AI is great at solving simple, well-defined problems but bad at integration and maintainability; that's why it'll never truly replace senior engineers.
Posted by enador@reddit | ExperiencedDevs | View on Reddit | 137 comments
Like in the title. I see a lot of doomposting regarding AI recently, but I think that AI development shouldn't really affect senior devs. It impacts them mostly indirectly, through misguided management. I didn't see this angle discussed (maybe I missed it), so I'll discuss it here.
It's difficult to argue that AI is great at quickly coming up with solutions to well-defined, self-contained problems. At the same time, if your prompt is generic enough and the problem complex enough, AI will build an ungodly monstrosity that's impossible to maintain. This is because a simple, well-defined problem becomes an open question, and here the hallucinations begin.
However, even complex issues can be divided into a lot of smaller, well-defined ones. To divide the complex problem into smaller ones, one needs an engineer. The most important part of being a senior engineer is being able to turn a complex issue into a finite number of maintainable and well-defined steps to solve it. This is something that AI is not good at and never will be because turning one task into countless smaller tasks increases the cost and complexity of reasoning exponentially. As long as AI tries to be cost-efficient, and it's forced to do it by competition, it will produce code that's just good enough for marketing but actually bad enough that the actual engineering effort is irreplaceable.
This is why senior engineers will never be replaced, and AI is a tool useful mostly for them. They can define the problem as a set of smaller subproblems that AI is good at solving, and they can use the generated parts to compose the sound product that's easy to maintain.
AI hits the juniors the hardest because before it was often their job. However, in the process it creates the new gap: it will become harder to become a senior engineer, so the value of one will increase in time. When it increases enough, the need for producing new engineers will return eventually. It's just that it will become a more prestigious profession, with an entirely different road and methods of education.
I think this is where we stand now. Personally, I enjoy AI because I always preferred the high-concept work rather than being a coding monkey. I'm glad that AI took this part away from me, but each one's situation is different, so I genuinely understand the uncertainty and fear. Though I think that whoever survives this test will be much better off long-term than before.
SplendidPunkinButter@reddit
We are waaaaaaaaaaaay past the point where the market should be flooded with AI generated killer apps if AI weren’t bullshit. Instead the market is flooded with slop. Why? Is the AI not smart enough to build good apps? If you need a skilled person to properly use AI in the first place, then it sure sounds like it’s not going to replace people.
OkLettuce338@reddit
Holy hell look around. The market is flooded with ai generated apps.
Tysonzero@reddit
Did you just not read their comment or…?
OkLettuce338@reddit
No I read it. The market IS flooded with vibe coded apps. In addition, their comment is loaded and is like saying “there should be lots of cars on the road instead we just see crap.” “Slop” is not defined. And we DO see the market flooded with AI generated apps.
Tysonzero@reddit
Their comment was perfectly reasonable and clear. If AI was genuinely good at building things fully autonomously without needing good human engineers, then we should see tons of new GOOD apps everywhere. We don’t. QED.
That’s not to say this won’t change in the next few years, maybe we’ll have a genuinely fun GTA6 competitor that was built fully by AI, maybe not. We shall see.
OkLettuce338@reddit
Seems like you didn’t read the comment
Iannelli@reddit
...and do they work?
OkLettuce338@reddit
Ever used “Claude”?
BobFellatio@reddit
«Will never», it doesnt matter what its bad at today, you gotta look at the trajectory and rate of change.
Gunny2862@reddit
This has been my mistake in judging new tech in the past. Letting my first impressions crystalize my long-term impressions.
EnderMB@reddit
You could've said the same thing about almost every technological advancement - the same was said for WYSIWYG editors replacing frontend dev, and look how that turned out.
As someone that spent 2-3 years working in one of the Nova models, every Applied Science team in AI, regardless of company, is aware of the fundamental limitations of the tech. These have been known of LLM's for decades, and despite a lot more research, we're nowhere near "solving" the hallucination problem.
IceMichaelStorm@reddit
I wouldn’t say hallucinating is the largest problem right now or what do you mean exactly?
tenthousandants44@reddit
Token generators are not truth generators
calvintiger@reddit
No and neither are humans, which is why we have things like unit tests and various other forms of verification.
OtaK_@reddit
Humans are accountable for their mistakes on the job. A LLM much less so.
Additionally I can fire a human that sucks at their job, and hire someone competent in their stead. Not something I can do with Claude. Getting rid of something mediocre to get something that is just strictly worse instead? Urgh.
tommyTurds@reddit
No one expects LLMs to just be doing things randomly. They’re acting at the behest of a human who is responsible for their mistakes.
realdevtest@reddit
I can tell that this comment was generated by something that just predicts what word comes next without any reasoning
Wonderful-Habit-139@reddit
Good ol "Humans are just like LLMs, they don't reason but just predict the next series of word to say"...
demosthenesss@reddit
I haven’t had a serious issue with hallucination for almost a year.
BobFellatio@reddit
No you couldnt, this is fundamentally different, and in the span of three years AI coding has gone from a amusing tech demo to almost all developers not developing by hand anymore. The adoption rate and rate of change is unlike anything we have seen before.
quentech@reddit
https://xkcd.com/605/
BobFellatio@reddit
Hehe, perhaps what i said, but not exactly what i meant. I did find it funny tho. Theres an xkcd for everything.
invest2018@reddit
Connecting the start and end points with a line and extending it to infinity is unlikely to be mathematically valid.
BobFellatio@reddit
Nobody said anything about infinity. Just better and cheaper. Having the line suddenly plateu at the current level is also unlikely.
another_dudeman@reddit
The innovation is now in the harness though
nacholicious@reddit
Also the trajectory of progress relative to the rate of spending. It's clear you get much better performance from a trillion dollars than hundred billion, but is it 10x better
BobFellatio@reddit
Not necessarily, its not unlikely that we will find ways to increase intelligence while significantly reducing size and training cost of models.
If you think of pretty much every IT technology we've invented, we have found ways to make it much better and much cheaper in parallel. Some examples that come to mind: Mobile phones in the 90s: Expensive, large, heavy and really shitty compared to what we have today. Computers back in the day: Filling a whole building while being dumber than your toaster. Internet when it first arrived: you had to pay per minute and the speeds was incredible slow. Gaming: At some point tetris was state of the art. Digital cameras in the early 2000s, 1 megapixel (1024x1024) was considered good, and it filmed in 320p. Now your phone outclass them to such a degree that they have gone extinct.
I read a paper 6 months ago about a chinese open source model that was almost as capable as Chat GPT 4o (but without the omni part), while costing 1/100th to develop, and roughly 1/10th to run. I dont remember the name of it, and it might have just been some over hyping, but its also not impossible. Im pretty sure it will happen anyhow, going forward.
akc250@reddit
Sure, but I wonder if the trajectory is really changing that fast anymore, when AI companies are now rushing to increase costs and shut down video generation models. Google sat on LLM technology for a decade. It only became huge as OpenAI decided to release it to the public and Google had to rush it out to consumers to compete.
notger@reddit
This sub is becoming more and more of a self-help group where people do group hugs and chant "we are irreplacable".
Any prediction about the future is likely to be wrong and the only way to know is to experience it. In the meantime, you are well-advised to plan for the worst and expect the best. Blinding yourself to a likely future is not a sane strategy.
Lopatron@reddit
It's a really exciting time. You have world class scientists and engineers saying it's already 10xed them, who are already 10x people themselves. On the other hand you have senior engineers who think the following is a true statement:
notger@reddit
I love that post. So much true stuff in it.
geon@reddit
Some people are predicting that AI will fail, others that it will succeed. Those are mutually exclusive, hence one group must be right.
fsk@reddit
There are several levels of success and fail.
geon@reddit
The marketing claim is that the agents are as good as or better than human developers. Phd level intelligence etc.
So the bar for success would be at least 1x standalone productivity.
Whitchorence@reddit
It seems like if you get automated out of a job it's cold comfort that the productivity gain was not quite as high as advertised.
geon@reddit
So far it seems the ai slightly impedes productivity.
tommyTurds@reddit
Phd level intelligence, in the context of writing actual production quality code, is always one of the funniest claims to me.
Every phd I’ve ever worked with writes absolute dogshit code. Most of my career is taking their ideas (which can be great and are things I’m not nearly smart enough to do) and turning that into something that can actually run in production.
I mean i get what you’re trying to say, but it’s just one of those things that kind of makes me chuckle.
All that aside, opus is genuinely better than at least 30-40% of the engineers at my megatech. That’s not to say it’s good. That’s more to say that there’s a lot of mediocre to bad engineers. Those people in can absolutely see being pushed out of the industry in the next 10 years (assuming they can get costs and all that shit to a reasonable level)
andrewharkins77@reddit
This needs to be higher. A post doctorate is either someone playing around with work visas, or couldn't find a job. It's rarely a requirement for Software Developers.
fsk@reddit
The AI agents are not completely replacing human developers. There usually still is a human driving them.
Example: If the human salary is $10k/month and the AI costs $1k/month, then the breakeven is a 1.1x productivity gain compared to the human working by themselves.
geon@reddit
I know. But that’s not what’s being marketed.
Big_Bed_7240@reddit
This might be the dumbest ”smart” thing I’ve read.
notger@reddit
Not really, as the predictions aren't covering all of the outcome space. The outcomes are not binary.
Also: I am not saying that neither group will be right, I am saying that following either side is a bad strategy as it leads to bad decision making. Do not join any church, prepare for either or neither of them being right.
12tone@reddit
Here's my prediction. It will succeed at some things and fail at others. I'm pretty sure that I'm guaranteed to be right.
geon@reddit
Exactly.
12tone@reddit
But yet I am in neither of the groups you mentioned
SplendidPunkinButter@reddit
Except that if you know anything about computer science theory, you know that coding is fundamentally not “solvable.” The problem of translating human language to code is not a thing that a computing machine can do in the general case, period. Having an LLM that’s good at producing plausible-sounding text does not change this provable fact.
Whitchorence@reddit
I like Dijkstra's essays as much as the next guy but we've been doing things that are theoretically unsound in that way since the field was invented.
another_dudeman@reddit
Someone will still need to make sure this software shit still technically works. I don't ever see the dumbass POs and managers that can't even operate a spreadsheet do that job.
Whitchorence@reddit
Then the work will still have substantially changed and become de-skilled
another_dudeman@reddit
Nope, there will still need to be skilled people to troubleshoot and verify
Etiennera@reddit
My work has been piloting AI heavily for a half year or so, and every post like this always claims AI fails at something I have it do every day. Sure, I need to operate and nudge it, but still it does the bulk.
Significant_Mouse_25@reddit
This and many other subs. I’m no AI bro. But clearly I’m in a different world than most. I’ve seen posts claiming it can’t even do unit tests. What models are people using? What do their prompts look like?
AI is here. Likely not going away. Get better at it. It’s pretty handy.
SplendidPunkinButter@reddit
Yeah it can “do unit tests” in that it creates functions that run as tests, and they often cover useful cases. It also writes redundant tests, and tests for cases that make no sense. It also misses crucial test cases that most humans wouldn’t.
“But humans make mistakes too.” Sure, but the entire point of using a computer is that it’s not supposed to make mistakes.
Significant_Mouse_25@reddit
The entire point of using a computer or any technology is efficiency to free up humans to do human work. Llms do that fine. And I’ve literally never seen opus do what you just said. It sometimes writes tests that fail. But generally the test cases are useful and unique.
Eskamel@reddit
No the entire point of a computer is to do something consistently otherwise you can never have freed up humans if they have to chase random output that even has 1% chance to fail.
Significant_Mouse_25@reddit
Incorrect. Computers use random number generation to useful ends for example. Your software isn’t bug free and doesn’t have 100% uptime. Automation exists to free up the human as much as it can. Full stop.
Eskamel@reddit
Random number generator is a niche usecase for things where the specific number's consistency is intentionally not important, yet the rest of the software around it is.
Also random number generators aren't fully non dererministic due to the way they are created themselves
No one talked about uptime or bugs, deterministic software can have infinite bugs yet keep on being deterministic. Same for uptime, there are many reasons why uptime can change, even due to things that aren't related to the software itself.
tommyTurds@reddit
Replies like this tell me that people just aren’t reading the shit that gets pooped out.
Opus can absolutely write useful tests. It will *always* create redundant cases that have to be cleaned up. I
Izkata@reddit
I recently ran across one where the test names were offset from the test bodies. They each described the wrong test.
Whitchorence@reddit
IME if you ask it to follow a TDD methodology (I basically use the one that comes from the Superpowers package with some modifications) the tests are much more likely to be meaningful and substantive than if you let it implement and then write tests (tends to have meaningless tests written to the implementation more than anything else).
Early_Rooster7579@reddit
Somehow everyone here works on truly novel projects every single day.
roger_ducky@reddit
For LLMs currently, anything not CRUD needs a lot of structure to work well.
The same kind of documentation that helps junior engineers helps the LLMs a lot, too.
Early_Rooster7579@reddit
What isn’t crud though? I’ve used it pretty heavily in TS, PHP and rust without much issues on a pretty broad spectrum of software
HealthyInstance9182@reddit
In addition now it’s so much easier to set up property-based tests and mutation tests which improves the quality of the unit tests
rlbond86@reddit
I would say it's more than a nudge, it often tries to do things suboptimally even if there's a correct example in that same file, then I ask why it did it that way instead of the same way as other functions in the same file and it gives that "oh you're right" comment.
I don't like when I get the bash safety prompt so I put "don't ever put a # in any bash command" in CLAUDE.md, it just completely ignores that anyway, and when I ask why it just apologizes.
Yes it's a super useful tool but I would say a junior engineer would easily not catch all the dumb mistakes it can make.
Will it get better? That's not totally clear. I could see a world where these are just short term issues but there's also the possibility that this is a fundamental limitation of LLM technology and some new breakthrough is necessary for a real revolutionary improvement.
ltdanimal@reddit
"Will it get better? That's not totally clear."
There is no way it doesn't get bette.
I think progress could slow some, but I don't see how someone can look at the strong up and to the right graph and not see how it is guaranteed to get better it's just by how much.
Even if almost nothing changed past a 20x inference speedup you'd have massively better system.
But the reality is the smartest people on the planet are now converging to this space. We're going to see huge leaps around areas that we didn't even think about a few years ago.
rlbond86@reddit
I think there's a difference between a series of marginal improvements (which is guaranteed) versus some kind of paradigm shift where it actually thinks for itself.
bupkizz@reddit
Yep. I 100% agree.
AI is an amazing tool. Like all tools it is used by a human.
Seniors aren’t replaceable if they get good at using the tools. Full stop.
Early_Rooster7579@reddit
For real. We just pulled off an incredibly complicated integration with a fairly low documented system in like 2 weeks with AI. This would’ve been a 6 month endeavor in 2022.
Affectionate_Link175@reddit
This sub is still trying to cope with reality
writesCommentsHigh@reddit
Here’s a safe prediction tho: computers get better and cheaper every year.
GPUs are growing even faster.
LLLMs are still evolving as well.
It’s pretty safe to say that AI will get better and better every year in some way shape or form.
If true than OPs premise is probably out the window
caprisunkraftfoods@reddit
I think the sun will rise tomorrow
Jeth84@reddit
Nailed it
ParticularBeyond9@reddit
This sub's posts would be accurate if they're basing their opinions on GPT 3
vooglie@reddit
I have my moral qualms with ai as much as the next guy but yes - blinding yourself to the truth doesn’t help anyone
cbusmatty@reddit
Well it helps most of us actually. There is likely to be a reduction in total jobs at the end of this, and having a sub full of people burying their head in the sand and not using these tools, building skillsets, can only help most of us who (like all good developers since the dawn of time) embrace change, adapt and grow
enador@reddit (OP)
The fact that the future is always uncertain to some degree doesn't mean that discussing it is pointless.
notger@reddit
Very true, but what is the 20-th iteration of the same argument which does not add anything new?
What use does it add except karma-farming from the echo-chamber and giving you a warm and fuzzy feeling? I imagine a horse-carriage driver in 1910 was having the same arguments in his horse-carriage driver pub after work.
Altruistic-Bat-9070@reddit
I think OP was just saying the sentiment of this post is pointless, not the discussion, which is probably a fair critique.
Havius@reddit
lol true
internetroamer@reddit
This missing the whole point which is impact on labor market
You don't need to replace senior engineers. You can just double output and then justify to lay off 10-25% of workers or just not hire new ones you normally would have
Whitchorence@reddit
I'm not saying we all need to become Chicken Littles, but a year ago "agentic engineering" was a total joke beyond very simple things and now, with enough guidance from an engineer, it's real and I am personally shipping lots of stuff to prod using those techniques. I am not going to rule out the technology advancing even further and making my role much less important. I mean, I hope not, and the last mile is often the hard part (look how hard it's been to get over the line with full self-driving), but there is a bit of whistling through the graveyard going on if you're thinking it's flat out impossible.
ContraryConman@reddit
Assuming everything in your premise is true, for AI to "never" replace senior engineers, it would:
Have to never get any better than it is now.
Not convince companies that code is cheap enough to generate that maintainability no longer matters
And you have no proof that some combination of the two won't happen in the near future.
bizcs@reddit
My current take on all this is that, when the true cost of tokens finally emerges, employers are going to be selective about how many tokens they give to which users. As unreliable stuff gets pumped out that starts costing something, employers will adopt a view that converges around token budgets being allocated to people that can make effective use of them.
But I could be completely wrong and that may be false hope.
call-the-wizards@reddit
It's bad at integration and maintainability now.
wrex1816@reddit
Totally original post. So glad someone finally had the courage to say this.
nomoreplsthx@reddit
Ah yes, I, a person with no relevant expertise, will confidently predict the trajectory of a technology indefinitely the most exceptional experts in the field cannot project the capabilities of six months out.
This is not a terrible characterization of the current state of AI programming tools. But the idea that this can be projected forever is laughable.
These newfangled biplanes may be the bees knees, but airplanes will never challenge the ocean liner!
Arquebuses are good at something, but archers will always have a place on the battlefield.
To be fair the mockery in the other direction works too.
The area of the airplane is over, future transport will be by rocket!
People who make confident predictions about techology are idiots or charletans. Universally.
aWalrusFeeding@reddit
Ai is good at those problems right now and won't replace dev right now
Navadvisor@reddit
Never is a long time buddy, the advances made in the last few years have already been a revolution and as far as I can tell there is nothing limiting the current techniques from going further. Maybe it will just even out and stay where it's at, but I'd personally bet on at least linear growth in capabilities which would be huge enough over the next 5 years with a side bet on exponential growth being a real possibility.
Snoo87743@reddit
Anytime a "simple" jira ticket, which has only the title, needs implementing AI can five you aomewhat solid solution. However these kind of tickets usually require deep investigation and often require refactoring some other code to even start the ticket. Thats where AI fails; when the scope is not well defined and you use it to save time and investigate instead of you
EatMoreKaIe@reddit
Actually, I had this exact situation happen to me yesterday and Claude really surprised me. I had the one liner ticket that had been created by an employee who no longer works here and no one else on the team knew what it was about. On a whim, I got Claude to use a few MCPs to go through all the slack threads and other doc storages and fill in the missing details and sure enough, it managed to piece it all together and summarize what needed to be done in a succinct way. It also did all this in less than a minute which is way faster than I could have done.
Then once it has all this context and combined it with a deep understanding of our code base (which is very old and very large) it came up with a refactoring plan that took into account business related edge cases that I would never have thought of.
I think context is the key. With enough context, it will figure out it's own scope. It's not enough that Claude can see all the source code but if it can combine that with knowledge of everything else that's going on in the company interesting things can happen. I haven't had hallucinations or slop generated for many months now.
The uncomfortable part is that at my company we're recording all meetings and conversations so that all of this becomes part of the knowledge base. Privacy is pretty much gone now.
another_dudeman@reddit
This is where AI really helps out.
Snoo87743@reddit
Oof that sounds good. In my current start up rarely anything is written down, which i hate, makes it hard to gather the data
coworker@reddit
But AI helps you do that investigation much faster
Snoo87743@reddit
Sometimes yes, but it can often make mistakes. Not updated docs, claude files if you share with team. Another teammate developing that ai does not have context etc...
Sp00ky_6@reddit
I think it comes down to what decisions we are willing to leave to LLMs. Agents can figure things out and take action, but where does responsibility live? How do we build trust and observability on these systems? Also the total costs (both $ and cognitive/organizational) are still mostly unknown, but they are likely to be significant.
The tools are here and pretty useful, but they can also be abused. Finding the balance is going to take a year or two as we all figure out what we want out of these things.
I’m definitely anxious that LLMs will shrink demand, but I’m finding I spend a lot more time thinking about approaches and organizational problems than I do code, but I’m a data engineer so it’s a little different.
Grounds4TheSubstain@reddit
One would be a fool to look at how far ML has come and think, "that thing is never going to replace me". Did you see that, having done so well on basic SWE benchmarks, that they have now introduced new benchmarks involving writing whole programs? This is where improvements will go in the next few years: https://arxiv.org/abs/2605.03546
another_dudeman@reddit
Sure bud
geon@reddit
The mens high jump world record went from 2 m in 1912 to 2.45 m in 1993.
https://en.wikipedia.org/wiki/Men%27s_high_jump_world_record_progression
Extrapolating from that, they will jump to the moon in about 854 million years.
coworker@reddit
With evolution at play, your extrapolation is actually reasonable though
Fair_Local_588@reddit
I think exit velocity for the Earth is like 16,000 mph so no, not realistic, which is the point.
coworker@reddit
First off, it’s 25,000 mph, so if you're going to kill my dream with physics, at least use the right math. Second, 854 million years ago, Earth’s atmosphere didn't even have enough oxygen to support a single insect, let alone a high jumper. If the atmosphere changes again, who's to say it won't become a super-dense hyper-fluid where a solid leg-kick floats you right into orbit? Don't put a ceiling on my evolution.
enador@reddit (OP)
It's true that it came far, but at the same time, what makes good design? I think an often underappreciated skill is a knowledge of one's own limitations. It makes me put more effort upfront so the future me won't suffer. AI doesn't have that because that would require self-reflection, and at this point, we have bigger problems than the job market. Without this, iterating over the same codebase multiple times will cause accelerated code degradation. Humans do it; AI will be even worse at it. It's very tempting to introduce unreadable micro-optimizations, and AI does this because it's less aware of consequences. I think that's its fundamental limitation.
Grounds4TheSubstain@reddit
To make progress on this benchmark in the future, it will train on existing, successful software while developing and being graded on its own design choices, meaning it will be able to absorb the architectural design lessons of successful software via reinforcement learning feedback.
tenthousandants44@reddit
That's already how it works
enador@reddit (OP)
So far, it has learned so much in so little time because it has had all the human knowledge readily available. So, now it will need to produce new knowledge? I think you underestimate how costly that would be.
Grounds4TheSubstain@reddit
First, I was talking about the ability to train on the architecture of existing software, so your question is not relevant.
Second, are we really still having this debate about whether LLMs can do new things? Do you use agentic coding tools and follow the field more generally? I used them to write a compiler for a complex language that's not in the training set. Or this article from this week, which has mathematicians sounding the alarm? https://openai.com/index/model-disproves-discrete-geometry-conjecture/
enador@reddit (OP)
Of course it can extrapolate the current knowledge to some degree. I meant that to learn from its mistakes, it would need to be put in a pipeline that would run AI and test its results constantly, which is unaffordable. It could be done for "solving" chess, but not for arbitrary complex business requirements. And as long as it cannot learn from its mistakes, there is a ceiling to its capability.
Grounds4TheSubstain@reddit
Dude, look up what reinforcement learning is. What you're describing is not "unaffordable", it's how reasoning LLMs are being trained currently , and have been being trained since the beginning.
enador@reddit (OP)
You are using these terms like I wouldn't know them. How would you exactly apply reinforcement to things that have no precedent in the current knowledge? Reinforcement works in a well-defined scope only.
Grounds4TheSubstain@reddit
By allowing them to do whatever they want, scoring the end result, and backpropagating the score across the the weights that lead to the decisions along the way. The same fundamental idea as everything else in machine learning. How do you think AlphaGo managed to beat humans at Go?
enador@reddit (OP)
Go has a limited set of well-defined rules. Nobody argues that AI is great at solving such problems.
i_do_floss@reddit
Ehh even those are well defined program specs that need to be turned into code.
Operating in a business environment where you dont even know the spec is still a level above that
IceMichaelStorm@reddit
I mean, the article says in the abstract that the models fail miserably, so it’s somewhat related but somewhat not. But also a typo in the abstract makes me not trust the thoroughness of the authors to be honest.
I agree with your first sentence, of course.
Grounds4TheSubstain@reddit
This is the next generation of benchmarks after SWE-Bench. It just came out. Nobody has yet put effort into it. The point is not "they don't do a good job at this now", it's "here's how they're going to evaluate the next generation of models when they do start putting effort into this".
IceMichaelStorm@reddit
I’m not sure.
This is about writing/copying apps from scratch. It can be useful but mostly I doubt it. I usually want to rewrite the AI app because it has POC code level and I want a proper one.
Where AI fails most for me is working on existing code bases that have grown into being complex. It also shows that these are badly architected but it happens over time with market-pressured managers.
But this also means, you need way more context to work within these large software systems. That’s one of the main weakpoints of it
alexs@reddit
If you say "AI will never X" you are probably wrong.
RoyalCultural@reddit
Literally. Look at what it can do now vs 12 months ago. Absolutely wild to rule anything out at this point.
pvgt@reddit
doesn't matter, we'll get laid off and re-hired at lower salaries unless we have unions. and even with unions they'll try to do that anyway.
Alfanse@reddit
do factor in the rate of capability increase. the AI of today is fastly inferior to the AI of 2 years from now.
Lucifernistic@reddit
I don't know. I find it's pretty incredible at doing relatively poorly defined problems that are almost entirely centered around heavy integration.
AI is able to build things in 3 days, with higher code quality, security, and maintainability, than I could do with a couple of senior / principle devs to help, in 6 months to a year. It doesn't even have to be that well defined.
Do I get wildly better / consistent results if I take the time to help it design a plan, requirements, and architecture? Sure. But even that isn't making as big of a deal anymore since it can just load up skills for all of that.
I think there will always be human devs in the loop but if you think senior engineers are safe you are fooling yourself. You need to find a way to adapt to the times and keep yourself relevant, not rest on the assumption that AI won't be able to do what you do.
Traditional-Dot2587@reddit
i won't say it won't replace every senior devs. I have seen many senior devs that are good at coding but not problem solving. What makes dev more valuable than AI is the problem solving. knowing what needs to be done and knowing that certain things can be flawed. AI is only as good as the instruction given to it.
Devs need to change their way of work, from coding to delegating, guiding, reviewing and setting things. We need to work together with AI rather than saying one or the other. It's just going to be that way moving forward.
At some point, companies will stop investing on the growth of AI and we will hit a sweet spot. It's always been like that with every piece of tech. It grows so fast in the first 5 years and it becomes stable with small improvement here and there.
demosthenesss@reddit
AI has taken away a bunch of the things I do and made them massively more time efficient.
PopularBroccoli@reddit
I don’t even find it useful as a tool. It’s like sitting at a slot machine that has a 30% chance of producing something okay. Why sit pulling the lever over and over again when I could just do it myself in less time?
Wide_Smoke_2564@reddit
Just sounds like you haven’t figured out how to leverage it consistently to be honest.
PopularBroccoli@reddit
Oh i spent a lot of time on it. Best results are from a layered modular architecture where you add just the relevant parts to a fresh project to then run the ai. This massively reduces the context window, allowing for much better results. Those better results are still poor quality
IceMichaelStorm@reddit
I mean, I am somewhat on your page but also somewhat on WideSmoke’s.
I think, these over generalizations are not helpful, the detail matters a lot. Writing lots of CRUD is dead simple, good MD files or skills and it works.
If you need to think of a sound design, AI is a very nice sparring partner, might even do some ground work, but if it implements all of it, horrible, even with best MD files.
But you can use agents to do small steps, which is often much faster than writing it yourself, although review remains critical
PopularBroccoli@reddit
I think im just a faster programmer than a lot of people
IceMichaelStorm@reddit
Maybe yeah
Rymasq@reddit
What I find with AI is, you really need a lot of care with the details.
AI will screw up details. AI added an unnecessary check to ensure that the "smallest object" exists. I reprompted and said "the smallest always exists" after denying the code.
Aggressive-Fix241@reddit
This is the most level-headed take I've seen on this. The decomposition angle is spot on — AI can solve the pieces but someone still needs to draw the map. I've noticed the same with my own workflow: the grunt work got faster but the architecture decisions got more critical, not less.
geon@reddit
I think you accidentally a negation.
enador@reddit (OP)
Good catch, thanks \^\^ . At least you now know it wasn't written by AI lol.
03263@reddit
Ok, convince the people doing hiring and paying for senior devs of this
obelix_dogmatix@reddit
Truly replace senior engineers … famous last words
Syntactico@reddit
I find it works great at integrating with existing codebases. You just have to produce comprehensive instruction documents in tandem with your code.
krsCarrots@reddit
Giving hope when hope is needed 🍻