It can write some types of code. It is fascinating that it can bootstrap moderately complex projects form a single shot. It does a better job at writing unit test (not perfect) then the fellow human programmer (few people like writing unit tests). It can even find bugs and point + correct broken code. But apart from that, AI cannot, or at least not yet, write the code - the full code.
If it could write the code, I do not see why not deploy it more effectively to write new types of operating systems, experiment with new programming languages and programming paradigms. The $3B is better spent on coming up with truly novel technology that these companies could monopolise with their models. Well, the can't, not yet.
My gut feeling tells me that this might be actually possible at some point but at enormous cost that will make it impractical for most intents and purposes. But even if it possible tomorrow, you would still need people that understand the systems because without them we are simply doomed.
In fact, I would go as much as saying that the demand for programmers will not plummet but skyrocket requiring twice as many programmer we have today. The world simply wont have enough programmers to supply. The reason I think this might actually happen is because the code produced by AI will be so vast overtime that even if humans need to handle/understand 1% that will require more than the 50M developers we have today.
If you’re writing simple code, it’s often a one-shot. With medium-complexity code, it gets the first 90% done in a snap. Easily faster than I could ever do it. The problem is that 90% is never the part that sucks up a bunch of time— it’s the final 10%, and in many cases for me, it’s been more hindrance than help. If I’d just taken the driver’s wheel, making heavy use of autocomplete, I’d have done better and with less frustration. Having to debug code I didn’t write that’s an integral part of what I’m building is an annoying context switch for anything non-trivial.
> Having to debug code I didn’t write that’s an integral part of what I’m building is an annoying context switch for anything non-trivial.
That's the problem I've been facing. The AI does 90% of the work, but the 10% that I have to do myself is 20x harder because I don't have as much knowledge of the codebase as I would if I had written all of it by myself, like in the old days (2 years ago).
Yeah that’s been my experience. The generators are shockingly good. But they don’t get it all the way, and then you are left looking at a mountain of code you don’t understand. And by the time you do, you could have just built it yourself.
You are trying to check the code for hallucinations. The AI will just hallucinate back a plausible answer for your questions. It’s entirely detached from the process that generated the code so it doesn’t have any more insight than you do on it.
> you are left looking at a mountain of code you don’t understand. And by the time you do, you could have just built it yourself.
SWEs that do not have (or develop) this skill (to fill-in the 10% that doesn’t work and fully understand the 90% that works and very, very quickly) will be plumbers in a few years if not earlier.
SWEs that don’t understand the strengths and limitations of their tools allowing them to choose the right one for the job won’t be software developers much longer, but they definitely won’t be plumbers. Maybe cycling through the gig platforms or working entry-level retail. Soft, arrogant, maladroit white collar workers make hilariously pathetic trade apprentices.
I’m very confused by this as in my domain space I’ve been able to nearly one-shot most coding assignments since this summer (really Sonnet3.5h) by pointing specific models at well-specified requirements. Things like breaking down a long functional or technical spec document into individual tasks, implementing, testing, deployment and change management. Yes, it’s rather straightforward scripting, like automation on Salesforce. That work is toast and spec-driven development will surge as people go more hands-off the direct manipulation of symbols representing machine instructions, on average.
There is vast difference between writing glue code and engineering systems. Who will come up with the next Spring Boot, Go, Rust, io_uring, or whatever, once the profession has completely reduced itself to pleasing short outcomes?
Maybe some day we'll collectively figure it out. I'm confused how people are getting so much success out of it. That hasn't been my experience. I'm not sure what I'm doing wrong.
Try using the brainstorming and execute plan loops with the superpowers plugin in Claude Code. It encapsulates the spec driven development process fairly well.
This misunderstands what LLM-based tools mean for complex software projects. Nobody expects that you should be able to ask them to write you a whole kernel or a web engine.
Coding agents in particular can be very helpful for senior engineers as a way to carry out investigations, double-check assumptions, or automate the creation of some parts of the code.
One key point is to use their initial output as a draft, as a starting point that still needs to be checked and iterated, often through pair programming with the same tool.
The mid-term impact of this transition is hard to anticipate. We will probably get a wide range of cases, from hyper-productive small teams displacing larger but slower ones, to AI-enhanced developers in organisations with uneven adoption quietly enjoying a lot more free time while keeping the same productivity as before.
But how is the senior engineer to get any work done, if they need to babysit the agent and every two minutes accept/reject it's actions? Genuine question. Letting that thing do "whatever" usually means getting insane multiple thousands lines long pull request that will need to be discarded and redone anyway.
Related - how do you get that thing to stop writing comments? If asked not to do so, it will instead put that energy into docstrings, debug logs and what not, poisoning the code for any further "AI" processing.
Stuff like (this is an impression, not an actual output):
// Simplify process by removing redundand operations
int sz = 100;
// Optimized algorithm, removed mapping:
return lookup(load(sz));
Most stuff in comments is actively misleading.
Also the horrible desire of writing new code and not reading docs, either in-project or on the web...
For writing ffmpeg invocations or single-screen bash scripts, great thing! For writing programs? Actively harmful
Yeah for me the three main issues are:
- overly defensive programming. In python that means try except everywhere without catching specific exceptions, hasattr checks, when replacing an approach by a new one adding a whole “backward compatibility” thing in case we need to keep the old approach etc. That leads to obfuscated errors, silent fails, bad values triggering old code etc
- plain editing things it is not supposed to. That is “change A into B” and it does “ok I do B but I also removed C and D because they had nothing to do with A” or “I also changed C in E which doesn’t cover all the edge cases but I liked it better”
- keep re-implementing logic instead of reusing
Oh, the defensive programming! That thing must have been trained on job interview code, or some enterprise stuff. Heaps of "improvements" and "corrections" that retry, stub, and simply avoid correctly doing stuff for no reason (fix deserialization bug that thing just caused? no, why! Let's instead assume API and docs are wrong and stuff is failing silently so let's retry all api calls N times, then insert some insane "default value in case API is unreachable" then run it, corrupt local db by writing that default everywhere, run some brain damaged test that checks that all values are present (they are, clonk just nuked them), claim extraordinary success and commit it with a message containing emoji medals and rockets).
And these "oh, I understand, C is completely incorrect" then proceeding to completely sabotage and invalidate everything.
Or assembling some nuclear python script like some McGyver and running it, to nuke even the repo itself if possible.
Best AAA comedy text adventure. Poor people who are forced to "work" like that. But cleanup work will be glorious. If companies will survive that long.
>One key point is to use their initial output as a draft, as a starting point that still needs to be checked and iterated, often through pair programming with the same tool.
This matches my experience. It's not useful for producing something that you wouldn't have been able to produce yourself, because you still need to verify the output itself and not just the behavior of the output when executed.
I'd peg this as the most fundamental difference in use between LLMs and deterministic compilers/transpilers/codegens.
The TDD community loves tests and finds writing code without tests more painful than writing tests before code.
Is your point that the TDD community is a minority?
> It does a better job at writing unit test (not perfect) then the fellow human programmer
I see a lot of very confused tests out of cursor etc that do not understand nor communicate intent. Far below the minimum for a decent human programmer.
I see tests as more of a test of the programmer's understanding of their project than anything. If you deeply understand the project requirements, API surface, failure modes, etc. you will write tests that enforce correct behaviour. If you don't really understand the project, your tests will likely not catch all regressions.
AI can write good test boilerplate, but it cannot understand your project for you. If you just tell it to write tests for some code, it will likely fail you. If you use it to scaffold out mocks or test data or boilerplate code for tests which you already know need to exist, it's fantastic.
LLMs can traverse codebases and do research faster. But I can see this one backfiring badly as structural slop becomes more acceptable since you can throw an LLM at it an fix the bug. Eventually you'll reach a stage of stasis where your tech debt is so high, you can't pay the interest even with an LLM.
>It is fascinating that it can bootstrap moderately complex projects form a single shot. It does a better job at writing unit test (not perfect) then the fellow human programmer (few people like writing unit tests). It can even find bugs and point + correct broken code. But apart from that, AI cannot, or at least not yet, write the code - the full code.
Apart from the sanitation, the medicine, education, wine, public order, irrigation, roads, the fresh water system, and public health ... what have the Romans ever done for us?
> It is fascinating that it can bootstrap moderately complex projects form a single shot.
Similar to "git clone bigproject@github.git"? There is nothing fascinating about creating something that has existed around the training set. It is fascinating that the AI can make some variations from the original content though.
> If it could write the code, I do not see why not deploy it more effectively to write new types of operating systems, experiment with new programming languages and programming paradigms.
This is where all the "vibe-coders" disappears. LLMs can write code fast but so does copy-paste. Most of the "vibe-coded" stuff I see on the Internet is non-functional slop that is super-unoptimized and has open supabase databases.
To be clear, I am not against LLMs or embracing new technologies. I also don't have this idea that we have some kind of "craft" when we have been replacing other people for the last couple decades.
I've been building a game (fully vibe-coded, that is the rule is that I don't write/read any lines of code) and it has reached a stage where any LLM is unable to make any change without fully breaking it (for the curious: https://qpingpong.codeinput.com). The end result is quite impressive but it is far from replacing anyone who has been doing serious programming anytime soon.
It doesn’t need to do all of a job to reduce total jobs in an area. Remove the programming part then you can reduce the number of people for the same output and/or bring people who can’t program but can do the other parts into the fold.
> If OpenAI believed GPT could replace software engineers, why wouldn’t they build their own VS Code fork for a fraction of that cost?
Because believing you can replace some or even most engineers leaves space for hiring the best. It increases the value of the best, and this is all assuming right now - they could believe they have tools coming in two years to replace many more engineers yet still hire them now.
> You sit in a meeting where someone describes a vague problem, and you’re the one who figures out what they actually need. You look at a codebase and decide which parts to change and which to leave alone. You push back on a feature request because you know it’ll create technical debt that’ll haunt the team for years. You review a colleague’s PR and catch a subtle bug that would’ve broken production. You make a call on whether to ship now or wait for more testing.
These are all things that LLMs are doing to various degrees of success though. They’re reviewing code, they can (I know because I had this with for 5.1) push back on certain approaches, they absolutely can decide what parts of code adds to change.
And as for turning vague problems into more clear features? Is that not something they’re unbelievably suited for?
Agree. I agree with many of the article's points, but not its conclusion.
> “You got way more productive, so we’re letting you go” is not a sentence that makes a lot of sense.
Actually, this sentence makes perfect sense if you tweak it slightly:
> You and your teammate got way more productive, so we’re letting (just) you go
This literally happens all the time with automation. Does anyone think the number of people employed in the field of accounting would be the same or higher without the use of calculators or computers?
> And as for turning vague problems into more clear features? Is that not something they’re unbelievably suited for?
I personally find LLMs to be fantastic for taking my thoughts to a more concrete state through robust debate.
I see AI turning many other folks’ thoughts into garbage because it so easily heads in the wrong direction and they don’t understand how to build self checking into their thinking.
It never was about code. What was the last time you said to yourself "I feel like grabbing some code right now!"?
Last time you struggled to accomplish something and then screamed "if only I had some code!!!!".
This sounds very, very stupid, doesn't it?
English is not my native language, but I find this wording very annoying.
People/business suffer for having needs unattended or requirements unsatisfied.
Latency too high? Have some code!
Database locked with long running query? Here, take some code!
You want to price exotic financial assets to calculate your risks? Have you tried to generate some code??
This is so strange.
I honestly do not think in terms of "code".
If your kid asked you about your job, would you say "I use the computer"?
AI accellerates my learning. It helps me understand, do more experiments and lured me into reading more code. I think AI also helps me be more productive, but I‘m less focussed and sooner exhausted. I also fear its addictive potential which is why I force myself to take breaks more often and try to not use it every day.
AI fundamentally changed the programming experience for me in a positive way, but I‘m glad that it’s not my full time job. I think it can also have bad effects which can not be easily avoided in fulltime roles under market conditions.
This is my experience as well. It's been great to make an application that's small in scope, doesn't require access to my main project repo and is basically a nice to have value add for the client.
I was already quite adept in the language and frameworks involved and the risk was very small so it wasn't a big time sink to review the application PR's. Had I not been it would have sucked.
For me the lesson learned wrt agentic coding is to adjust my expectations relative to online rhetoric and it can be sometimes be useful for small isolated one-offs.
Also it's once in a blue moon I can think of a program suitable for agentic coding so I wouldn't ever consider purchasing a personal license.
being able to properly deal with scale and security. Being able to be confident that if I am capturing PII data into my application, its as secure as it can be and as secure as if a principal developer had put the architecture together etc.
Mass market SAAS will generally just use other products to handle this stuff. And if there does happen to be a leak, they just say sorry and move on, there are very few consequences for security failures.
I'm thinking a lot about this currently as a recent convert (as of Opus 4.5).
I think this post is on the right track, but like much of this discourse, it isn't really addressing how the technology will grow and the disciplines will adapt.
I'm by no means a doomer, but its obviously a huge change.
Generative coding models will never be 100% perfect. The speed of their convergence to acceptable solutions will decline in complex and novel systems, and at some point there will be diminishing returns to increasing investment in improving their performance.
The cost of software will fall precipitously and it seems unlikely that the increase in the value of programmers / engineers as they currently practice will offset the decline in the price in software. However, following the law of supply and demand, the supply and the amount of software produced will surely grow, and I think someone has to use the models to build software. I expect being trained in software engineering will be very helpful for making effective use of these tools, but such training may not sufficient for a person to succeed in the new labor market.
The scope of problem that a valuable engineer is expected to manage will grow enormously, requiring not only new skills in using generative coding/language models, but also in reasoning about the systems they help create. I anticipate growth in crossover PM / engineering roles. I guess that people who generalize across the stack and current sub-disciplines will thrive and valuable specialties and side-disciplines will include software architecture, electrical engineering, robotics, communication, and business management.
Some people will thrive in this new field, but it may be a difficult transition for many. I suspect that confusion about model capabilities and how to make the most of them and which people are doing valuable things will put a lot of friction and inefficiency into the transition time-frame.
Last thought, given how great models are at coding compared to general of knowledge, administrative, and bureaucratic work, I expect models are widely used to build systems that are supply shocks on such work. I don't think my argument above applies to such workers. I'm worried most about them.
Ok but isn't it a matter of time until AI is good enough at the rest also?
Meanwhile it's depressing that AI is doing the fun part of the job (for me of course, some people quite like the other aspects of the job; and I do like some of them as well, but writing code is still the most fun I have)
I used AI to write some python code and some bazel rules to generate some python code around that to do some new workflow system I wanted to prototype. It just did it, it would make mistakes but since I had it running tests it would fix the code after running the tests.
The big issue is that I didn’t know the APIs very well, and I’m not much of a Python programmer. I could have done this by hand in around 5 days with a ramp up to get started, but it took less than a day just to tell the AI what I wanted and then to iterate with more features.
It's decent at explaining my code back to me, so I can make sure my intent is visible within code/comments/tracing messages. Not too bad at writing test cases either. I still write my code.
I'm saying it's not good enough to write code yet, but good at explaining code. If it can't that means I messed up and I need to make it clearer. Once it's explanation and my meaning line up, the code is good enough and I move on. Meta point is that people are using AI for the wrong thing. It's great at consuming, not creating.
> It’s like saying calculators replaced accountants. Calculators automated arithmetic, but arithmetic was never the job. The job was understanding financials, advising clients, making judgment calls, etc. The calculator just made accountants faster at the mechanical part.
Mechanical and later electical calculators replaced human calculators. Accountants switched from having to delegate computation to owning a calculator.
I spent about 9 years in web/frontend development and when all the AI coding things started coming out, I was admittedly concerned and defensive. I thought the output was gross, it couldn't help me with what I was working on, etc. I recently switched over to the backend, so now I'm dealing with an entirely different set of technologies, architecture, and problems to solve. I've found that, so far, about 80% of my time is spent planning, discussing requirements, determining how to integrate new features into the system or how fixing a bug may affect something else, etc. I use AI for the coding part because the code is sort of an afterthought. I've fully leaned into it and the experience has been fantastic. I ask Claude Code questions and bounce ideas off it. I ask it to write some code for me and I review it and make minor adjustments. The TLDR is I'm not really worried about it taking my job anymore, but I recognize that _not_ using it is going to become very obvious in the near future (it already is). This article pretty much nailed it. Engineering isn't just about pumping out lines of code, the code is just the output of your skills which provide the most value. Once I started looking at things that way, my stress level dropped considerably.
Reasons why the attempted cursor acquisition might not be about replicating cursor (with or without human help):
shutting down competition; market share; understanding user behavior; training data
I disagree with much of this. Programming isn't just a tool we use in pursuit of being an “Engineer” or whatever aggrandizing title is applied. I cant help but smile at the pretension of it.
Currently AI models are inconsistent and unpredictable programmers, but less so when applied to non-novel small and focused programming tasks. Maybe that will change resulting in it being able to do your job. Are you just writing lines of code, organized into functions and modules using a “hack it till it works” methodology? If so, I suggest be open to change.
Professionally programming is just a means to an end that is solving a business problem. Just like hammering nails or turning screws is not a job in itself. The craft is important, but ultimately what people pay for is a final product (a house, a bridge, a software that does X, etc.), and our job is to build something that match their constraints. What you describe is the equivalent of complex novel construction projects, not what the vast majority of the construction industry is about.
Building a house is like a CRUD app, it already was a solved problem technically. AI is like prefabs or power tools. If your job and what you were interested in was building houses AI is great. If you were a brick layer not so much.
Engineer is not an aggrandizing title, it’s the job. Being paid for the hobby of writing code was just an anomaly that AI will close in the majority of the industry IMO.
There are many different ways to write code. The more code there is, the more possible versions of the system could have existed to solve that same set of problems; each with different tradeoffs.
The challenge is writing code in such a way that you end up with a system which solves all the problems it needs to solve in an efficient and intuitive way.
The difference between software engineering and programming is that software engineering is more like a discovery process; you are considering a lot of different requirements and constraints and trying to discover an optimal solution for now and for the foreseeable future... Programming is just churning out code without much regard for how everything fits together. There is little to no planning involved.
I remember at university, one of my math lecturers once said "Software engineering? They're not software engineers, they're programmers."
This is so wrong. IMO, software engineering is the essence of engineering. The complexity is insane and the rules of how to approach problems need to be adapted to different situations. A solution which might be optimal in one situation may be completely inappropriate for a slightly different situation due to a large number of reasons.
When I worked on electronics engineering team projects at university, everyone was saying that writing the microcontroller software was the hardest part. It's the part most teams struggled with, more so than PCB design... Yet software engineers are looked down upon as members of an inferior discipline... Often coerced into accepting the lesser title of 'developer'.
I'm certain there will be AIs which can design optimal PCBs, optimal buildings, optimal mechanical parts, long before we have AI which can design optimal software systems.
Whenever someone gets into Important Reasons why Software Engineering is Different from Programming, I hear a bunch of things that should just be considered “competent programming”.
> Programming is just churning out code without much regard for how everything fits together
What you’re describing is a beginner, not a programmer
> There is little to no planning involved
> trying to discover an optimal solution for now and for the foreseeable future
I spend so much time fighting against plans that attempt to take too much into account and are unrealistic about how little is known before implementation. If the most Important Difference is that software engineers like planning, does that mean that being SE makes you less effective?
I disagree with the first point because I worked for quite a few companies on many different projects and less than 10% were doing it even approximately right... And these were all well regarded companies. I even worked for a company backed by YCombinator for over 1 year as a contractor.
I agree that you shouldn't plan too much, but my experience is that anticipating requirements is possible and highly valuable. It doesn't require planning but it requires intuition and an ability to anticipate problems and adhere to certain principles.
For example, in my current job, I noticed a pattern that was common to a lot of tasks early on. It was non-obvious. I implemented a module which everyone on my team ended up using for essentially every task thereafter. It could have been implemented in 100 different ways, it could have been implemented at a later time, but the way I implemented it meant that it saved everyone a huge amount of time since early on in the project.
Also, we didn't have to do any refactoring and were later able to add extra capabilities and implement complex requirement changes to all parts of the code retroactively thanks to this dependency.
One time we learned that we had to calculate the target date/time differently and our requirements engineer was very worried that this would require a large refactoring to all our processes. It didn't; we changed it in one place and didn't have to update even a single downstream process.
It was a relatively complex module which required some understanding of the business domain but it provided exactly the right amount of flexibility. Now, all my team members know how to update the config on their own. We haven't yet encountered a case it couldn't handle easily.
I have similar stories to tell about many companies I worked for. When AI can replace me, it will be able to replace most entrepreneurs and managers.
I think it’s undeniable as an engineering profession when performance and algorithm choice come into play, that or safe control of embedded or industrial devices. There are other contexts where its engineering too, it’s just engineering in the sense that someone designing commodity headphones is doing electronics engineering.
Software engineering will get automated, all of the issues with the current models will get worked out in time. People can beg and wish that its not true, but it is. We have a few more good years left and then this career is over
I think this is a not insane prediction, but much like truck driving and radiology the timeline is likely not that short.
Waymo has been about to replace the need for human drivers for more than a decade and is just starting to get there in some places, but has had basically no impact on demand yet, and that is a task with much less skill expression.
So far there are mainly horrible "AI-coded" websites that look like they were produced by the Bootstrap framework in 2014. They use 100% CPU in Firefox despite having no useful functionality.
It is the Mc Donalds version of programming, except that Mc Donalds does not steal the food they serve.
When we talk about code, you think it's about code, but it's communication _about solving problems_ which happens to use code as a language.
If you don't understand that language, code becomes a mystery, and you don't understand what the problem is we're trying to solve.
It becomes this entity, "the code". A fantasy.
Truth is: we know. We knew it way before you. Now, can you please stop stating the obvious? There are a lot of problems to solve and not enough time to waste.
To be fair, AI cannot write the code!
It can write some types of code. It is fascinating that it can bootstrap moderately complex projects form a single shot. It does a better job at writing unit test (not perfect) then the fellow human programmer (few people like writing unit tests). It can even find bugs and point + correct broken code. But apart from that, AI cannot, or at least not yet, write the code - the full code.
If it could write the code, I do not see why not deploy it more effectively to write new types of operating systems, experiment with new programming languages and programming paradigms. The $3B is better spent on coming up with truly novel technology that these companies could monopolise with their models. Well, the can't, not yet.
My gut feeling tells me that this might be actually possible at some point but at enormous cost that will make it impractical for most intents and purposes. But even if it possible tomorrow, you would still need people that understand the systems because without them we are simply doomed.
In fact, I would go as much as saying that the demand for programmers will not plummet but skyrocket requiring twice as many programmer we have today. The world simply wont have enough programmers to supply. The reason I think this might actually happen is because the code produced by AI will be so vast overtime that even if humans need to handle/understand 1% that will require more than the 50M developers we have today.
If you’re writing simple code, it’s often a one-shot. With medium-complexity code, it gets the first 90% done in a snap. Easily faster than I could ever do it. The problem is that 90% is never the part that sucks up a bunch of time— it’s the final 10%, and in many cases for me, it’s been more hindrance than help. If I’d just taken the driver’s wheel, making heavy use of autocomplete, I’d have done better and with less frustration. Having to debug code I didn’t write that’s an integral part of what I’m building is an annoying context switch for anything non-trivial.
Same... And the errors are often really nonsensical and nested in ways that a human or thinking brain simply would never do
> Having to debug code I didn’t write that’s an integral part of what I’m building is an annoying context switch for anything non-trivial.
That's the problem I've been facing. The AI does 90% of the work, but the 10% that I have to do myself is 20x harder because I don't have as much knowledge of the codebase as I would if I had written all of it by myself, like in the old days (2 years ago).
Yeah that’s been my experience. The generators are shockingly good. But they don’t get it all the way, and then you are left looking at a mountain of code you don’t understand. And by the time you do, you could have just built it yourself.
Yeah, but you can ask ai questions about it so you can understand it faster
You are trying to check the code for hallucinations. The AI will just hallucinate back a plausible answer for your questions. It’s entirely detached from the process that generated the code so it doesn’t have any more insight than you do on it.
Works great if you’re using a very common language. I wasted more time than I care to admit trying this with a pascal code base.
You are absolutely right.
> you are left looking at a mountain of code you don’t understand. And by the time you do, you could have just built it yourself.
SWEs that do not have (or develop) this skill (to fill-in the 10% that doesn’t work and fully understand the 90% that works and very, very quickly) will be plumbers in a few years if not earlier.
SWEs that don’t understand the strengths and limitations of their tools allowing them to choose the right one for the job won’t be software developers much longer, but they definitely won’t be plumbers. Maybe cycling through the gig platforms or working entry-level retail. Soft, arrogant, maladroit white collar workers make hilariously pathetic trade apprentices.
I’m very confused by this as in my domain space I’ve been able to nearly one-shot most coding assignments since this summer (really Sonnet3.5h) by pointing specific models at well-specified requirements. Things like breaking down a long functional or technical spec document into individual tasks, implementing, testing, deployment and change management. Yes, it’s rather straightforward scripting, like automation on Salesforce. That work is toast and spec-driven development will surge as people go more hands-off the direct manipulation of symbols representing machine instructions, on average.
There is vast difference between writing glue code and engineering systems. Who will come up with the next Spring Boot, Go, Rust, io_uring, or whatever, once the profession has completely reduced itself to pleasing short outcomes?
Same was asked of many transitions to a higher abstraction level. Who will know how to use Libraries if they google everything?
Maybe some day we'll collectively figure it out. I'm confused how people are getting so much success out of it. That hasn't been my experience. I'm not sure what I'm doing wrong.
Try using the brainstorming and execute plan loops with the superpowers plugin in Claude Code. It encapsulates the spec driven development process fairly well.
This misunderstands what LLM-based tools mean for complex software projects. Nobody expects that you should be able to ask them to write you a whole kernel or a web engine.
Coding agents in particular can be very helpful for senior engineers as a way to carry out investigations, double-check assumptions, or automate the creation of some parts of the code.
One key point is to use their initial output as a draft, as a starting point that still needs to be checked and iterated, often through pair programming with the same tool.
The mid-term impact of this transition is hard to anticipate. We will probably get a wide range of cases, from hyper-productive small teams displacing larger but slower ones, to AI-enhanced developers in organisations with uneven adoption quietly enjoying a lot more free time while keeping the same productivity as before.
But how is the senior engineer to get any work done, if they need to babysit the agent and every two minutes accept/reject it's actions? Genuine question. Letting that thing do "whatever" usually means getting insane multiple thousands lines long pull request that will need to be discarded and redone anyway.
Related - how do you get that thing to stop writing comments? If asked not to do so, it will instead put that energy into docstrings, debug logs and what not, poisoning the code for any further "AI" processing.
Stuff like (this is an impression, not an actual output):
Most stuff in comments is actively misleading.Also the horrible desire of writing new code and not reading docs, either in-project or on the web...
For writing ffmpeg invocations or single-screen bash scripts, great thing! For writing programs? Actively harmful
Yeah for me the three main issues are: - overly defensive programming. In python that means try except everywhere without catching specific exceptions, hasattr checks, when replacing an approach by a new one adding a whole “backward compatibility” thing in case we need to keep the old approach etc. That leads to obfuscated errors, silent fails, bad values triggering old code etc - plain editing things it is not supposed to. That is “change A into B” and it does “ok I do B but I also removed C and D because they had nothing to do with A” or “I also changed C in E which doesn’t cover all the edge cases but I liked it better” - keep re-implementing logic instead of reusing
Oh, the defensive programming! That thing must have been trained on job interview code, or some enterprise stuff. Heaps of "improvements" and "corrections" that retry, stub, and simply avoid correctly doing stuff for no reason (fix deserialization bug that thing just caused? no, why! Let's instead assume API and docs are wrong and stuff is failing silently so let's retry all api calls N times, then insert some insane "default value in case API is unreachable" then run it, corrupt local db by writing that default everywhere, run some brain damaged test that checks that all values are present (they are, clonk just nuked them), claim extraordinary success and commit it with a message containing emoji medals and rockets).
And these "oh, I understand, C is completely incorrect" then proceeding to completely sabotage and invalidate everything.
Or assembling some nuclear python script like some McGyver and running it, to nuke even the repo itself if possible.
Best AAA comedy text adventure. Poor people who are forced to "work" like that. But cleanup work will be glorious. If companies will survive that long.
>One key point is to use their initial output as a draft, as a starting point that still needs to be checked and iterated, often through pair programming with the same tool.
This matches my experience. It's not useful for producing something that you wouldn't have been able to produce yourself, because you still need to verify the output itself and not just the behavior of the output when executed.
I'd peg this as the most fundamental difference in use between LLMs and deterministic compilers/transpilers/codegens.
Without arguing with your main point:
> (few people like writing unit tests)
The TDD community loves tests and finds writing code without tests more painful than writing tests before code.
Is your point that the TDD community is a minority?
> It does a better job at writing unit test (not perfect) then the fellow human programmer
I see a lot of very confused tests out of cursor etc that do not understand nor communicate intent. Far below the minimum for a decent human programmer.
I see tests as more of a test of the programmer's understanding of their project than anything. If you deeply understand the project requirements, API surface, failure modes, etc. you will write tests that enforce correct behaviour. If you don't really understand the project, your tests will likely not catch all regressions.
AI can write good test boilerplate, but it cannot understand your project for you. If you just tell it to write tests for some code, it will likely fail you. If you use it to scaffold out mocks or test data or boilerplate code for tests which you already know need to exist, it's fantastic.
It's very worrying that this comment was downvoted
> it can even find bugs
This is one of the harder parts of the job IMHO. What is missing from writing “the code” that is not required for bug fixes?
LLMs can traverse codebases and do research faster. But I can see this one backfiring badly as structural slop becomes more acceptable since you can throw an LLM at it an fix the bug. Eventually you'll reach a stage of stasis where your tech debt is so high, you can't pay the interest even with an LLM.
>It is fascinating that it can bootstrap moderately complex projects form a single shot. It does a better job at writing unit test (not perfect) then the fellow human programmer (few people like writing unit tests). It can even find bugs and point + correct broken code. But apart from that, AI cannot, or at least not yet, write the code - the full code.
Apart from the sanitation, the medicine, education, wine, public order, irrigation, roads, the fresh water system, and public health ... what have the Romans ever done for us?
...brought peace?
https://en.wikipedia.org/wiki/Pax_Romana
> It is fascinating that it can bootstrap moderately complex projects form a single shot.
Similar to "git clone bigproject@github.git"? There is nothing fascinating about creating something that has existed around the training set. It is fascinating that the AI can make some variations from the original content though.
> If it could write the code, I do not see why not deploy it more effectively to write new types of operating systems, experiment with new programming languages and programming paradigms.
This is where all the "vibe-coders" disappears. LLMs can write code fast but so does copy-paste. Most of the "vibe-coded" stuff I see on the Internet is non-functional slop that is super-unoptimized and has open supabase databases.
To be clear, I am not against LLMs or embracing new technologies. I also don't have this idea that we have some kind of "craft" when we have been replacing other people for the last couple decades.
I've been building a game (fully vibe-coded, that is the rule is that I don't write/read any lines of code) and it has reached a stage where any LLM is unable to make any change without fully breaking it (for the curious: https://qpingpong.codeinput.com). The end result is quite impressive but it is far from replacing anyone who has been doing serious programming anytime soon.
It doesn’t need to do all of a job to reduce total jobs in an area. Remove the programming part then you can reduce the number of people for the same output and/or bring people who can’t program but can do the other parts into the fold.
> If OpenAI believed GPT could replace software engineers, why wouldn’t they build their own VS Code fork for a fraction of that cost?
Because believing you can replace some or even most engineers leaves space for hiring the best. It increases the value of the best, and this is all assuming right now - they could believe they have tools coming in two years to replace many more engineers yet still hire them now.
> You sit in a meeting where someone describes a vague problem, and you’re the one who figures out what they actually need. You look at a codebase and decide which parts to change and which to leave alone. You push back on a feature request because you know it’ll create technical debt that’ll haunt the team for years. You review a colleague’s PR and catch a subtle bug that would’ve broken production. You make a call on whether to ship now or wait for more testing.
These are all things that LLMs are doing to various degrees of success though. They’re reviewing code, they can (I know because I had this with for 5.1) push back on certain approaches, they absolutely can decide what parts of code adds to change.
And as for turning vague problems into more clear features? Is that not something they’re unbelievably suited for?
Agree. I agree with many of the article's points, but not its conclusion.
> “You got way more productive, so we’re letting you go” is not a sentence that makes a lot of sense.
Actually, this sentence makes perfect sense if you tweak it slightly:
> You and your teammate got way more productive, so we’re letting (just) you go
This literally happens all the time with automation. Does anyone think the number of people employed in the field of accounting would be the same or higher without the use of calculators or computers?
> And as for turning vague problems into more clear features? Is that not something they’re unbelievably suited for?
I personally find LLMs to be fantastic for taking my thoughts to a more concrete state through robust debate.
I see AI turning many other folks’ thoughts into garbage because it so easily heads in the wrong direction and they don’t understand how to build self checking into their thinking.
It never was about code. What was the last time you said to yourself "I feel like grabbing some code right now!"? Last time you struggled to accomplish something and then screamed "if only I had some code!!!!". This sounds very, very stupid, doesn't it? English is not my native language, but I find this wording very annoying. People/business suffer for having needs unattended or requirements unsatisfied. Latency too high? Have some code! Database locked with long running query? Here, take some code! You want to price exotic financial assets to calculate your risks? Have you tried to generate some code?? This is so strange. I honestly do not think in terms of "code". If your kid asked you about your job, would you say "I use the computer"?
AI accellerates my learning. It helps me understand, do more experiments and lured me into reading more code. I think AI also helps me be more productive, but I‘m less focussed and sooner exhausted. I also fear its addictive potential which is why I force myself to take breaks more often and try to not use it every day.
AI fundamentally changed the programming experience for me in a positive way, but I‘m glad that it’s not my full time job. I think it can also have bad effects which can not be easily avoided in fulltime roles under market conditions.
I used AI to build an app just for myself that parses data (using pandas, python etc, not LLM but an LLM coded it) for a report that i need to produce
it's purely for myself, no one else.
I think this is what AI can do at the moment, in terms of mass market SaaS vibe codes, it will be harder. Happy to be proven wrong.
This is my experience as well. It's been great to make an application that's small in scope, doesn't require access to my main project repo and is basically a nice to have value add for the client.
I was already quite adept in the language and frameworks involved and the risk was very small so it wasn't a big time sink to review the application PR's. Had I not been it would have sucked.
For me the lesson learned wrt agentic coding is to adjust my expectations relative to online rhetoric and it can be sometimes be useful for small isolated one-offs.
Also it's once in a blue moon I can think of a program suitable for agentic coding so I wouldn't ever consider purchasing a personal license.
Aka its the next stage of stackoverflow/google search
What is it that you feel is missing that would take AI from “just for myself” to “mass market SaaS vibes”?
being able to properly deal with scale and security. Being able to be confident that if I am capturing PII data into my application, its as secure as it can be and as secure as if a principal developer had put the architecture together etc.
Mass market SAAS will generally just use other products to handle this stuff. And if there does happen to be a leak, they just say sorry and move on, there are very few consequences for security failures.
You're right
but guess who advises that architecture and implements it... the principal developer/architect.
You can use good security tools, badly.
What use is privacy and security when all our data lives in a DC in us-east-1?
Yes i cross referenced the data to ensure it was producing the correct numbers vs my manual methods
I'm thinking a lot about this currently as a recent convert (as of Opus 4.5). I think this post is on the right track, but like much of this discourse, it isn't really addressing how the technology will grow and the disciplines will adapt.
I'm by no means a doomer, but its obviously a huge change.
Generative coding models will never be 100% perfect. The speed of their convergence to acceptable solutions will decline in complex and novel systems, and at some point there will be diminishing returns to increasing investment in improving their performance.
The cost of software will fall precipitously and it seems unlikely that the increase in the value of programmers / engineers as they currently practice will offset the decline in the price in software. However, following the law of supply and demand, the supply and the amount of software produced will surely grow, and I think someone has to use the models to build software. I expect being trained in software engineering will be very helpful for making effective use of these tools, but such training may not sufficient for a person to succeed in the new labor market.
The scope of problem that a valuable engineer is expected to manage will grow enormously, requiring not only new skills in using generative coding/language models, but also in reasoning about the systems they help create. I anticipate growth in crossover PM / engineering roles. I guess that people who generalize across the stack and current sub-disciplines will thrive and valuable specialties and side-disciplines will include software architecture, electrical engineering, robotics, communication, and business management.
Some people will thrive in this new field, but it may be a difficult transition for many. I suspect that confusion about model capabilities and how to make the most of them and which people are doing valuable things will put a lot of friction and inefficiency into the transition time-frame.
Last thought, given how great models are at coding compared to general of knowledge, administrative, and bureaucratic work, I expect models are widely used to build systems that are supply shocks on such work. I don't think my argument above applies to such workers. I'm worried most about them.
Ok but isn't it a matter of time until AI is good enough at the rest also?
Meanwhile it's depressing that AI is doing the fun part of the job (for me of course, some people quite like the other aspects of the job; and I do like some of them as well, but writing code is still the most fun I have)
I used AI to write some python code and some bazel rules to generate some python code around that to do some new workflow system I wanted to prototype. It just did it, it would make mistakes but since I had it running tests it would fix the code after running the tests.
The big issue is that I didn’t know the APIs very well, and I’m not much of a Python programmer. I could have done this by hand in around 5 days with a ramp up to get started, but it took less than a day just to tell the AI what I wanted and then to iterate with more features.
It's decent at explaining my code back to me, so I can make sure my intent is visible within code/comments/tracing messages. Not too bad at writing test cases either. I still write my code.
Are you saying that you literally write the features by yourself and that you only use LLMs to understand old code and write tests?
Or a more meta point that “LLMs are capable of a lot”?
I'm saying it's not good enough to write code yet, but good at explaining code. If it can't that means I messed up and I need to make it clearer. Once it's explanation and my meaning line up, the code is good enough and I move on. Meta point is that people are using AI for the wrong thing. It's great at consuming, not creating.
> It’s like saying calculators replaced accountants. Calculators automated arithmetic, but arithmetic was never the job. The job was understanding financials, advising clients, making judgment calls, etc. The calculator just made accountants faster at the mechanical part.
Mechanical and later electical calculators replaced human calculators. Accountants switched from having to delegate computation to owning a calculator.
Calculators never threatened to be able to approximate human above-median reason and taste though.
Likely won't for a while. The race to get all of the memory is likely a squeeze attempt against startups and not consumers. Side effect consumers.
We need regulations to prevent such large scale abuse of economic goods especially if the final output is mediocre.
I spent about 9 years in web/frontend development and when all the AI coding things started coming out, I was admittedly concerned and defensive. I thought the output was gross, it couldn't help me with what I was working on, etc. I recently switched over to the backend, so now I'm dealing with an entirely different set of technologies, architecture, and problems to solve. I've found that, so far, about 80% of my time is spent planning, discussing requirements, determining how to integrate new features into the system or how fixing a bug may affect something else, etc. I use AI for the coding part because the code is sort of an afterthought. I've fully leaned into it and the experience has been fantastic. I ask Claude Code questions and bounce ideas off it. I ask it to write some code for me and I review it and make minor adjustments. The TLDR is I'm not really worried about it taking my job anymore, but I recognize that _not_ using it is going to become very obvious in the near future (it already is). This article pretty much nailed it. Engineering isn't just about pumping out lines of code, the code is just the output of your skills which provide the most value. Once I started looking at things that way, my stress level dropped considerably.
Why do you think “not using it is going to become very obvious”?
Reasons why the attempted cursor acquisition might not be about replicating cursor (with or without human help): shutting down competition; market share; understanding user behavior; training data
I disagree with much of this. Programming isn't just a tool we use in pursuit of being an “Engineer” or whatever aggrandizing title is applied. I cant help but smile at the pretension of it.
Currently AI models are inconsistent and unpredictable programmers, but less so when applied to non-novel small and focused programming tasks. Maybe that will change resulting in it being able to do your job. Are you just writing lines of code, organized into functions and modules using a “hack it till it works” methodology? If so, I suggest be open to change.
Professionally programming is just a means to an end that is solving a business problem. Just like hammering nails or turning screws is not a job in itself. The craft is important, but ultimately what people pay for is a final product (a house, a bridge, a software that does X, etc.), and our job is to build something that match their constraints. What you describe is the equivalent of complex novel construction projects, not what the vast majority of the construction industry is about.
Building a house is like a CRUD app, it already was a solved problem technically. AI is like prefabs or power tools. If your job and what you were interested in was building houses AI is great. If you were a brick layer not so much.
Engineer is not an aggrandizing title, it’s the job. Being paid for the hobby of writing code was just an anomaly that AI will close in the majority of the industry IMO.
There are many different ways to write code. The more code there is, the more possible versions of the system could have existed to solve that same set of problems; each with different tradeoffs.
The challenge is writing code in such a way that you end up with a system which solves all the problems it needs to solve in an efficient and intuitive way.
The difference between software engineering and programming is that software engineering is more like a discovery process; you are considering a lot of different requirements and constraints and trying to discover an optimal solution for now and for the foreseeable future... Programming is just churning out code without much regard for how everything fits together. There is little to no planning involved.
I remember at university, one of my math lecturers once said "Software engineering? They're not software engineers, they're programmers."
This is so wrong. IMO, software engineering is the essence of engineering. The complexity is insane and the rules of how to approach problems need to be adapted to different situations. A solution which might be optimal in one situation may be completely inappropriate for a slightly different situation due to a large number of reasons.
When I worked on electronics engineering team projects at university, everyone was saying that writing the microcontroller software was the hardest part. It's the part most teams struggled with, more so than PCB design... Yet software engineers are looked down upon as members of an inferior discipline... Often coerced into accepting the lesser title of 'developer'.
I'm certain there will be AIs which can design optimal PCBs, optimal buildings, optimal mechanical parts, long before we have AI which can design optimal software systems.
Whenever someone gets into Important Reasons why Software Engineering is Different from Programming, I hear a bunch of things that should just be considered “competent programming”.
> Programming is just churning out code without much regard for how everything fits together
What you’re describing is a beginner, not a programmer
> There is little to no planning involved > trying to discover an optimal solution for now and for the foreseeable future
I spend so much time fighting against plans that attempt to take too much into account and are unrealistic about how little is known before implementation. If the most Important Difference is that software engineers like planning, does that mean that being SE makes you less effective?
I disagree with the first point because I worked for quite a few companies on many different projects and less than 10% were doing it even approximately right... And these were all well regarded companies. I even worked for a company backed by YCombinator for over 1 year as a contractor.
I agree that you shouldn't plan too much, but my experience is that anticipating requirements is possible and highly valuable. It doesn't require planning but it requires intuition and an ability to anticipate problems and adhere to certain principles.
For example, in my current job, I noticed a pattern that was common to a lot of tasks early on. It was non-obvious. I implemented a module which everyone on my team ended up using for essentially every task thereafter. It could have been implemented in 100 different ways, it could have been implemented at a later time, but the way I implemented it meant that it saved everyone a huge amount of time since early on in the project.
Also, we didn't have to do any refactoring and were later able to add extra capabilities and implement complex requirement changes to all parts of the code retroactively thanks to this dependency.
One time we learned that we had to calculate the target date/time differently and our requirements engineer was very worried that this would require a large refactoring to all our processes. It didn't; we changed it in one place and didn't have to update even a single downstream process.
It was a relatively complex module which required some understanding of the business domain but it provided exactly the right amount of flexibility. Now, all my team members know how to update the config on their own. We haven't yet encountered a case it couldn't handle easily.
I have similar stories to tell about many companies I worked for. When AI can replace me, it will be able to replace most entrepreneurs and managers.
I think it’s undeniable as an engineering profession when performance and algorithm choice come into play, that or safe control of embedded or industrial devices. There are other contexts where its engineering too, it’s just engineering in the sense that someone designing commodity headphones is doing electronics engineering.
Software engineering will get automated, all of the issues with the current models will get worked out in time. People can beg and wish that its not true, but it is. We have a few more good years left and then this career is over
I think this is a not insane prediction, but much like truck driving and radiology the timeline is likely not that short.
Waymo has been about to replace the need for human drivers for more than a decade and is just starting to get there in some places, but has had basically no impact on demand yet, and that is a task with much less skill expression.
If you are so sure AI is more competent than you at your job, who am I to disagree
So far there are mainly horrible "AI-coded" websites that look like they were produced by the Bootstrap framework in 2014. They use 100% CPU in Firefox despite having no useful functionality.
It is the Mc Donalds version of programming, except that Mc Donalds does not steal the food they serve.
like half of engineers are vibe coding full time in their jobs, wake up
If you think you are right, shouldn't you tell them to wake up? They would be the ones promoting their irrelevance.
They’re promoting their irrelevance because they don’t see the alternative. Get yours while you still can attitude. And they’re probably right.
Why not say that the sky will fall on ou heads while you’re at it? /s
Yet? Always be prudent and include this little word.
The reason why there are meetings is due to existing org layers.
Thus, the root cause of the meetings' existence is BS mostly. That's why you have BS meetings.
The fastest way to drive AI adoption is thus by thinning out org layers.
Not all meetings are status reports. There is also the working meeting where people figure stuff out.
When we talk about code, you think it's about code, but it's communication _about solving problems_ which happens to use code as a language.
If you don't understand that language, code becomes a mystery, and you don't understand what the problem is we're trying to solve.
It becomes this entity, "the code". A fantasy.
Truth is: we know. We knew it way before you. Now, can you please stop stating the obvious? There are a lot of problems to solve and not enough time to waste.
[dead]