- 164comments
- 12comments
- 6comments
- —discuss
- 39comments
- 112comments
- 453comments
- 7comments
- —discuss
- 139comments
- 13comments
- 38comments
- 1comments
- 288comments
- 79comments
- 446comments
- 367comments
- 32comments
- 128comments
- 392comments
- 366comments
- 172comments
- 47comments
- 29comments
- 350comments
- 37comments
- —discuss
- 35comments
- 21comments
- 27comments
Sad to say, but this is no different from human written code. Human written code just takes even longer to realize the mistakes because the pace is slower.
I think at the end of the day, it is not impossible to have AI write "good" or "high quality" code. If anything, once the patterns are established, AI will be more likely to adhere to the patterns and rules than any human team. It requires the most experienced engineers on the team to split their time writing the core patterns and documenting them in references/skills.
But it takes a lot of "taste" and a willingness to slow down a bit with AI (to create necessary artifacts), something teams find hard to do when you can ship so fast now.
My experience has been that there is a camp of very senior engineers that are unwilling to adapt to reality and focus on documentation and writing (effectively producing skills and agent guidance which multiplies their effectiveness); they will cling to their knowledge thinking coding a sacred art.
To me the difference is humans (ideally) will learn when they build something in a non-optimal way, and so will improve over time to become a competent engineer / architect. We cannot be perfect but to me a huge part of life is learning from failure and improving yourself, something that LLMs short-circuit and cannot replace.
LLMs cannot truly learn and so are destined to produce whatever the "average" software looked like at their training cutoff, or worse to produce code based on _other_ LLM generated code.
Ouroboros eat your heart out
While I mostly agree, I think this is something we need to assume the Pareto principle applies to: likely 20% of humans will improve but 80% will not.
LLMs learn, and in two main ways: in-context and in training stages, release to release. The former is quick and sample efficient - perfect for adjusting AI behavior on the fly, and for enabling AI's own problem-solving capabilities. The latter modifies the "behavior defaults" and gives you performance gains that stick.
Why do you think that "write maintainable code" is somehow impossible to learn for an AI? We already have AI storming the frontiers of research math - way beyond the "average" of the field. If you can RL for "better at math", I see no reason why "better at maintaining code" would be somehow impossible.
You can construct an RL env where a codebase is presented as a "tree", and the AI is given one change to make at a time - and the per-change reward is not just whether the change itself has been evaluated as "made successfully", but also whether it made future changes down the line more or less likely to be successful, and harder or easier to make.
This is a formulation already used by some "maintainable code" benchmarks, so I expect something like it to make is way into frontier lab RL pipelines some time between "next week" and "a couple months ago".
Yes but the ceiling is still higher, and that's the author's point. If you vibe code, without code review, code becomes a mess quickly. If humans write code by hand, then this is often the case too, but crucially, this is not unavoidable. Sure, most codebases are a terrible mess, but some are not. AIs unfortunately got trained on all of them (+ reinforcement-learned stuff) and therefore their quality standard is about as low as that of the average codebase, ie pretty damn bad.
But there are plenty examples of acceptably decent yet long-lived codebases, both in OSS and inside companies. You simply couldn't get that quality by vibe coding. (unless you review every line of code and every design decision, at which point you're about as fast as you would be writing it all by hand, assuming some seniority)
I really don't think so, poor written human code IME is rarely overly complex, where as the AI code is almost always vastly over complex. Naturally complexity can be an issue because it leads to more surface area for failures and challenges to diagnose, but where I am REALLY seeing an issue is the complexity hiding an issue. Something that should normally fail or produce an error is covered up by something multiple layers deep in the code that returns an incorrect value instead of an error when something goes off the rails.
AI written code is a function of the human created constraints around it.
That is why I believe the most senior engineers on the team with the most scars and most experience need to shift into writing those constraints instead of writing code.
In writing those constraints, they can multiply their effect across a tireless fleet of agents that generally want to copy existing patterns and can be guided to use skills.
When the pace is slower you can notice mistakes earlier because you have time to reflect. It also allows you to detect when it’s becoming hard to maintain and you can correct course, rather than after it has become an unworkable mess.
It doesn't mean that people do. This is a false narrative we tell ourselves. Yes, there are craft-oriented devs and teams, but these are the exception rather than the rule because in the end, it is the GTM and business teams that define what, when, how and rarely the engineering teams.
There is no team without tech debt because there is no "golden" project where every decision has been made right because of reflection on decisions made wrong.
I formulate this idea as "Clean code never survives first contact with users"
After a certain point, people would be forced to refactor, because they find themselves unable to handle the complexity.
With LLMs, there is no such friction. So the complexity get piled upon complexity in the form of a million best practices that is indiscriminately followed...
This is a fallacy; this is why legacy code exists that teams just work around. They lack the tests to verify it, the person that wrote it is long gone, it's handling some mission critical dataflow so no one touches the code and just builds around it.
What you say here is not always the case.
Yes, "code rot" is not in any way an AI-unique problem. Codebases like Flash Player or Bethesda Engine have been deep in decay long before AI was capable of contributing to them.
Historically, this was caused by hiring the cheapest developers one can find, having high turnover, outsourcing, pushing to ship at any cost and more. AI just lets you get there faster, and without having to hire bargain bin Indians.
The thing is, today's AI is already far better at "code rot per feature shipped" than the worst of developers - and I struggle to believe that we're at the limit there.
I've already seen benchmarks that test for AI's ability to make incremental changes and tweaks to code continuously - thus, tracking whether earlier changes make the latter changes harder. This makes for a clear target to RL for.
AI is not better nor worse at producing code rot; just faster at it.
AI produced code is a function of the team driving and instructing the agents along with the scaffolding produced by the team (skills, examples, docs, comments); same with human teams.
A team that cannot guide a human team to produce better code will not be able to guide an AI team to produce better code because it's the same skillset: being able to write good docs, create constraints structurally in code, produce core architecture that enforces good behavior.
Not entirely wrong, but there's a very big hole: "the scaffolding produced by the team" also includes the scaffolding produced by past AIs.
An AI that knows how to keep the documentation accurate and up to date, and does it by default, would, all other things equal, rot your codebase less. An AI that changes the code without checking whether it obsoleted a bunch of examples in the docs would rot your codebase more.
While I think that you can reduce "AI-induced code rot" with good prompting and steering, you could also make headway against it at model level, by making the AI "well-behaved" by default.
This statement is also true of humans. Everything you've stated here is also true for human engineers.
But the agent can be instructed reliably to keep documentation accurate and up to date and will then do so dutifully. Put it in AGENTS.md that it must always update the /docs directory by creating a new doc or updating an existing doc and it will do it. (Yes, adherence may be 95% of the time, but that is likely several points higher than with most non-NASA human teams)
Better yet, extract docs from code comments. Even better when the docs are spatially co-located and line of sight as the agent crawls through code.
I don't think so. It's true that human also write shitty code but the key difference is we actually remember what is the intention behind those crappy implementations so someone can fix it later. aka it is the matter of long term memory that currently LLM architecture is not capable of.
You can argue that claude can read the whole linux codebase and report bugs, but they can only report local bugs, not systematic one. 1M context windows seems like huge, but the effective range is actually pretty limited, and it still does not equal to human insight.
Long term memory is easier than you think when you consider what an agent has to do when it is reading and editing code: instruct the agent to leave comments on its rationale and reasoning directly in the code. This is infrastructure free memory that every agent that then sees the code will read. Your code review agent will see the reasoning and decision making your coding agent formulated. When an agent comes and refactors this code in 6 months, the comments will be there (and it will update it!). When an agent is trying to troubleshoot an issue, it will read the comment. No infrastructure needed! Don't overthink it; use comments.
Code comments are line-of-sight for agents and one of the cheapest, highest leverage ways to get better coding performance from AI because unlike skills that may or may not activate, comments end up in context as long as they are well placed and carry the right instructions.
Best places to have it leave comments: 1) start of the file because it frequently uses `sed -n 1,200p` to read files and 2) inside the body of the method because it may find by keyword and read a few lines past. If your harness is set up with an LSP, language standard comments are also useful because then it can read comments on the member.
Tips for comments: point it to other, related members or artifacts; point it to external canonical docs; point is to a specific issue number or PR; have examples directly in the comment using your language's example markers; point it to example, reference usages in code. Use AGENTS.md to tell your agents how you want it to leave comments and to specifically read, follow, and maintain comments.
You don't need infrastructure or special architecture; Every coding agent is text-in, text-out. You need comments that get carried with text-in and a bit of guidance to the agent on how to use comments effectively.
No, developers definitely do not remember what they did two months ago. If you are busy, even two weeks is a problem. That is why we discuss documentation so much, self-documenting code, tickets and tests.
Well, and "intention" is a mine field of its own.
I've seen LLMs "connect the dots" across complex systems many times before. When it works, it's shocking how quickly it can pin down a bug that spans across the software stack.
1M context window is plenty. Once it's skimmed the code and come up with a theory for the problem, it can spin up a subagent that has a whole fresh context window and it can dedicate the whole thing to that one hunch.
Taste and smell still apply. You have to know the art and have comparative priors in order to judge the output.
We're a few years into a new technology that is still improving. This is a point-in-time critique.
Mostly stalled now...
Expecting big progress to be made every month is a bit silly.
If the models are as capable in a years time as they are today you could say they have stalled.
Improvement can still be happening, and it can be happening slower than it was a year ago. Both can be true.
And the observation that it's happening slower than it was means, not that it's stalled, but at least that it's not exponential anymore. The exponential phase is already over. That has implications for where we think the plateau is likely to be.
Also AI is fine at creating maintainable software, you just have to nag it to and not accept its first attempt at it, and subject it to peer review. This is plenty similar to human developers.
But people don't use AI the way they use human beings because it's a technology and it's being consistently overhyped by its own makers as superhuman intelligence
It creates maintainable looking code that will not stand the test of time.
Some humans don't even manage to do that.
Yeah and they’re typically fired for incompetence, or their startup fails, or their product flounders but that’s okay because the company doesn’t need the product to actually be successful.
Meanwhile, the people who work on actual products that matter have a valid criticism that can’t be dismissed with “some humans don’t even manage to do that.”
The dirty secret is there's a lot of people employed to do things that are not "products that matter" according to your definition.
Yes, including myself in the past. I’m grateful to be out of that situation.
Probably a vast majority of software written never actually gets interacted with by a user. A vast majority of VC-funded software probably never gets interacted with by a user.
Feels pretty great to work on a product that actually has a user, and that user isn’t actually the product (and that user isn’t being manipulated or exploited). Tiny, tiny minority of the software industry I work in.
That does not sound right, I would guess the most code is in enterprise and business applications and people have to use it no matter how bad it is. I would guess only a tiny fraction of code - besides hobby projects - ends up in the trash without ever having users.
If you would fire all incompetent [1] software developers, there would be almost nobody left.
[1] There is probably a more fitting word, I am just reusing incompetence here, but that is not a really fitting description, I think. I would maybe say carelessness or something like that, but a single word is not going to capture the issue accurately.
Same with pretty much any other profession. Hunting for reliable and reasonably-priced HVAC, plumbing, handymen, etc. is a part-time job.
But the point OP was trying to make is that because LLM harnesses output code that is at least better than the worst, say, 20% of developers then we should be fine with it. Meanwhile, I’ve worked in shops where you have to be way above the worst 20% in order to keep your job or even be hired in the first place. And it isn’t some crazy lose-sleep-over-it, stressful requirement. It’s just that literally if you’re not good enough your work will be identified as a liability and you’ll be let go. It’s only happened to 3 people fwiw.
I think crappy code that looks like crap, is better than good looking code that is actually crap if someone cares to look..at least it can be fixed early.
LLMs commit crap, and read the "pattern" back, and consider it as gospel and repeat it all across the code base...
Nobody does this though, that’s the problem. If most people don’t use a tool the “correct way” at what point do we blame the tool?
For starters, LLM’s need to stop being our friends. But that won’t happen because the dopamine loop is baked in on purpose.
Why would I want to nag it and do all that when I could just do it correctly myself in the first place?
That's like asking why senior/staff engineers tend to review more code than write themselves. In certain environments there's a scaling constraint, and you can generally have wider impact through oversite than typing at a keyboard yourself. That's not to say one is more worthy of your time or not, just that it's not unique to AI.
Generally though, you are also investing your time into leveling up junior engineers to take over responsibilities from you. I just never really see that happening with AI. Even as it gets "better" technically, there's no real growth pattern to its work and it doesn't understand ownership or responsibility.
But if scaling isn't a problem, then sure just write it yourself.
Same reason you’d hire engineers rather than expecting the CTO to do all the programming?
It doesn’t take much effort to setup cross-agent reviews and automatic reviews for slop and accretion, while directing design decision questions back to the human to consider. I have had a considerable increase in throughput of code that I designed and made the important decisions about, and that I’m pleased with the quality of, although as always in these discussions, someone will be a long shortly to tell me that that implies I must be a terrible engineer.
I did hear an argument a couple times saying "this is messy now but a newer model will come and clean this up in the future", hope it's true.
It's not true, a newer model can do better to clean things up but you can't brute force engineer your way out of a shit design.
I predict there's always going to be a Wirth's law or Peter principle of slopcoding — it will always be easier to tack on a new feature than to understand a whole system and clean it up.
Alternatively stated: codebases will expand to the limit of an organization's ability to manage them, so the equilibrium will remain at the point of near, but not total, incomprehensibility.
It's definitely true. I didn't use AI at all until this January because the code it wrote was too bad. Now I'm comfortably using AI for less important programming tasks and it's okay as long as you steer it carefully. I see no reason why it wouldn't get better up to a point that AI written code is better than what any human programmer could produce. It's also constantly getting better at refactoring.
2 years away:tm:
For this to work long-term there has to be continuous improvement.
My belief is that paying off tech debt requires a better model than creating it. At the same time there are people who will create tech debt no matter the tool.
Should the models stop improving, the debt will pile up.
I don't see context size improving.
It was mostly 256k, then it went to 1M and now it has stalled there.
There are already consequences.
We (collectively) were unprepared for a machine that presents itself in human forms. We were the frogs that boiled ourselves. We built a world of images and words on a screen. And then we built a machine that can (increasingly) mirror that world; it does so in a way which most of us are incapable of disambiguating.
It feels like there is indeed a ghost in the machine.
And there is, but that ghost is us. And that ghost is fading surprisingly quickly.
too doom and gloom.
has the author not started to develop instincts with regard to ai usage and pitfalls?
apparently it needs spelled out.
you won’t lose your expertise if you continue to develop it, and blaming ai is like blaming macros or installers or…
the game hasn’t changed, really, but many are fretting that it has fallen apart.
Then you have to write code. Understanding without doing is not successful IMO.
Tbf, at this point, this has been said ad-nauseam.
At least I did not find a new thought in that (granted, relatable) rant.
"This is bad and you are bad" requires people to not defend their reality through rationalization, but the point we're at with AI right now is driven by exactly that. So this is at best highly ineffective at reaching the people it claims to want to reach.
That said, the underlying emotion of "you all suck and I hope you lose your jobs you frauds" is relatable and worth screaming from the rooftops of Linkedin dot com for the catharsis alone.
There's a continuum between "vibe coded by someone with no technical knowledge or inclination" and "hand written domain driven design development". You can absolutely use coding agents AND have maintainable code. But yes, the coding agents will not magically make everything maintainable if you don't tell them to.
"Code maintainability and good architecture don’t have good measurements that we can apply"
Who has no wisdom? There are dozens of ways to measure code maintainability. Cyclomatic complexity is just one.
Nothing stops you from wiring up something like SonarQube metrics to your agentic coding workflow.
The author admits to a measurement: maintainability. There are quantitative and qualitative ways to measure this, the biggest signal being outright abandonment
That's terrible as a signal. It's too late for anything, and there's many reasons things get abandoned.
One way of understanding "good" here is "actionable", imho.
There's the rub. It requires knowing about and caring about maintainability. And a lot of the people who "haven't written a line of code since 2025" don't care
“Claude, make this code maintainable.”
I’m beginning to believe that if this was a “solvable” problem then the billions of dollars poured into coding agents would have solved it by now.
Humans aren't magical, so if people can figure out what maintainable code is and how to write it, so can a non-human. Furthermore, humans are rather unlikely to be the optimal form of intelligence for software architecture and engineering, so it's almost certainly the case that if humans can figure it out, some non-human could do it better. That it's not yet solved by models suggests that it's not easy and maybe even not practically achievable with current tools. But that's not the same thing as not solvable.
I like this framing. Humans invented software (and engineering in general) as a means to solve problems with methods that work best for us. There may be entirely different, and parallel, problem-solving methodologies outside of human best-practices.
I have not found it difficult to get these systems to write maintainable code. You're probably correct that most people don't care about this so they don't bother to do so. But IMO, this doesn't seem like a problem that is crying out for a solution.
The trillion dollar question is how you do this, if your employees do not care (they are optimising for salary & time spent not code quality) and you have no way of telling apart AI slop vs. good maintainable code. (If you could you would just train the AI.)
Before AI there was at least some way to tell apart good programmers from bad, because there was some human effort involved in coding. Now with AI and slop generation there is almost now way to do this.
From what I see is it's mainly managers and higher brass who doesn't care about code quality and sustainability, and aims to drive time to market metrics down aggressively with AI.
Any employee who cares about code quality will become a poor performer with a red luddite label because they dare to change what the AI has emitted for them.
I'd love to be wrong, very wrong about this, actually.
I think that's true but it's a special case. AI is here to stay and with AI coding IS faster and quality is better than ever before. Ideally you want your luddite fired along with the slop generators and keep the ones who are using AI and taking their time to deliver a maintainable code.
Chances are high the luddite is the one with the deepest understanding about the product and the code base - the one who is actually passionate about the project and actually cares.
And unless you have ownership level stake in the company, you *shouldnt* care about what happens downstream of selling labor to the company. Yes, I highly recommend 8 and skate.
Ive seen soooo many people burnt out, or "ive given years to the company and i got hit with layoffs", or "$200 software would have saved $1000000 when I brought it up to them". And companies will throw you away the MOMENT your usefulness is gone, even if just perceived. So, use them just as much as they use you.
And that idea of slacker is ALSO a way to generate more money for you, by slyly withholding or slowing work. I didnt get my paltry 3% last year. Inflation up 15% or whatever stupid number. But I can control how much work I do, so my effective wage/hour stays with inflation.
Save your caring for your personal projects, nonprofits you help at, your and family/friends labor you help with.
I'll disagree with you, but probably what we do and our life philosophies are pretty different, so that's OK.
Well, what part do you disagree with?
1. Caring about the company when you are a worker and not owner?
2. Companies will throw away/layoff people with no notice?
3. Work slowage (work-to-rule) as a counter to low/no pay raises in accordance to general inflation
4. Invest emotional and physical labor in ventures you gain completely out of
That could be the case. But if other companies figure out a way to deliver good code with AI, then the company with the luddites is going bankrupt. Luddites are better than slop generators, but both are worse against a developer with AI who actually cares and has a deep understanding of the code base and programming.
I don't think that's true.
I don't think companies are living and dying, by how much code they can produce. Even tech companies.
They absolutely live and die by how quickly they can produce working solutions.
Maybe? This does not seem to be in evidence to me. In my experience thus far, this seems to be way more of a personality and ideology split than a depth of understanding split. On my teams thus far, I've seen the deepest experts land on opposite sides of this question.
That's true, but it doesn't have to stay in this form.
Citation needed, because the last study I read about was painting a completely different picture about code quality. Also, just because the AI pulling and remixing code from a known repository with high quality doesn't mean your code will be at the same quality automatically. Passing tests is not enough.
The thing is it's not possible to see who generates slop and who generates code, and if you fire the only people who knows about the codebase intimately, you'll be on a very exciting, possibly fatal ride. I don't recommend this. AI doesn't know your history and trade-offs. These guys do, and can guide you to clear.
AI can't.
Believing that AI will create bug-free code from start is believing that Rust is the silver bullet.
There are no silver bullets.
This is exactly what I said in my top comment. This is a huge problem.
...and then you said:
The luddite is the person holding everything together.
Like the office maid who looks like doing nothing but keeping morale high and everybody's sanity intact.
You can hold everything together and not be a luddite.
No, the luddite is the person in one bad location on the continuum, opposite the AI psychotic on the other bad end of the spectrum. The person holding everything together is the one applying wisdom to the use of tools for the things they are good at while avoiding the things they are bad at, rather than blindly following one ideology or its opposite.
The reason that we can (not that everyone, or even most people will) create higher quality software now is that it is now much easier to try out a bunch of edge and corner cases, which would have previously often required a prohibitive amount of time to set up and run. It's a dream! I can certainly believe that people aren't taking advantage of this, but they should be!
But I totally agree with you about the measurement problem. I think it's a very difficult time to be a hiring and firing manager.
"Luddites" can just be the ones pointing out the Emperor's new clothes have a big hole in them. Removing them is just ensuring that whatever mistakes you're making get reinforced.
Is this claim based on something?
I'm not against or "for" AI (whatever that means), I try to use it as effectively I can, but for me it's not at all obvious that quality is better than ever before.
Speed I can buy, especially in new projects and utilities, but quality? At least I haven't seen this in practice, if anything I'm just seeing more code, issues, PR's and pressure ==> more slop, more bugs, less quality.
You can always say "skill issue" and "process issue", but that's partly my point here, AI doesn't magically solve this.
Presumably it is based on that person's personal experiences, like your own comment and all the other comments here?
Speaking for myself, based on my own personal experiences, quality is by far the bigger advantage of these tools. It has never ever been easier to write automated tests and to automate tedious manual validation. I'm running my code through like 10x more paces than I ever did before, because I can just say "hey try running this in these twenty different ways" (including with browser automation, if that's relevant), without needing to either do the tedious steps to run all that or to take the time to write a script to do it, and to compile and attach the findings to the PR. This saves me hours to days of work on validation, but the reality is that I just wouldn't have spent that time in the past, I just stopped at a lower bar for quality, because I couldn't justify the ROI for spending all that time on it. But now the ROI is huge, so it's a no brainer.
If people are not taking advantage of this, then yes, that is literally a skill issue.
...or domain. You said "including with browser automation", so you do web or web-adjacent development.
Not all of us are doing that. What I work on doesn't have any UI or output besides a log file most of the time, but it connects to many places and does many things like an octopus, but nobody sees that, but feels that it's there because their environment keeps on working.
I said "if that's relevant" and you jumped to "this guy just makes websites". I do many things.
The octopus you just described sounds to me like an excellent example of what having the ability to more easily do tedious validation is most useful for. If you know that the "environment keeps on working", there must be some way for you to observe that fact. And if it is an octopus, it is likely difficult and/or to change the conditions and observe the correctness with respect to those changes. I find it so much easier to do this exact kind of thing now. Or, "easier" really isn't the right word. It's that the activation energy is low enough now that I'm able to do a lot of things up front that I used to rely on runtime monitoring to validate.
I guess YMMV, and it's not magic, but for me it totally changes the calculation on when it makes sense to automate something (like that chart from the old xkcd about how many times you'll do the thing and how long it takes to automate) in a way that means I'm doing a bunch of things that are useful for quality that just would never have passed the bar in the past.
Good use case.
For some reason this reminds me A LOT of past discussions about microservices, most wonderful on paper and forever debated, but I've never seen it work out perfectly in practice, for me it's mostly been a cluster F in most companies that adopted them.
Currently I see AI similarly, in theory perfect, in practice I don't see the claimed effects. So yes, skill issue, but skills are relevant and your company probably can't hire a rockstar team (if that matters in the future).
But your comment on personal experiences was very good! Spot on, we are all biased, easy to forget. Thank you for that.
I agree with your analogy as well, and I'm very open to the idea that this will mostly end up being good in theory but bad in practice.
I fully agree with this point about quality. I am doing so much more testing than I used to, because I have so much more time to do it, and it's so much easier to automate the more tedious kinds of validation.
Maybe it's true that lots of people aren't taking advantage of this and are shipping trash, but that's their own problem, and there have always been people who do the job poorly.
I often have to ask myself if I’m asking for something different because of preference or need. I don’t really know what others are doing but I see this comment a lot about needing to always correct agents. I can’t figure out if it’s an exaggeration or not because once I’ve planned how I want something done I pretty much have zero need to intervene.
Impossible to know without doing a detailed comparison. Are you using the same LLMs? Are your criteria for correcting the output the same? Are you working on similar code? Are your plans and prompts the same?
Without any comparison I’m pretty sure the ones trying to correct LLM code are the same ones that incessantly bike shed human code decisions too.
It's fine to have preferences. The biggest problem with bikeshedding is the time spent (wasted) debating. But you don't have to debate the AI, you just tell it your preferences, and ideally encode them so that they are repeatable.
They don't care about maintainability because they're not going to be debugging it at 1am on Friday night but mainly because if something is wrong they have lots of people below them to blame for it.
Developers, however, are still responsible for the code! We must review the AI....all 80k lines of code it generated yesterday. If we don't then we are at fault. And we must go full throttle of course. So ....not be picky and retrograde about accepting what is generated.....
IOW we know who is going to get screwed and it isn't them.
We have all become underwriters and should get paid accordingly for that fact.
How are underwriters paid?
Badly.
Yeah but I mean, what is their compensation based on? I'm genuinely pretty curious about the point you're making here. I don't know anything about underwriting or how that profession works or how it is analogous to what you see as the future for software developers. But it seems interesting and possibly insightful.
Trends over time will drive more observable changes. If a whole generation of programmers picks up bad habits that their managers don't care about (think very junior), that will take some time to play out. It's like children's literacy. You don't notice overnight, but a decade of neglect and you have a reading problem in kids.
Yes, this is why it's good that we're having these debates. We should keep having them. Personally, I am not yet convinced that this is a real problem. But I also haven't yet worked on a team with people who have entered the field in the last few years.
The real problem is all the managers pushing so hard for velocity instead of letting the employees pick up the tools at their convenience.
Go fast and break things has been a mantra for how long?
I think a lot of the laments about "good" code are really about "ownership" - and as someone who spent most of my working career in OTHER peoples code bases I have seen some things. There are a lot of you who think that your code bases are "great" when they are NOT. Personal understanding is not a good measure of quality.
The increased cadence from AI is just speed running to the legacy code base.
The answer: express the concern, and reiterate it after every issue that arises because of increased complexity. Start building a plan on how to "unravel" the mess, how to migrate things in place, how to start drawing boundaries in your systems. The system is designed to reward heroes who fix problems - you want to be super man who stops the bridge from falling apart, not the engineer who pushes the costly fixes it before it does.
Practically since eternity, but just as we learnt to manage current rate of "fast" and "breakage", somebody attached a solid booster behind us. So we're trying to understand what happened and what's happening and what will happen.
People owning what they did, have responsibility and initiative about doing better is always a good thing, yes.
I can understand that, I'm sorry you had to go through this.
My codebases are as great as my knowledge. I love when someone reads my code and points where I f'ed up. I also love to show what I did has achieved something I was aiming for and discuss how to achieve it betterer.
And this is the problem. Because I work silently and diligently build something looking unimpressive while working like an atomic clock without any problems.
You pay someone and give them the explicit responsibility for code quality in your system, empowering them to gate check-in with any static analysis and adversarial agent review they feel like. It’s not a silver bullet but you absolutely can do better than just giving up.
Anything involving an AI won't work. If it did AI companies would already train their AI with it. Review solutions during training, generate synthetic data etc., or as a budget solution just route the requests through more models before giving you an answer.
This isn't true, there's a limit to how much feed-forward processes can correct themselves and well-poisoning of context is far too common.
Try using an LLM to rewrite an LLM output without the slop (vs asking for no slop to begin with) or sandboxed subagents that critique a parent's draft.
There is absolutely a step-function improvement in quality but: 1) not everyone wants to explode their cost by adding extra calls 2) this can't just be "trained in" to a system as obviously they have attempted this but the technique still provides an uplift.
If you can do adversarial agent review, or use an LLM to fix the slop and it works, then technically you already have an LLM trained which knows what to do. You can just use them to improve the other model. The issue is that we don't have these at the moment.
There were only bad ways, and the best way was to just find people who were both good programmers and cared about quality to keep an eye on the rest. Nothing much has changed in that respect.
There is still lots of human effort involved. If there isn't, you've found the bad programmer
I know how to distinguish good maintainable code from garbage. I have known for quite a few years. But knowing how to train someone, or an AI? I'm a good coder, not necessarily a good teacher. And there are things about code that I _feel_, not that I can rationally explain.
I completely agree with you on this. I was talking about the employer perspective. They have no way to tell apart the two and thus have no way to set the incentives.
You might write good, maintainable code, but they will prefer the slop generator who delivers quicker.
Sounds like there is a compensation problem then.
Disagree, throwing more money at devs hoping it will improve quality while continuing to take away their authority will only perpetuate the slop machine.
It's not that hard: treat people with dignity and take their contributions seriously, not as a disposable meat mass. In fact, not only will this improve code quality, it's likely to improve employee retention too.
Sounds like a culture problem. I’ve never had a team I lead not give a shit. Even if the project is dumb (some projects are dumb), there’s enough pride in competence and desire not to dump garbage on the people around you to ensure they at least want to try.
One way to get workers to care about the quality of the product they are being paid to build (for someone else's profit) is for the management to show some loyalty and build trust over the course of a decades-long career by treating them fairly and avoiding layoffs.
SonarQube
Lol no. Just get your company quality gates and put it in the Claude.md.
I guarantee you will pass the gate, and the code will still be dumb, except now you'll spend 10x the tokens.
That's why companies were interviewing people on tasks that had nothing to do with writing maintainable software. /s
I understand the concern and it should be addressed and researched. But, simply saying "humans were writing code themselves" doesn't provide any evidence for better quality.
I don't think it's that hard of a question to answer. I've noticed on my team, our thinking has shifted from how do you directly solve a problem, to how you get an agent to effectively solve the problem and not produce slop in the process.
One thing that we have done that's probably made the biggest impact is alot more upfront architecture with the knowledge that pretty soon agents will be running wild all over the code. Having worked with these agents for a while now, you get a very good sense of how they will go about solving a problem and the various footguns they will encounter along the way. Editing an AGENTS.md file or building a skill is not nearly as fun as coding by hand but it will pay dividends over and over if you do it right.
Another big thing is doing refactoring passes. Early on in our projects our agents generated ALOT of slop and we had to go back and fix alot of it. But every time we did one of these passes, a major aspect was improving agent instructions / skills / etc so it doesn't happen again. It can be a painful process at first but I found that over time, the amount of slop the agent produces goes down by orders of magnitude.
I feel like we're still very much programming, but we're now doing it at a "higher level" where we are not writing the code ourselves but instructing the agent to. And IMHO, properly instructing an agent on a production codebase is not a trivial task.
You can, but nobody does. Of course, there's a lot of non-AI slop too. (Hello systemd)
I for one, have far more rigorous quality checks in my hobby projects (where AI coded), than I ever could justify when I hand-coded them.
I'm not claiming to be everybody, but surely a good portion of the population are using these technologies similarly.
We're not talking about unit tests here. We're talking about data structure and code complexity.
The comment you replied to did not say anything about unit tests. And "software quality" is much broader than "data structure and code complexity".
I do think this gets to the heart of the matter. I think many programmers have missed the forest for the trees on why things like data structure and code complexity matter. They do matter, but they don't matter in and of themselves. They matter because they are the best techniques we have for making software that is of high quality (the software, that is, not the code) and which remains so over time, while continuing to be developed and adapted.
I strongly believe that it is now much easier to create software that is of high quality and adaptability, orthogonally to the data structure and code complexity concerns. Those concerns remain relevant, but it's a mistake to think of them as the primary thing rather than things that support the primary thing
Yet.
I'm sure in few years, as new criteria enter benchmarks, AI will be creating the clearest and smartest code people every seen, by default.
Plenty of measurements exist, but the trick, as the article said, is finding good measurements.
Cyclomatic complexity has been pretty solidly discredited within the maintainability research community for decades.
Sonar's cognitive complexity metric is a bit better, but here's a study that found that it still only has about a 0.5 correlation with how much difficulty programmers actually had reading code as measured by multiple methods.
They found that the most accurate way to measure code complexity that didn't involve something like an eye tracker or EEG is still basically just vibes - asking programmers if they thought it was hard to understand.
https://www.frontiersin.org/journals/neuroscience/articles/1...
Halstead Effort came in second, and scored pretty well, but here's another one where it doesn't do so well, either. And it scores the SonarQube metrics even worse, with only a 0.35 correlation: https://www.sciencedirect.com/science/article/abs/pii/S01641...
Yeah, we have this rule in our SQ and I absolutely hate it because of the number of times I'm forced to break something up. Complexity that matters is usually at a higher, more architectural level.
Yeah. Looking at the code samples in that first page was pretty interesting. They're cherry-picked, of course, but still it's a little galling to see which ones SQ gave a lower score.
I definitely wouldn't want AI to be autonomously using that as a guide without doing some fairly serious internal A/B testing first. Kind of like for cyclomatic complexity, it's just too easy to find ways to maliciously comply. And if that's what you ask AI to do then that's likely what you're going to get.
//There's a continuum between "vibe coded by someone with no technical knowledge or inclination" and "hand written domain driven design development".
Somewhere in the middle of that continuum sits - "domain driven specifications, described using high level english concepts(that are well defined) from the domain , combined with a selection of a few standard architectures"
But that's not actually a change. Humans wouldn't magically make everything maintainable if you don't tell them to (and maybe not even if you do). You have to monitor them and train them, carefully, basically forever.
I find it funny how you can pick examples on one side of the spectrum and then generalize to the entire spectrum.
Induction fallacy at its best
The author seems to be confusing "I didn't write any code" with "I don't care about software design and maintainability". There exist maintainable and thoughtfully designed software systems for which the designer did not write any code and didn't read most of it. It's not the median, but the median software project has always been unmaintainable before agents.
Sure but many many developers who are willing to outsource every modicum of thought and effort to AI don't care about design and maintainability. And the supposed "super intelligence" of AI gives them a cop out for that
Absolutely! How did we get horrible code in the past anyhow?
Here's the one simple trick that fixes everything:
- Use the code that your agents write in anger.
There you go. Do I know when my agents fuck up? Yes, I absolutely do -- because I'm a user of the code I have my agents write, and I ask things like "why is it taking 50 ms to start this program ..." and then I go in and find stupidity, and excise it. I do this over and over again.
Is it faster than writing it out by hand? Maybe! It's definitely a different perspective.
Start behaving like a baby "why, why, why" and then do a bit of reading, and you'll be fine.
A lot of these blog posts seem like they're aimed at software written by B2B companies who don't even use their own software ...
We've had strong coding agents for less than a year. Anyone making such a definitive statement about how vibe-coded projects progress over time is basing it on guesswork, not evidence.
Or, evidence from a sample of poorly designed projects.
So there's evidence that it can be done badly enough that the problems show up in less than a year. That doesn't prove that more evidence will show up after a couple years, or five, but it does make one suspect that it will.
In my experience, the most difficult part of a software project is getting the specifications correct. To abuse Harold Abelson famously quote: "Programs must be written for people to read, and only incidentally for machines to execute."
"A Project must have proper tests and specs, and only incidentally for a working program that executes"
Somewhat, but I've for sure seen new projects become grindy messes in a few weeks.
As much as I'd love to believe it, it is now a conservative take. Sure, having a solid architecture in mind still matters right now, but manually writing code is completely unnecessary and, before long, even designing the architecture is going to be completely automated.
But then what will I do for work? :(
Our only hope is that work will cease to exist.
[Looks at people running AI companies]
Hmm, I have deep concerns about what and who is going to cease to exist.
Well technically my take is still valid
And we will all live happily ever after? What a load of naive nonsense.
Does writing table schemas count as writing code? Is that architecting? Both?
I personally don’t trust coding agents to have enough context to write domain-specific table schemas, and I don’t have the patience to transcribe all of the context into a natural language prompt. If I ask it to, it’ll write something for sure, and maybe that can be a jumping point for me, but at some point I have to physically write what the columns will be.
I repeat, I hope you are right as I enjoy coding a lot. I just wish that people who keep not using LLMs too much won't have troubles because of that.
You should definitely try it if you haven't already. Most of the models are deeply familiar with domain specific areas in ways that most of us aren't. In fact, I would say the problem is generally the opposite. If model size and effort is is high, it will over architect what an app needs. I find myself reigning in a giant email notification system with subscriptions and "channels" and other such nonsense when sometimes you just want a simple one off.
Yes, the LLM knows the definitions of the words in my domain, that’s clear and obvious.
No, the LLM doesn’t know our product strategy and why certain things matter and certain things don’t. That’s very much what I get paid to do. There’s not even agreement within our team about what path we should take through the domain-product space, there’s no chance in hell that the LLM will choose a profitable random walk through that domain-product space.
If it were able to do that, then AGI would have already been achieved and we’re only compute power away from OpenAI or Anthropic making the marginal utility of any piece of code $0.
Yeah, it's not just over-engineering, it is just that generic systems that look like everything else usually don't solve a business problem, but I think I would be less dismissive of its world model in general.
LLMs are not a "random walk", they take in information and they explore the space according to the way they've been instructed.
I’m taking this view now as well. If you’re reading code, you’re probably doing it wrong. You should absolutely be setting criteria that can be objectively measured and rejecting code that doesn’t meet those criteria or perform as specified. We are all senior software engineering managers now, with a fleet of cheap and ambitious young engineers doing all the authoring.
But reading code? What does that accomplish, other than to slow your dev process down enormously? Serious question.
This is the classic "make no mistakes".
On a serious note, I might set as criteria "avoid code duplication". Does that mean that the model/agent will actually follow it?
I am an OSS developer and I often see PRs (i.e. from the general public) that look correct, pass all CI checks, are heavily documented and they are still wrong.
Most of the times either they duplicate code that already exists somewhere else, or they implement a "feature" by opening a can of worms for subsequent "features" in the same area.
Here I am starting to think a central part of knowledge work is the knowledge gained.
Maybe you are right, maybe not, let's see. Unfortunately I kinda agree, because humans are really good at being lazy and going in the path of least resistance (including me). It's genuinely difficult to not use AI even if it makes my work worse, as long as it's easier and faster.
I'm not quite a doomer. I use AI every single day. I also take a fairly negative view of AI generated code, but I think it's unavoidable while remaining employed these days. A fact I find pretty depressing.
I absolutely agree with the author that humans need to be in the loop reviewing and understand the code they're merging, and generally take a "Hey, build X like Y utilizing Z" approach when using AI to build instead of the "Hey, solve this problem" approach. Our PE overlords actually mandate the latter, but I'm not doing it.
However, a point the author misses is that with AI, major refactors become relatively quick. Hours instead of months/years.
Yes, AI can and probably will land you with major foundational and architectural problems, but your architecture isn't set in stone anymore. Your entire codebase bends like a leaf in the wind.
Yes but if you don't care enough to think about, analyze, and decide how to improve your architecture, it doesn't matter. And I'm seeing that a lot with full vibe code code bases
Yes, and I'm arguing you can absolutely vibe code your way out of a corner you vibe coded your way into. Especially with enough vibe coded functionality tests. Your codebase is entirely mutable soup.
Caring about the architecture only matters if you intend to build on top of it, where it become hard to mutate for needs. Big refactors are quick and (relatively) cheap if you don't care about the code.
Don't get me wrong, I am not in any way a fan of vibe coding but the "you're going to vibe code yourself into a corner you can't get out of" argument doesn't hold water.
Yea, probably with a more expensive model.
Yep great point, it works both ways. It can dig you a hole faster, but also help you get out of it with less effort.
This will probably maintain the problem in a different form.
Counter prediction: using AI attractive even for employees. Do you really think you’d wanna join such a company? No way.
I would. I'm already seeing people label their docs etc. as not made with AI so someone will actually read them
This is the labor theory of value; consumers don't care if the code is hand-made, they want the cheap goods (software) that are the output.
AI labs should produce AI that can take support calls. But let it stay away from the fun jobs like coding.
I personally loathe interacting with AI support
I have more than 15 years in the industry. I have worked in some of the most horrendous codebase someone has ever conceived. Honestly? Humans can do worst than AI.
Let's stop overvaluing human work. Sure, I don't want AI to write the entire codebase without I know what the fuck it did.
But AI is on an equal footing with an average dev.
I promise you no one in this industry is overvaluing human work.
To me, the problem is not about what AI is good or bad at, but about institutional knowledge. If AI is doing the coding, writing, designing, or whatever else, you slowly lose the ability to a) learn from others in the org because nobody knows what you need to learn anymore and b) actually improve stuff because there's no more "what good looks like"
which begs the question - will institutional knowledge not matter before long?
it will not. Dead Internet Theory - eventually everything will be slop. And even if you care to avoid the slop, there is no place left for anyone to differentiate from the market slop offering because the 10-20% that actually care about not being fed slop are not enough to sustain a market.
Really it seems more like late stage capitalism and things that were already occurring before gen AI in all markets, not just internet based things.
My companies current national marketing campaign strategy, structure and creative work has been created by an MBA from outside our industry who did a "deep dive" using AI. The project manager they brought in to run the tickets is also adding AI made decisions on brand and ad copy.
The short version of the result is that we're completely copying the leaders in our market that have 100x our marketing budget because they're the only ones with documented strategies for AI to cheatsheet from, with no sense or irony or concern that perhaps their scale is a core part of their strategy.
It's not just institutional knowledge. We're suffering the death of the specialist. Now every generalist is pulling triggers with no sense of limitations.
Man.. just code, let people build, design, adjust. Who cares? Who are these people writing these posts? Why should we give anything they have to say warrant? These posts are getting old, very quick.
I agree, but the problem is when we outsource understanding.
It's a product of the times. You have some folks who are really passionate about the craft of software, some folks to whom it's just a job, and an entire culture that promotes and incentivizes engagement through raw emotional connection (outrage, passion, pick an emotion).
What you don't see from most perspectives are the silent masses who simply don't engage, don't care about the discussion, and/or are too busy doing what they enjoy.
I think too many people (very much including me) who are "passionate about the craft of software" allowed themselves to become too focused on the code itself as if that were the first order concern of the craft. But the craft has always been about the quality of the software, including how its quality changes as a function of time and adaptation. Code quality is only in support of that concern, it is not itself the primary concern.
It has been a fairly painful experience for me to shift my thinking on this, but it's a much better mindset. I still care about many of the same code quality concerns I always have, but I'm thinking a lot more about why I care than I once did.
It is a warning to developers who don't understand the trade-offs involved.
Also a confirmation to people who have the same inner thoughts and are ashamed to admit in public that they think the exact same thing.
I think we need such kind of posts to combat the influx of AI news.
To analogize coding to sculpting - you're in the "toddler with playdough" phase. At some point you might want to make something other people will find pleasing and maybe even worth putting on display, the rules of composition, the details of materials, and the techniques for not wasting expensive materials suddenly become important.
If you're just writing code to fuck around or automate a small part of your life, whatever. But if you're making a big system or wanting other people to use your product, these things about how to make good software become more relevant.
I work in an industry where software bugs can cause real harm to real people. I care that my coworkers are producing more bugs than ever because they don't take the time to think about the code anymore.
Then where are your industrial controls to catch bugs before they are introduced to the product? Where is the feedback loop to the developers running the AI to control the situation?
LLMs should not touch these technologies. Keep your slop factories in webdev. We have missles, planes and pacemakers to make still. If you ever feel like using your brain again, theres plenty of work to be done that llms cannot touch.
Isn't this a testing and validation problem?
Anyone could say the same about any post they don't agree with, doesn't seem very helpful.
I don't know, we spent the last few years with tons of posts telling us "coding is dead" and the biggest companies in the world telling us our jobs are going to go extinct.
That mindset has also deteriorated my working environment due to some coworkers buying into it.
So it's kind of nice to see some sanity checks that align with my beliefs too. I can share articles like this with my teammates. I can see that I'm not alone in thinking most LLM code is slop.
I think too junior devs NEED to see this. My team had a couple of promising juniors who are now completely brain rotted by AI and can't even write "Hello World" without consulting Claude anymore.
Yeah I agree. We're still swinging this pendulum. We haven't found the stable equilibrium yet. It's useful for people to keep having this debate.
Forgetting about AI for the moment I do like this paragraph very much:
.... because I have found the same thing - that there's nobody more zealous about some paradigm than those who are recently converted to it and who haven't come to find that everything has its trade-offs. Design is always about evaluating the trade-offs and seeing which ones most suit the given situation.
"There will be consequences"
And, so what? Software as an engineering discipline has long lacked standardization and regulation to be on par with other engineering disciplines, and the fact that code and all its surrounding ecosystems are not "visibile" or "malleable" makes this extremely hard.
You can use terraform and yaml to define infrastructure that literally spins up machines _somewhere_ in the internet. With all its issues, bugs and associated consequences mostly being ignored.
I just don't understand the difficulty in KNOWING what needs to be done: NO LLM usage in university/grad/high schools, NO LLM usage in the first 3 years of your professional career.
Once the basics are solidly grasped, then they can use it at will.
The problem has never been about wisdom, knowledge or LLMs writing good, bad, maintainable or terrible code. It has always been about the skill level of people using it AND on the fact that people start off-loading basic things to these models that they wouldn't before.
If you have the knowledge and "suffered" through experience to learn the fundamentals, than not using LLMs becomes more deterimental than beneficial.
You just CAN NOT skip the trial by fire of learning and absorbing knowledge on your own. That's all.
The only plausible thing to enforce in practice is "no LLM usage in tests", ie using pen and paper or a fully managed digital device.
I'm going to make my own prediction: this isn't going to happen
You’re both sorta right imo. Some companies are going to try and virtue-signal artisanal AI-free code, and some will just deliver sorta what you asked for, and neither will be perfect.
I'll go one further: We'll see companies proudly boasting "no humans" as a competitive advantage and arguing letting humans do certain types of work is unsafe.
Not a lot of shops these days are bragging about hand-writing machine code and a complete lack of automated testing. We had already automated so much before LLMs came into the picture.
Neither of those mentioned is an equivalent in any way.
Seems like a pretty large claim. Why does a binary exist here and not a continuum?
Why would you brag abput this. Most places make perf products do not use llms to write all their code. This isnt rare. You people dont know much outside of webdev. There isnt a single defense outfit doing it lol.
Yep.
Just look at the growing gap in traffic accident rates between human and AI driven cars. Humans are losing.
We should just throw in the towel now then, I guess.
I don't disagree with you, but now what? Seriously - if we accept humans are just generally losing to AI now, what's the right thing for us to do now?
I dunno. Lobotomy? Heroic amounts of debauchery?
Any suggestions?
I think we'll probably get both kinds of companies, but that they will both be fringe approaches, and the predominant choice will be to have humans working with AI tools.
Luddites have existed since the dawn of times:
this has some pretty extreme selection bias though. You could just as easily say
"We're not going to wear Google Glass"
"We're not going to use Blockchain for every single transaction"
"We're not going to connect every single object and device we have to IoT"
Im assigning you some reading homework today (or have you totally offloaded your ability to read to LLMs too)?
https://archive.nytimes.com/www.nytimes.com/books/97/05/18/r...
And why are the people calling people Luddite some of the least intelligent people I meet, and seem to be complete sheep fighting some psychological war on behalf of their billionaire lords who own the machinery.
Not wanting to hand off all your labor to a machine does not make you a luddite, nor should it be acceptable to call people that because you dont know how to have a real conversation.
I created a summarizer for that
https://briff.site
Okay. I prefer to read Thomas Pynchon's writing. Y
Thats a really bad website too btw, design and functionality. Maybe use your brain before you lose it.
For a 100x engineer with 40 yrs exp that you claim to have, you make some real basic high schooler projects lol.
I already know how to code proficiently in more than 20 languages, that's the result of 40 years working in the field and being a 100X programmer. The summarizer is just a tool for myself, I build tools every day to simplify my life, not for sale, not for monetization, like tracking gym schedules or pickleball rallies with my friends, like kanban boards for my wife, they solve problems I have, not you, so that's enough for me and rewarding in itself
And AI does it all for me while I sip coffee and play mahjong (built by AI of course)
It seems like "luddite" is a reasonable description for a strict "NO-AI" policy. You softened it to "not wanting to hand off all your labor" in this comment, but the word "any" would be a more apt way to describe the quote in the comment you replied to.
Pet peeve - luddites had never been against technology. They had been against using low paid inexperienced grunt workers to displace well paid experts.
See: https://www.smithsonianmag.com/history/what-the-luddites-rea...
Those people will never reach mastery, because they no longer make choices, they no longer take responsibility for mistakes in coding and no longer learn from those mistakes. It’s the AI that’s making mistakes now, the AI doesn’t learn from those mistakes, and neither are the people relying on AI for coding.
Just to give a hot take, it's funny to look at his builtwith.com. As a developer you have a static site that depends on Cloudflare, Mailchimp, Postmark, Isso ...
Twenty years ago any self-respecting dev would have run the equivalent of all that themselves on their own metal. In 2026 elite neckbeard practice is write the "never reach mastery, because they no longer make choices, they no longer take responsibility" post, hit "publish" and it's magically deployed around the global internet for you. Like a child.
In the future we will see more and more companies proudly boasting their “NO-AI” policy as a competitive advantage. And they will be right.
Yes, a "NO-CLOUD" policy was already so popular, surely this will happen too.
https://builtwith.com/alexn.org
If you have ever looked at a kite and said to yourself, "that is as good as a hang-glider. maybe better.", you might be subject to the perils of one-shot AISDLC
As much as I agree with this article, I feel that there's a logical flaw here. The author admits that it's difficult to measure bad code - but continues with the assertion that it exists. If the only negative to "bad code" is that it's difficult to maintain once the author has left or difficult to refactor, then the question really is whether or not LLMs will continue to be able to maintain their spaghetti code.
Just because it's bad for a human doesn't necessarily mean everything will fall apart - unless a human has to maintain it unaided.
I think this kind of blog post reflects a fear of infantilization by AI. By asserting that only humans know what good code looks like, the author is attempting to mend his own ego.
Build the systems around the code and let the agents do their work.
I think the key distinction is that it's difficult to determine whether a single commit or pull request is bad code at a glance, but it's easy to determine when the codebase as a whole has gone bad due to many low quality commits.
The latter often doesn't even require looking at the code, you can usually feel it just by using the software. From my experience, all software primarily written by AI is full of little bugs and inconsistencies that reflect bad code architecture (such as two very similar pieces of functionality in two different places behaving in wildly different ways, due to the AI being unaware of the first when asked to implement the second and writing the code twice)
I totally agree, and I agree with the article in general. I just don't think it follows that code that's bad because humans had to maintain it is necessarily bad by definition.
I've vibecoded loads of AI apps for myself and almost none of them are still in use. Not because I didn't really want them or they didn't work, but because the more I used them the dirtier I felt, as though I could feel the bad decisions and the bugs underneath just by interacting with it.
(In the future we will see more and more companies proudly boasting their “NO-AI” policy as a competitive advantage. And they will be right.)
dumbest take ever. AI is here and not going away, any company that does so will not survive or will be a niche thing for hippies.
It is not going away, just like OOP has not gone away..
I think this is a form of path dependency. We talk about unmaintainable messes, but honestly, if you look at Kairosoft or famous game codebases, you'll find 30,000 lines in a single file or just completely chaotic code. I am actually in a position where I see bad code all the time.
The truth is, "good code" is relative. It is determined by the specific domain and the composition of the team. Is incomprehensible FP (Functional Programming) code good? No, it isn't. A programmer must assess the team's capabilities and adapt accordingly. Good code is ultimately something that morphs based on the shape of the organization. Once defined this way, good code might share certain commonalities (like readability or a shared mental model), but its actual form varies wildly.
So, what is good code? That definition is missing. To be blunt, the Hacker News posts insisting that we must write "good code" are essentially a form of self-hypnosis.
Just look at paradigms. The mechanics of OOP have changed significantly, FP approaches have evolved, and DOD or DDD are fundamentally different from their early days. Whenever paradigms are discussed, someone claims, "That problem was solved in the past, and nowadays we do X," only for someone else to reply, "I don't think that's actually solved," leading to a fragmented breakdown in consensus. Ultimately, which knowledge remains as tacit knowledge is entirely dependent on the organization's capability.
You could argue that AI is terrible at simplifying code. However, I am skeptical that AI coding needs to be identical to human coding. When you actually code with AI, it often produces structures humans would call anti-patterns, including God Objects. Yet some of those structures can be faster or simpler for machines to navigate. There is no reason to assume that the optimal modularity for AI maintainers must be identical to the optimal modularity for human maintainers.
Of course, I am not denying that the rewards of good architecture are delayed, or that there comes a point where maintenance becomes impossible. But as the AI era ushers in an age of overproduction, software could become disposable, strictly personal, highly tailored to small niches, or ultimately, heavily polarized.
Realistically, programming domains fall into two major categories: "ship it and forget it" (one-offs) and continuous services. I agree with the OP's point that AI struggles to understand boundary delineations. But honestly, you can enforce those boundaries by injecting them into the spec. How those boundaries are drawn in the first place, however, is purely a matter of personal experience.
Personally, I define "good code" as code that allows the entity responsible for the software to achieve its purpose with a sufficiently low cost and error rate, factoring in the software's expected lifespan and future changes.
If you ask an AI to generate work based on this standard of what level of code is "adequate," you might get entirely different results. The biggest problem with discussions around AI is not just that ideological identities prevent proper evaluation (as seen in that article), but that the AI itself scales proportionally to its input. It is an incredibly difficult issue to judge because you don't know an individual's workflow or exactly how they are utilizing the tool.
I do think the value of reading code is important. However, much of what we are discussing in the AI era is actually rooted in the path dependency of how to become a good human senior developer.
Instead, the core focus of AI-driven development might shift toward defining broader abstractions: data semantics, invariant external contracts, and migration strategies.
Ultimately, I believe the paradigm shift of our era should lead us to ask: "How do we write code most economically in a system where AI is the primary maintainer?" The OP might think differently, but at least, that is where I stand.
i wouldn't know a single engineer who'd want to work at a place like that
I would.
Nothing new to read here..
I still can't believe that anyone who has used AI and understands development thinks that "No AI" policies are going to exist anywhere but the smallest of niches. Maybe they've been blessed to only ever work with the top 1% of the industry?
At this point AI is a better developer than most mid-level SMEs I've worked with. I hate this fact, but I don't have a shred of evidence to dispute it anymore. I work on a 20 year old codebase that thousands of developers use and it finds bugs in pre-AI code daily.
The LLMs are great for reviews. The No AI" policies would be probably restricted to coding..
Indeed - it's great at finding mistakes.
IMO reviews are more useful for communicating some change to the rest of your team so they can maintain it later than as bug finders. AI is better as a bug finder.
I believe many of us here knows that the idea of a simple prompt to make something more than a sketch or a prototype don’t really work.
Unless you steer and understand what an LLM will produce, you will end up with something that possible ”works” that has no future plans baked in. Suno generated music has a very unpleasant feeling of sounding like competent music with nothing to say.
I’d say that vibecoded software is similar. My speculation is that current breed of LLMs do not have an I, and I really don’t exactly knows what goes on in those vast arrays of numbers. There’s something there perhaps, but no person.
Still even in the short term someone wants to run a company that expects responsibility of its organisation, how are you going to exact that responsibility if no one actually understands how the thing the organisation makes works.
Maybe a simple crud system can be made fast and loose. But a bank settlement? A pacemaker? Deletion of sensitive data?
I know some companies are betting on that the agent can fix what the agent breaks. It may be true, but up until now everytime I try to relax on strict steering of an agent it tends to go badly rather fast.
Again I don’t know, but I think as long as we don’t invent synthetic persons with their own ideas on what they want to do, which btw opens a massive can of worms, the current situation will persist. However clever the current breeds of systems are.
I do want to state that a find the current trajectory fascinating. I use LLMs daily, it expands the number solutions I can explore. But in order to make something I feel is mine. There’s a choice and the buck stops with me.
Imagine AI making your favorite book or film. Could it be prompted into making Lord of the Rings or whatever it is that you enjoy? I think those who only watched the film might say yes :-) Those who love the book might say no.
There’s a few things on this. I’ve read the trilogy many many times during my life. Every reread I found something new, something I hadn’t considered before. Still it was eerie to see Peter Jackson’s rendition because it was very close to how I picture the book in my mind.
So, I think if you know just a little bit of film making and writing and LLMs you know you could not prompt any of them into existence with just saying ”write lord of the rings”
But if you have an idea for a creative project, using an LLM to explore ideas can definitely be a fruitful endeavour.
The gods don’t give gifts by ghost goblins cannot be dismissed as slop. I’ve personally used LLM to explore large sets of photographs to use as a backdrop for a DJ set, essentially to have the set tell a story.
So, as most things, it’s complicated. But at the same time I’m curious what will happen next
I think this is a bad take... if you are going that long without noticing, you are (hopefully) delivering end user value all that time. That is the main driver of code. You can normally dig/hack/rearchitect your way out of an ugly code situation. If you've been building value for the last year based on the hacky code, that's a win.
Hard disagree. Usually I am wiser than the AIs but there have been times where the AI has pushed back and made me see the light on some poor design I was about to pursue
Sometimes, a rubber duck is good enough. But sometimes, you can benefit a lot from a rubber duck that can say "actually, your entire line of reasoning is wrong". Even if the latter is more frustrating.
Its not a rubber duck anymore when it talks back to you.
I think the issue here is the nebulous definition of vibecoding.
You, as an experienced engineer, are doing a lot of hand holding and review of LLM-generated output, maybe even(?) using it as purely a check on your own work. There are others, though, that are essentially outsourcing the entire process to a basic, underspecified chat prompt.
Honestly, I'm not clear where the line is as far as what is reasonable, and it seems to me that no one is really addressing that question. It's easy to wave your hands and say "vibecoding is bad", without really defining what vibecoding is.
I suppose it'll be a few years before we see the true volume of technical debt catch up with the worst offenders, but even then, without the original LLM conversations associated with it, it'll be difficult to assess that in a structured way. Mind you, that doesn't even address the ever-improving models.
It's a mix. There are some projects where I let the LLM mostly run free and I focus on architecture. Even through some scary parts. Recently I picked a parallelization scheme for this and let the LLM cook. I have no idea how it implemented it, except that it (supposedly) "does it the way I asked:
https://github.com/ityonemo/bpa
For my projects that I use day to day in prod there is considerably more hand-holding and code review.
No. Most rules are there before and will stay after the expert enter the field within its career path, unless the field is bright new territory no one fooled before, which is rare.
Not only experts have to know the rules, otherwise they wouldn’t be expert, but good experts also ideally know why the rules were set, and at least have a fairly well aligned representation of what it would likely lead to to follow or not each rule in this or that situation, which rules are in conflicts and what the tradeoffs are when favoring one on the other.
Everything is context dependant, yes. And LLMs can help to leverage on far wider contexts that a single individual would be able to do on its own. It’s require interest to reach some goal in some social context, and not everyone will use LLMs with the same creativity.
This week-end I was discussing with a friend about our respective use of LLMs. At some point they told me they no longer read the MR, as LLMs can also do great job on that matter now, which is in sharp contrast with what I do. Not that I don’t auto-review with LLMs, but I use that as a first step, be it mine or some other colleague. And then I ask an LLM to prepare me a reading plan to check the MR, taking into account the activity scope and the implementation architecture. To me it was an obvious way to go, but for them it was something they never considered. That’s random sample, of course they would certainly be cases we would switch the "I wouldn’t have thought about it" role.
So yes, LLMs can be used to produce faster giant piles of unmaintainable codebases. Or they can be used to strengthen processes that lead to code quality. The nail won’t prevent us to knock our thumb, to use a nail in reverse side, or to smash our coworker.
I get the paperclip plant issue, but that’s some extreme scenario (which is of course the point of the allegory), and most bad uses will be far more mundane in how they look and what the consequences are. And most good uses will look more and more transparent to users to the point they won’t even wonder about it at each use. Like, most people open the tap, see water fall and they don’t get a sense of wonder, not giving a thought to this masterpiece of engineering and gratitude for all people that works daily in the shadow for this miracle to happen. They don’t leave the toilets thinking "how freaking amazing such a complex system of wastewater is something I can benefit from everyday, unlike so many other of the 100G humans that walked this earth."
Next time you go to WC or tap some water, think about it.
There are always new people to believe in the rules without question so rules survive.
Not all rules survive by this fact alone, which obviously is as true as the fact that there are people that don’t care about any rule and people that will deliberately break rule for the mere sake of the thrill. Rules can also stop being used, they can keep going on charming sufficient fan base willing to fight for their application in an environment which changed upside down and were every other new rules are directly conflicting with this old one.
The problem imo is the slow deterioration of institutional knowledge that offloading the mental task of wisdom gathering to AI is causing.
One interesting comparison is to the history of manufacturing. West/America decided one day that manufacturing would be cheaper to outsource and better (short term) profit was to be made by outsourcing it all to China. The institutional expertise started to deteriorate, to the point that America simply didn't even have the capacity, or expertise anymore to produce stuff (such as grill brush [1])
I feel like you could take all the handwavy comment that are made today to dismiss this caution, and find equal dismissal back then when companies were actively outsourcing the manufacturing.
"I'm coding 10x faster" "look at the output velocity per employee"
"we are producing much more (in China)" "look at profit / number of (manufacturing) employers"
Seems ok if you're American / Chinese but I'm struggling to understand how the rest can be OK with allowing institutional knowledge to deteriorate while having an active dependency to the former two. We already see this with the tech dependency towards USA and manufacturing competition from China.
[1] https://youtu.be/3ZTGwcHQfLY
It is not unsurprising to find dev lacking knowledge in domain stuff that would be actually required if they were doing the development themselves, rather than outsource it to an LLM.
So the client who writes the ticket understand the domain, the LLM that implement it understand it too.
The dev is the only one that is clueless.
There is more to development than domain knowledge. There are decades of wisdom about development best practices on how to keep a large codebase maintainable, performant, testable, etc. that LLMs frequently don't follow. And now humans aren't practicing those skills either or passing down lessons learned to the next generation of programmers.
Maybe it will be no big deal, and nobody will read code anymore, but it is understandable why somebody might be concerned about it.
Without domain knowledge, the dev will miss critical simplifications. That is one way the code base accrue complexity.
The LLM that implements the ticket understands the domain? I'm rather skeptical of that claim...
The outsourcing of manufacturing was largely influenced by the externality of environmental pollution. People wanted to live in clean habitats. We had rivers on fire for weeks, extinction of species, and horrible pollution causing cancer and birth defects. Giving "the commies" the dirty, dangerous jobs, and moving everyone up into middle management and white collar knowledge work was seen as an evolution of the economy.
The political move to put all of that death and destruction onto China, where environmental regulation is willfully ignored in the interest of economics, was a smart move on their end.
We lost so much intellectual knowledge and other "tribal" technology in these processes to this outsourcing, which is unfortunate. We're almost having to rebuild our manufacturing from first principles, which may not be a bad thing.
You don't have to completely abolish manufacturing to clean up pollution better.
it wasn't completely abolished. but the economics forced out the low-value-added work, only high-value manufacturing (and politically important manufacturing) remained.
No, you don't, and I am not saying I feel that was reasonable, just that it was an emotional appeal used to sell it to the layfolk.
Clean Air Act was passed in 1963. the Clean Water Act was passed in 1972. The China Shock began in 2000 after PNTR.
Avoiding environmental regulation was one reason to move factories to China, avoiding unions and high wages was another.
But the local communities weren't demanding that their factories be moved overseas so they could have a clean if impoverished towns. Your entire causality is completely backwards.
It isn't 100% causality, but it was definitely one of the ways it was sold to people through PR/propaganda.
Free trade / globalism was the larger "true" move, and yes getting away from unions and high wages to maximize profits was a big part.
Corporations made a change from valuing stakeholder value (employees, communities, customers, suppliers, the nation) to pure shareholder value (profits over everything else, despite the long term result that profits by any and all means hurts all stakeholders and eventually can cannibalize the company unless you have monopolistic moats).
I promise you the people in the towns where the factories were closing were 100% against it. And the corporations never tried to claim that "at least your town will be cleaner" probably because they were mostly leaving behind superfund sites they were trying to avoid cleaning up.
Was the New Yorker crowd clinking glasses together talking about how good it was to get the dirty steel mills out of the country? Sure maybe but this argument never resonated with anyone on the ground.
Were companies actively moving factories offshore so they could be dirtier because they didn't have to comply with American environmental laws? Absolutely.
In Europe we lead the regulation :)
De-tracked video link: https://youtu.be/3ZTGwcHQfLY
Do not share videos with si parameters. It links together the accounts of the sender and receiver.
Thanks. Edited
I recommend https://sharecleaner.com/ for cleaning links of trackers when sharing from mobile. No affiliation, just a user on iOS.
@mitxela: Thank to for taking a moment of your time to educate those who didn't know.
Absolutely people were extremely dismissive to anyone saying that we're losing the ability to make things in this country!
There were all these theories like Comparative Advantage that people would trot out to point out that, if you don't like outsourcing, not only are you ignorant and backwards you're also probably racist.
I specifically remember seeing the ricardian tables, featuring 8th grade math proving how competitive advantage means everyone wins. This was very persuasive to me as a college sophomore, all my professors were on board and I was young.
I guess we were all lacking wisdom.
I remember being a young guy, blue collar background, but in the left-wing anti-globalization movement (think Seattle WTO protester types).
When I was in college or going to parties with New Yorker Reader types I'd try to argue against globalization and they would start smugly dropping their theories on me. I was too young and naive to refute them in any kind of convincing way, but my life experience told me that it was wrong.
But it was pervasive in the zeitgeist. Basically anyone trying to argue against it was backwards and stupid, or shrill if they were lefty.
I wonder if the exact date of anti-globalization switching from being a "left populism" to "right populism" thing could be determined.
However I'm not sure it could really be stopped or whether it was inevitable from containerization. Can you really prevent people from getting cheaper products forever?
Again containerization happened in the 1970s and 1980s. The China Shock was 2000-2010.
What technology was invented in 2000 that allowed everything to go to China? No technology. That was the year the US adopted Permanent Normalized Trade Relations with China. It was a government policy change!
By the way the US already had trade relations with China. All PNTR did was promise not to change the policy in the future. That was the starting gun that signaled to all American companies that they should start investing in production in China.
So would globalization happen on some scale just because of the march of technology? Sure. But the way it happened, the speed at which it happened and the extent to which it happened, were the result of policy choices.
It was so very weird. Even in Ireland (which was the target of the first wave of US outsourcing), people seemed to believe that everyone could be "knowledge workers" and that industry was unimportant.
I still don't really get why everyone believed this, but they did. Something something Upton Sinclair I guess.
And they were right. American manufacturing is $3 trillion/year, second only to China, and 1/5th of the entire world. American built airliners and American built engines fly all over the world. Millions of American cars come out of factories every year. American rockets put more into orbit than everyone else combined.
The idea that the US lost its manufacturing is not based in reality. One reason people think that is because people think cheap plastic crap and consumer electronics when they think “manufacturing.” Another reason is that US manufacturing has become highly efficient and automated, so a fairly small portion of the population works in it.
Yeah, your iPhone wasn’t built in the US. But the plane that got it here probably was.
You're arguing that American manufacturing is healthy because of Boeing?
Among others, yes. Nearly $100 billion in revenue last year. From your snark I guess we’re not supposed to count a manufacturer with revenue equaling the GDP of a small country because they’ve had some quality issues?
If you can’t see the direction and velocity of China you’re beyond hope
Americans cars are falling behind, they rely heavily on American cultural exports which relies on the world not thinking Americas are twats because the voted for Trump. Germany alone is over twice the size of America in terms of car exports. Including domestic destinations then china is three times the size of America.
America planes fall out of the sky far too often as Boeing pivoted from engineering to politics
Musk does good rockets. Not America. China is catching up and lost to overtake unless starship proves itself; and if it does China will copy that too.
Chinese planes are increasing very year, comac now about 15% of traffic, and heading towards about 20% in a decades time. and Boeing is about the same size as airbus at 40% each which will shrink as china catches up, won’t surprise me in 2050 is the largest provider of passenger planes is Chinese.
And do you know where the tools and dies to make those things come from? Where the robots that work at those lines come from? Where the chips in those robots at those lines come from?
Hint: not here. Even basic machining is cooked here. Go on Xometry and compare the pricing of any simple design made in USA vs. China and you’ll see, we can’t competitively do the basics anymore (you conflate basic with crap).
This also applies to the UK: major manufacturer of turbines, i.e. Rolls Royce in Derbyshire. More electronics than you'd expect, but all short-run highly specialized things like satellites and scientific instruments and Martin-Baker ejection seats.
Massive numbers of people on production lines is not coming back. The rest of the economy is too well paid for that and automation too advanced. VW will retain car manufacturing in Europe by automating away its workforce, at the cost of high redundancies.
What many people don't realize is why is America suddenly caring about manufacturing everything inhouse when Elites decided earlier Globalization was good, because China wants to command respect more than America. That is also the reason why Gold is pushing higher. If China was nobody, no nation would care for gold. So all these unnecessary downstream reasons distort what is actually going on.
If there was no China, Globalization would spread very strongly with Bangladesh manufacturing clothes and Milan doing the fashion show. All those calculations are coming undone. So please keep this main reason front and center when we discuss this matters.
How come we cannot keep/define "specs" of how to manufacture such "mundane" objects such as a grill brush, everything from what raw materials are required, how should the assembly line(s) look like, which industrial robots or their human professional alternatives are required, etc. etc.
It just seems to me that fundamentally this is a knowledge organization problem.
Grill brush found out that the problem is chinese products are so much cheaper (without ascribing cause of that) that north american producers can only compete on native advertising (B2C) or purchasers that require a north american supply chain (B2B).
I think you missed their conclusions. The problem wasn't that China was a lot cheaper. The problem was that substantial subcomponent or process vendors only exist in china. Like, they could find American sheet metal stamping companies but they were all using custom tooling made in China because American tooling doesn't exist anymore.
Tooling is a high-skill trade. America used to be amazing at tooling. What's more, tooling isn't super cost sensitive because one tool can make thousands of parts.
When outsourcing to China began the tools would be made in the USA and shipped to China for the low-skill work. But over time those Chinese manufacturers figured out the tooling. What's more they realized that controlling the tooling would let them control the whole process.
They started doing things like, for instance, including the tooling in the price of the product so you don't even see it, if you use their tools. Just send them the cad file and they'll do the rest. So American tool making went away. But this was a conscious decision on the part of the Chinese manufacturer and an unconscious decision on the part of American importers who didn't really value their in-house expertise.
To abuse an old phrase about engineers and bridges:
“Any idiot can run a business, but it takes an MBA to run a business that barely functions.”
I don't see how that is relevant here. And no, not "any idiot" can run a business.
This analogy really only works assuming that maintaining [insert company]'s institutional expertise requires N engineers to maintain individual lines of code manually.
This is already showing not to be true with AI. Institutional knowledge is not the same as knowing how to implement low-level software details. Even today, every company does not need engineers to remember git cli syntax by memory, how to write parsers for JSON, or write the large amount of boilerplate from scratch that is at every software company. Most company's already hire engineers who have zero experience in the existing code base, yet they are productive despite this lack of institutional knowledge. For a company to maintain institutional knowledge they may only need N/K engineers.
This breaks down at a sectorial level. If one company offloads that knowledge to the wider economy/world, fine. But if <<all> the companies in a domain/sector do it at the same time, at some point the destruction is extremely hard to undo.
We were already there at places with a lot of contractors, capped at 1 or 2 year stays. You saw teams of 10, 20, with 2 people with serious institutional knowledge. Now the contractors are just Claude
What really doesn't make sense to me is setting uniform token targets for all engineers. I'd be more interested in setting token ranges with hard limits for certain roles that see diminishing returns from the overuse of coding agents and where maintaining institutional knowledge, architectural awareness... is more valuable than moving as fast as possible. Like everything else, these limits should be adjusted over time as tooling for those roles improves but it really feels like there are human limits that organizations should avoid exceeding.
Americans produce the highest technology equipment in the world. Our machining base is structurally sound, but its all making weapons, so you don't hear about it.
It's true that China benefitted immensely from outsourcing, but they took the jobs Americans didn't want. It's the same with immigration today - folks cross the Rio Grande to do chores that Americans won't, or others fly in and work for nothing in academia while they wait for their PhD.
Have you considered the institutional knowledge could actually be actively preserved and distributed with AI? The kind of tacit knowledge that is situated and not readily preserved in a book might be absorbed by thinking machines and proliferated to the next person who needs it. The caveats would be trade secrets, skill differentiators, that people might not be willing to discuss, and manufacturing secrets of national importance. Maybe you can think of others.
Well that's what this whole debate comes down to. And, to my mind at least, it's a rare case of an actually interesting question about AI, because like the article says, it can plausibly deteriorate exactly that kind of knowledge. But as you note, it can also maintain it.
I think there's a sense in which it might do both at the same time. AI's version of on call tacit knowledge might be something like lazy-loading just-in-time tacit knowledge, but at the cost of who knows what cognitive paths we might have by keeping that knowledge resting in-house. We would gain real efficiency but we wouldn't know we wouldn't know.
It's funny that you credit AI with this (and I don't disagree), because tacit knowledge was exactly the thing Hubert Dreyfus spent a career insisting computers would never have, and his wisdom was taught to generations of undergraduates across the country and world who treated it as received wisdom and still is regarded as such in certain academic corners.
This is probably true.
Many artisan fabric techniques were lost when industrial looms displaced those jobs. But heavy industry enabled cloth to be manufactured at a rate that people were free to spend their time and money on other priorities - and now (centuries later) the society is in a sufficiently advanced stage of development that the old techniques are being rediscovered.
Perhaps the example underscores the importance of thoughtful preservation of insitutional knowledge.
Could you elaborate on this? It sounds like something any anthropologist would laugh at; there's always an oral culture.
I edited that sentence to try and clarify a bit, but Dreyfus was a continental philosopher out of Berkley active basically from the 1950s to the 2010s. He engaged in decades of friendly debates with other philosophers about the limits of computers.
When he described those limits, he always frustrated computer scientists and analytic philosophers because he spoke in a kind of informal philosophical vocabulary and didn't really formalize his ideas. So he would say computers didn't have things like "tacit knowledge" or "insight" and he railed against "symbol manipulation". Famously he declared chess would never surpass human expert play. He wrote a book called "What Computers Can't Do" and another called "What Computer's Still Can't Do".
I personally think his argument was laughably wrong if well intentioned. But some people think it was respectable. I think in the present day, he's often rehabilitated with a kind of apologetic reinterpretation, such that things like transformers, weights, vectors, etc were what he really meant all along.
I think he was not wrong that some higher layer of sophistication would prove to be necessary, but he was wrong, I think definitively, to think that "symbol manipulation" of computers was a kind of category error. Even today's best models are still running on logic gates over 1's and 0's, and it was his failure of imagination to doubt that those could be the conceptual bedrock for AI, tacit knowledge and all.
He passed away in 2017, which is too bad because I would have loved to have seen his interpretation of things like GPT-6 Astra.
It's also very automated so the jobs went away, but the US continues to manufactures a lot of things.
They took jobs that Americans would have wanted higher pay to do, higher benefits, higher safety standards...
You don't develop "institutional knowledge" from a few prompts. It takes years to develop them.
Right. And the spin I would put on your point: while the U.S. could never compete with outsourced labor, institutional manufacturing expertise could (you would think) still be valuable to startups finding some kind of niche in the manufacturing space.
With all this new 3D printing infrastructure, and constantly improving robotics, maybe some manufacturing efficiencies in some spaces could emerge that compete on cost and outcompete the cost of shipping stuff across an ocean. The availability of institutional expertise could be one of the necessary ingredients for mixing and matching the way to a new efficiency.
Yeah it's bleak in terms of automating away labor, but I'd like to think robotics is the next automated frontier after LLMs and we want to get there first.
Peace through global interdependence wasn't a bad side effect at all. The analogy only holds so much, but I'd wholly agree that any profound deterioration of skills across the board would create dependencies on AI infra that are on a whole new level.
At the very latest, when "AGI robots" will be commonplace and taking on the most crucial labour on our behalf.
I believe this is what Socrates said (verbatim) about the invention of writing.
Be fair to Socrates though.
We have significantly degraded in our ability to memorize e.g. epic poems or preserve oral traditions since the advent of writing.
I know that sounds trite in retrospect, given the benefits, and that's half the point. But there are real tradeoffs too! For example, culture has a LOT more generation-to-generation turnover as a result of literacy being common, making it unstable.
I think (?) my point was that Socrates equated Memory with Wisdom, and the invention of writing would impact Memory and thus Wisdom. I believe he was just plain wrong. Equating (manual) coding/code-review with Wisdom makes the same category error.
Was Socrates even wrong, though? There have certainly been millions of words written that were lacking in wisdom and have convinced people of falsehoods to their detriment! When’s the last time you saw a debate be reasonable? devoid of cheap jokes or hot takes? actually change someone’s mind? or the whole audience’s mind?
In the world where we didn’t outsource Western manufacturing to China, our homes would not be overflowing with disposable junk; in the world where the Luddites won, our wardrobes would not be stuffed full of disposable clothes; perhaps in the world where Socrates won our minds would not be jammed full of nonsense.
It's not really a case of 'winning' though is it? There is no agreed criteria by which to judge which is going to pass muster for a majority of people (both in Socrates time and ours).
One affects memorization and is a trade-off for writing, a process that arguably allows someone to work through their own knowledge of a subject and find or realize gaps in understanding.
Having something else do all of that for you is completely different and essentially tossing any gain outside of the creation of some gray slurry of an output in the trash.
To build a $100M software company, you need 6 engineers and 6 laptops.
To build a $100M hardware company, you need 60 engineers and $100M.
Everyone decided on the most logical choice. It gets even worse if you jump into profit margins, as software is the unambiguous winner there too.
Until software becomes commodity and generated on the fly by large language models.
That until was like 50 years though -- an entire career span for those who started at the start.
I think this is a good point but we should see AI at the person level. For your example, not all companies/people manufacture, but everyone has access to AI. Soon a lot of people will lose skills to AI and at some point we would need them back and there won't be many skilled workers with it
I've been taking a "help me help myself" approach to interactions with LLMs. I might run my plan on how I'm going to build something past an AI asking it to critique it. I then incorporate that info into my original plan, or not. However, I am responsible for the design, I write the code. I started doing this because for decades I've been building my own programs since I learned to code just for the sheer fun of it (like many here) and I don't want to give that up to LLMs. I might as well let them play my video games for me.
I wish more people would at least consider this idea.
That's the whole point of Global managerial class even without AI. The idea is just by creating metrics one can measure how teams, groups, organization and whole industries are performing.
Before the current AI complex the previous one Agile Industry Complex One can see we starting producing 10x more code, 50x more JIRAs resolved, 100x more network bandwidth used. All these metrics tremendous gain in productivity.
Also it may not have been seen in US but in many places a change in political regime does want to burn down previously collected institutional knowledge because new dispensation have their own idea about what is knowledge, who will preserve it and how it will be preserved.
It is definitely seen in US.
Bogus nonsense.
1. Manufacturing output in West/America is higher than it ever was in history.
2. What has reduced over time is the number of people employed in manufacturing, not manufacturing output. And that's an effect of automation, same as it happened over centuries in agriculture.
3. West/America did not decide anything, the realities of capitalism did.
You're either a capitalist and embrace efficiency of producing where it makes sense (for many different reasons) or you're fine with inefficient third-tier industries.
The tax payer may help a handful of sectors that would otherwise die survive if they are truly critical for national security, but just a handful, not all of them.
Most of us can't hunt either. Somethings go away in time because they're no longer needed (by the masses). That is essentially what society does -- it allows us to multiply a few for the benefit of many so the many can focus on other things.
The resilience is more in ensuring archival process to ensure it's never lost forever, and a continuity plan to ensure enough (but few) people still know.
It's appalling that we still hold on to such things that are no longer necessary. Code maintainability is not a problem when you don't have to open a file and inspect how something works anymore. You use english to add to it. You sit on chairs everyday where you don't give a shit how they were created. They fulfill their purpose. hopefully the same can be said for your software.
What exactly is appalling? Are you referring to something in the sentence you quoted?
I am dumb on chair making, as the average chair user. I wouldn't trust myself building a commercial-grade chair.
If you are as dumb with software as I am with chair making, you should refrain from building software for others. This is regardless of LLM.
Even if you never read/write code anymore you still need to care. LLMs also suffer from a bad code. LLMs very quickly lose track of their own shit and start producing more bugs.
Hilarious and unhinged junior dev and/or outsourced dev and/or expert beginner take here.
Not everything is baby's first React app for an internal business or B2B SaaS startup with 3 users. Sometimes your software is actually used by people, and bugs happen, and you need to figure out why bugs happen, and quickly. You do this by reading the code. Yes, AI is very helpful with this--sometimes. But even SOTA agents cannot solve every bug, particularly when the person directing them has no clue what they're doing, as in your case, so they aren't given good constraints or starting points.
So have you never had a problem the LLM couldn't solve? What do you do in that scenario?
How do you guide the LLM away from making a horrible choice if you don't understand the internals?
You can still read the code, ask the AI questions about it, and ask it to fix things. Coding agents are a power tool for cleaning up codebases, but you have to ask for the right cleanups, which means figuring out what it is based on your understanding. (You can also ask the AI to suggest ideas.)
This takes time away from implementing new features, but that’s true of all code health maintenance.
This only works in small projects. For large projects, it is close to impossible. Everybody talks about how new models appear all the time and nobody comments on the fact that context size has almost stalled.
Can't everything be broken down into smaller pieces? And isn't that actually a good principle in general?
Sure. But in order to do central refactoring and actually improve stuff you need to keep in context all the small pieces.
You can split a small program into piece A, B, C All of them look correct on their own. But they duplicate something in 3 different ways and person/agent who can "see" all of them can see the duplication and refactor.
Current model context is simply not enough for large projects.
Same problem for letting AI review code. A PR might look correct on its own and be small enough to fit into context. But somebody who has access to the whole code of the project again sees the duplication.
I am an OSS developer and when reviewing PRs I actually look at how the same problem was solved in other popular OSS projects. No AI can check this today because there is simply not enough context.
Basically if we had unlimited context what you said might be true. But context size is limited today.
I suppose it's not enough to break everything down into smaller pieces, your architecture also needs to be composable. That would address the issue of duplication. Think a plugin system.
As far as maintainability is concerned, have we all forgotten the 12 Factor App in the age of AI?
https://12factor.net/ https://en.wikipedia.org/wiki/Twelve-Factor_App_methodology
Kevin Hoffman expanded on that with the 15-factor app: https://developer.ibm.com/articles/15-factor-applications/
Mind you I may be dating myself as I was first introduced to this paradigm in 2015 working as a Java SpringBoot engineer on an enterprise project that I then migrated (57 microservices) all to Scala, after onboarding two weeks to Scala fresh from no prior Java experience.
I feel like there is so much "wisdom" encoded in books and writings from some of the most prolific engineers and architects over the last several decades.
Look at Matt Pocock's skills with simple primitives like grilling the human, researching through wayfinder maps (a Godsend to my workflow prior to Cursor Projects and orchestrator patterns), and having a solid Domain Driven Design through defining a shared glossary and breaking up work around proper seams.
12 factor is good but it is a very low bar.
It is perfectly possible to vibe-code a badly designed app that still passes those 12, 15 or whatever points you define.
Could you share some higher bars that may make up a better "rubric" for this issue? I am actively trying to do so. Shy of just condensing core Manning publications that cover domains of interest, I am struggling to find a good bar to have my clankers validate against outside of minimizing cyclomatic complexity.
I do not think, the intention in using AI, will ever be gaining wisdom. It is rather moving computations to a larger computing system.
Source?
False. AI is evolving by leaps and bounds and the outputs are better and better by the day
Personally, repeating what the article says, I haven't coded since may 2026, not a single line, and AI has been delivering exceptional results, improving by the day
Any AI related article necessarily needs to consider future improvements as they will come not by surprise but by steady refinements, and posts like this will look just like the same early AI slop they complain about
Coding IS solved
Maybe I've just given up, or maybe I'm a realist? But I fully believe AI will just...catch up with everything?
There's so much money in it right now. There's such a momentum. There are zero incentives to slow down for those that are in charge.
I've accepted that in 5-10 years, the vast majority of human devs. and engineers will not touch a single line of code. It'll be small increments, with a couple of big ones here and there.
And there will not be any triumph for those that hold steadfast to the principle of human coding. They'll be tiny boutique shops that do custom stuff, in the same way cobblers are to the mega shoe factories.
We've been able to make watches with machines for 50 years. Handmade Watchmakers, and their producuts are more valuable and in demand than any guy running a conveyer belt. They still make 20x more if not 50x. People look up to them.
You can make a choice not to become a button pusher and still do things by hand. You dont have to fry your brain. You're falling for a massive trap to strip you of your value.
the difference is that code is not something the average person will marvel at over the intricate hand built details. You might get a few nerds interested if it's something super novel (like rewriting a major modern web app without any 3rd party libraries that is 30x more performant), but most people dont care. Extra true for businesses, I can't think of a business other than maybe a mom&pop shop or a place where showing wealth is important where they wouldnt happily issue casio watches far before they would consider a rolex.
Regular businesses are not going to build their own slop software because of AI, this is delusional. People will be hired because they love to build. Only a small percentage of the world loves to build things.
The most passionate people with the best reputation will suceed, they will continue to be in the most demand and will be the most valuable, like a swiss watchmaker.
Sorry LLMs arent replacing anyone who's got strong skills and cultivates them. You're so wrong here. These more passionate builders will however replace the lazy people who offload all their skill and brain capability to llms, and they'll use AI to help them.
Being lazy and letting llms do everything for you is not a good strategy.
Yeah, I have a really hard time buying the "AI code isn't maintainable long term" because if you look at the quality increase from January of this year to now, it's insane. We've gone from Opus 4.6 and GPT 5.4 to Fable 5.1 and GPT 6 in less than a year, where we went from 0 mathematics problems being solved by AI to a Millennium prize problem being solved in a few days of compute. Obviously not the same as coding skills, but it shows the growth in raw reasoning and abstraction.
I understand why people are resistant to this from an emotional perspective, but I really don't see a plateau in sight. RLVR is clearly still cooking and narrow RSI seems to be on the horizon.
But I'm also a realist. If the technology exists, it will be used to the maximum economical extent.
Why would companies employ human engineers then? What is the value addition to justify high human salaries. If AI is going to get so good (and I am not saying it won’t happen, that’s a separate debate), why can’t AI figure out the prompts itself?
I'm in the same boat. When I started thinking about LLM development as multi-year iterative training adjustments, it felt like the majority of shortcomings can be addressed.
The deeper issue is American business culture.
Everything is reactionary. The focus is always on short term profits. No one cares about long term effects --- until they start impacting short term profits.
This is an inherent vulnerability that is easily exploited by someone willing to engage in some simple, long term planning --- like China has already done with manufacturing.
For decades, the USA gladly shifted manufacturing to China without a second thought. Now, we lack the expertise to make our own and have no choice but to use China.
Replace "manufacturing" with "software" and "China" with "AI" and it's deja vu all over again.
I will very happily take the other side of this bet. Maybe if LLMs stayed as September 2026 LLMs for the next 20 years, I'd grant it's possible. But that's not what's going to happen.
Are LLMs getting much better at coding? The new models are arguably better at understanding context and coming up with a solution that works but they still write just as much hilariously bad or hacky code as the models did 2 years ago. I e.g. ask Fable 5.1 to make changes to my software and if it's something that's not a standard CRUD architecture on top of a well-specified library it will come up with pretty wild stuff, i.e. producing shell scripts with a "cat" command that contains a 1,000 line Golang program that it compiles on the fly to load some data, or a whole Javascript tool it crams into a string in a server-side web app to make some UI element work instead of figuring out how to do it "right". Another model wrote its own (buggy) YAML parser for reading translations and I had to figure out why it broke my whole pipeline (can't even say why it chose to write that instead of just reusing a standard library, maybe I told it before to be careful when adding untrusted dependencies so it interpreted it as writing everything itself). Seems these systems are always trained to produce the desired output but I think it's very difficult to get training data that teach the systems how to keep the code maintainable when making hundreds or thousands of such individual edits. Honestly not sure if that will get better fast, so far it seems not!
I guess there's no reason to believe these models can't be as smart as a great software architect / engineer or team of such people that build an elegant and maintainable software solution over many years together based on customer feedback, then again the models are appallingly bad at some forms of reasoning, I mean they will "understand" something once you make them aware of it like e.g. a flaw in the software architecture, but when asking them to audit the code and check for issues they will often have a blind spot to finding such problems. It's interesting, like they have very high ability but very little awareness or self-directed thinking outside of the prompts they receive.
In the beginning of August 2026, I started a hobby project: building a MongoDB-like database. I have 20 years of industry experience and a master's degree in computer science, so I followed a disciplined, spec-driven development model using Claude, Kiro, Qwen Coder, and Cursor.
The first version was built in about two weeks of part time work. Then I started exploring. I learned relational algebra, researched almost every kind of database, reworked the internals, built a small relational algebra layer, a query planner, and an executor, covering everything from the backend storage to the query language. I learned more in those two months than in the previous 20 years.
Did I care what code the agents wrote? No. I read zero lines of generated code. What I cared about was correctness, verified through tests, and the high-level product features. For the first time in my career, I acted as a senior product manager, steering the project along the right roadmap. Without AI, I wouldn't have been able to do that.
When you have superpowers in your hands, you don't need to worry about the laundry. For the first time in my career, I can produce code in C, C++, Java, .NET, or any other language. Sometimes it takes me longer than a senior developer in that language, but does that really matter? Absolutely not. Writing documentation and code by hand in 2026 is like driving a horse and buggy. It doesn't matter how skilled you are with the reins; you'll never compete with a car. My hobby db project isnt opened source yet.
"My project isn't open sourced yet" underlines the problem. There has been a massive explosion of software that works, but badly. The average user experience suffers as a result.
Call me cynical, but I dont remember interacting with well engineered software, outside of pure, mainstream opensource repos.
Most of proprietary software is just crap, and always has been. The agents are not producing worse code than typical, demotivated, i-dont-care-what-i-am-building-i-wont-try-using-it corporate development teams have over the years. I'd even bet that because now making changes and fixes is so much easier, the user perceived quality will trend upwards for popular stuff.
I would bet every single dollar I have that if the source was shared, there would be 15 "big brains" showing up to point out how it's actually "bad".
No one wants your hobby db
Exactly. If you try to understand, really strive to understand what the AI is producing, then a lot of these arguments fall flat.
Okay, you're going to need to go into a lot more detail about exactly how you learned more in 2 months than you did in 20 years having read zero generated code.
Code examples are literally bread and butter when it comes to learning.
How can you learn without looking at code? That's like saying that you can learn to be an architect without looking at drawings...
I didnt know what is relational algebra is? why query plannering execution done in certain ways. What are backend storage system available in open source or commercial products. What are the traitts of each backend storage and how does that effects the replication and failover scenarios. I know all of this, I implemented a small subset of sql's relational algebra in my project and its correct and concise. I expored different backend storage systems, first tried one, then move to next and then 3rd in line, all that in span of less than a month without ever looking what code the coding agents generates.
To be as frank as possible--you "understand" these topics in the same way someone who watches pop-science videos about astronomy "understands" astronomy. That is to say, you have a surface level understanding at best.
This is the sort of behavior people have been talking about over the last few years. People confuse what the AI "understands" with what they understand, and it makes them say delusional stuff like "I learned more in the last 2 months than in 20 years."
To make this concrete, it takes someone about 4 months to go through a university course.
They might take six courses at a time.
So you are saying that you learned more than 400 times as much content as a university course?
I'm sorry but this just sounds hilarious And unlikely to most people, and it comes off a bit crazy sounding... For example, do you actually think you could even pass a single university course exam on introductory relational algebra without AI I assistance at all?
I think what is more likely is that you feel confident trying to solve problems with AI's help.
That is not learning.
To learn something, it has to be in your brain, not the AI.
Learning is not the same as becoming capable.
There's an ocean of difference between those two
You need to learn the concepts not implementation detail. If you know what is relational algebra and AI teaches you this in hours then you dont need to know the implementation details any longer. If you can borrow intelligence to solve a problem, you should do that. We, the most IT industry is already working in high level languages which is not how the computer works but abstractions over abstractions that very very few people know the entire cycle of a program how it runs. So AI is not a future abstracton, it is already here and its "intelligent" as opposed to other abstractions. Writing code or documentation by hand is gone, its past. All programming languages, framworks are implementation details and AI has already mastered them, we need to now move on to high goals than learning how to write a loop.
or how to be a doctor without knowing how to cut open a person and look at the internals. looking forward to the future doctors
Yet millions of people code react slop without ever touching a line of assembley
vibecoder 9/10 times, claim 10x improvement, project not open source
and 9/10 times project open source, project bad slop
We're in the era of centaur chess.
Those with 20 years of experience are crushing it.
My worry is not about us. The worry is about the kids. I'm tech lead. I don't think the juniors at my company are learning anything from me. I'm not certain they're learning anything about _software_. Idk, I could be wrong. We barely talk because everyone has become so siloed.
Not trying to come at you, but what is stopping you from educating the juniors at your company?
I am trying. The main issue is it would be extremely annoying for the juniors, who have work to get done, and prefer to throw problems at Claude rather than ask me questions.
In general, I believe learning usually needs to be driven by the learner.
Here is one my project is opened source. Highway https://github.com/System-D-AB/highway
It is fault tolerant, distributed broker which gurantees durable queues, pub/sub and RPC all into one easy to use programming model. The application is in production and passing millions of messages every day with sub-millisecond performance, you can crash a server and replica set invokes within seconds without losing any messages. The entire project is created in less than a month with part time working, just because of AI. Its in production and already proven.
I'm sorry, I just don't buy the hyperbole. Something with your previous approach was deeply wrong if you condensed 20 years of learning into 2 months.
Databases, programming languages, compilers, operating system internals, networks these things are pretty complicated in itself. With your day jobs you can't master any of this without spending years in your part time. But now at least you can start something of your interest and learn on the way. Learning a new concepts, implements it, test it and reproduce cycle is so fast that it is only possible because of AI. You need to find the process which works for you. AI already knows every programming language, its best practices, entire computer science knowledge is already mastered by current state of AI. There isn't any single database internals which Ai doesnt know. If you dont believe just start implementing JavaScript or C compiler yourself without even having knowledge of compilers and in a few weeks you will where you stand. You will not only having a functional compiler but you will know the internals of compiler constructions, all the concepts.
"When you have superpowers in your hands, you don't need to worry about the laundry" - love this
But why though?
lol
I also don't care the exact cpu instructions my compiler emits. In fact I have no idea what the "stuff" it emits means (beyond the most elementary)... That doesnt stop me from creating very useful software that maybe even billions of people rely on (once you include the users of products that use our software)
A bit offtopic, but this seems to be an AI pattern everywhere:
Similarly, for self-driving, the AI for driving learns from rulebooks meant for beginners and observes patterns of driving from ordinary drivers in the wild and let’s be honest, most ordinary drivers in the wild are pretty bad. The best the AI self drivers trained in this way can hope to get is the average slop driver minus the catastrophic errors.
Similarly, in legal specialties, there are a few highly expert and wise practitioners, and hordes of average practitioners, and quite a few near-malpractice practitioners, and it is the latter group who write the most stuff on the web because that is how they do marketing, trying to make their ignorance look better — and mostly louder — than what the average lay-person knows. The AIs can NOT tell the difference and 'learns', and dispenses both the good and the actively harmful advice. Acdg to relative who is a top expert in their specialty, AIs can both find key relationships between laws in complicated situations but also readily dispense the actively harmful advice, and you MUST already be a top expert to recognize which is which.
They are great averaging machines, but when average is not good enough, you must be on top of your game.
There is a cluster of users posting nonsensical AI-hate articles such as these.
Oh boy do I have some news for you about legacy codebases.
I’m not even sure that his axiom is true. A project started with a 2024/5 model can subsequently be worked on by more capable models (who don’t, unlike some humans, have a deep aversion to paying down technical debt). I’ve seldom seen a human written legacy application spontaneously acquire a better engineering team every six months.
Wisdom is very rare. I constantly have to tell people "we did it like that in 2012 because it seemed like a good idea and there were no better options - but this is 2026 and there are now better options and we shouldn't do things the old way instead.
Too often people get an idea and don't stop to ask if there is an even better idea. (I'm guilty of this myself). Often it takes a while to figure out what the good ideas are, but people want an answer now.
One thing I think it’s still easy for a human to beat AI at is a good PR summary.
They are often too verbose, or miss capturing an important concept or purpose, or add bullets for parts of the changes that no one cares about.
AI lacks wisdom because it processes patterns, not reality. It excels at statistical synthesis but cannot understand the meaning, stakes, or consequences of its outputs.The real danger is cognitive atrophy. If we outsource our critical thinking to automated averages, we stop doing the messy, high-latency work of building our own internal world models—effectively redefining "wisdom" as whatever the model outputs.Would you like to turn this short summary into a social media post, a quick debate rebuttal, or a thesis statement?
"New technology: cool but the supply chain is vulnerable". A tale as old as time.
Every large code base goes to shit unless you have a very strict bdfl at the top.
AI is nothing special or new in this regard. It just gives a single guy the velocity to ruin a codebase at the rate of a full enterprise team at double the speed.
A shitty code base still makes money, and that's all that will ever count for the majority of the employed developers, those who don't blog.
Will be downvoted for sure, but here we go:
They can't think that a software can be written, designed and maintained in English or higher constructs.
It is binary of either vibe coded app, or worshipping the code that they worship (and I spent 20+ years on)..the industry is moving on faster than most can wrap their heads on, but for those who understand software engineering was always beyond and above code, they will adapt, meanwhile, those who built their entire identity and skillset around managing code will struggle. There will be a place for those people, but it will be niche and specific, and we won't need as many.
HN is on a comical doomer binge the last couple of years.
We should start building our Dyson Sphere now because the only solution to bad computation is more computation. Evolution through competition is the only way forward I see. It won't necessarily improve human legibility without feedback, though. And in the end, we can't care TOO much about legibility.
Imagine breeding a new tomato plant. One is robust and tasty, the other is neither, but the genetic code is easier to understand. Picking the legible plant, because legibility is important to you, and may make future tinkering easier, is a dead end. Being able to create a billion different tomato plants and applying selection pressure is more effective, if you have the resources.
The idea that we are applying all this effort to make creating nauseating ads and heinous travel booking sites easier disgusts me, but hey, whatcha gonna do?!? :-)
Solid points. I will add a few points.
1) A lot of the time i spent deciding on interfaces (methods, classes, etc.) for humans. E.g. should this be two methods or one, should this method be in this class or moved to utility. Those problems went away. 2) What about performant code? This can be prompted away and when the measurements in your performance tests do not go down, then you can step in. 3) Sad to say but the AI has always been better than me at code-reviews. Maybe this is just me and if so I own that, but to the articles point, it might be harder to fix now. 4) "vibe-coded projects devolve over time into an unmaintainable mess". Preventing and managing this mess is the new skill sets we need to develop as software engineers. 5) Another skill-set we will need to master is how to maintain and grow our coding skills. Some ideas are: a) every once in a while implement a feature yourself. b) no AI Tuesdays! c) Have the AI quiz you on the code base. d) Have the AI develop HTML docs about how the code works.
1) Those problems did not go away. They only "went away" if you assume no one will ever need to read code again. And if you believe that, you clearly aren't reading the code already. Which I guess is why you think naming and architecture problems went away. AI is the absolute worst at architecture decisions.
2) not even sure what you mean by this
3) probably shouldn't admit that. It implies the reviewer has a less than average understanding of the code they're reviewing.
4) preventing and managing vibe code devolving into a pile of slop requires programmers not use AI. 80% accuracy repeated in more and more layers === more and more failures. In other words, the skill required is exactly the skill of being a good programmer without AI.
5) e) no AI all the time or only use AI as search. You're almost there with a and b. With c, it just doesn't understand well enough to "quiz you". With d, how are you going to know if the docs are correct if you aren't reading the slop?
I've found it helpful to make an explicit delineation in the codebase for which interfaces I'm going to agonize over and refine and which I'm going to ignore.
One set goes into haxe files and I use reflaxe macros to write tiny compilers that generate docs, clients, servers, cli's, test cases, serializers and deserializers, etc in whatever language is appropriate. That leaves gaps, which the AI can then fill in.
So I iterate on the haxe stuff. If the AI is struggling to "draw the rest of the owl," I change the source of truth until it has enough guidance re: type related errors, failing tests, and documentation. As requirements change, these things change.
As for the rest of the owl, it's disposable. Every few months I'll delete it and have an AI rewrite it from scratch (now with a smarter model and better docs and API specs and tests to guide it). This keeps the cruft from accumulating while preserving human contact with the code.
Most software is not good, using these rules. And that is written by human developers. The problem is that most software developers have less than 5 years experience. The community doubles every 5 years. Experience is scarce. So AI, learning from all the stuff online, does not learn great code, it learns from available code. You can make it produce better code, if you do the hard work of defining better in terms the AI understands. I've been trying to find a way forward with AI generated code, using my own definition of 'good architecture', and results with chatgpt 6 are promising. Not perfect, but good enough. I used a book I was writing as input, you can read an unfinished version of it here: https://programming-for-wizards.dev.muze.nl/ (Still working out the kinks of the underlying software) The other approach is to explicitly keep the whole decision tree as a causal chain in a repository: https://github.com/muze-labs/spiral-developer Still testing that one out.
We are teaching machines to do the work.
Capital always prefers machines.
I've read a lot of posts similar to this recently and keep asking myself if the past 20 years of my professional experience was that atypical.
I’m sure there are companies who are writing “perfectly maintainable and highly scalable code”. However, for half of my career I’ve been brought into startups to clean up the mess created by engineering teams.
While AI may create an unmaintainable mess (I’m not totally convinced), from my perspective many (not all) engineering teams have been doing that all along. #v2 #refactor
Ah yeah, light bulb moment here. I do think the disconnect I sense in reading these comments and discussing this might be in my level of expectations for how good "code quality" has ever been or even really can be in non-trivial applications.
To me, AIs seem to write code with a normal (low) level of quality, but I now have so many more tools to make sure the software works nonetheless, to refactor quickly, and to encode better practices that (mostly) stick in the future.
So to me, this is all a huge net win. But I guess if I'd ever worked on a pristine perfectly engineered system, I might see this all differently.
To me, it's always been a mess, and I'm just ecstatic that I have more ways to manage the mess now.
Is it a fact though? Maybe my projects aren't ambitious enough, but about 6 months ago I stopped hitting the point where AI can't maintain what it has written. It's probably unmaintainable by humans, but that might be an increasingly irrelevant quality.
A book that has shaped my thinking on AI, and much else besides, has been Horkheimer's Eclipse of Reason.
This book gives a conceptual groundwork for what I believe to be both the promise, and crisis of AI. AI excels at, and either will soon or already has surpassed humans at a kind of reasoning that Horkheimer calls "instrumental reason." This is reason as a tool for achieving ends. AI will, I think, surpass humans at writing maintainable code, as this falls within that domain. The writer is thoughtful, but the obstacles listed in the article are technical, and they will likely fall.
A second kind of reasoning, which he calls "objective reason," is reasoning about which ends are worth pursuing. Our contemporary pragmatic, positivist culture in the West has a hard time reasoning about ends. We tend to just see different perspectives, competing power networks, etc. We are skeptical of capital-J Justice, for example. We tend to see it as historically and culturally contingent, or else as a mask for powerful interests. Horkheimer believed this learned relativism was responsible for new forms of domination and control that emerged in the 20th century in both totalitarian and putatively free societies.
I think the author's thesis would be correct if it had focused on AI's inability to discern ends. But that then raises the question: can we discern ends? Horkheimer never settled on a way to rehabilitate objective reason. He focused on critique, and his critics suggest he ended up with a tangle of negations that couldn't hold up any objective ends. I think rehabilitating objective reason is the task that faces us whether we like it or not, given our moment in history with the advent of AI. It's something practitioners need to think about. Either we recover the philosophical tools of objective reason in light of the critique which caused them to be discarded in the first place, or ends will be imposed arbitrarily by any and everyone with the means to do so.
you don't need to read or write code directly to prevent it from becoming an unmaintainable mess. they way you prompt, the loops and skills you build for your harness and the way you set structure around it can get you most of the benefits of reading & writing code manually
In traditional software development process two artifacts are created. The obvious one is software. The second, is knowledge of the problem domain in the development team heads.
We could take the Hofstadter idea and say that there is an isomorphism between these two artifacts. Software is a manifestation of the team's knowledge the way an organism is a manifestation of a genome.
With AI, the second artifact is no longer necessarily produced. We don't need to have a team that, as the project progresses, slowly becomes a group of domain experts. Experts that can drive the project direction. Experts that, with time, can see the flaws in their first project and start new breakthrough projects to fix these flaws.
Very interesting post. Feels like my comment from two days ago [1], elaborated. I must thank the author for verbalizing all the concerns.
So now we have a fork: red-pillers who want to regenerate software all the time (esp those with unlimited token budgets) and blue-pillers who want to maintain more code mass per developer-head. Plus, we have software artisans.
That will be an interesting horse race.
[1]: https://news.ycombinator.com/item?id=49774239
The entire field of software engineering has plenty of depth without requiring you to be the person actually writing and reading the code. The author falsely conflates reading and writing the code to be the same as understanding the structure and design and architecture of the system. These two things aren't the same and never were.
Well said. The cost of bad architecture imposes a heavy penalty on the project down the line.
Long hours of grinding with the problem, the solution, the code and the machine all makes for an intuitive understanding of what a good architecture is in the first place.
LLMs are a tool. A useful tool unlike any we have seen so far. But still a tool. Treating it as such and leveraging it should be prioritized.
Effort is all. It was before AI, it is the same with AI.
Putting in massive effort with AI leads to quality exactly like before hand coding.
Vast majority of work was junk before AI because vast majority of people put the minimal possible effort.
The difference now is that AI can make low effort work look high effort at a surface level.
True, but we'll also get higher standards I think. It's become easier to do things, so skilled people can do more complicated things. We compare to what humans can achieve.
Little boxes, little boxes, little boxes made out of ticky tacky. Little boxes made out of ticky tacky and they all look just the same.
The agents are trained by contained data. All agents. Similar data sets. Human knowledge packaged into little boxes that all look just the same.
When the boxes are doing all the work for us, the differences between them will come down to color.
There’s yellow ones, and blue ones…. And, they’re all made out of ticky tacky and they all cost just the same…
I'm weirdly attached to this line. Physical automation a.k.a. factories are not portable. The manufacturing techniques are, but to "spin up" a new factory requires expertise, effort and capital. In contrast to software, replication can be done quite effortlessly, with containers and such tools, it's easily a one-man job.
Writing, coding and art are not meant to be repetitive, scalable tasks. I think that's what's driving the core of the backlash. If these forms of human expressions can be automated, then why would we need to apply our mental capacity at all?