▲ 76 points
back
105 comments
People are absolutely insane with their takes on AI replacement theory. The complexity of our stacks has grown exponentially since the 70s. Very few people actually comprehend how many layers of indirection, performance, caching, etc. are between their CRUD web app and bare metal these days.
AI is going to increase the rate of complexity 10 fold by spitting out enormous amounts of code. This is where the job market is for developers. Unless you 100% solve the problem of feeding every single third party monitoring tool, logging, compiler output, system stats down to the temperature of RAM, and then make it actually understand how to fix said enormous system (it can't do this even if you did give it the context by the way), then AI will only increase the amount of engineers you need.
> AI is going to increase the rate of complexity 10 fold by spitting out enormous amounts of code.
This is true, and I am (sadly, I'd say) guilty of it. In the past, for example, I'd be much more wary about having too much duplication. I was working on a Go project where I needed to have multiple levels of object mapping (e.g. entity objects to DTOs, etc.), and with LLMs it just spit out the answer in seconds (correct I'd add), even though it was lots and lots of code where in the past I would have written a more generic solution to prevent me from having to write so much boilerplate.
I see where the evolution of coding is going, and as a late middle aged developer it has made me look for the exits. I don't disagree with the business rationale of the direction, and I certainly have found a ton of value in AI (e.g. I think it makes learning a new language a lot easier). But I think it makes programming so much less enjoyable for me personally. I feel like it's transformed the job more to "editor" from "author", and for me, the nitty gritty details of programming were fun.
Note I'm not making any broad statement about the profession generally, I'm just stating with some sadness that I don't enjoy where the day-to-day of programming is heading, and I just feel lucky that I've saved up enough in the earlier part of my career to get out now.
I don't always programming in the small, and still feel that AIs provide plenty of chance for architecture, design, refactoring. For me it's been an absolute boon, I'm enjoying build more than ever. At any rate it's undeniably transformative and I can see many people not enjoying the end state.
Really? I sort of feel the opposite. I am a mid-career as well and HIGHLY TIRED of writing yet another set of boilerplate to do a thing or chase down some syntax error in the code and the fact that AI will now do this for me has given me a lot more energy to focus on the higher level thinking about how it all fits together.
I do not look forward to the amount of incompetence and noise that increasing adoption of these tools will usher in. I've already had to deal with a codebase in which it was clear that the author fundamentally misunderstood what a trie data structure was. I was also having an difficult time trying to talk to them about the implementation and their misconceptions. lo and behold I eventually find out the reason they chose this data structure was because they asked ChatGPT what to do and they never actually understood, conceptually, what they were doing or using. This made the whole engagement with the code and process of fixing things way harder. Not only did I now have to fix the bunk code, I also had to spend significant time disabusing the author of their own misunderstandings...
So, AI created a job opportunity for you?
>Not only did I now have to fix the bunk code, I also had to spend significant time disabusing the author of their own misunderstandings...
People with your attitude will be the first to be replaced.
Not because you code isn't as good as an AI; maybe it's even better. But because your personality makes you a bad teammate.
That’s called consultancy and you can bill chunky rates by the hour. You should be rubbing your hands with glee!
And then work out how to do code review and fixing using AI, lightly supervised by you so that you can do it all whilst walking the dog or playing croquet or something.
I've yet to see an LLM response or an LLM generated diff that suggests removing or refactoring code. Every AI solution is additive; new functions, new abstractions added in every step. Increased complexity is all but baked into the system.
Software engineering jobs involve working in a much wider solution space - writing new code is but one intervention among many. I hope the people blindly following LLM advice realize their lack of attention to detail and "throw new code at it" attitude comes across as ignorant and foolish, not hyper-productive.
I agree they cannot handle a complex codebase at all at this moment in time.
But I think I would rather just end my career instead of transitioning into fixing enormous codebases written by LLMs.
The complexity has grown but not the quality. We went from writing ADA code with contracts and all sorts of protections with well thought architectures, to random crap written in ReactJS in web sites that now weigh more than a full install of Windows 95.
I’m really ashamed of what SWE has become and AI will increase that tenfold as you say. We shouldn’t cheer up on that, especially if I will have to debug all that crap.
And if it increases the number of engineers, they won’t be good due to a lack of education (I already experience this at work). But anyway I don’t believe it, managers will not waste more money on us, that would go against modern capitalism.
I wonder if anyone is applying AI to cobol…
I expect pretty much the opposite to happen: it makes sense for languages, stacks and interfaces to become more amenable to interfacing with AI. If a machine can act more reliably by simplifying its inputs at a fraction of the cost of the equivalent human labour, the system has always adjusted to accommodate the machine.
The most obvious example of this already happening is in how function calling interfaces are defined for existing models. It's not hard to imagine that principle applied more generally, until human intervention to get a desired result is the exception rather than the rule as it is today.
I spent most of the past 2 years in "AI cope" mode and wouldn't consider myself a maximalist, but it's impossible not to see already from the nascent tooling we have that workflow automation is going to improve at a rapid and steady rate for the foreseeable future.
I'd really like to know the parameters are. I hear claims like, "it saves me an hour a day," or, "I'm 30% more productive with AI." What do these figures mean? They seem like proxies for fuzzy feelings.
When I see boring, repetitive code that I don't want to look at my instinct isn't to ignore it and keep adding more boring, repetitive code. It's like seeing that the dog left a mess on your carpet and pretending you didn't see it. It's easier than training the dog and someone else will clean it... right?
My instinct is to fix the problem causing there to be boring, repetitive code. Too much of that stuff and you end up with a great surface area for security errors, performance problems, etc. And the fewer programmers that read that code and try to understand it the more likely it becomes that nobody will understand it and why it's there.
The idea that we should just generate more code on top of the code until the problem goes away is alien to me.
Although it makes a lot more sense when I probe into why developers feel like they need to adopt AI -- they're afraid they won't be competitive in the job market in X years.
So really, is AI a tool to make us more productive or a tool to remove our bargaining power?
> So really, is AI a tool to make us more productive or a tool to remove our bargaining power?
Don't you notice how it makes you more productive, that you can solve problems faster? It would be really odd if not.
And regarding the bargaining power: that's not the other side of the scale, it's a different problem. If your code monkey now gets as good as your average developer, the average developer will have lost some relative value, unless he also upped his game by using AI.
If everyone gets better, why would you see this as something bad, which makes us lose "bargaining power"? Because you no longer can put the least effort which your employer expects from you? Even then: it's not like AI makes things harder, it makes them better. At least for me software development has become more enjoyable.
While 5 years ago I was asking myself if I really want to do this for the rest of my career, I now know that I want to do this, with this added help, which takes away much of the tedious stuff like looking up solution-snippets on Stack Overflow. Plus, I know that I will have to deal less and less with writing code, and more and more with managing code solutions offered to me.
>it makes a lot more sense when I probe into why developers feel like they need to adopt AI -- they're afraid they won't be competitive in the job market in X years.
Amazing. You think that the only reason people are using AI is because it's being forced on them?
I honestly feel kinda bad for some people in this thread who don't see the freight train coming.
Well to give to a concrete example.
I use it to write test cases for the CRUD applications that I sometimes have to work on. Some test cases already exist and I feed the tests and actually code including additional instructions into a model and get relatively decent output. We also use a code review bot that we feed repository relevant instructions to and get decent basic PR comments. It even caught an edge case that 3 other developers didn't consider.
I think AI can be yet another tool that takes some repetitive tasks off my hands. I still obviously check all the code it generated.
Sort of off-topic, but is there any generative AI for code? From my limited understanding, the code is trained on human written code, and the model adapts it to what most closely matches.
What I'm curious about is, can it find innovative ways to solve problems? Like the infamous Quake 3 inverse-sqrt hack? Can it silently convert (read: optimize) a std::string to a raw char* pointer if it doesn't have any harmful side effects? (I don't mean "can you ask it to do that for you?" , I mean can it think to do that on its own?) Can it come up with trippy shit we've never even seen before to solve existing problems? That would truly impress me.
Take a bloated electron app, analyze the UI, and output the exact same thing but in C++ or Rust. Work with LLVM and find optimizations a human could never see. I remember seeing a similar concept applied to physical structures (like a small plane fuselage or a car) where the AI "learns" to make a lighter stronger design and it comes out looking so bizarre, no right angles, lots of strange rounded connections that almost like a growth of mold. Why can't AI "learn" to improve the state of the art in CS?
> Take a bloated electron app, analyze the UI, and output the exact same thing but in C++ or Rust. Work with LLVM and find optimizations a human could never see. I remember seeing a similar concept applied to physical structures (like a small plane fuselage or a car) where the AI "learns" to make a lighter stronger design and it comes out looking so bizarre, no right angles, lots of strange rounded connections that almost like a growth of mold. Why can't AI "learn" to improve the state of the art in CS?
So such things already exist, and for me, the most frustrating thing about LLMs is that they just suck the oxygen out of the room for talking about anything AI-ish that's not an LLM.
The term for what you're looking for is "superoptimization," which tries to adapt the principles of mathematical nonconvex optimization that AI pioneered to the problem of finding optimal code sequences. And superoptimization isn't new--it's at least 30 years old at this point. At this point, it's mature enough that if I were building a new compiler framework from scratch, I'd design at least the peephole optimizer based around superoptimization and formal verification.
(I kind of am putting my money where my mouth is there--I'm working on an emulator right now, and rather than typing in the semantics of every instruction, I'm generating them using related program synthesis techniques based on the observable effects on actual hardware.)
> To do so, Mr. Giorgi has his own timesaving helper: an A.I. coding assistant. He taps a few keys and the software tool suggests the rest of the line of code. It can also recommend changes, fetch data, identify bugs and run basic tests. Even though the A.I. makes some mistakes, it saves him up to an hour many days.
> Still, nearly two-thirds of software developers are already using A.I. coding tools, according to a survey by Evans Data, a research firm.
> So far, the A.I. agents appear to improve the daily productivity of developers in actual business settings between 10 percent and 30 percent, according to studies. At KPMG, an accounting and consulting firm, developers using GitHub Copilot are saving 4.5 hours a week on average and report that the quality of their code has improved, based on a survey by the firm.
We're in for a really dire future where the worst engineers you can imagine are not only shoveling out more garbage code but the ability to assess it for problems or issues is much more difficult.
> We're in for a really dire future where the worst engineers you can imagine are not only shoveling out more garbage code but the ability to assess it for problems or issues is much more difficult
It will probably still be more productive. IDEs, Stack Exchange...each of these prompted the same fears and realised some of them. But the benefits of having more code quicker and cheaper, even if more flawed, outweighed those of quality. The same way the benefits of having more clothes and kitchenware and even medicine quicker and cheaper outweighed the high-quality bespoke wares that preceded them. (Where it doesn't, and where someone can pay, we have artisans.)
In the mean time, there should be an obsolescence premium [1] that materialises for coders who can clean up the gloop. (Provided, of course, that young and cheap coders of the DOGE variety stop being produced.)
[1] https://www.sciencedirect.com/science/article/abs/pii/S01651...
> At KPMG, an accounting and consulting firm, developers using GitHub Copilot are saving 4.5 hours a week on average and report that the quality of their code has improved, based on a survey by the firm.
I don't have any specific experience with KPMG, but considering the other "big name" firms' work I've encountered, there's, uh, lots of room for improvement.
I made good money cleaning up after the 2000s outsourcing boom.
It was lucrative cleaning up shit code from Romania and India.
I'm hoping enough people churn out enough hot garbage that needs fixing now that I can jack up my day rate.
I remember when the West would have no coders because Indian coders are cheaper.
I remember when nocode solutions would replace programmers.
I remember.
The increase in productivity means you need fewer inexperienced and/or bad engineers to a project. On the other hand, they may be retained to go after bolder, more numerous targets.
I don’t think that future will happen, because eventually someone will realize there is a competitive advantage in building a truly good product with people who actually know what they’re doing, and when other companies catch on they will start doing that and bad prompt kiddy engineers will be gone.
I'm a big user of LLM tools.
The problem, so far, is that they're still...quite unreliable, to say it least. Sometimes I can feed the model files, and it will read and parse the data 100 out of 100 times. Other times, the model seems clueless about what to do, and just spits out code on how to do it manually, with some vague "sorry I can't seem to read the file", multiple times, only to start working again.
And then you have the cases where the models seem to dig themselves into some sort of terminal state, or oscillate between 2-3 states, that they can't get out off - until you fire up a new model, and transfer the code to it.
Overall they do save me a ton of time, especially with boilerplate stuff, but very routinely even the most SOTA models will have their stupid moments, or keep trying to do the same thing.
Are you including the time you spend fighting the model?
You could be describing the performance of me and most of my friends and colleagues over the past five years.
It’s insane how similar non-deterministic software systems already are to biological. Maybe I’ve been wrong and consciousness is a computation.
I always thought hacking scenes in sci-fi were unrealistic, but if you're cooking up AI-fortified code lasagna at your endpoints, there are going to be a mishmash of vulnerabilities: Expert robust thought will be spread very thin by the velocity that systemic forces push developers to.
> Mark Zuckerberg, Meta’s chief executive, stirred alarm among developers last month when he predicted that A.I. technology sometime this year would effectively match the performance of a midlevel software engineer
Either Meta has tools an order of magnitude more powerful than everyone else, or he's drinking his own koolaid.
At some point in the past, tools like Wordpress et al made it easy for the average person to roll out their own website.
This probably increased the overall demand for professional website makers and messed-up-Wordpress-fixers.
Now the argument goes that the average business will roll out their own apps using ChatGPT (amusing / scary), or that big software co's will replace engineers with LLMs.
For this last point, I just don't see how any of the current or near-future models could possibly load enough context to do actual engineering as opposed to generating code.
I've found that AI has saved me time consulting Stack Overflow. It combines thorough knowledge of the documentation with a lot of practical experience gleaned from online forums
It has also saved time producing well-defined functions, for very specific tasks. But you have to know how to work with it, going through several increasingly complex iterations, until you get what you want.
Producing full applications still seems a pipedream at this stage.
>Producing full applications still seems a pipedream at this stage.
Do you mean like: "write me an app that does XYZ?"
Well, it's a pipedream because you probably couldn't even get a room of developers to agree on how to do it. There are a million ways.
But this isn't really how programmers are expecting to use AI, are they?
That's kinda still how I get the most out of it - search, more or less. Claude gives me great starting points from which I can do some refining/confirming searches and documentation lookups. _Starting_ with search feels like a drag now. But the information and code I get is unreliable at least 20 % (just a guess, frankly, did no statistics) of the time, so I treat the output as things to try or investigate, rather than things to ship.
You'll probably get a few responses from folks that happily tab complete their software and don't sweat the details. Some get away with that, I'm generally not in a position where it's OK to not fully understand the system I'm building. There's a lot of stuff that's better to find out during development than in a late night production system debugging session.
New to AI assisted coding, but I'm finding myself spending a lot of time debugging its convincingly wrong output.
I've been choosing not to use most of the AI code assistant stuff for a while, I try it every now and then. Each time it's the same outcome, it actively reduces my productivity by a fair amount. I suspect this is due to a mix of the majority of my programming being non-trivial (library building, complex-ish algos), and that I'm a bit of a perfectionist coder who enjoys programming.
LLMs are useful tools for programming, as a kind of search engine and squeaking rubber duck. AI as a programmer is worse than a junior, it's the junior that won't actively learn and improve. I think current AI architecture limits it from being much more than that.
It seems like AI will generate opportunities for fixing code. Both in reducing internal technical debt ("code maintenance", which is a specialized skill already) and external technical debt (architecture, which is being built by AI also). Eventually AI will be good enough for both of these things as well, and then we may just become the priests of the Temples of Syrinx.
Our great computers fill our hollow halls.
Business wants short term-solutions. The long-term effects it doesn't care, even if it clearly bites them in the ass.
Microsoft uses the ability to replace software engineers to sell their own AI.
In my Python programming, I have found that ChatGPT makes me something like 10x more productive. From learning to use a new API to tracking debugging, and especially things finding errors in my code from stack traces. Getting results goes SO MUCH FASTER.
However, it has not alleviated any responsibility from me to be a good coder because I have question literally everything little dang thing suggests. If I am learning a new API, it can write code that works, but I need to go read the reference documentation to make sure that it is using the API with current best practices, for example. A lot of code I have to flat out ask it why it did things in a certain way because they look buggy inefficient, and half the time it apologizes and fixes the code.
So, I use the code in my (personal) projects copiously, but I don't use a single line of code that it generates that I don't understand, or it always leads to problems because it did something completely wrong.
Note that, at work, for good reasons, we don't use AI generated code in our products, but I don't write production code in my day job anyway.
Nice cope from programmers. But reality hits hard.
There's always one in the comments. Even way before AIs, with any discussion on low code, hiring tech turns, etc.
What did programmers do to you to trigger such deep-felt insecurities thusly?
People got wildly out of hand thinking that AI would do what we currently do without us. The real truth is AI is gonna do 10x what we currently do with us.