RE: https://hachyderm.io/@mitchellh/116580433508108130

I'm not an infra guy but 'leaving the code to AI' seems very very dangerous to me.

It reminds me of the arguments I've had with 'nobody cares about the code' people in the past (surprisingly many of them programmers), meanwhile we were dealing with bugs that were a direct result of poorly engineered code.

And the nature of that is the code base becomes terribly broken past a surprisingly early point of no return.

If you 'leave it to' LLMs, no matter how good they are, they will inevitably write code that goes far beyond this point, they can just keep adding complexity inhumanly.

This results in non-linear issues managers rarely understand - bug rate increases, features become impossible to add, security issues abound, performance tanks, etc.

But people end up thinking 'that's just how it is' because they don't realise it's because of the mess.

LLMs will also hallucinate + introduce subtle bugs unlike anything a human would write, deep, deep within the spaghetti.

Unwise :)

@ljs i'm leaving programming to awk oneliners and copy pasting from stack overflow

@lkundrak Sorry, did sir imply he had thumbs there? Did I read that right?

*Screeches*

@ljs lol no thumbs sir

i have not upgraded my fedora 43 to fedora 44

i upgraded to debian

i can paste with middle click

no thumbs

sir

no thumbs

@lkundrak FEDORA??? WHY IS SIR NOT USING SLACKWARE???

AI tools are useful, but they NEED a human-in-the-loop.

Unfortunately every time technology like this emerges, the aristocratic class (i.e. upper management) see it as a cue to eliminate jobs.

And maybe even from the very dawn of software they've wanted to get rid of programmers.

We're awkward, we tell them their badly formulated ideas won't work, we dare to have specialised skills that cost them more.

So, so many attempts throughout the history of computing to get rid of us, this is the most serious of course, but the dream of getting rid of us blind them to the limitations of this technology.

Even if AI did live up to every claim, it'd be developers wielding them, of course they don't see that :)

@ljs is this a trick question?

not enough active neurons

@lkundrak good, good, now reduce more

(hail satan)

@ljs @lkundrak sir lacks the floppy disk drive to install slackware and is unrepentantly old school

@wil @lkundrak A floppy disk MUST FLOP sirs

@ljs Long before AI, while i was doing some stuff for Nokia, we sometimes had to make PoC's look worse so that some random manager would not think that the product is complete. And now it's like... well, you know how it is like.

Further, I've never been in a project, worked on an upstream open source project feature, or anything, where initial the building of "the thing" would have taken more than a few weeks. After that follows countless months of tuning, adjusting and most importantly discussing with other people about the nuts and bolts.

IMHO, the bubble in AI is really related the misconception that building some functional version of software would be the main contributing cost factor for shipping quality software.

@jarkko @ljs the main issue in many cases isn't so much about doing the thing, it's doing the thing *well*

@jarkko @ljs What amazes me in this is that it’s often software development professionals who seem to think the code generation has been the bottleneck. Many of them seem to get defensive or elusive when questioned was this really the bottleneck.

Of course these people also seem to have been more on the managerial side for years, so maybe they are not that familiar with the development side.

@pare @jarkko @ljs

I also don't buy the "it's just another abstraction layer" storyline. Like assembly -> high-level languages -> natural language prompts.

We're not at a stage where we could do review on prompts and expect them to yield deterministic results.

@jani @jarkko @ljs I think using LLMs means that you accept the non-determinism inherent in the system. If you want to make deterministic abstractions, by all means, but think if a non-deterministic tool is the right one for that.

@jarkko @ljs Yeah that tracks for sure. At work the cost associated with most big tasks isn't "make it work" but "make it work in a state that's shippable".

@ljs fully agree!
Reviewed-by: Sashiko <sashiko-bot@kernel.org>

@jarkko @ljs surely you just tell the LLM to get the users to show it what the problem is? Or explain to the customer that their demand that "it just work" needs more detail before code can be written?

That's why we have graduate developers now.

@HuguesRoss @ljs Yeah, and I mean if this "make it work as PoC" does not really time that much with a talented team.

Almost never AI cost savings are not weighted with customer satisfication metrics, competitive advantage reached by using LLMs for when writing apps and other software to some highly compted market such as e.g., iOS App Store. When I buy stuff I don't care about at all how expensive it was produce or manufacture. Nobody does.