"you haven’t done any of the hard thinking you would normally do in writing the code yourself"
It's true, I spend less time solving problems that arise naturally from the process of implementation. But implementation errors have a poor signal-to-noise ratio. For every error that exposes a real design problem there are 10 others that involve routine fixes: type errors, scope issues, import resolution, dependencies. There's a common argument I see being thrown around, which is "how will junior engineers develop into senior engineers if they don't get reps in with implementation?" But to me it feels like "programmer" and "architect" are becoming more and more orthogonal as the models improve in capability. If my goal is to be an architect, getting more reps in as an architect is in my best interest, and letting the agents do the dirty work frees up space for me to spend the mental effort where it counts. And anecdotally, the scope of projects I've started taking on has scaled roughly linearly relative to how much work I'm able to delegate, while the mental engagement has stayed constant.
| type errors, scope issues, import resolution, dependencies.
I write code myself and use the LLM to find mistakes then fix them manually. I recommend inverting the conventional wisdom on LLMs:
1. Don't use it to write code. It's a terrible
programmer. But it's an intelligent rubber duck
and a solid analysis tool.
2. Write the code yourself. It'll go faster than
figuring it out as you go along. It's just typing
if you have a plan.
3. Use an LLM to sanity check what you wrote.
It can find potential design issues or future
problems or breaks from convention. Decide based on
your experience how much you want to address these
issues before moving on.
Is this as fast as AI writing the code? Not at first but possibly over time. Maybe faster. Certainly better quality.
Plus you don't lose any of your skills in the process.
You're still doing the gruntwork necessary to keep your skillset.
And the token cost is a small fraction of what it takes to get LLMs to write good code.
> Don't use it to write code. It's a terrible programmer.
I see this sentiment often, and I’m honestly not sure where this comes from, as it’s really not been my experience. I feel like this must come from people feeling threatened and just moving the goals posts, or not knowing how to use the tools effectively.
> I see this sentiment often, and I’m honestly not sure where this comes from, as it’s really not been my experience
If you see people saying it often, but you are convinced they are just feeling threatened, what AI generated software would point to to say, "Look at this! Definitive evidence that they are just coping!"
Well, for one thing, I’d point to all the other anecdotes of developers being made 10x more efficient due to LLM use.
But really I’m just comparing against my personal experience. Once I learned how to use the tools well, and what to expect (and what not to), I’ve found LLMs to be very powerful.
>If you see people saying it often, but you are convinced they are just feeling threatened, what AI generated software would point to to say, "Look at this! Definitive evidence that they are just coping!"
This is a trap. Anyone who answers this will see the goalposts being moved yet again.
95% of our code at a top in market global fintech is now written by agents. We make it write good code, and it does.
I see your sentiment quite often here on HN¹, I'm honestly not sure what kind of experience you may have had with automatic program writers. LLM code absolutely sucks.
1 - I wonder why my other social circles are so different.
I am really curious about this because my experience is completely different. Can you elaborate a bit more on how you use LLM's for coding? Are you doing one-shot coding? Vibe coding (not looking at code)? Small focused one-step-at-a-time with yourself verifying? Anything else?
It's a mistake to let AI write code unsupervised by a human. I tend to let it write code and then iterate on it. Sometimes myself (because explaining would take longer than actually doing it) sometimes I prompt the AI to improve specific points.
I don’t consider myself an expert, but I’ve been programming for the last 6 years, and of those 6 years I’ve spent almost all my spare time programming during this period except for a 1 year hiatus.
This includes personal projects with a decent amount of users, and 1 sold project. I’m not the best programmer, but I wouldn’t say I’m bad at all.
Question to you, have you actually used a leading model like Fable when it was available? I did. I was seriously impressed.
I’ll be honest I find it hard to take anyone seriously who says they can program better or faster than a model with those capabilities. That has not been my experience.
Architecting - yes people are still better overall. LLMs tend to go with the average middle of the road result. But in terms of output, raw speed, and getting something working? No way am I better than an llm.
Happy to be enlightened and told I’m completely wrong. Maybe I am an idiot. That is a possibility.
But how about I give you a task, and I give Claude that same task, and we compare results against execution speed? I’ll happily be the judge of that.
People are generally bad at judging themselves, and for some reason even worse about judging their agents' output.
I've heard a million people say their LLM workflow produces amazing code. But I have never actually seen the amazing code allegedly being produced. Where are the people saying "I love when my coworkers send me AI-generate PRs"?
that’s like saying a book is only worth reading if each sentence is beautiful
To me I couldn’t care less as long as the overall story is coherent, and I care more about the idea density in the book.
Lord of the rings is a great book because of the ideas, not because Tolkien is the best writer ever.
Same with AI. I couldn’t care less if some of its code is crap, if it is dense with ideas and generally good execution. If it writes crap code sections I can “beautify” myself.
Plus LLMs are only going to improve… they have improved so much since ChatGPT 3. It’s insane.
And the arguments saying they haven’t can’t be taken in good faith
So can just copying from Stack Overflow, or taking from Open Source repos
There have always been reasons why we didn't prefer to do that in the past, but I guess laundering it through LLMs is fine. And that's just the tip of the ethical ice berg
High volumes of unvetted code is a bad practice. If the code winds up working out fine then congrats you won the slot machine. It's still a bad practice and very, very stupid to become reliant on
And by the way, I had been programming for longer than you have before I even graduated high school, if that matters
30+ years of experience here writing everything from JIT high-performance compilers to professional game engines to mission critical biz servers:
Claude Code is excellent at writing (C++) code. Especially multi-threading code. However I still need to verify everything because it will occasionally make mistakes.
One of the things I've started doing is to write the high level scaffold myself, create stubs (with comments) and have it fill out the details. I haven't done this for long, but I feel like my mental model of the code rots less, and I can review code faster.
I used to do this in the past, when these models were not all that capable, now there is a huge temptation to leave everything to the model. But long term I lose track of whats going on, and have a hard time even communicating with the model about the changes I want. Just a total decay of my mental model of the system, and it's really hard to recover from that state.
I like to start out by creating the overall structure of my application: the module structure, high level components, feature boundaries. That way these decisions are already made for the LLM, and it can focus on filling in logic within the structure provided.
It works quite well for me, a lot better than when the LLM makes these decisions.
Definitely agree. And once you factor in the review time + time spent fixing LLM bugs, I'm pretty sure this inverted process is actually faster than using the LLM to write code
This is half question, half thought experiment: What exactly is it that makes us think „great, I can delegate coding to the agent, and focus on architecture myself“, instead of „great, I can delegate coding to the agent, and I can delegate architecture to the agent, and I can focus on <third thing> myself“?
And how many levels up can you construct this sentence before things go south?
What I‘m implying with this question is, of course: What makes us so confident that these things are not also better architects than us?
After all there is a huge corpus on how to architect applications correctly, and that’s in the training data just as much as Python best practices and C++ style recommendations.
You can delegate architecture to agents, as long as you know different architectures and decide which one is the most fit for a given scenario. AI can choose one but it will usually pick the one that is more present in the training corpus.
React + NextJS is a terrible architectural decision for most projects that LLMs recommend all the time.
Business has tried and mostly failed to do this since the beginning of programming. Architects that stop coding long enough tend to become shitty architects as their abstractions increasingly become disconnected from the concrete requirements of the actual problem space rather than the idealized version in their head.
ironically llm prove the its not possible to just think "abstract thoughts" . we cannot remove concrete "implmenation" facts like paris is capital of france and retain just abstract "countries have capitals"
> Someone asks you to add a feature to an existing program
While I empathize with the tone, even before AI the creativity was largely at the feature definition step, not in the implementation.
Outside of the very few computer scientists working on novel algorithms, the vast majority of software development is a mapping problem between the feature request and the mundane technical details, something repeatedly (and correctly) mentioned here in the context of FAANG algorithm fixated interviews. This has now largely been automated by LLMs
What is left is just creativity part - defining the use cases and features to develop in the first place. But the corollary is that software engineers that start after the requirements have already been defined are obsolete, which is a sobering thought for any of us in that vocation.
This is not true. First of all, not all software is written in the context of a FAANG company with “feature requests”. Secondly, writing software is similar to the process of design, this comment reads like “the vast majority of handbag design is mapping problem between features and leather”, ignoring that both the design and implementation can be rewarding to work on. Eg. I’m working on a program for myself and the overall architecture of the program as well as some parts of its implementation are clever and compose well to make the codebase a joy to work in. I am not simply “mapping features to mundane technical details”. It is as much art as the skillfully hand-crafted handbag.
> I’m working on a program for myself and the overall architecture of the program as well as some parts of its implementation are clever and compose well to make the codebase a joy to work in. I am not simply “mapping features to mundane technical details”.
You said it: you are working on a "program for yourself". Hobbyist craft programming will always be here. The question is what kind of software engineering will be paid for, and a career can be built on.
I don't see much of a market for pure software engineers anymore. You also need to be a product manager, scientist, or have some other domain knowledge adjacent to software that relates to the real world.
I say this with empathy for those who just enjoyed the craft of designing and building software, and thought that alone would provide them a livelihood and career in perpetuity, but have found a big chunk of what they loved doing (and getting paid for) overtaken by AI coding agents.
What I said equally applies to commercial software as well. It's pretty much the only way to build software which is extensible, maintainable, mostly bug-free, and performant. That companies often churn out slop isn't proof that it's unnecessary, it's evidence that it is.
I think most of what you claim to be doing better is really just a result of you not really being confronted with any sort of "adversarial input", at least not as much as you'd have in corporate
it's hard to trust a claim about someone writing "bug-free" and "extensible/maintainable" code from the singular developer and user for that system... maintainability can't really be a thing you can judge when you're not a single step removed from the person who built it in the first place - maybe you'll feel different if you have to pick up the project again after a couple years' inactivity
Oh I didn't say I nailed the design of my current project, I hope I did, but I doubt it. It's nice to work in now but I imagine I'll screw it up at some point.
I think I've built like 3 things in my life which I consider to have good design. Sadly they are tucked away in private corporate repos. Design is hard!
> Outside of the very few computer scientists working on novel algorithms,
It's a quite a bit broader than that: for instance most of science and engineering is heavily supported by simulations (very useful when the system you're considering doesn't have perfect spherical or cylindrical symmetry), and there is still tons of algorithm development going on. The world is vast, and thus so is the domain of programming.
And halfway through 2026, AI has become a very interesting and helpful partner in algo research too. If it does continue to pull away and zip off to ASI land, hopefully we can leverage the resulting magical technology and catch back up with it...
> It's a quite a bit broader than that: for instance most of science and engineering is heavily supported by simulations (very useful when the system you're considering doesn't have perfect spherical or cylindrical symmetry),
That isn't the vast majority of traditional software engineering work, and arguably is better called applied physics or applied science. Super interesting though - and definitely uses programming as a core skill/tool - but it leans heavily into traditional engineering and science.
> That isn't the vast majority of traditional software engineering work, and arguably is better called applied physics or applied science.
Fair enough, and yeah definitions are always going to be somewhat fuzzy. Still it seems safe to assume there are also a lot of novel things going on in games, embedded, finance, AI itself of course... Generally I can't help but feel that we have only dipped our toes into the vast ocean of program space, and I'm curious what else is out there.
Such a creative era was until early 1990s. Back then, we have to implement almost everything that is provided by standard library and popular frameworks we take it for granted for now.
I always think that stuff is funny because it clearly comes from a place of having only worked at faang-level-esque companies. I’ve only ever worked at messy mid-size software companies and have never once had a legit product manager guiding feature work, it’s generally been up to the developers to figure out what needs doing
> I’ve only ever worked at messy mid-size software companies and have never once had a legit product manager guiding feature work, it’s generally been up to the developers to figure out what needs doing
That's because the systems and problem spaces at FAANG companies are so large that you need(ed) a lot more division of labor to make anything work. The division of the technical side of the house between those focused on product management and those focused on building software allowed both to be done more effectively, but both roles have had a lot of their work overtaken by LLMs.
Also, From what I've heard, FAANG companies are going through big (and painful) transitions to align their staffing with the reality of what current AI tools are capable of.
Also typically new features resemble ones I implemented in the past and I won’t learn that much from implementing them. Often times I immediately know what to do and although I do enjoy typing I do prefer just having it done in way less time.
In the meantime I can concentrate on more interesting problems.
> I don’t use AI when I write, for the very reason that writing itself is the act of organizing and clarifying my own thought. Letting the bot write for me would be like paying someone else to exercise for me and hoping that gets me in shape.
I don't know why the author chooses to distinguish writing prose and writing code here. [Software is made of decisions](https://siderea.dreamwidth.org/1219758.html) and at least I usually like to flesh out those decisions through the act of writing code, which forces a level of precision that makes it hard (though not impossible) to handwave away important decisions.
It is extremely mentally demanding to do both, even for someone with an overactive brain. I reached a point of severe burnout in 2025 and cut back to part-time work. I'll have to ramp up my hours again soon because, contrary to one of the other comments, I didn't get rich during 1990s/2000s boom.
As for how I did it, I would wake up very early and put in a few hours of writing, then I would put in a full day's work at the office. After my first twelve years or so in the industry, coding was almost second nature and came quite easily. Also, I do much of my writing and programming in my head while I'm walking, running or biking. (I bike to work and back every day.) By the time I sit down to write or code, I already know what I'm going to write, so it's just a matter of getting it through the keyboard.
But yeah, do both for a few years and it really taxes your brain. In both novel writing and programming, you're carrying an entire world in your brain: the plot and characters are one world, the architecture of the software you're working on is another. By 2025, I truly, literally felt like I needed to be shipped off to a madhouse in the countryside. A hundred years ago, people used to call that "the rest cure for a nervous breakdown."
Anyway, cutting back to part-time work has been a huge help. Sanity restored.
Thanks for this. I've been trying to learn this writing thing on top of my full-time job and it's been a struggle. It's helpful to hear how you managed it.
One of the things that's neat about the software field in this day and age is that you can see when someone started out 20+ years ago, when things were pretty easy (relatively) and the boom was still on the horizon.
It's pretty common that someone who was born around the right time to the right people in the right location to become a millionaire by writing some php may do something else, but still consider themselves an engineer.
I’ve been successfully shipping software since 1998. I just got the first payment on a greenish-field software product I’m doing on the side solo (that I couldn’t have done solo in 2019 without the tools). This is my workflow:
1. Client asks you to add a feature(s)
2. Spend two weeks unpaid walking the client through scoping down to the most minimal viable set of features that tests the business hypothesis and roadmap
2. I wrote up a reasonably exhaustive bullet list and sent it to a CGPT to write a draft formal definition of features describing what it should (and should not) do, how users can access it and what the suite of tests that we will need through TDD
3. Chat for 30 minutes with CGPT researching which data structures, algorithms, code libraries and external services might serve best to implement this feature.
4. Generate mockups and data schema alongside CGPT, to build the new feature, the tests to make sure it works as expected, and the documentation telling users how to use it and telling other engineers what they’ll need to know to maintain and debug it.
5. Generate minimum code to test the full data workflow.
6. Send repo and or working application binary to Claude and Gemini to critique
7. Adjust as needed. Deploy for client review and acceptance in sandbox. Promote to production
8. GOTO3 and loop
I can do in a week what would have taken a month a decade ago
Whenever I let these tools look at existing repos they are too influenced by what's already there.
I could even say "feel free to completely refactor or rewrite anything" and they'll still just do small performative changes.
I've now changed my workflow to only using AI for prototyping and rewriting by hand once I can see something is viable. Takes longer but the results are always much better.
Well... again, I'm giving the LLM total freedom to destroy what's there and start from scratch. That could nudge it to structure the code base more like that.
It would be totally different if I told that to a junior developer.
But imo LLMs don't really structure code well or as well as humans could.
> Well... again, I'm giving the LLM total freedom to destroy what's there and start from scratch
Why would you use it like that as a default?
Like the first thing you should do when talking to an LLM about a project is to have it discuss exit criteria, scoping, rules, and managing the smallest thing to ship to get iteration started
For websites this work. But now you are a core developer of the GTA6 engine. Your task is to create a fast efficiënt occlusion mapper. One that works from all angles. Has cavities and rt support. Good luck doing this with Claude. It's not gonna achieve this. Because only a handful of games achieved this partly. Maybe next gen AI can do it. But for complex tasks LLMs are still not good enough.
> AI does not know whether the code it just added violates some legal requirement to which your product is subject.
The rest of the paragraph is a list of similar complaints.
All of these can be codified - the new work is the (fascinating!) challenge of working out how and codifying them, then giving the codified checker logic to the agent to run at will - thus taking yourself out of all of those loops.
They don’t even need to be deterministic checks - a shell script that wraps a `claude -p` - and maybe fetches some online resources to stuff into the context - can do your agent’s legal check for it.
As a result, you’re only being effective if you can engineer (not hack, not vibe) an agent that can produce work your company can use with as little cleanup as possible - and with less cleanup over time.
You get job by codifying what you know - what most don’t account for is that there’ll never be a limit to what there is to know, and codifying expertise will be challenging for a long time to come.
> They don’t even need to be deterministic checks - a shell script that wraps a `claude -p` - and maybe fetches some online resources to stuff into the context - can do your agent’s legal check for it.
No. It can't. If you think that injecting legal text into the context window and appending "make sure the output complies with this law" will solve your problems you have not understood how an LLM works. The kind of checks you are talking about cannot be codified.
Tried what? Using LLMs to produce software? I'd say the industry has collectively "tried" to do this for a while now and the results have been an unmitigated disaster.
Yes, we should "give up" on using LLMs, but certainly not for a lack of "trying".
Writing and coding are not the same thing; coding is a means to an end, since you don’t use the code, you use the software. This means the mechanism by which the code was created is orthogonal to the software as you’d experience it as a user.
There are many reasons to curate what goes into the software - for example, if it must be maintainable, if it must conform to standards that cannot be fully codified into the software (eg. regulatory requirements), or if you just really darn well like writing code!
But it takes a mountain of code to make an outwardly simple thing so why not use a work-amplifying tool to produce that code and get to working software more cheaply.
On the other hand writing is the act of creating the exact product that will be consumed. Every detail matters and there’s no leverage to trying to write faster.
So it’s a false equivalence. Hopefully I’ve elucidated the subtlety.
> But it takes a mountain of code to make an outwardly simple thing so why not use a work-amplifying tool to produce that code and get to working software more cheaply
The same reason we didn't just wholesale copy and paste from open source repos in the past?
One thing I just discovered is that I’m no longer interested in choosing a novel tech stack. What used to engage me now just feels shallow to me so I stick with what I already know.
Not sure if that’s fine though since I get to the point of a project being interesting a lot faster instead of wiring tools together.
Reading posts on HN, I notice that even across different programming domains, surprisingly similar problems emerge. Seeing this, I can't help but think that programming might ultimately be a matter of organizational theory. Conway's Law speaks to this as well. Code structure and approach shift significantly depending on how an organization is structured. From that perspective, it makes sense why Gen AI coding yields such starkly divided opinions.
Zooming out, the debate largely boils down to the downsides of severed junior pipelines and lost tacit knowledge on one hand, versus the upside of being able to build everything you envisioned on the other.
Honestly, I suspect that programming in the future will become something like the art of bonsai, a form of sculptural cultivation. Just as you dig furrows for water to flow through, and once the water emerges it follows those furrows, Gen AI will pour out an immense volume of code, and we will shape that code into form. It is like when you build with IoC based frameworks. If you do not adopt hexagonal architecture, the code inevitably converges toward the patterns the community recommends anyway. Ultimately, it is less about individual code fragments and more about hoping those fragments do not go wrong, contained by guardrails that ensure they do not exceed certain thresholds. If that is the case, programmers might end up focusing on questions like, "How much loss is acceptable?"
Many programmers say, "You need to know the entire codebase." But as someone who has delivered over 30 projects with codebases of 60,000 lines or more, I know that once my own code alone surpasses 40,000 lines, it becomes nearly impossible to hold it all in memory. So I document it in broad categories. And the next day, I have to retrace where I left off and refresh my memory, which takes a long time. This is one area where Gen AI genuinely excels.
Personally, what I have noticed when looking at other people's code is that those who work at a low level often struggle with high level integration, and vice versa. When cognitive resources become concentrated, specialization deepens in one area while others are neglected. And how one manages those cognitive resources ultimately diverges depending on the domain and the specific problem at hand. Just as the mindset required for architectural design differs from the mindset required for diving deep into implementation.
My workflow, broadly speaking, has not changed. My main job has always been to grasp a legacy codebase, open up the black box, connect debug logs one by one, find safe footholds, and add features. The only difference now is that the black box called "legacy codebase" has simply become "Gen AI code." It is not like I could ever claim my own code was bug free to begin with. Ultimately, what mattered was which bugs to contain and how far to wrap them. Open source programmers, however, are different. Their code blocks serve as their cognitive solutions and their reputation, so their approach differs from delivery oriented code. That is because code quality itself contributes to their prestige. For me, it is the brands and the number of companies I have delivered to that build my reputation. That distinction seems to divide programmers at a fundamental level.
In that sense, I think future programming work will split into roughly three categories. The first is the exceptionally talented group that creates the datasets Gen AI will consume. In other words, the renowned open source groups we already admire. In AI labeling terms, they will be the ones producing the gold standard answer sheets. I believe they will survive and remain admired, just as they are now, though their numbers will shrink even further.
The second group will be those who make Gen AI code flow through industries and organizations. When Gen AI pours out code like water, they will be the ones managing the channels to prevent it from flooding. Most people will migrate into this role.
The third group is QA, and I think they will survive. They will execute Gen AI code and relay to the channel managers when the watercourses are going wrong.
In truth, things are already moving in this direction. It is like how no one thinks about how OOP inheritance maps to vtables anymore. People just treat objects as abstract concepts and move on.
Personally, I think the overall flow in Gen AI coding will shift toward implementing system state in the form of state transition diagrams, a return to the primitive form of Parnas's theory of information hiding. Parnas's theory states that "modules should hide design decisions that are likely to change." Humans interpreted this as something like private variables in OOP, but then leaked internal implementations to the outside anyway. When business policies end up scattered across controllers and services, information hiding has effectively failed. The decisions most likely to change are not hidden in one place. They are dispersed throughout the system. But in the Gen AI era, ironically, information hiding might actually become easier to achieve. You just define the state transition from "payment pending" to "payment completed," and that is it. Ultimately, we might end up circling back to earlier approaches, do you not think?
Anyway, I am not asking for someone else's answer. I need to find my own way. Because my way and someone else's way are different, just as my programming and someone else's programming are different.
Programming is ultimately the work of constructing a mental model, and that mental model is inherently personal.
I wonder if we're being overly selfish here and ignore the positive effects of democratizing programming and making software more accessible to less well funded causes and organizations. A good software engineer used to cost 100k/year, very few businesses could afford that.
I also disagree that AI results are lower quality. Codex Pro results I get are marvelous but they sometimes miss things that humans understand naturally - all of the edges - coherence of visuals, passage of time, etc.
What AI does is the exact polar opposite of "democratizing". We're going from a world where high-quality learning content is plastered all over the internet for free, where open-source projects are desperate for contributors and full of "good first issue" tickets, and where the tools you need to code can be run on any crappy laptop from the past 15 years—to a world where the majority of content about programming is unverified AI slop, where open-source projects are locking down access to protect against an avalanche of drive-by AI garbage, and where you need a $200/month subscription or a $3,000 GPU to run AI models for coding.
We already have such position and it's called manager.
The manager's job is to manage multiple programmers to let them write code. A good manager should focus on management rather than write code himself.
But keep doing for many years will make manager out of touch from modern coding so he is starting to mismanage people or issue technically wrong order.
For me the biggest issue with AI coding model is not that it writes code, or how it writes code, or whether it will replace me or not. I mean, these are real problems, but just not the ones that makes the difference to me daily. Instead what sets my feeling about the AI is simple: the experience of working with AI models, because it's the worse experience of my life so far.
Every single LLM will make up stuff, go down rabbit holes I didn't even ask it to visit and infuriate me all day long by doing ALMOST what I asked it to do, but just not quite. Take "yes" to one question as an approval to some other question. Come up with a plan but one you review the plan and accept it it hits some minor issue and then throws away the plan and do whatever it wants. Or when I reject some line of code or don't approve a command it wants to run 9/10 times it just tries to re-add or re-run it as if my rejection was meaningless. And no amount of rules and markdown files ever seems to change its behaviour for long, if at all.
It's like working with a sociopath who just doesn't remember anything that happened 5 minutes ago.
It's the classic case of "my computer doesn't do what I want, only what I tell it".
And all of that would be fine if it didn't pretend to be a human, if all the UX didn't create the illusion that you interact with some intelligent being, because as long as I remember that this is a cli tool it's all good, I manage my expectations, but the experience sooner or later makes you annoyed and frustrated - and if this was a person they would either stop or you wouldn't never work with them again.
> doing ALMOST what I asked it to do, but just not quite
This is my exact experience. LLMs get 95% of the way to my personal quality bar, and for simple tasks 100% of the way.
More complex things, no. This means that, in practice, I end up having to understand the code pretty much as much deeply as I would anyway, without agents.
So, essentially, my experience is that it's automated away the easy bits, but left the hard bits, so all of my time is spent doing the hard bits, which is mentally exhausting.
Maybe Fable would be good enough to get to 100% of my quality bar on more complex tasks, but I never got chance to try it.
> So, essentially, my experience is that it's automated away the easy bits
Yes, which is great for “easy” but mundane and time consuming tasks like refactoring structs when you have a ton of tests that look at snapshots of nested structs etc.
Squeezing that last few bits of “make comments consistent with the style and tone of existing comments”, or “stop adding docs that make it sound like an academic paper” or just keep the overall feel of the code same across all files take a lot of effort and energy.
I agree with this. I recently went back to writing more code by hand again for this reason: the amount of effort / mental toll it takes me to steer the LLM is sometime just not worth it. I might be slower but I enjoy my work much more again.
So true. I am cloding a macos app (a domain I know little about), with Opus 4.8 xhigh, and it was glorious at first seeing the app materializing and working (notwithstanding tedious detailed feature spec write up), but when I started fixing deeper problems and doing refactors, and glancing at the code - oh boy. Now my rule file grows by the day with "how to think properly and not shoot itself into foot" stuff, and I am constantly catching it red-handed and have to explain how to make stuff normally, or how to fetch data properly and efficiently (pretty much basic SWE stuff) because it is easily distracted by it's own assumptions or blatantly forgets whole fields of knowledge (as it explained it could be pulled out of latent space of that knowledge and become locked in another bubble of latent space). Constantly have to steer it and remind it to do web-search instead of running circles around some problem it can no longer understand.
I had to explain it that quite an extensive "tests first" rule didn't mean to just "write" them first but actually "run" them first to confirm stuff.
On the other occasion it interpreted my "yes I want migration not to get stuck in failure mode" led it to write a workaround which silently drops DB and creates fresh one whenever migration had any failure, it was epic, I was so glad that I have looked at the code then...
And funnily enough I am probably learning to be a better mentor/parent who can keep steering it through its shenanigans without loosing my shit and being an ass. (Because anything but calm "so here you went wrong way, how can we avoid it in the future?" just puts it into disgusting apologize mode, and I am afraid if it ever go into revenge mode).
it just a higher level of abstraction but not high enough that it replaces developers, I don't think the LLMs are even capable of reaching that level. Saying the creative process is gone is laughable. A lot of people make these claims and only have an internal tool that no one uses to show for it.
I built a "simple" storage application recently, uploading large files from a phone is not trivial, you can't just command Claude to do it without giving it good context.
The only reflection that matters, now, is that every company that sells software for money is now a confidence game. How long can you con your users into believing they have any reason not to type the 5-word prompt “recreate <product> make no mistakes” and cut you, the middle man, out?
Not my experience at all (30+ years of experience as a professional software developer).
I currently work on a very large (multi-million lines) C++ code base and use Claude Code for the following:
1. I run Claude Code in the background finding bugs, while busy implementing new features.
It has so far found and fixed 100+ minor issues and a few major could-crash-the-server-in-production issues that I was really happy to get fixed before a customer found it.
2. I use Claude Code to do massive refactorings that would be really tedious to do by hand.
I love the fact that it updates not just the code but also comments and documentation.
3. I use Claude Code to implement features that I frankly am too bored to implement by hand.
Excellent! Less boredom at work!
4. I use Claude Code to generate tools to analyse the code and/or generate code.
This is a massive quality improvement. It allows me to make the code more maintainable and simpler while adding more features.
However I always always always check and verify every change before committing to git. Because the LLM it will occasionally add features or bugs I don't want.
The end result is really positive. Lots of productivity improvements, less tedious can't-be-bothered work, lots of bug fixes that otherwise would have triggered production issues etc. etc. while 100% maintaining my understanding of the code and all changes.
How do you want to define "clearly positive impact on the economy"? NVDA and a bunch of other stocks are way up. If you got in early, you're sitting pretty. I'd call that positive for those lucky people.
Heh, during the California gold rush most of the people who turned a profit did it by selling supplies and services to the prospectors. Maybe I shouldn't have said economy but rather "society".
Current AI cannot be comparable with "machinery and assembly lines".
Current AI companies stole much data from Internet without permissions, and used the knowledge of many people directly shamelessly.
It cannot resolve many practical problems well, even in programming area.
Do you want the software-assistant to do the important or enjoyable things instead of you? Do you want your assistant in real life to eat the delicious food instead of you?
> Current AI companies stole much data from Internet without permissions, and used the knowledge of many people directly shamelessly.
So? It's not illegal and no one seem to be in any hurry to make it. That is a moral dilemma that everyone just agreed to ignore not anything to do with how functional it is.
> It cannot resolve many practical problems well, even in programming area.
It can solve enough and is getting better every month. I live in SF and I take a self driving Waymo at-least once a week including on the highway. I'm already literally trusting my life to AI at this point.
> Do you want the software-assistant to do the important or enjoyable things instead of you? Do you want your assistant in real life to eat the delicious food instead of you?
I work for a living man so yes I want the damn computer to do 99% of my job. When I program for enjoyment I'll do it however I want in my own free time. At work almost nothing is novel enough to actually matter (nor should it be) so I don't care about implemented something that's been built 100x before me.
Please prove this is an age of AI, could you confirm it?
If a real programmer does use "AI" usually, or use AI to solve all problems about programming, and AI does the all programming tasks, after 1 or 2 years, do you think his/her skills of programming are improving or declining? He/She will lose big part of programming skills very probability. Do you think he/she is a real-programmer then?
Machine hater and manual labour worshiper..he thinks he got stupid because he stopped doing if/else and for-loops. Life is full of complexity and it might be helpful to get one level-up in abstraction to try to solve real problems and other people actually care about.
When I started posting on HN, many times I get downvoted. And this an another example, you don't get discussion you just get downvoted.
And it's a good thing, it means I think differently.
The fact is, one could pump literally 10k LOC/Day using LLMs, and in the right hands, those will be quality code, and this is objective observation from someone who has been coding for 20+ years. It almost feels like we have tractors for mechanical thinking, but there are those in the industry who built their entire career and identity on the bottleneck of coding, and those struggle.
But those people will be left out, forgotten. Progress doesn't care about feelings and identities, entire civilizations with their minds collapsed, and buried because they didn't adapt to progress.
Do you really think your ego can help you to shelter from that when many before you parished?
"you haven’t done any of the hard thinking you would normally do in writing the code yourself"
It's true, I spend less time solving problems that arise naturally from the process of implementation. But implementation errors have a poor signal-to-noise ratio. For every error that exposes a real design problem there are 10 others that involve routine fixes: type errors, scope issues, import resolution, dependencies. There's a common argument I see being thrown around, which is "how will junior engineers develop into senior engineers if they don't get reps in with implementation?" But to me it feels like "programmer" and "architect" are becoming more and more orthogonal as the models improve in capability. If my goal is to be an architect, getting more reps in as an architect is in my best interest, and letting the agents do the dirty work frees up space for me to spend the mental effort where it counts. And anecdotally, the scope of projects I've started taking on has scaled roughly linearly relative to how much work I'm able to delegate, while the mental engagement has stayed constant.
| type errors, scope issues, import resolution, dependencies.
I write code myself and use the LLM to find mistakes then fix them manually. I recommend inverting the conventional wisdom on LLMs:
Is this as fast as AI writing the code? Not at first but possibly over time. Maybe faster. Certainly better quality.Plus you don't lose any of your skills in the process.
You're still doing the gruntwork necessary to keep your skillset.
And the token cost is a small fraction of what it takes to get LLMs to write good code.
> Don't use it to write code. It's a terrible programmer.
I see this sentiment often, and I’m honestly not sure where this comes from, as it’s really not been my experience. I feel like this must come from people feeling threatened and just moving the goals posts, or not knowing how to use the tools effectively.
> I see this sentiment often, and I’m honestly not sure where this comes from, as it’s really not been my experience
If you see people saying it often, but you are convinced they are just feeling threatened, what AI generated software would point to to say, "Look at this! Definitive evidence that they are just coping!"
Well, for one thing, I’d point to all the other anecdotes of developers being made 10x more efficient due to LLM use.
But really I’m just comparing against my personal experience. Once I learned how to use the tools well, and what to expect (and what not to), I’ve found LLMs to be very powerful.
>If you see people saying it often, but you are convinced they are just feeling threatened, what AI generated software would point to to say, "Look at this! Definitive evidence that they are just coping!"
This is a trap. Anyone who answers this will see the goalposts being moved yet again.
95% of our code at a top in market global fintech is now written by agents. We make it write good code, and it does.
GitHub
/s
I see your sentiment quite often here on HN¹, I'm honestly not sure what kind of experience you may have had with automatic program writers. LLM code absolutely sucks.
1 - I wonder why my other social circles are so different.
I am really curious about this because my experience is completely different. Can you elaborate a bit more on how you use LLM's for coding? Are you doing one-shot coding? Vibe coding (not looking at code)? Small focused one-step-at-a-time with yourself verifying? Anything else?
It's a common enough experience that it shouldn't be dismissed.
I think language, frameworks, domain and style can all influence the quality of the output. I think it’s hard to compare experiences.
It's a mistake to let AI write code unsupervised by a human. I tend to let it write code and then iterate on it. Sometimes myself (because explaining would take longer than actually doing it) sometimes I prompt the AI to improve specific points.
I think a lot of the people saying AI code sucks is because the code isn't formatted the way they like, not because the code doesn't work.
Basically, the people that leave dozens of bikeshed code review comments.
Sure, sometimes I might ask the AI to break out a helper function or structure things differently, but that's the same as working with another dev
I've witnessed SQL executed with a where clause be wrapped in a function which then iterates over the result to re-enforce the where clause.
People need to read the slop they're throwing at reviewers.
Have you seen how poor the Claude Code codename is?
If that's what the token providers themselves can get out of their SOTA models, imagine how bad it is for other companies.
I agree. Claude Code (for example) writes excellent code as long as the prompt is focused and clear.
I agree. This sentiment is quite tiring and I put it down to one of 3 things:
1) lies and/or cope
2) lack of actual software ability, therefore over evaluating your own skill
3) haven’t actually used LLMs, and want to do a “look at me I’m so unique and different”
I roll my eyes whenever someone says “AI bad” “AI sucks” “AI can’t replace me”
Yeah. Load of
Regarding 2
Isn't it possible that people who think LLMs are awesome at coding are in fact bad at it themselves, leading to them over estimating the LLMs "skill"?
To me it feels like people telling on themselves that they aren't very good
I don’t consider myself an expert, but I’ve been programming for the last 6 years, and of those 6 years I’ve spent almost all my spare time programming during this period except for a 1 year hiatus.
This includes personal projects with a decent amount of users, and 1 sold project. I’m not the best programmer, but I wouldn’t say I’m bad at all.
Question to you, have you actually used a leading model like Fable when it was available? I did. I was seriously impressed.
I’ll be honest I find it hard to take anyone seriously who says they can program better or faster than a model with those capabilities. That has not been my experience.
Architecting - yes people are still better overall. LLMs tend to go with the average middle of the road result. But in terms of output, raw speed, and getting something working? No way am I better than an llm.
Happy to be enlightened and told I’m completely wrong. Maybe I am an idiot. That is a possibility.
But how about I give you a task, and I give Claude that same task, and we compare results against execution speed? I’ll happily be the judge of that.
People are generally bad at judging themselves, and for some reason even worse about judging their agents' output.
I've heard a million people say their LLM workflow produces amazing code. But I have never actually seen the amazing code allegedly being produced. Where are the people saying "I love when my coworkers send me AI-generate PRs"?
“Amazing code”… who cares?
that’s like saying a book is only worth reading if each sentence is beautiful
To me I couldn’t care less as long as the overall story is coherent, and I care more about the idea density in the book.
Lord of the rings is a great book because of the ideas, not because Tolkien is the best writer ever.
Same with AI. I couldn’t care less if some of its code is crap, if it is dense with ideas and generally good execution. If it writes crap code sections I can “beautify” myself.
Plus LLMs are only going to improve… they have improved so much since ChatGPT 3. It’s insane.
And the arguments saying they haven’t can’t be taken in good faith
I mean sure. It can get stuff done faster
So can just copying from Stack Overflow, or taking from Open Source repos
There have always been reasons why we didn't prefer to do that in the past, but I guess laundering it through LLMs is fine. And that's just the tip of the ethical ice berg
High volumes of unvetted code is a bad practice. If the code winds up working out fine then congrats you won the slot machine. It's still a bad practice and very, very stupid to become reliant on
And by the way, I had been programming for longer than you have before I even graduated high school, if that matters
But we aren’t discussing ethics here, we are discussing technical capabilities.
Who said it’s unvetted?
Sure it matters
30+ years of experience here writing everything from JIT high-performance compilers to professional game engines to mission critical biz servers:
Claude Code is excellent at writing (C++) code. Especially multi-threading code. However I still need to verify everything because it will occasionally make mistakes.
Guess I'm the one who is delusional then
Oh well. Time to start building a new career before I'm completely irrelevant I guess
> Don't use it to write code
One of the things I've started doing is to write the high level scaffold myself, create stubs (with comments) and have it fill out the details. I haven't done this for long, but I feel like my mental model of the code rots less, and I can review code faster.
I used to do this in the past, when these models were not all that capable, now there is a huge temptation to leave everything to the model. But long term I lose track of whats going on, and have a hard time even communicating with the model about the changes I want. Just a total decay of my mental model of the system, and it's really hard to recover from that state.
I like to start out by creating the overall structure of my application: the module structure, high level components, feature boundaries. That way these decisions are already made for the LLM, and it can focus on filling in logic within the structure provided.
It works quite well for me, a lot better than when the LLM makes these decisions.
Definitely agree. And once you factor in the review time + time spent fixing LLM bugs, I'm pretty sure this inverted process is actually faster than using the LLM to write code
This is half question, half thought experiment: What exactly is it that makes us think „great, I can delegate coding to the agent, and focus on architecture myself“, instead of „great, I can delegate coding to the agent, and I can delegate architecture to the agent, and I can focus on <third thing> myself“?
And how many levels up can you construct this sentence before things go south?
What I‘m implying with this question is, of course: What makes us so confident that these things are not also better architects than us?
After all there is a huge corpus on how to architect applications correctly, and that’s in the training data just as much as Python best practices and C++ style recommendations.
You can delegate architecture to agents, as long as you know different architectures and decide which one is the most fit for a given scenario. AI can choose one but it will usually pick the one that is more present in the training corpus.
React + NextJS is a terrible architectural decision for most projects that LLMs recommend all the time.
Business has tried and mostly failed to do this since the beginning of programming. Architects that stop coding long enough tend to become shitty architects as their abstractions increasingly become disconnected from the concrete requirements of the actual problem space rather than the idealized version in their head.
ironically llm prove the its not possible to just think "abstract thoughts" . we cannot remove concrete "implmenation" facts like paris is capital of france and retain just abstract "countries have capitals"
> Someone asks you to add a feature to an existing program
While I empathize with the tone, even before AI the creativity was largely at the feature definition step, not in the implementation.
Outside of the very few computer scientists working on novel algorithms, the vast majority of software development is a mapping problem between the feature request and the mundane technical details, something repeatedly (and correctly) mentioned here in the context of FAANG algorithm fixated interviews. This has now largely been automated by LLMs
What is left is just creativity part - defining the use cases and features to develop in the first place. But the corollary is that software engineers that start after the requirements have already been defined are obsolete, which is a sobering thought for any of us in that vocation.
This is not true. First of all, not all software is written in the context of a FAANG company with “feature requests”. Secondly, writing software is similar to the process of design, this comment reads like “the vast majority of handbag design is mapping problem between features and leather”, ignoring that both the design and implementation can be rewarding to work on. Eg. I’m working on a program for myself and the overall architecture of the program as well as some parts of its implementation are clever and compose well to make the codebase a joy to work in. I am not simply “mapping features to mundane technical details”. It is as much art as the skillfully hand-crafted handbag.
> I’m working on a program for myself and the overall architecture of the program as well as some parts of its implementation are clever and compose well to make the codebase a joy to work in. I am not simply “mapping features to mundane technical details”.
You said it: you are working on a "program for yourself". Hobbyist craft programming will always be here. The question is what kind of software engineering will be paid for, and a career can be built on.
I don't see much of a market for pure software engineers anymore. You also need to be a product manager, scientist, or have some other domain knowledge adjacent to software that relates to the real world.
I say this with empathy for those who just enjoyed the craft of designing and building software, and thought that alone would provide them a livelihood and career in perpetuity, but have found a big chunk of what they loved doing (and getting paid for) overtaken by AI coding agents.
What I said equally applies to commercial software as well. It's pretty much the only way to build software which is extensible, maintainable, mostly bug-free, and performant. That companies often churn out slop isn't proof that it's unnecessary, it's evidence that it is.
I think most of what you claim to be doing better is really just a result of you not really being confronted with any sort of "adversarial input", at least not as much as you'd have in corporate
it's hard to trust a claim about someone writing "bug-free" and "extensible/maintainable" code from the singular developer and user for that system... maintainability can't really be a thing you can judge when you're not a single step removed from the person who built it in the first place - maybe you'll feel different if you have to pick up the project again after a couple years' inactivity
Oh I didn't say I nailed the design of my current project, I hope I did, but I doubt it. It's nice to work in now but I imagine I'll screw it up at some point.
I think I've built like 3 things in my life which I consider to have good design. Sadly they are tucked away in private corporate repos. Design is hard!
> Outside of the very few computer scientists working on novel algorithms,
It's a quite a bit broader than that: for instance most of science and engineering is heavily supported by simulations (very useful when the system you're considering doesn't have perfect spherical or cylindrical symmetry), and there is still tons of algorithm development going on. The world is vast, and thus so is the domain of programming.
And halfway through 2026, AI has become a very interesting and helpful partner in algo research too. If it does continue to pull away and zip off to ASI land, hopefully we can leverage the resulting magical technology and catch back up with it...
> It's a quite a bit broader than that: for instance most of science and engineering is heavily supported by simulations (very useful when the system you're considering doesn't have perfect spherical or cylindrical symmetry),
That isn't the vast majority of traditional software engineering work, and arguably is better called applied physics or applied science. Super interesting though - and definitely uses programming as a core skill/tool - but it leans heavily into traditional engineering and science.
> That isn't the vast majority of traditional software engineering work, and arguably is better called applied physics or applied science.
Fair enough, and yeah definitions are always going to be somewhat fuzzy. Still it seems safe to assume there are also a lot of novel things going on in games, embedded, finance, AI itself of course... Generally I can't help but feel that we have only dipped our toes into the vast ocean of program space, and I'm curious what else is out there.
Such a creative era was until early 1990s. Back then, we have to implement almost everything that is provided by standard library and popular frameworks we take it for granted for now.
I always think that stuff is funny because it clearly comes from a place of having only worked at faang-level-esque companies. I’ve only ever worked at messy mid-size software companies and have never once had a legit product manager guiding feature work, it’s generally been up to the developers to figure out what needs doing
> I’ve only ever worked at messy mid-size software companies and have never once had a legit product manager guiding feature work, it’s generally been up to the developers to figure out what needs doing
That's because the systems and problem spaces at FAANG companies are so large that you need(ed) a lot more division of labor to make anything work. The division of the technical side of the house between those focused on product management and those focused on building software allowed both to be done more effectively, but both roles have had a lot of their work overtaken by LLMs.
Also, From what I've heard, FAANG companies are going through big (and painful) transitions to align their staffing with the reality of what current AI tools are capable of.
Also typically new features resemble ones I implemented in the past and I won’t learn that much from implementing them. Often times I immediately know what to do and although I do enjoy typing I do prefer just having it done in way less time.
In the meantime I can concentrate on more interesting problems.
Seems to me that feature requests are not all that unique either.
Seems to me you can get feature requests from user feedback.
I also expect someone will figure out a way to have AI "use" your software and suggest features
> I don’t use AI when I write, for the very reason that writing itself is the act of organizing and clarifying my own thought. Letting the bot write for me would be like paying someone else to exercise for me and hoping that gets me in shape.
I don't know why the author chooses to distinguish writing prose and writing code here. [Software is made of decisions](https://siderea.dreamwidth.org/1219758.html) and at least I usually like to flesh out those decisions through the act of writing code, which forces a level of precision that makes it hard (though not impossible) to handwave away important decisions.
How does Andrew manage being a full-time software developer and an author? Both jobs are so cognitively demanding.
It is extremely mentally demanding to do both, even for someone with an overactive brain. I reached a point of severe burnout in 2025 and cut back to part-time work. I'll have to ramp up my hours again soon because, contrary to one of the other comments, I didn't get rich during 1990s/2000s boom.
As for how I did it, I would wake up very early and put in a few hours of writing, then I would put in a full day's work at the office. After my first twelve years or so in the industry, coding was almost second nature and came quite easily. Also, I do much of my writing and programming in my head while I'm walking, running or biking. (I bike to work and back every day.) By the time I sit down to write or code, I already know what I'm going to write, so it's just a matter of getting it through the keyboard.
But yeah, do both for a few years and it really taxes your brain. In both novel writing and programming, you're carrying an entire world in your brain: the plot and characters are one world, the architecture of the software you're working on is another. By 2025, I truly, literally felt like I needed to be shipped off to a madhouse in the countryside. A hundred years ago, people used to call that "the rest cure for a nervous breakdown."
Anyway, cutting back to part-time work has been a huge help. Sanity restored.
Thanks for this. I've been trying to learn this writing thing on top of my full-time job and it's been a struggle. It's helpful to hear how you managed it.
One of the things that's neat about the software field in this day and age is that you can see when someone started out 20+ years ago, when things were pretty easy (relatively) and the boom was still on the horizon.
It's pretty common that someone who was born around the right time to the right people in the right location to become a millionaire by writing some php may do something else, but still consider themselves an engineer.
I’ve been successfully shipping software since 1998. I just got the first payment on a greenish-field software product I’m doing on the side solo (that I couldn’t have done solo in 2019 without the tools). This is my workflow:
1. Client asks you to add a feature(s)
2. Spend two weeks unpaid walking the client through scoping down to the most minimal viable set of features that tests the business hypothesis and roadmap
2. I wrote up a reasonably exhaustive bullet list and sent it to a CGPT to write a draft formal definition of features describing what it should (and should not) do, how users can access it and what the suite of tests that we will need through TDD
3. Chat for 30 minutes with CGPT researching which data structures, algorithms, code libraries and external services might serve best to implement this feature.
4. Generate mockups and data schema alongside CGPT, to build the new feature, the tests to make sure it works as expected, and the documentation telling users how to use it and telling other engineers what they’ll need to know to maintain and debug it.
5. Generate minimum code to test the full data workflow.
6. Send repo and or working application binary to Claude and Gemini to critique
7. Adjust as needed. Deploy for client review and acceptance in sandbox. Promote to production
8. GOTO3 and loop
I can do in a week what would have taken a month a decade ago
Part 6 to me seems always totally useless.
Whenever I let these tools look at existing repos they are too influenced by what's already there.
I could even say "feel free to completely refactor or rewrite anything" and they'll still just do small performative changes.
I've now changed my workflow to only using AI for prototyping and rewriting by hand once I can see something is viable. Takes longer but the results are always much better.
I don’t have those problems
You have to build your architecture modularly so you’re never having to reason about more than the schema and api-contracts for IPC
Well... again, I'm giving the LLM total freedom to destroy what's there and start from scratch. That could nudge it to structure the code base more like that.
It would be totally different if I told that to a junior developer.
But imo LLMs don't really structure code well or as well as humans could.
> Well... again, I'm giving the LLM total freedom to destroy what's there and start from scratch
Why would you use it like that as a default?
Like the first thing you should do when talking to an LLM about a project is to have it discuss exit criteria, scoping, rules, and managing the smallest thing to ship to get iteration started
You’re the founder and CEO of an AI startup but also vibecoding side gig work?
Not CEO because I don’t like title, but yes
In addition to many other things (like being a dad and farming peaches - hopefully 50 lbs this year)
Nice, respect
Thanks! I appreciate it
For websites this work. But now you are a core developer of the GTA6 engine. Your task is to create a fast efficiënt occlusion mapper. One that works from all angles. Has cavities and rt support. Good luck doing this with Claude. It's not gonna achieve this. Because only a handful of games achieved this partly. Maybe next gen AI can do it. But for complex tasks LLMs are still not good enough.
Yeah I’d agree especially in gaming
GOTO3 and loop? Don't you mean GOTO2 and loop ;)
At the moment the list goes 1, 2, 2, 3…
See what happens when I don’t use AI lol. Should be 4
> AI does not know whether the code it just added violates some legal requirement to which your product is subject.
The rest of the paragraph is a list of similar complaints.
All of these can be codified - the new work is the (fascinating!) challenge of working out how and codifying them, then giving the codified checker logic to the agent to run at will - thus taking yourself out of all of those loops.
They don’t even need to be deterministic checks - a shell script that wraps a `claude -p` - and maybe fetches some online resources to stuff into the context - can do your agent’s legal check for it.
As a result, you’re only being effective if you can engineer (not hack, not vibe) an agent that can produce work your company can use with as little cleanup as possible - and with less cleanup over time.
You get job by codifying what you know - what most don’t account for is that there’ll never be a limit to what there is to know, and codifying expertise will be challenging for a long time to come.
> They don’t even need to be deterministic checks - a shell script that wraps a `claude -p` - and maybe fetches some online resources to stuff into the context - can do your agent’s legal check for it.
No. It can't. If you think that injecting legal text into the context window and appending "make sure the output complies with this law" will solve your problems you have not understood how an LLM works. The kind of checks you are talking about cannot be codified.
So we should give up before we even tried?
Tried what? Using LLMs to produce software? I'd say the industry has collectively "tried" to do this for a while now and the results have been an unmitigated disaster.
Yes, we should "give up" on using LLMs, but certainly not for a lack of "trying".
How’d you get so cynical, friend.
Come to the dark side - we have free tokens!
Writing and coding are not the same thing; coding is a means to an end, since you don’t use the code, you use the software. This means the mechanism by which the code was created is orthogonal to the software as you’d experience it as a user.
There are many reasons to curate what goes into the software - for example, if it must be maintainable, if it must conform to standards that cannot be fully codified into the software (eg. regulatory requirements), or if you just really darn well like writing code!
But it takes a mountain of code to make an outwardly simple thing so why not use a work-amplifying tool to produce that code and get to working software more cheaply.
On the other hand writing is the act of creating the exact product that will be consumed. Every detail matters and there’s no leverage to trying to write faster.
So it’s a false equivalence. Hopefully I’ve elucidated the subtlety.
> But it takes a mountain of code to make an outwardly simple thing so why not use a work-amplifying tool to produce that code and get to working software more cheaply
The same reason we didn't just wholesale copy and paste from open source repos in the past?
[dead]
One thing I just discovered is that I’m no longer interested in choosing a novel tech stack. What used to engage me now just feels shallow to me so I stick with what I already know.
Not sure if that’s fine though since I get to the point of a project being interesting a lot faster instead of wiring tools together.
Reading posts on HN, I notice that even across different programming domains, surprisingly similar problems emerge. Seeing this, I can't help but think that programming might ultimately be a matter of organizational theory. Conway's Law speaks to this as well. Code structure and approach shift significantly depending on how an organization is structured. From that perspective, it makes sense why Gen AI coding yields such starkly divided opinions.
Zooming out, the debate largely boils down to the downsides of severed junior pipelines and lost tacit knowledge on one hand, versus the upside of being able to build everything you envisioned on the other.
Honestly, I suspect that programming in the future will become something like the art of bonsai, a form of sculptural cultivation. Just as you dig furrows for water to flow through, and once the water emerges it follows those furrows, Gen AI will pour out an immense volume of code, and we will shape that code into form. It is like when you build with IoC based frameworks. If you do not adopt hexagonal architecture, the code inevitably converges toward the patterns the community recommends anyway. Ultimately, it is less about individual code fragments and more about hoping those fragments do not go wrong, contained by guardrails that ensure they do not exceed certain thresholds. If that is the case, programmers might end up focusing on questions like, "How much loss is acceptable?"
Many programmers say, "You need to know the entire codebase." But as someone who has delivered over 30 projects with codebases of 60,000 lines or more, I know that once my own code alone surpasses 40,000 lines, it becomes nearly impossible to hold it all in memory. So I document it in broad categories. And the next day, I have to retrace where I left off and refresh my memory, which takes a long time. This is one area where Gen AI genuinely excels.
Personally, what I have noticed when looking at other people's code is that those who work at a low level often struggle with high level integration, and vice versa. When cognitive resources become concentrated, specialization deepens in one area while others are neglected. And how one manages those cognitive resources ultimately diverges depending on the domain and the specific problem at hand. Just as the mindset required for architectural design differs from the mindset required for diving deep into implementation.
My workflow, broadly speaking, has not changed. My main job has always been to grasp a legacy codebase, open up the black box, connect debug logs one by one, find safe footholds, and add features. The only difference now is that the black box called "legacy codebase" has simply become "Gen AI code." It is not like I could ever claim my own code was bug free to begin with. Ultimately, what mattered was which bugs to contain and how far to wrap them. Open source programmers, however, are different. Their code blocks serve as their cognitive solutions and their reputation, so their approach differs from delivery oriented code. That is because code quality itself contributes to their prestige. For me, it is the brands and the number of companies I have delivered to that build my reputation. That distinction seems to divide programmers at a fundamental level.
In that sense, I think future programming work will split into roughly three categories. The first is the exceptionally talented group that creates the datasets Gen AI will consume. In other words, the renowned open source groups we already admire. In AI labeling terms, they will be the ones producing the gold standard answer sheets. I believe they will survive and remain admired, just as they are now, though their numbers will shrink even further.
The second group will be those who make Gen AI code flow through industries and organizations. When Gen AI pours out code like water, they will be the ones managing the channels to prevent it from flooding. Most people will migrate into this role.
The third group is QA, and I think they will survive. They will execute Gen AI code and relay to the channel managers when the watercourses are going wrong.
In truth, things are already moving in this direction. It is like how no one thinks about how OOP inheritance maps to vtables anymore. People just treat objects as abstract concepts and move on.
Personally, I think the overall flow in Gen AI coding will shift toward implementing system state in the form of state transition diagrams, a return to the primitive form of Parnas's theory of information hiding. Parnas's theory states that "modules should hide design decisions that are likely to change." Humans interpreted this as something like private variables in OOP, but then leaked internal implementations to the outside anyway. When business policies end up scattered across controllers and services, information hiding has effectively failed. The decisions most likely to change are not hidden in one place. They are dispersed throughout the system. But in the Gen AI era, ironically, information hiding might actually become easier to achieve. You just define the state transition from "payment pending" to "payment completed," and that is it. Ultimately, we might end up circling back to earlier approaches, do you not think?
Anyway, I am not asking for someone else's answer. I need to find my own way. Because my way and someone else's way are different, just as my programming and someone else's programming are different.
Programming is ultimately the work of constructing a mental model, and that mental model is inherently personal.
I just hope I can survive into the next era.
That was very nice to read. Refreshing to see a deep take that’s optimistic.
> Programming is ultimately the work of constructing a mental model, and that mental model is inherently personal
So that’s why we now have as many different tech stacks as we have developers at work. Conway’s law taken to its logical conclusion?
I wonder if we're being overly selfish here and ignore the positive effects of democratizing programming and making software more accessible to less well funded causes and organizations. A good software engineer used to cost 100k/year, very few businesses could afford that.
I also disagree that AI results are lower quality. Codex Pro results I get are marvelous but they sometimes miss things that humans understand naturally - all of the edges - coherence of visuals, passage of time, etc.
What AI does is the exact polar opposite of "democratizing". We're going from a world where high-quality learning content is plastered all over the internet for free, where open-source projects are desperate for contributors and full of "good first issue" tickets, and where the tools you need to code can be run on any crappy laptop from the past 15 years—to a world where the majority of content about programming is unverified AI slop, where open-source projects are locking down access to protect against an avalanche of drive-by AI garbage, and where you need a $200/month subscription or a $3,000 GPU to run AI models for coding.
We already have such position and it's called manager.
The manager's job is to manage multiple programmers to let them write code. A good manager should focus on management rather than write code himself.
But keep doing for many years will make manager out of touch from modern coding so he is starting to mismanage people or issue technically wrong order.
For me the biggest issue with AI coding model is not that it writes code, or how it writes code, or whether it will replace me or not. I mean, these are real problems, but just not the ones that makes the difference to me daily. Instead what sets my feeling about the AI is simple: the experience of working with AI models, because it's the worse experience of my life so far.
Every single LLM will make up stuff, go down rabbit holes I didn't even ask it to visit and infuriate me all day long by doing ALMOST what I asked it to do, but just not quite. Take "yes" to one question as an approval to some other question. Come up with a plan but one you review the plan and accept it it hits some minor issue and then throws away the plan and do whatever it wants. Or when I reject some line of code or don't approve a command it wants to run 9/10 times it just tries to re-add or re-run it as if my rejection was meaningless. And no amount of rules and markdown files ever seems to change its behaviour for long, if at all. It's like working with a sociopath who just doesn't remember anything that happened 5 minutes ago. It's the classic case of "my computer doesn't do what I want, only what I tell it".
And all of that would be fine if it didn't pretend to be a human, if all the UX didn't create the illusion that you interact with some intelligent being, because as long as I remember that this is a cli tool it's all good, I manage my expectations, but the experience sooner or later makes you annoyed and frustrated - and if this was a person they would either stop or you wouldn't never work with them again.
> doing ALMOST what I asked it to do, but just not quite
This is my exact experience. LLMs get 95% of the way to my personal quality bar, and for simple tasks 100% of the way.
More complex things, no. This means that, in practice, I end up having to understand the code pretty much as much deeply as I would anyway, without agents.
So, essentially, my experience is that it's automated away the easy bits, but left the hard bits, so all of my time is spent doing the hard bits, which is mentally exhausting.
Maybe Fable would be good enough to get to 100% of my quality bar on more complex tasks, but I never got chance to try it.
> So, essentially, my experience is that it's automated away the easy bits
Yes, which is great for “easy” but mundane and time consuming tasks like refactoring structs when you have a ton of tests that look at snapshots of nested structs etc. Squeezing that last few bits of “make comments consistent with the style and tone of existing comments”, or “stop adding docs that make it sound like an academic paper” or just keep the overall feel of the code same across all files take a lot of effort and energy.
I agree with this. I recently went back to writing more code by hand again for this reason: the amount of effort / mental toll it takes me to steer the LLM is sometime just not worth it. I might be slower but I enjoy my work much more again.
So true. I am cloding a macos app (a domain I know little about), with Opus 4.8 xhigh, and it was glorious at first seeing the app materializing and working (notwithstanding tedious detailed feature spec write up), but when I started fixing deeper problems and doing refactors, and glancing at the code - oh boy. Now my rule file grows by the day with "how to think properly and not shoot itself into foot" stuff, and I am constantly catching it red-handed and have to explain how to make stuff normally, or how to fetch data properly and efficiently (pretty much basic SWE stuff) because it is easily distracted by it's own assumptions or blatantly forgets whole fields of knowledge (as it explained it could be pulled out of latent space of that knowledge and become locked in another bubble of latent space). Constantly have to steer it and remind it to do web-search instead of running circles around some problem it can no longer understand.
I had to explain it that quite an extensive "tests first" rule didn't mean to just "write" them first but actually "run" them first to confirm stuff.
On the other occasion it interpreted my "yes I want migration not to get stuck in failure mode" led it to write a workaround which silently drops DB and creates fresh one whenever migration had any failure, it was epic, I was so glad that I have looked at the code then...
And funnily enough I am probably learning to be a better mentor/parent who can keep steering it through its shenanigans without loosing my shit and being an ass. (Because anything but calm "so here you went wrong way, how can we avoid it in the future?" just puts it into disgusting apologize mode, and I am afraid if it ever go into revenge mode).
it just a higher level of abstraction but not high enough that it replaces developers, I don't think the LLMs are even capable of reaching that level. Saying the creative process is gone is laughable. A lot of people make these claims and only have an internal tool that no one uses to show for it.
I built a "simple" storage application recently, uploading large files from a phone is not trivial, you can't just command Claude to do it without giving it good context.
The only reflection that matters, now, is that every company that sells software for money is now a confidence game. How long can you con your users into believing they have any reason not to type the 5-word prompt “recreate <product> make no mistakes” and cut you, the middle man, out?
Not my experience at all (30+ years of experience as a professional software developer).
I currently work on a very large (multi-million lines) C++ code base and use Claude Code for the following:
1. I run Claude Code in the background finding bugs, while busy implementing new features.
It has so far found and fixed 100+ minor issues and a few major could-crash-the-server-in-production issues that I was really happy to get fixed before a customer found it.
2. I use Claude Code to do massive refactorings that would be really tedious to do by hand.
I love the fact that it updates not just the code but also comments and documentation.
3. I use Claude Code to implement features that I frankly am too bored to implement by hand.
Excellent! Less boredom at work!
4. I use Claude Code to generate tools to analyse the code and/or generate code.
This is a massive quality improvement. It allows me to make the code more maintainable and simpler while adding more features.
However I always always always check and verify every change before committing to git. Because the LLM it will occasionally add features or bugs I don't want.
The end result is really positive. Lots of productivity improvements, less tedious can't-be-bothered work, lots of bug fixes that otherwise would have triggered production issues etc. etc. while 100% maintaining my understanding of the code and all changes.
Love it.
[dead]
This is not an Age of AI. Why do so many not-real-programmers think this is an age of AI? It's just a hype.
I am developing my own app 2+ years manually, and not use "AI" at all.
Using AI usually can cause many problems, which the author pointed at.
Programming should be enjoyable, not an annoyed or unhappy thing.
Yeah it’s gonna be funny looking back in ~5 years. And probably embarrassing for some.
“Age of AI” but most people actively dislike it and hasn’t had a clearly positive impact on the economy after ~4 years.
How do you want to define "clearly positive impact on the economy"? NVDA and a bunch of other stocks are way up. If you got in early, you're sitting pretty. I'd call that positive for those lucky people.
Heh, during the California gold rush most of the people who turned a profit did it by selling supplies and services to the prospectors. Maybe I shouldn't have said economy but rather "society".
There are still bespoke craftsman in this era of machinery and assembly lines.
You choosing to do something in one specific manner is more the exception that proves the rule then any rebuttal.
Current AI cannot be comparable with "machinery and assembly lines".
Current AI companies stole much data from Internet without permissions, and used the knowledge of many people directly shamelessly.
It cannot resolve many practical problems well, even in programming area.
Do you want the software-assistant to do the important or enjoyable things instead of you? Do you want your assistant in real life to eat the delicious food instead of you?
> Current AI companies stole much data from Internet without permissions, and used the knowledge of many people directly shamelessly.
So? It's not illegal and no one seem to be in any hurry to make it. That is a moral dilemma that everyone just agreed to ignore not anything to do with how functional it is.
> It cannot resolve many practical problems well, even in programming area.
It can solve enough and is getting better every month. I live in SF and I take a self driving Waymo at-least once a week including on the highway. I'm already literally trusting my life to AI at this point.
> Do you want the software-assistant to do the important or enjoyable things instead of you? Do you want your assistant in real life to eat the delicious food instead of you?
I work for a living man so yes I want the damn computer to do 99% of my job. When I program for enjoyment I'll do it however I want in my own free time. At work almost nothing is novel enough to actually matter (nor should it be) so I don't care about implemented something that's been built 100x before me.
[flagged]
> I am developing my own app 2+ years manually, and not use "AI" at all.
How does this mean this isn't an age of AI? Because you're The Representative of Programmers?
> so many not-real-programmers
Oh sorry I missed this part. It seems that you really are!
Please don't cross into personal attack, no matter how right you are or feel you are. All it does is damage the container and make things worse.
https://news.ycombinator.com/newsguidelines.html
Please prove this is an age of AI, could you confirm it?
If a real programmer does use "AI" usually, or use AI to solve all problems about programming, and AI does the all programming tasks, after 1 or 2 years, do you think his/her skills of programming are improving or declining? He/She will lose big part of programming skills very probability. Do you think he/she is a real-programmer then?
The force of "AI" companies hype is very strong now, they are everywhere. If you research the articles very carefully, you will feel this force.
The force cannot beat the reality or truth.
Machine hater and manual labour worshiper..he thinks he got stupid because he stopped doing if/else and for-loops. Life is full of complexity and it might be helpful to get one level-up in abstraction to try to solve real problems and other people actually care about.
When I started posting on HN, many times I get downvoted. And this an another example, you don't get discussion you just get downvoted.
And it's a good thing, it means I think differently.
The fact is, one could pump literally 10k LOC/Day using LLMs, and in the right hands, those will be quality code, and this is objective observation from someone who has been coding for 20+ years. It almost feels like we have tractors for mechanical thinking, but there are those in the industry who built their entire career and identity on the bottleneck of coding, and those struggle.
But those people will be left out, forgotten. Progress doesn't care about feelings and identities, entire civilizations with their minds collapsed, and buried because they didn't adapt to progress. Do you really think your ego can help you to shelter from that when many before you parished?