I’ve managed, for most of my adult life, to avoid getting into online arguments. I feel strongly about this and I have many years of experience on the subject of software engineering. I’m also unemployed and struggling to get back to work in a tech job market that is in shreds so yes, I’m a bit touchy about it.
I’m just going to exit this discussion with this: -
Most of my friends and colleagues that are also experienced software engineers share the same concerns that I do. Many of them openly admit that they are forgetting how to do things and are increasingly delegating and becoming more reliant on AI tools, and in many cases are being aggressively pushed towards that by management.
Some of them are pushing back against their management and warning of the increase in bugs, reduction in codebase knowledge and understanding etc. They can see it happening already.
Basically, the people shouting the loudest about how great this revolution is are the managers and leaders whose job it is to make more things happen faster and cheaper. Those who are exhibiting scepticism are the actual “doers” who write code and build systems.
the bottom line rule remains more actual than ever in these days…
IF U DON’T USE IT…U LOOSE IT…
we’re far away from any next level of natural and healthy BALANCE…
this 4th industrial revolution, the 2nd age of machinery will bring us the biggest transition phase human mankind has ever faced…
and so far, EVERY transition was a pretty rough period in history…
Ahem. Heavily subsidised for the time being. It’s in fact way more expensive to have ML do things for you than hiring a human professional to do the work. Money is burning, baby! And they are banking on breaking the mastery chain (and addiction) so the suckers using it won’t have any other choice but to stick to using it forever, regardless of cost. Oh yeah
@Phillip I know, I’ve already mentioned that here on the forum in more than one thread I think. It’s the world’s most aggressive loss leader strategy and everyone is buying into it.
Personally I think there’s a wider question here about work patterns and reliance on (or addiction to) technology - what sort of world do we actually want to live in?!
This particular doer is enthusiastic about what AI is allowing me to do. Which is, more things this month than last month, which was more than the month before that.
There will no doubt be some consequences for the industry as a whole. Some of which will be driven by management short-sightedness. But you can’t blame crappier code on AI. It’s just a better tool, like high level languages were a better tool than assembler. The quality of output rests with the user of the tool. And if management isn’t allowing time for the necessary coder/QA practices to maintain quality, then it’s a management issue, not AI.
Yes, and what constitutes good quality is determined by years of learning and writing code by hand. What happens when nobody does that any more?
The AI can’t be the source of truth forever can it? We need human knowledge and experience to train the source of truth. The source of truth can’t train itself.
It’s called marking your own homework.
Nobody is answering the question of what happens when the human expertise dies out.
But it isn’t better and that’s part of the problem. It’s an approximation of expertise, an impression of expertise. Guessing a string of words based on patterns isn’t expertise. LLMs gets things wrong all the time, and a human needs to correct it. LLMs hallucinate and plainly makes things up. It’s well documented and many of us have seen it ourselves.
Again - when the human experience dies out, who corrects it?!
Others have said more or less the same, but one of the coolest aspects of AI assisted coding for musicians (and other primarily non-technical people) is just the ability to make your own bespoke tools. I got rid of my entire eurorack setup a couple of years ago, but there are modules that I’ve missed the sound or functionality of. Many of those modules are extensively well documented, and I bring my own knowledge and experience of the modules as well, and between those two things, I’ve been able to code reasonably good approximations of those modules that I can now use as VSTs in Ableton. I would never release them commercially, but for my own uses, this is magical.
It’s a better tool than writing code directly, when driven correctly. It can work faster than you do, not make syntax errors, write tests without you asking and spot the little edge cases you missed. Write a commit messages that fills in all the background and remember ALL the context when it’s time to write a pull request. If it doesn’t do the job to your satisfaction, it’s up to you to make up the difference. Either on a case by case basis, or else in guides for AI consumption, which it automatically finds every time without prompting.
Like I said, it’s a tool, you have to know how to use it.
No. The skills are the same. We’re just trying to shortcut all the processes, which will inevitably lead to disaster.
People are laughing off how coding is easy, AI can replace that. Oh and since AI can do all the coding now we can focus more on the problem. But now we have a ton of skills, so AI can take care of that too. So now we can just yeet shit at the wall as fast as AI can spec it and code it (agents)
And the code is shit. And so is the reasoning. But it looks solid, sounds well thought out. But there’s subtle bugs and weird decisions littered all over, and if you’re not reading all the code (gets hard when you yeet a ton of it at the wall) and really really criticising it (humans are terrible at this with big PRs) then you’re building a castle on quick sand
I use AI a lot, but I also limit how I use it heavily. Verify everything. Read and understand all the code. It’s not a quick fix, and as others have pointed out the amount of bugs + future token cost is going to wreak havoc on companies building their future on it
I can assure you that you will discover later how wrong you are
It’s non deterministic, one persons amazing prompt can have an entirely different outcome in a new clean session. You can’t prompt your way to greatness.
It’s built to please, make you addicted. It will LOOK good, but there’s no way it can fit all your situational context into it’s session
@Xylr I can’t argue with that and I’ve built some bits with AI myself. It does feel like magic.
I’m certainly not arguing with how impressive this stuff is, but I’m very strongly arguing with how sustainable it is, how ethical it is and how dangerous it is to become reliant on it.
We’re heading towards a global workforce that’s going to be using this stuff for everything, and yet again we have giant tech companies with too much control.
Also the impact on the environment is terrible, we’re already seeing that.
I think we’ll learn another hard lesson once the tools are no longer freely available. Anyone can use anything in these halcyon days of availability, but at some point someone is going to want a return on their massive investment. We’ll have a pile of unsupported, non-migrated, unmaintained app because everyone’s genie will require a whole lot of money for tokens.
@hermbot Some of my engineer friends are basically waiting to get paid £1000 a day on contract to clean up all the sludge mess that everyone is making.
I understand and respect that you’re in a weird spot right now. I just don’t see the issue here the way you do. If some people want to waste their money on sloppy vibe code, what’s it to us? People were already wasting money on useless gym music that sounds worse than AI slop.
Vibe coders aren’t going to replace real coders. There will probably be an adjustment period in terms of the job market. For building large, reliable systems, you need experience. For building a game of Tic Tac Toe, less so. There are just so many things inbetween those two extremes. And this is today, who knows what’s going to be the case in just a year?
You’re wrong, in the case of the apps I’m dealing with. It’s very good at pulling in all the context that’s relevant to me for the job at hand because the context is built into the app, not the AI, and is revealed to AI through progressive disclosure.
I dropped a working NEON SIMD optimized polyBlep implementation into Claude (latest Opus model at the time, 4.6 I believe) and asked it to convert it to a regular polyBlep. It knew all the context, I explained exactly what it was. I was just too lazy to convert/look it up myself.
It outputted 99% of the function, it missed to add one multiply, sounded like shit. This is a recipe that is widely available on the internet, it’s what 5 lines of code?
Yes, and where does that experience and those “real” coders come from if everyone stops learning how to do things properly? In the future, we won’t have that expertise if we farm it out to LLMs. The people who have the real expertise won’t exist, because we’ve stopped training them.
You need to appreciate that there are already whole swathes of entry level software engineers who have graduated from university only to find that the job market doesn’t exist or doesn’t want them because of a belief that we don’t need juniors any more.
There are people with PHDs that can’t get an entry level role.
The seniors of today were the juniors of yesterday.
I’d like to make it clear that I’m not trying to “pull rank” on @bibenu or anyone else here. I make no claim to be the top expert or have the most years of experience.
The debate here isn’t just about what’s possible with AI and what isn’t, or about code quality. I know that good stuff can be made as well as crap.
The conversation is also about having a sustainable job market and making a decision about what sort of future we want to have. Society has wandered into all sorts of shit with it’s eyes open many times before.
I’m merely suggesting we apply a bit of critical thinking. Both us and our governments!
