess I should have finished reading before commenting, this is obviously not true and I don't think a human would make such a mistake... I fell for vibeslop yet again.
Although I agree they're better developers than they are engineers for similar reasons to you, and would also make a distinction on the headline* between "do you mean 'smart' as in 'learned a lot' or as in 'learns from few examples'?" (because the answer is different):
> Once you can fire off "Build a hacker news clone" and you actually end up with a maintainable and properly built thing, then I'd agree they're suitable for software engineering.
I've not seen the HN source code, but superficially it seems quite bare-bones (which I appreciate!). Are you sure you wouldn't prefer a harder target?
* The body shows that Noah Smith himself is well aware of, if not this specific argument, then at least the general point
> I've not seen the HN source code, but superficially it seems quite bare-bones (which I appreciate!). Are you sure you wouldn't prefer a harder target?
Yeah, I think it'd suffice. It's less about if it's possible or not (as apparently building something resembling of a browser "from scratch" isn't that hard), and more about exactly how it ends up being built, as that's the "software engineering" part that I'm arguing is still the hard part, and LLMs aren't that great at it (yet?).
A smaller test case would also be faster to execute, rather than spending 3 days of an engineers time, but I guess ultimately that's less important, just a happy side effect.