32
The two worlds of programming: why developers who make the same observations about LLMs come to opposite conclusions (www.baldurbjarnason.com)
submitted 2 months ago by to c/programming@programming.dev
29 comments fedilink hide all child comments
 

cross-posted from: https://programming.dev/post/52842255

Have been working my way through this author's essays, thought this one was a unique observation.

you are viewing a single comment's thread
view the rest of the comments
[–] 5 points 2 months ago (2 children)

Yeah, same here

I needed to calculate some statistics from log files and that would have taken me ages - mostly because my sed-/awk-fu is lacking

But with AI I got a script in half an hour - after correcting it a few times - and it is reasonable ok enough for the job

I tried to refactor a really large spaghetti state machine with AI, but they're all failing, because it's too large

It's nice to get boiler plate code written, but when it comes to overview of logic on a large project, they obviously can't keep up

As you said, it's a tool. But just like with any tool, you shouldn't use it for every use case

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 4 points 2 months ago (1 child)

    Your statements are supposed to be moderate, but given the physical realities of actually running the massive LLMs it is terrible. Maybe you saved some time at the expense of learning to write your log script, but you also helped boil the oceans a bit in the process. All the things you tried and failed helped boil the oceans for no benefit at all. You're more optimistic about the utility of it than I am, but even given your perspective, it is a disaster when you consider the externalities of the AI data centers.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 2 points 2 months ago*

    That's all correct and true

    I was on a business trip, very sleep deprived and needed a fast solution

    So, I guess, this was the tool for the job

    Probably it got lucky, because with other things it just fucked up after that
    So, my trust into it is nil
    But for quick scripts, that would take me much longer to build, it's working

    But I'm also usually using local models

    I'm aware of their power and water consumption, and you're right in that

    Edit: and besides the environmental impact, which I don't use them often or only locally, I'm still a tech guy and want to know, what they're actually capable of
    Which still isn't much and I don't see LLMs getting better with more complexity

    There are to many problems
    We have programming languages for reason, because they're mathematically sound and not ambiguous
    Human speech just isn't and that's already the point, why LLMs can never really work for that

    But I was surprised, that it once came up with a solution to filter my log files and calculate some statistics out of it
    And I have to give it that

  • permalink
  • fedilink
  • source
  • parent