logoalt Hacker News

rphv • yesterday at 7:42 PM • 2 replies • view on HN

This is the usual reaction when people run into an uncomfortable truth. We're all bad at objectivity - we preferentially look for evidence that supports the outcome we want (or the one that's least uncomfortable). In this case it's pretty obvious that AI is better at coding than humans. Lower failure rate per task is better code. We're just shipping more of it. Nobody's claiming AI code is bug-free. Just that it's better and faster than how we used to do it.


Replies

rsyring • yesterday at 7:47 PM

> Just that it's better and faster than how we used to do it.

I think only one of those is true. It takes a lot of effort to get quality architecture and not-function-soup code (at least in Python) from LLMs. I can't imagine the pile of crap code bases these software factory type solutions are producing. Well, you can see a lot of them on GitHub I guess, so no need to imagine.

Further, one of the negative impacts I've seen from being able to go a lot of faster is that it translates to product teams being able to ship faster which is not always better. We've lost the "back pressure" we used to have in software development which forced product teams to consider, then consider again, then consider again as each new sprint requires reevaluating priorities. This resulted in a lot of requests stalling out and never being built that should have never been suggested in the first place. Or, to be more charitable, the delay gave the product team time to get over the initial excitement of a new idea and really consider if it was warranted.

Now we are building all-the-things and I think that's only better in some cases. I think it's proving worse in most other cases. Especially when you consider that every line of code that ships has liability attached to it.

That results in large bloated mess projects that no human can reasonably understand. That means we are dependent on LLMs to understand it and, the larger the code base, the more LLM tokens are required and the worse the LLMs are at understanding all the relevant parts and making wise architecture decisions. This creates a downward spiral that is harder and harder to manage and means features take longer and longer to ship (although, still maybe less time than pre-LLM dev...we'll see).

sdevonoes • yesterday at 9:00 PM

Do we want “faster” development? Like I couldn’t care less if my bank app or netflix didn’t ship a single new feature in the next 12 months. Users don’t need constant stream of features. But faster development is “better”. Progress and all they say. Like yeah, why I would like a car if my horse is just fine

➕ show 1 reply