I'm a CS teacher, so this is where I see a huge danger right now and I'm explicit with my students about it: you HAVE to write the code. You CAN'T let the machines write the code. Yes, they can write the code: you are a student, the code isn't hard yet. But you HAVE to write the code.
This is the ultimate problem with AI in academia. We all inherently know that “no pain no gain” is true for physical tasks, but the same is true for learning. Struggling through the new concepts is essentially the point of it, not just the end result.
Of course this becomes a different thing outside of learning, where delivering results is more important in a workplace context. But even then you still need someone who does the high level thinking.
It's not a perfect analogy though because in this case it's more like automated driving - you should still learn to drive because the autodriver isn't perfect and you need to be ready to take the wheel, but that means deliberate, separate practice at learning to drive.
I think that's a bit of a myth. The Greeks and Romans had weightlifting and boxing gyms, but no forklifts. Many of the most renowned Romans in the original form of the Olympics and in Boxing were Roman Senators with the wealth and free time to lift weights and box and wrestle. One of the things that we know about the famous philosopher Plato was that Plato was essentially a nickname from wrestling (meaning "Broad") as a first career (somewhat like Dwayne "The Rock" Johnson, which adds a fun twist to reading Socratic Dialogs or thinking about relationships as "platonic").
Arguably the "meritocratic ideal" of the Gladiator arena was that even "blue collar" Romans could compete and maybe survive. But even the stories that survive of that, few did.
There may be a lesson in that myth, too, that the people that succeed in some sports often aren't the people doing physical labor because they must do physical labor (for a job), they are the ones intentionally practicing it in the ways to do well in sports.
They don’t go to the gym, they don’t have the energy; the job shapes you. More or less the same for the farmers in the family.
Perhaps this was less so in the industrial era because of poor nutrition (source: Bill Bryson, hopefully well researched). Hunter gatherer cultures that we still study today have tremendous fitness (Daniel Lieberman).
We may not have any evidence that they had forklifts but we also can't rule out the possibility entirely :)
Why do you think that? It's definitely true. You can observe it today if you want to visit a country where peasants are still common.
From Bret Devereaux's recent series on Greek hoplites:
> Now traditionally, the zeugitai were regarded as the ‘hoplite class’ and that is sometimes supposed to be the source of their name
> but what van Wees is working out is that although the zeugitai are supposed to be the core of the citizen polity (the thetes have limited political participation) there simply cannot be that many of them because the minimum farm necessary to produce 200 medimnoi of grain is going to be around 7.5 ha or roughly 18 acres which is – by peasant standards – an enormous farm, well into ‘rich peasant’ territory.
> Of course with such large farms there can’t be all that many zeugitai and indeed there don’t seem to have been. In van Wees’ model, the zeugitai-and-up classes never supply even half of the number of hoplites we see Athens deploy
> Instead, under most conditions the majority of hoplites are thetes, pulled from the wealthiest stratum of that class (van Wees figures these fellows probably have farms in the range of ~3 ha or so, so c. 7.5 acres). Those thetes make up the majority of hoplites on the field but do not enjoy the political privileges of the ‘hoplite class.’
> And pushing against the ‘polis-of-rentier-elites’ model, we often also find Greek sources remarking that these fellows, “wiry and sunburnt” (Plato Republic 556cd, trans. van Wees), make the best soldiers because they’re more physically fit and more inured to hardship – because unlike the wealthy hoplites they actually have to work.
( https://acoup.blog/2026/01/09/collections-hoplite-wars-part-... )
---
> Many of the most renowned Romans in the original form of the Olympics and in Boxing were Roman Senators
In the original form of the Olympics, a Roman senator would have been ineligible to compete, since the Olympics was open only to Greeks.
> In the original form of the Olympics, a Roman senator would have been ineligible to compete, since the Olympics was open only to Greeks.
I did debate how to word that mixing of Greek and Roman things in the same sentence. I had emotional context I wanted to convey and considered a word like Decathlon there as more technically correct, but then fought the modern context that of the people that even know what the Decathlon is they know it in the context of it being a smaller event in the modern Olympics, from which perspective Olympics remains more technically correct as the modern English word for both.
As to the text you are quoting, I think it as much supports my claims as you think it doesn't. Ignoring the subject change from "weightlifting" (and sports more generally) to farming and soldiering, it mostly describes the general state of armies and feudalism in general through much of time: you have the rank and file from blue collar classes, and you have the officer corps from white collar classes. The wealthier class is fewer, but given more charge and importance. The lower class does more of the grunt work. The Romans had rich Officers and blue collar "enlisted".
The myth that I was referring to was that weightlifting is somehow a new invention because no one labors physically anymore. There have always been leisure classes that needed to lift weights as a hobby to get good at sports (and that class was also more often awarded medals in sports or important commands in armies, if we want to also connect to the blog post you quoted). As far as I'm aware there was never a period in recorded history where "everyone" was equally fit from physical labor and there was no such thing as training and gyms and needing leisure time to do that.
[Further tangent: Even "pre-history" and the modern (mis)conception of the "paleo ideal" idea of tribes of equally buff hunter-gatherers starts to fall apart when you ask questions about family units or what they think the "gatherer" side of the equation meant (and manage to divorce it from modern ideas of agriculture being highly intense labor) or what those societies would look like if more people lived to old age or how those societies survived things like the Ice Age (fattier and more hibernatory, because we are a mammalian species, we cannot escape that).]
> The ability of skinny old ladies to carry huge loads is phenomenal. Studies have shown that an ant can carry one hundred times its own weight, but there is no known limit to the lifting power of the average tiny eighty-year-old Spanish peasant grandmother.
My favorite historic example of typical modern hypertrophy-specific training is the training of Milo of Croton [1]. By legend, his father gifted him with the calf and asked daily "what is your calf, how does it do? bring it here to look at him" which Milo did. As calf's weight grew, so did Milo's strength.
This is application of external resistance (calf) and progressive overload (growing calf) principles at work.
[1] https://en.wikipedia.org/wiki/Milo_of_Croton
Milo lived before Archimedes.
Alexander Zass (Iron Samson) also trained each day: https://en.wikipedia.org/wiki/Alexander_Zass
"He was taken as a prisoner of war four times, but managed to escape each time. As a prisoner, he pushed and pulled his cell bars as part of strength training, which was cited as an example of the effectiveness of isometrics. At least one of his escapes involved him 'breaking chains and bending bars'."
Rest days are overrated. ;)
Training volume of Bulgarian Method is not much bigger than that of regular training splits like Sheiko or something like that, if bigger at all. What is more frequent is the stimulation of muscles and nervous system paths and BM adapts to that - one does high percentage of one's current max, essentially, one is training with what is available to one's body at the time.
Also, ultra long distance runners regenerate cartilages: https://ryortho.com/2015/12/what-ultra-long-distance-runners...
Our bodies are amazing.
Like many educational tests the outcome is not the point - doing the work to get there is. If you're asked to code fizz buzz it's not because the teacher needs you to solve fizz buzz for them, it's because you will learn things while you make it. Ai, copying stack overflow, using someone's code from last year, it all solves the problem while missing the purpose of the exercise. You're not learning - and presumably that is your goal.
People used to get strong because they had to survive. They stopped needing strength to survive, so it became optional.
So what does this mean about intelligence? Do we no longer need it to survive so it's optional? Yes/No informs on how much young and developing minds should be exposed to AI.
Here's the thing -- I don't care about "getting stronger." I want to make things, and now I can make bigger things WAY faster because I have a mech suit.
edit: and to stretch the analogy, I don't believe much is lost "intellectually" by my use of a mech suit, as long as I observe carefully. Me doing things by hand is probably overrated.
I’ve worked with plenty of self taught programmers over the years. Lots of smart people. But there’s always blind spots in how they approach problems. Many fixate on tools and approaches without really seeing how those tools fit into a wider ecosystem. Some just have no idea how to make software reliable.
I’m sure this stuff can be learned. But there is a certain kind of deep, slow understanding you just don’t get from watching back-to-back 15 minute YouTube videos on a topic.
But if they actually spent time trying to learn architecture and how to build stuff well, either by reading books or via good mentorship on the job, then they can often be better than the folks who went to school. Sometimes even they don't know how to make software reliable.
I'm firmly in the middle. Out of the 6 engineers I work with on a daily basis (including my CTO), only one of us has a degree in CS, and he's not the one in an architecture role.
I do agree that learning how to think and learn is its own valuable skill set, and many folks learn how to do that in different ways.
Yeah I just haven’t seen this happen. I’ve seen plenty of people graduate who were pretty useless. But … I think every self taught programmer I’ve worked with had meaningful gaps in their knowledge.
They’d spend a week in JavaScript to save them from 5 minutes with C or bash. Or they’d write incredibly slow code because they didn’t know the appropriate algorithms and data structures. They wouldn’t know how to profile their program to learn where the time is being spent. (Or that that’s even a thing). Some would have terrible intuitions around how the computer actually runs a program, so they can’t guess what would be fast or slow. I’ve seen wild abstractions to work around misunderstandings of the operating system. Hundreds of lines to deal with a case that can’t actually ever happen, or because someone missed the memo on a syscall that solves their exact problem. There’s also hairball nests of code because someone doesn’t know what a state machine is. Or how to factorise their problem in other ways. One guy I worked with thought the react team invented functional programming. Someone else doesn’t understand how you could write programs without OO inheritance. And I’ve seen so many bugs. Months of bugs, that could be prevented with the right design and tests.
I’ve worked with incredibly smart self taught programmers. Some of the smartest people I’ve ever worked with. But the thing about blind spots is you don’t know you have them. You say you’re self taught, and self taught people can be better than people who went to school. In limited domains, yeah. Smart matters a lot. But you don’t know what you don’t know. You don’t know what you missed out on. And you don’t know what problems in the workplace you could have easily solved if you knew how.
Also, to clarify, I'm not arguing that self-taught vs CS grad is mutually exclusive to smart/not smart. There are plenty of not-smart self-taught engineers and plenty of smart grads.
> In limited domains
I'd argue that many, if not most, teams operate in limited domains.
That gets expensive, fast. There's just so much to cover already, between communication skills, programming skills, debugging skills, architecture / "whiteboarding problems", data structures and algorithms, general problem solving ("interview problems"). A job interview can never be a fully rigorous test of someone's actual skills. Most don't cover even a fraction of that stuff already.
> I'd argue that many, if not most, teams operate in limited domains.
It depends what you consider yourself responsible for. If you think of your job (or your team's job) as shipping features X, Y and Z within this react based web app, then sure - you operate in a limited domain. But if your job is "solve the user's actual problems" then things can get pretty broad, pretty fast. Sometimes you write code. Sometimes you're debugging a hard problem. Or talking to the users. Or identifying and tracking down a performance regression. Or writing an issue for a bug in 3rd party code. Or trawling through MDN to figure out a workaround to some browser nonsense. Or writing reliable tests, or CI/CD systems. And so on.
Its only really junior engineers who have the luxury of a limited scope.
Am I an outlier or am I missing something here?
I've worked with PhDs on projects (I'm self-taught), and those guys absolutely have blind spots in how they approach problems, plenty of them. Everyone does. What we produce together is better because our blind spots don't typically overlap. I know their weaknesses, and they know mine. I've also worked with college grads that overthink everything to the point they made an over-abstracted mess. YMMV.
>you just don’t get from watching back-to-back 15 minute YouTube videos on a topic.
This is not "self taught". I mean maybe it's one kind of modern-ish concept of "self taught" in an internet comment forum, but it really isn't. I watch a ton of sailing videos all day long, but I've never been on a sailboat, nor do I think I know how to sail. Everyone competent has to pay their dues and learn hard lessons the hard way before they get good at anything, even the PhDs.
1. contacts - these come in the form of peers who are interested in the same things and in the form of experts in their fields of study. Talking to these people and developing relationships will help you learn faster, and teach you how to have professional collegial relationships. These people can open doors for you long after graduation.
2. facilities - ever want to play with an electron microscope or work with dangerous chemicals safely? Different schools have different facilities available for students in different fields. If you want to study nuclear physics, you might want to go to a school with a research reactor; it's not a good idea to build your own.
And I'd argue for:
3. Realisation of the scope of computing.
IE Computers are not just phones/laptop/desktop/server with networking - all hail the wonders of the web... There are embedded devices, robots, supercomputers. (Recent articles on HN describe the computing power in a disposable vape!)
There are issues at all levels with all of these with algorithms, design, fabrication, security, energy, societal influence, etc etc - what tradeoffs to make where. (Why is there computing power in a disposable vape?!?)
I went in thinking I knew 20% and I would learn the other 80% of IT. I came out knowing 5 times as much but realising I knew a much smaller percentage of IT... It was both enabling and humbling.
If you weren't even "clever enough" to write the program yourself (or, more precisely, if you never cultivated a sufficiently deep knowledge of the tools & domain you were working with), how do you expect to fix it when things go wrong? Chatbots can do a lot, but they're ultimately just bots, and they get stuck & give up in ways that professionals cannot afford to. You do still need to develop domain knowledge and "get stronger" to keep pace with your product.
Big codebases decay and become difficult to work with very easily. In the hands-off vibe-coded projects I've seen, that rate of decay was extremely accelerated. I think it will prove easy for people to get over their skis with coding agents in the long run.
That's kinda how I see vibe coding. It's extremely easy to get stuff done but also extremely easy to write slop. Except now 10x more code is being generated thus 10x more slop.
Learning how to get quality robust code is part of the learning curve of AI. It really is an emergent field, changing every day.
The key difference with LLMs is that React was written very intentionally by smart engineers who provided a wealth of documentation to help people who need to peek under the hood of their framwork. If your LLM has written something you don't understand, though, chances are nobody does, and there's nowhere you can turn to.
If (as Peter Naur famously argued) programming is theory building, then an abstraction like a framework lets you borrow someone else's theory. You skip developing an understanding of the underlying code and hope that you'll either never need to touch the underlying code or that, if you do, you can internalize the required theory later, as needed. LLM-generated code has no theory; you either need to supervise it closely enough to impose your own, or treat it as disposable.
Agreed! And I think that's what I'm getting at. Adding what they're now calling "skills," or writing your own, is becoming crucial to LLM-assisted development. If the LLM is writing too much slop, then there probably wasn't sufficient guidance to ensure that slop wouldn't be written.
The first step of course is to actually check that the generated code is indeed slop, which is where many people miss the mark.
With all respect, that's nonsense.
Absolutely no one gains more than a superficial grasp of a skill just by observing.
And even with a good grasp of skills, human boredom is going to atrophy any ability you have to intervene.
It's why the SDCs (Tesla, I think) that required the driver to stay alert to take control while the car drove itself were such a danger - after 20+ hours of not having to to anything, the very first time a normal reaction time to an emergency is required, the driver is too slow to take over.
If you think you are learning something reviewing the LLM agent's output, try this: choose a new project in a language and framework you have never used, do your usual workflow of reviewing the LLMs PRs, and then the next day try to do a simple project in that new language and framework (that's the test of how much you learned).
Compare that result to doing a small project in a new language, and then the next day doing a different small project in that same language.
If you're at all honest with yourself, or care whether you atrophy or not, you'd actually run that experiment and control and objectively judge the results.
I'd agree, if my goal was "to be a great and complete coder."
I don't. I want just enough to build cool things.
Now, that's just me.
That being said, I'd also venture to say that your attitude here might be a tad dinosaurish. I like it too, but also, know that to a large extent, especially in the market -- this "quality" that you're striving for here may just not happen.
> I don't. I want just enough to build cool things.
That's great, but that wasn't the point I was responding to. I was specifically addressing your specific point of:
>>> I don't believe much is lost "intellectually" by my use of a mech suit, as long as I observe carefully.
I still think that's nonsense; no one learns much by observing.
Paraphrasing the old joke, you aren't going to get to Carnegie Hall by observing violinists.
There are other fictional variants: the giant mech with the enormous support team, or Heinlein's "mobile infantry." And virtually every variantion on the Heinlein trope has a scene of drop commandos doing extensive pre-drop checks on their armor.
The actual reality is it isn't too had for a competent engineer to pair with Claude Code, if they're willing to read the diffs. But if you try to increase the ratio of agents to humans, dealing with their current limitations quickly starts to feel like you need to be Tony Stark.
Let's not mince words here, what you mean is that you don't care to learn about a craft. You just want to get to the end result, and you are using the shiny new tool that promises to take you from 0 to 100% with little to no effort.
In this way, I'd argue what you are doing is not "creating", but engaging in a new form of consumption. It used to be you relied on algorithms to present to you content that you found fun, but the problem was that algorithm required other humans to create that content for you to later consume. Now with LLMs, you remove the other humans from the loop, and you can prompt the AI directly with exactly what you wish to see in that moment, down to the fine grained details of the thing, and after enough prompts, the AI gives you something that might be what you asked for.
You are rotting your brain.
In true HN fashion of trading analogies, it’s like starting out full powered in a game and then having it all taken away after the tutorial. You get full powered again at the end but not after being challenged along the way.
This makes the mech suit attractive to newcomers and non-programmers, but only because they see product in massively simplified terms. Because they don’t know what they don’t know.
Thinking through the issue, instead of having the solve presented to you, is the part where you exercise your mental muscles. A good parallel is martial arts.
You can watch it all you want, but you'll never be skilled unless you actually do it.
Or "An [electric] bicycle for the mind." Steve Jobs/simonw
You need to be strong to do so. Things of any quality or value at least.
Now compare this to using the LLM with a grammar book and real world study mechanisms. This creates friction which actually causes your mind to learn. The LLM can serve as a tool to get specialized insight into the grammar book and accelerate physical processes (like generating all forms of a word for writing flashcards). At the end of day, you need to make an intelligent separation where the LLM ends and your learning begins.
I really like this contrast because it highlights the gap between using an LLM and actually learning. You may be able to use the LLM to pass college level courses in learning the language but unless you create friction, you actually won’t learn anything! There is definitely more nuance here but it’s food for thought
There is more than one kind of leverage at play here.
That's the job of the backhoe.
(this is a joke about how diggers have caused quite a lot of local internet outages by hitting cables, sometimes supposedly "redundant" cables that were routed in the same conduit. Hitting power infrastructure is rare but does happen)
Regardless of whose fault it was, the end result was the bucket snagged the power lines going into the datacentre and caused an outage.
Unless you happen to drive a forklift in a power plant.
> expose millions to fraud and theft
You can if you drive forklift in a bank.
> put innocent people in prison
You can use forklift to put several innocent people in prison with one trip, they have pretty high capacity.
> jeopardize the institutions of government.
It's pretty easy with a forklift, just try driving through main gate.
> There is more than one kind of leverage at play here.
Forklifts typically have several axes of travel.
The activity would train something, but it sure wouldn't be your ability to lift.
There are enthusiasts who will spend an absolute fortune to get a bike that is few grams lighter and then use it to ride up hills for the exercise.
Presumably a much cheaper bike would mean you could use a smaller hill for the same effect.
If you practice judo you're definitely exercising but the goal is defeating your opponent. When biking or running you're definitely exercising but the goal is going faster or further.
From an an exercise optimization perspective you should be sitting on a spinner with a customized profile, or maybe do some entirely different motion.
If sitting on a carbon fiber bike, shaving off half a second off your multi-hour time, is what brings you joy and motivation then I say screw it to further justification. You do you. Just be mindful of others, as the path you ride isn't your property.
It's the whole "journey vs destination" thing.
Currently AI seems to be the rocket you strap to your back as you put on VR glasses and enjoy the entertainment. You'll get there fast or blow up in the middle.
The True Artisanal Coders are the ones running the whole way, enjoying the scenery and the physical conditioning they get.
And there are people in between with bikes, cars etc. (different stages of AI use)
Analogies are fun =)
But if you're a junior developer, you won't be able to gain the most vital skills.
[0] https://eazypilot.com/blog/automation-dependency-blessing-or...
I think forklifts probably carry more weight over longer distances than people do (though I could be wrong, 8 billion humans carrying small weights might add up).
Certainly forklifts have more weight * distance when you restrict to objects that are over 100 pounds, and that seems like a good decision.
So the idea is that you should learn to do things by hand first, and then use the powerful tools once you're knowledgeable enough to know when they make sense. If you start out with the powerful tools, then you'll never learn enough to take over when they fail.
Indeed, usually after doing weightlifting, you return the weights to the place where you originally took them from, so I suppose that means you did no work at in the first place..
OK but then why even use Python, or C, or anything but Assembly? Isn't AI just another layer of value-add?
There has to be a base of knowledge available before the student can even comprehend many/most open research questions, let alone begin to solve them. And if they were understandable to a beginner, then I’d posit the LLM models available today would also be capable of doing meaningful work.
We have a decent sized piece of land and raise some animals. People think we're crazy for not having a tractor, but at the end of the day I would rather do it the hard way and stay in shape while also keeping a bit of a cap on how much I can change or tear up around here.
Unfortunately, many sdevs don't understand it.
I wouldn't want to write raw bytes like Mel did though. Eventually some things are not worth getting good at.
When I stand still for hours at a time, I end up with aching knees, even though I'd have no problem walking for that same amount of time. Do you experience anything like that?
Forklift operators don't lift things in their training. Even CS students start with pretty high level of abstraction, very few start from x86 asm instructions.
We need to make them implement ALU's on logical gates and wires if we want them to lift heavy things.
> We need to make them implement ALU's on logical gates and wires
Things must have certainly changed since I was a CS student :-/ We did an assembler course in second year, and implemented a basic adder in circuitry in a different course.
This was in the mid-90s, when there was definitely little need for assembly programmers outside of EE (I was CS).
Though I also wonder what advanced CS classes should look like. If they agent can code nearly anything, what project would challenge student+agent and teach the student how to accomplish CS fundamentals with modern tools.
As an added bonus, being able to discuss your code with another engineer that wasn't involved in writing it is an important skill that might not otherwise be trained in college.
Well, whether we like it or not, we are all eventually going to find out if "developing a product that adds value to its users" can be done when you have no more skill than aforementioned users.
Skills atrophy is a real thing.
It was a bizarre disconnect having someone be both highly educated and yet crippled by not doing.
The students had memorized everything, but understood nothing. Add in access to generative AI, and you have the situation that you had with your interview.
It's a good reminder that what we really do, as programmers or software engineers or what you wanna call it, is understanding how computers and computations work.
> The first principle is that you must not fool yourself and you are the easiest person to fool.
It's probably also worth reading Feynman's Cargo Cult Science: https://sites.cs.ucsb.edu/~ravenben/cargocult.html
Makes me wonder if C programmers look at JS programmers and shrug, “they don’t understand what their programs are actually doing.”
I’m not trying to be disingenuous, but I also don’t see a fundamental difference here. AI lets programmers express intent at a higher level of abstraction than ever before. So high, apparently, that it becomes debatable whether it is programming at all, out whether it takes any skill, out requires education or engineering knowledge any longer.
Star Trek or Idiocracy.
Just checking I have that right... is that what you meant?
I think that's what you were implying but it's just want to check I have that right? if so
... that ... is .... wow ...
Knowing the complexity of bubble sort is one skill, being able to write code that performs bubble sort is a second, and being able to look at a function with the signature `void do_thing(int[] items)`and determine that it's bubble sort and the time complexity of it in terms of the input array is a third. It sounds like they had the first skill, used an AI to fake the second, but had no way of doing the third.
A good analogy here is programming in assembler. Manually crafting programs at the machine code level was very common when I got my first computer in the 1980s. Especially for games. By the late 90s that had mostly disappeared. Games like Roller Coaster Tycoon were one of the last ones with huge commercial success that were coded like that. C/C++ took over and these days most game studios license an engine and then do a lot of work with languages like C# or LUA.
I never did any meaningful amount of assembler programming. It was mostly no longer a relevant skill by the time I studied computer science (94-99). I built an interpreter for an imaginary CPU at some point using a functional programming language in my second year. Our compiler course was taught by people like Eric Meyer (later worked on things like F# at MS) who just saw that as a great excuse to teach people functional programming instead. In hindsight, that was actually a good skill to have as functional programming interest heated up a lot about 10 years later.
The point of this analogy: compilers are important tools. It's more important to understand how they work than it is to be able to build one in assembler. You'll probably never do that. Most people never work on compilers. Nor do they build their own operating systems, databases, etc. But it helps to understand how they work. The point of teaching how compilers work is understanding how programming languages are created and what their limitations are.
People learn by doing. There's a reason that "do the textbook problems" is somewhat of a meme in the math and science fields - because that's the way that you learn those things.
I've met someone who said that when he get a textbook, he starts by only doing the problems, and skipping the chapter content entirely. Only when he has significant trouble with the problems (i.e. he's stuck on a single one for several hours) does he read the chapter text.
He's one of the smartest people I know.
This is because you learn by doing the problems. In the software field, that means coding.
Telling yourself that you could code up a solution is very different than actually being able to write the code.
And writing the code is how you build fluency and understanding as to how computers actually work.
> I never did any meaningful amount of assembler programming. It was mostly no longer a relevant skill by the time I studied computer science (94-99). I built an interpreter for an imaginary CPU at some point using a functional programming language in my second year.
Same thing for assembly. Note that you built an interpreter for an imaginary CPU - not a real one, as that would have been a much harder challenge given that you didn't do any meaningful amount of assembly program and didn't understand low-level computer hardware very well.
Obviously, this isn't to say that information about how a system works can't be learned without practice - just that that's substantially harder and takes much more time (probably 3-10x), and I can guarantee you that those doing vibecoding are not putting in that extra time.
The brave new world is that you no longer have to do “coding” in our sense of the word. The doing, and what exercises you should learn with have both changed.
Now students should build whole systems, not worry about simple Boolean logic and program flow. The last programmer to ever need to write an if statement may already be in studies.
Notice how I also talked about coding being a way that you learn how computers work.
If you don't code, you have a very hard time understanding how computers work.
And while there's some evidence that programmers may not need write all of their code by hand, there's zero evidence that either they don't need to learn how to code at all (as you're claiming), or that they don't need to even know how computers work (which is a step further).
There's tons of anecdotes from senior software engineers on Hacker News (and elsewhere) about coding agents writing bad code that they need to debug and fix by hand. I've literally never seen a single story about how a coding agent built a nontrivial program by itself without the prompter looking at the code.
I don't know that it's all these things at once, but most people I know that are good have done a bunch of spikes / side projects that go a level lower than they have to. Intense curiosity is good, and to the point your making, most people don't really learn this stuff just by reading or doing flash cards. If you want to really learn how a compiler works, you probably do have to write a compiler. Not a full-on production ready compiler, but hands on keyboard typing and interacting with and troubleshooting code.
Or maybe to put another way, it's probably the "easiest" way, even though it's the "hardest" way. Or maybe it's the only way. Everything I know how to do well, I know how to do well from practice and repitition.
Indeed, a lot of us looked with suspicion and disdain at people that used those primitive compilers that generated awful, slow code. I once spent ages hand-optimizing a component that had been written in C, and took great pleasure in the fact I could delete about every other line of disassembly...
When I wrote my first compiler a couple of years later, it was in assembler at first, and supported inline assembler so I could gradually convert to bootstrap it that way.
Because I couldn't imagine writing it in C, given the awful code the C compilers I had available generated (and how slow they were)...
These days most programmers don't know assembler, and increasingly don't know languaes as low level as C either.
And the world didn't fall apart.
People will complain that it is necessary for them to know the languages that will slowly be eaten away by LLMs, just like my generation argued it was absolutely necessary to know assembler if you wanted to be able to develop anything of substance.
I agree with you people should understand how things work, though, even if they don't know it well enough to build it from scratch.
Maybe the world didn't fall apart, but user interactions on a desktop pc feel slower than ever. So perhaps they should.
Software got significantly worse in that time period, though
Even while vibe-coding, I often found myself getting annoyed just having to explain things. The amount of patience I have for anything that doesn't "just work" the first time has drifted toward zero. If I can't get AI to do the right thing after three tries, "welp, I guess this project isn't getting finished!"
It's not just laziness, it's like AI eats away at your pride of ownership. You start a project all hyped about making it great, but after a few cycles of AI doing the work, it's easy to get sucked into, "whatever, just make it work". Or better yet, "pretend to make it work, so I can go do something else."
The idea was to develop a feel for cutting metal, and to better understand what the machine tools were doing.
--
My wood shop teacher taught me how to use a hand plane. I could shave off wood with it that was so thin it was transparent. I could then join two boards together with a barely perceptible crack between them. The jointer couldn't do it that well.
This kind of workload was a shock to me. It more than a year to adapt to it.
The progression from basic arithmetic, to complex ratios and basic algebra, graphing, geometry, trig, calculus, linear algebra, differential equations… all along the way, there are calculators that can help students (wolfram alpha basically). When they get to theory, proofs, etc… historically, thats where the calculator ended, but now there’s LLMs… it feels like the levels of abstractions without a “calculator” are running out.
The compiler was the “calculator” abstraction of programming, and it seems like the high-level languages now have LLMs to convert NLP to code as a sort of compiler. Especially with the explicitly stated goal of LLM companies to create the “software singularity”, I’d be interested to hear the rationale for abstractions in CS which will remain off limits to LLMs.
I've hired and trained tons of junior devs out of university. They become 20x productive after a year of experience. I think vibe coding is getting new devs to 5x productivity, which seems amazing, but then they get stuck there because they're not learning. So after year one, they're a 5x developer, not a 20x developer like they should be.
I have some young friends who are 1-3 years into software careers I'm surprised by how little they know.
Recently in comments people were claiming that working with LLMs has sharpened their ability to organize thoughts, and that could be a real effect that would be interesting to study. It could be that watching an LLM organize a topic could provide a useful example of how to approach organizing your own thoughts.
But until you do it unassisted you haven’t learned how to do it.
With how rapidly the world has been changing lately, it has become difficult to estimate which of those more specific skills will remain relevant for how long.
I think learning these algorithms was useful not because we use these tools ourselves but because it taught us useful concepts: very practical step-by-step algorithms, and a hands-on introduction to decimal places and powers of ten.
There are other ways to learn these things but I learned them in elementary school by doing hundreds of arithmetic problems over 4-5 years. That knowledge stays with me even if I don’t pull out long division for anything practical.
And plenty of people will still come along who love to code despite AI's excelling at it. In fact, calling out the AI on bad design or errors seems to be the new "code golf".
https://www.slater.dev/2025/08/llms-are-not-bicycles-for-the...
I do Windows development and GDI stuff still confuses me. I'm talking about memory DC, compatible DC, DIB, DDB, DIBSECTION, bitblt, setdibits, etc... AIs also suck at this stuff. I'll ask for help with a relatively straightforward task and it almost always produces code that when you ask it to defend the choices it made, it finds problems, apologizes, and goes in circles. One AI (I forget which) actually told me I should refer to Petzold's Windows Programming book because it was unable to help me further.
> grug once again catch grug slowly reaching for club, but grug stay calm
I actually fear more for the middle-of-career dev who has shunned AI as worthless. It's easier than ever for juniors to learn and be productive.
Without the clarity that comes from thinking with code, a programmer using AI is the blind leading the blind.
The social aspect of a dialogue is relaxing, but very little improvement is happening. It's like a study group where one (relatively) incompetent student tries to advise another, and then test day comes and they're outperformed by the weirdo that worked alone.
It's the difference between the employee who copy-pastes all of their email bodies from ChatGPT versus the one who writes a full draft themselves and then asks an LLM for constructive feedback. One develops skills while the other atrophies.
Though also in the 90's the standard library was new and often had bugs
I learned more about programming in a weekend badly copying hack modules for Minecraft than I learned in 5+ years in university.
All that stuff I did by hand back then I haven't used it a single time after.
You write sorting algorithms in college to understand how they work. Understand why they are faster because it teaches you a mental model for data traversal strategies. In the real world, you will use pre-written versions of those algorithms in any language but you understand them enough to know what to select in a given situation based on the type of data. This especially comes into play when creating indexes for databases.
What I take the OPs statement to mean are around "meta" items revolved more around learning abstractions. You write certain patterns by hand enough times, you will see the overlap and opportunity to refactor or create an abstraction that can be used more effectively in your codebase.
If you vibe code all of that stuff, you don't feel the repetition as much. You don't work through the abstractions and object relationships yourself to see the opportunity to understand why and how it could be improved.
I only had to do this leg work during university to prove that I can be allowed to try and write code for a living. The grounding as you call it is not required for that at all,since im a dozen levels of abstraction removed from it. It might be useful if I was a researcher or would work on optimizing complex cutting edge stuff, but 99% of what I do is CRUD apps and REST Apis. That stuff can safely be done by anyone, no need for a degree. Tbf I'm from Germany so in other places they might allow you to do this job without a degree
If AI is used by the student to get the task done as fast as possible the student will miss out on all the learning (too easy).
If no AI is used at all, students can get stuck for long periods of time on either due to mismatches between instructional design and the specific learning context (missing prereq) or by mistakes in instructional design.
AI has the potential to keep all learners within an ideal difficulty for optimal rate of learning so that students learn faster. We just shouldn't be using AI tools for productivity in the learning context, and we need more AI tools designed for optimizing learning ramps.
Edit: I expect it wouldn't be super hard to create though, you'd just have to hook into the editor's change event, probably compute the diff to make sure you don't lose anything, and then append it to the end of the json.
It does seem like they’re going the wrong way, repelling tech to keep things easy instead of embracing new tech by updating their teaching methods.
But I also think we’ve collectively fallen flat in figuring out what those methods are.
The one requirement I think is dumb though is we're not allowed to use the language's documentation for the final project, which makes no sense. Especially since my python is rusty.
Since you mentioned failure to figure out what better teaching methods are, I feel it's my sworn duty to put a plug for https://dynamicland.org and https://folk.computer, if you haven't heard about them :)
Making students fix LLM-generated code until they're at their wits' end is a fun idea. Though it likely carries too high of an opportunity cost education-wise.
People said this about compilers. It depends what layer you care to learn/focus on. AI at least gives us the option to move up another level.
Your curriculum may be different than it is around here, but here it's frankly the same stuff I was taught 30 years ago. Except most of the actual computer science parts are gone, replaced with even more OOP, design pattern bullshit.
That being said. I have no idea how you'd actually go about teaching students CS these days, considering a lot of them will probably use ChatGPT or Claude regardless of what you do. That is what I see in the statistic for grades around here. For the first 9 years I was a well calibrated grader, but these past 1,5ish years it's usually either top marks or bottom marks with nothing in between. Which puts me outside where I should be, but it matches the statistical calibration for everyone here. I obviously only see the product of CS educations, but even though I'm old, I can imagine how many corners I would have cut myself if I had LLM's available back then. Not to mention all the distractions the internet has brought.
In my experience, people who talk about business value expect people to code like they work at the assembly line. Churn out features, no disturbances, no worrying about code quality, abstractions, bla bla.
To me, your comment reads contradictory. You want initiative, and you also don't want initiative. I presume you want it when it's good and don't want it when it's bad, and if possible the people should be clairvoyant and see the future so they can tell which is which.
What I read from GP is that they’re looking for engineering innovation, not new science. I don’t see it as contradictory at all.
The word you’re looking for is skill. He wants devs to be skilled. I wouldn’t thought that to be controversial but hn never ceases to amaze
That includes understanding risk management and knowing what the risks and costs are of failures vs. the costs of delivering higher quality.
Engineering is about making the right tradeoffs given the constraints set, not about building the best possible product separate from the constraints.
Sometimes those constraints requires extreme quality, because it includes things like "this should never, ever fail", but most of the time it does not.
If it's firmware for a solar inverter in Poland, then quality matters.
That's typical misconception that "I'm an artist, let me rewrite in Rust" people often have. Code quality has a direct money equivalent, you just need to be able to justify it for people that pay you salary.
My son is in a CS school in France. They have finals with pen and paper, with no computer whatsoever during the exam; if they can't do that they fail. And these aren't multiple choice questions, but actual code that they have to write.
This was 30 years ago, though - no idea what it is like now. It didn't feel very meaningful even then.
But there's a vast chasm between that and letting people use AI in an exam setting. Some middle ground would be nice.
I wrote assembler on pages of paper. Then I used tables, and a calculator for the two's-complement relative negative jumps, to manually translate it into hex code. Then I had software to type in such hex dumps and save them to audio cassette, from which I could then load them for execution.
I did not have an assembler for my computer. I had a disassembler though- manually typed it in from a computer magazine hex dump, and saved it on an audio cassette. With the disassembler I could check if I had translated everything correctly into hex, including the relative jumps.
The planning required to write programs on sheets of paper was very helpful. I felt I got a lot dumber once I had a PC and actual programmer software (e.g. Borland C++). I found I was sitting in front of an empty code file without a plan more often than not, and wrote code moment to moment, immediately compiling and test running.
The AI coding may actually not be so bad if it encourages people to start with high-level planning instead of jumping into the IDE right away.
The only way to learn when abstractions are needed is to write code, hit a dead end, then try and abstract it. Over and over. With time, you will be able to start seeing these before you write code.
AI does not do abstractions well. From my experience, it completely fails to abstract anything unless you tell it to. Even when similar abstractions are already present. If you never learn when an abstraction is needed, how can you guide an AI to do the same well?
> Hell, I'd even like developers who will know when the code quality doesn't matter because shitty code will cost $2 a year but every hour they spend on it is $100-200.
> Except most of the actual computer science parts are gone, replaced with even more OOP, design pattern bullshit.
Maybe you should consider a different career, you sound pretty burnt out. There are terrible takes, especially for someone who is supposed to be fostering the next generation of developers.
In the US education has been bastardized into "job training"
Good workers don't really need to think in this paradigm.
> you give it a simple task. You’re impressed. So you give it a large task. You’re even more impressed.
That has _never_ been the story for me. I've tried, and I've got some good pointers and hints where to go and what to try, a result of LLM's extensive if shallow reading, but in the sense of concrete problem solving or code/script writing, I'm _always_ disappointed. I've never gotten satisfactory code/script result from them without a tremendous amount of pushback, "do this part again with ...", do that, don't do that.
Maybe I'm just a crank with too many preferences. But I hardly believe so. The minimum requirement should be for the code to work. It often doesn't. Feedback helps, right. But if you've got a problem where a simple, contained feedback loop isn't that easy to build, the only source of feedback is yourself. And that's when you are exposed to the stupidity of current AI models.
> There should be a TaskManager that stores Task objects in a sorted set, with the deadline as the sort key. There should be methods to add a task and pop the current top task. The TaskManager owns the memory when the Task is in the sorted set, and the caller to pop should own it after it is popped. To enforce this, the caller to pop must pass in an allocator and will receive a copy of the Task. The Task will be freed from the sorted set after the pop.
> The payload of the Task should be an object carrying a pointer to a context and a pointer to a function that takes this context as an argument.
> Update the tests and make sure they pass before completing. The test scenarios should relate to the use-case domain of this project, which is home automation (see the readme and nearby tests).
If that's the input needed, then I'd rather write code and rely on smarter autocomplete, so meanwhile I write the code and think about it, I can judge whether the LLM is doing what I mean to do, or straying away from something reasonable to write and maintain.
To me this reads like people have learned to put up with poor abstractions for so long that having the LLM take care of it feels like an improvement? It's the classic C++ vs Lisp discussion all over again, but people forgot the old lessons.
It's not that hard, but it's not that easy. If it was easy, everyone would be doing it. I'm a journalist who learned to code because it helped me do some stories that I wouldn't have done otherwise.
But I don't like to type out the code. It's just no fun to me to deal with what seem to me arbitrary syntax choices made by someone decades ago, or to learn new jargon for each language/tool (even though other languages/tools already have jargon for the exact same thing), or to wade through someone's undocumented code to understand how to use an imported function. If I had a choice, I'd rather learn a new human language than a programming one.
I think people like me, who (used to) code out of necessity but don't get much gratification out of it, are one of the primary targets of vibe coding.
I can spend all the time I want inside my ivory tower, hatching out plans and architecture, but the moment I start hammering letters in the IDE my watertight plan suddenly looks like Swiss cheese: constraints and edge cases that weren't accounted for during planning, flows that turn out to be unfeasible without a clunky implementation, etc...
That's why Writing code has become my favorite method of planning. The code IS the spec, and English is woefully insufficient when it comes to precision.
This makes Agentic workflows even worse because you'll only your architectural flaws much much later down the process.
You can't accurately plan every little detail in an existing codebase, because you'll only find out about all the edge cases and side effects when trying to work in it.
So, sure, you can plan what your feature is supposed to do, but your plan of how to do that will change the minute you start working in the codebase.
If you think through a problem as you're writing the code for it, you're going to end up the wrong creek because you'll have been furiously head down rowing the entire time, paying attention to whatever local problem you were solving or whatever piece of syntax or library trivia or compiler satisfaction game you were doing instead of the bigger picture.
Obviously, before starting writing, you could sit down and write a software design document that worked out the architecture, the algorithms, the domain model, the concurrency, the data flow, the goals, the steps to achieve it and so on; but the problem with doing that without an agent is then it becomes boring. You've basically laid out a plan ahead of time and now you've just got to execute on the plan, which means (even though you might even fairly often revise the plan as you learn unknown unknowns or iterate on the design) that you've kind of sucked all the fun and discovery out of the code rights process. And it sort of means that you've essentially implemented the whole thing twice.
Meanwhile, with a coding agent, you can spend all the time you like building up that initial software design document, or specification, and then you can have it implement that. Basically, you can spend all the time in your hammock thinking through things and looking ahead, but then have that immediately directly translated into pull requests you can accept or iterate on instead of then having to do an intermediate step that repeats the effort of the hammock time.
Crucially, this specification or design document doesn't have to remain static. As you would discover problems or limitations or unknown unknowns, you can revise it and then keep executing on it, meaning it's a living documentation of your overall architecture and goals as they change. This means that you can really stay thinking about the high level instead of getting sucked into the low level. Coding agents also make it much easier to send something off to vibe out a prototype or explore the code base of a library or existing project in detail to figure out the feasibility of some idea, meaning that the parts that traditionally would have been a lot of effort to verify that what your planning makes sense have a much lower activation energy. so you're more likely to actually try things out in the process of building a spec
The way I develop mirrors the process of creating said design document. I start with a high level overview, define what Entities the program should represent, define their attributes, etc... only now I'm using a more specific language than English. By creating a class or a TS interface with some code documentation I can use my IDEs capabilities to discover connections between entities.
I can then give the code to an LLM to produce a technical document for managers or something. It'll be a throwaway document because such documents are rarely used for actual decision making.
> Obviously, before starting writing, you could sit down and write a software design document that worked out the architecture, the algorithms, the domain model, the concurrency, the data flow, the goals, the steps to achieve it and so on;
I do this with code, and the IDE is much better than MS Word or whatevah at detecting my logical inconsistencies.
To me, reading the prompt example half a dozen levels up, reminds me of Greenspun's tenth rule:
> Any sufficiently complicated C++ program contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of Common Lisp. [1]
But now the "program" doesn't even have formal semantics and isn't a permanent artifact. It's like running a compiler and then throwing away the source program and only hand-editing the machine code when you don't like what it does. To me that seems crazy and misses many of the most important lessons from the last half-century.
[1]: https://en.wikipedia.org/wiki/Greenspun%27s_tenth_rule (paraphrased to use C++, but applies equally to most similar languages)
the problem is that you actually have to implement that high level DSL to get Lisp to look like that, and most DSLs are not going to be able to be as concise and abstract as a natural language description of what you want, and then just making sure it resulted in the right thing — which then I'd want to use AI for, to write that initial boilerplate, from a high level description of what the DSL should do.
And a Lisp macro DSL is not going to help with automating refactors, automatically iterating to take care of small compiler issues or minor bugs without your involvement so you can focus on the overall goal, remembering or discovering specific library APIs or syntax, etc.
I get my dopamine from solving problems, not trying to figure out why that damn API is returning the wrong type of field for three hours. Claude will find it out in minutes - while I do something else. Or from writing 40 slightly different unit tests to cover all the edge cases for said feature.
I still write my code in all the places I care about, but I don’t get stuck on “looking up how to enable websockets when creating the listener before I even pass anything to hyper.”
I do not care to spend hours or days to know that API detail from personal pain, because it is hyper-specific, in both senses of hyper-specific.
(For posterity, it’s `with_upgrades`… thanks chatgpt circa 12 months ago!)
But this is exactly what LLMs help me with! If I decide I want to shift the abstractions I'm using in a codebase in a big way, I'd usually be discouraged by all the error, lint, and warning chasing I'd need to do to update everything else; with agents I can write the new code (or describe it and have it write it) and then have it set off and update everything else to align: a task that is just varied and context specific enough that refactoring tools wouldn't work, but is repetitive and time consuming enough that it makes sense to pass off to a machine.
The thing is that it's not necessarily a bottleneck in terms of absolute speed (I know my editor well and I'm a fast typist, and LLMs are in their dialup era) but it is a bottleneck in terms of motivation, when some refactor or change in algorithm I want to make requires a lot of changes all over a codebase, that are boring to make but not quite rote enough to handle with sed or IDE refactoring. It really isn't, for me, even mostly about the inconvenience of typing out the initial code. It's about the inconvenience of trying to munge text from one state to another, or handle big refactors that require a lot of little mostly rote changes in a lot of places; but it's also about dealing with APIs or libraries where I don't want to have to constantly remind myself what functions to use, what to pass as arguments, what config data I need to construct to pass in, etc, or spend hours trawling through docs to figure out how to do something with a library when I can just feed its source code directly to an LLM and have it figure it out. There's a lot of friction and snags to writing code beyond typing that has nothing to do with having come up with a wrong abstraction, that very often lead to me missing the forest for the trees when I'm in the weeds.
Also, there is ALWAYS boilerplate scaffolding to do, even with the most macrotastic Lisp; and let's be real: Lisp macros have their own severe downsides in return for eliminating boilerplate, and Lisp itself is not really the best language (in terms of ecosystem, toolchain, runtime, performance) for many or most tasks someone like me might want to do, and languages adapted to the runtime and performance constraints of their domain may be more verbose.
Which means that, yes, we're using languages that have more boilerplate and scaffolding to do than strictly ideally necessary, which is part of why we like LLMs, but that's just the thing: LLMs give you the boilerplate eliminating benefits of Lisp without having to give up the massive benefits in other areas of whatever other language you wanted to use, and without having to write and debug macro soup and deal with private languages.
There's also how staying out of the code writing oar wells changes how you think about code as well:
If you think through a problem as you're writing the code for it, you're going to end up the wrong creek because you'll have been furiously head down rowing the entire time, paying attention to whatever local problem you were solving or whatever piece of syntax or library trivia or compiler satisfaction game you were doing instead of the bigger picture.
Obviously, before starting writing, you could sit down and write a software design document that worked out the architecture, the algorithms, the domain model, the concurrency, the data flow, the goals, the steps to achieve it and so on; but the problem with doing that without an agent is then it becomes boring. You've basically laid out a plan ahead of time and now you've just got to execute on the plan, which means (even though you might even fairly often revise the plan as you learn unknown unknowns or iterate on the design) that you've kind of sucked all the fun and discovery out of the code rights process. And it sort of means that you've essentially implemented the whole thing twice.
Meanwhile, with a coding agent, you can spend all the time you like building up that initial software design document, or specification, and then you can have it implement that. Basically, you can spend all the time in your hammock thinking through things and looking ahead, but then have that immediately directly translated into pull requests you can accept or iterate on instead of then having to do an intermediate step that repeats the effort of the hammock tim
And then you just rm -rf and repeat until something half works.
You don't even have to be as organised as in the example, LLMs are pretty good at making something out of ramblings.
I actually don't like _writing_ code, but enjoy reading it. So sessions with LLM are very entertaining, especially when I want to push boundaries (I am not liking this, the code seems a little bit bloated. I am sure you could simplify X and Y. Also think of any alternative way that you reckon will be more performant that maybe I don't know about). Etc.
This doesn't save me time, but makes work so much more enjoyable.
I think this is one of the divides between people who like AI and people who don't. I don't mind writing code per se, but I really don't like text editing — and I've used Vim (Evil mode) and then Emacs (vanilla keybindings) for years, so it's not like I'm using bad tools; it's just too fiddly. I don't like moving text around; munging control structures from one shape to another; I don't like the busy work of copying and pasting code that isn't worth DRYing, or isn't capable of being DRY'd effectively; I hate going around and fixing all the little compiler and linter errors produced by a refactor manually; and I really hate the process of filling out the skeleton of an type/class/whatever architecture in a new file before getting to the meat.
However, reading code is pretty easy for me, and I'm very good at quickly putting algorithms and architectures I have in my head into words — and, to be honest, I often find this clarifies the high level idea more than writing the code for it, because I don't get lost in the forest — and I also really enjoy taking something that isn't quite good enough, that's maybe 80% of the way there, and doing the careful polishing and refactoring necessary to get it to 100%.
> I think this is one of the divides between people who like AI and people who don't. I don't mind writing code per se, but I really don't like text editing — and I've used Vim (Evil mode) and then Emacs (vanilla keybindings) for years, so it's not like I'm using bad tools; it's just too fiddly.
I feel the same way (to at least some extent) about every language I've used other than Lisp. Lisp + Paredit in Emacs is the most pleasant code-wrangling experience I've ever had, because rather having to think in terms of characters or words, I'm able to think in terms of expressions. This is possible with other languages thanks to technologies like Tree-sitter, but I've found that it's only possible to do reliably in Lisp. When I do it in any other language I don't have an unshakable confidence that the wrangling commands will do exactly what I intend.
When I code, I mostly go by two perspectives: The software as a process and the code as a communication medium.
With the software as a process, I'm mostly thinking about the semantics of each expressions. Either there's a final output (transient, but important) or there's a mutation to some state. So the code I'm writing is for making either one possible and the process is very pleasing, like building a lego. The symbols are the bricks and other items which I'm using to create things that does what I want.
With the code as communication, I mostly take the above and make it readable. Like organizing files, renaming variables and functions, modularising pieces of code. The intent is for other people (including future me) to be able to understand and modify what I created in the easiest way possible.
So the first is me communicating with the machine, the second is me communicating with the humans. The first is very easy, you only need to know the semantics of the building blocks of the machine. The second is where the craft comes in.
Emacs (also Vim) makes both easy. Code has a very rigid structure and both have tools that let you manipulate these structure either for adding new actions or refine the shape for understanding.
With AI, it feels like painting with a brick. Or transmitting critical information through a telephone game. Control and Intent are lost.
On the other hand, most AI zealots (Steve Yegge comes readily to mind) don't care about what the code looks like. They never even see it.
Vehement agreeing below:
S-expressions are a massive boon for text editing, because they allow such incredible structural transformations and motions. The problem is that, personally, I don't actually find Lisp to be the best tool for the job for any of the things I want to do. While I find Common Lisp and to a lesser degree Scheme to be fascinating languages, the state of the library ecosystem, documentation, toolchain, and IDEs around them just aren't satisfactory to me, and they don't seem really well adapted to the things I want to do. And yeah, I could spend my time optimizing Common Lisp with `declare`s and doing C-FFI with it, massaging it to do what I want, that's not what I want to spend my time doing. I want to actually finish writing tools that are useful to me.
Moreover, while I used to have hope for tree-sitter to provide a similar level of structural editing for other languages, at least in most editors I've just not found that to be the case. There seem really to be two ways to use tree-sitter to add structural editing to languages: one, to write custom queries for every language, in order to get Vim style syntax objects, and two, to try to directly move/select/manipulate all nodes in the concrete syntax tree as if they're the same, essentially trying to treat tree-sitter's CSTs like S-expressions.
The problem with the first approach is that you end up with really limited, often buggy or incomplete, language support, and structural editing that requires a lot more cognitive overhead: instead of navigating a tree fluidly, you're having to "think before you act," deciding ahead of time what the specific name, in this language, is for the part of the tree you want to manipulate. Additionally, this approach makes it much more difficult to do more high level, interesting transformations; even simple ones like slurp and barf become a bit problematic when you're dealing with such a typed tree, and more advanced ones like convolute? Forget about it.
The problem with the second approach is that, if you're trying to do generalized tree navigation, where you're not up-front naming the specific thing you're talking about, but instead navigating the concrete syntax tree as if it's S-expressions, you run into the problem the author of Combobulate and Mastering Emacs talks about[1]: CSTs are actually really different from S-expressions in practice, because they don't map uniquely onto source code text; instead, they're something overlaid on top of the source code text, which is not one to one with it (in terms of CST nodes to text token), but many to one, because the CST is very granular. Which means that there's a lot of ambiguity in trying to understand where the user is in the tree, where they think they are, and where they intend to go.
There's also the fact that tree-sitter CSTs contain a lot of unnamed nodes (what I call "stop tokens"), where the delimiters for a node of a tree and its children are themselves children of that node, siblings with the actual siblings. And to add insult to injury, most language syntaces just... don't really lend themselves to tree navigation and transformation very well.
I actually tried to bring structural editing to a level equivalent to the S-exp commands in Emacs recently[2], but ran into all of the above problems. I recently moved to Zed, and while its implementation of structural editing and movement is better than mine, and pretty close to 1:1 with the commands available in Emacs (especially if they accept my PR[3]), and also takes the second, language-agnostic, route, it's still not as intuitive and reliable as I'd like.
[1]: https://www.masteringemacs.org/article/combobulate-intuitive...
You can’t deny the fact that someone like Ryan dhal creator of nodejs declared that he no longer writes code is objectively contrary to your own experience. Something is different.
I think you and other deniers try one prompt and then they see the issues and stop.
Programming with AI is like tutoring a child. You teach the child, tell it where it made mistakes and you keep iterating and monitoring the child until it makes what you want. The first output is almost always not what you want. It is the feedback loop between you and the AI that cohesively creates something better than each individual aspect of the human-AI partnership.
The only thing I would change about what you said is, I don’t see it as a child that needs tutoring. It feels like I’m outsourcing development to an offshore consultancy where we have no common understanding, except the literal meaning of words. I find that there are very, very many problems that are suited well enough to this arrangement.
I care about making stuff. "Making stuff" means stuff that I can use. I care about code quality yes, but not to an obsessive degree of "I hate my framework's ORM because of <obscure reason nobody cares about>". So, vibe coding is great, because I know enough to guide the agent away from issues or describe how I want the code to look or be changed.
This gets me to my desired effect of "making stuff" much faster, which is why I like it.
In real engineering disciplines, the Engineer is accountable for their work. If a bridge you signed off collapses, you're accountable and if it turns out you were negligent you'll face jail time. In Software, that might be a program in a car.
The Engineering mindset embodies these principles regardless of regulatory constraints. The Engineer needs to keep in mind those who'll be using their constructions. With Agentic Vibecoding, I can never get confident that the resulting software will behave according to specs. I'm worried that it'll scewover the user, the client, and all stakeholders. I can't accept half-assed work just because it saved me 2 days of typing.
I don't make stuff just for the sake of making stuff otherwise it would just be a hobby, and in my hobbies I don't need to care about anything, but I can't in good conscience push shit and slop down other people's throats.
If vibe coding delivers in one day, + an additional 2 days to solve stupid bugs, what you deliver with utter perfection in 3 months, then the industry doesn't give a shit about slop.
Is it maintainable? Well it's AI that's going to maintain it.
I think the future will turn into one where source code is like assembly code. Do you care about how your automated compiler system is spitting out assembly? Is the assembly code, neat and organized and maintainable? No. You don't care about assembly code. The industry is shifting in the direction where they don't care about ALL source code.
That's what's currently not possible, it might work in a small webapp or similar. But in a large system, it absolutely falls apart when having to maintain it. Sure, it can fix a bug, but it doesn't understand the side effects it creates with the fix, yet.
Maybe in the future that will also be possible. I do agree with you about business/management not caring about long term impacts if short term gains are possible.
Software Developers have long been completely disconnected from the consequences of their work, and tech companies have diluted responsibility so much that working software doesn't matter anymore. This field is now mostly scams and bullshit, where developers are closer to finance bros than real, actual Engineers.
I'm not talking about what someone os building in their home for personal reasons for their own usage, but about giving the same thing to other people.
In the end it's just cost cutting.
Who are you people who spend so much time writing code that this is a significant productivity boost?
I'm imagining doing this with an actual child and how long it would take for me to get a real return on investment at my job. Nevermind that the limited amount of time I get to spend writing code is probably the highlight of my job and I'd be effectively replacing that with more code reviews.
And maybe child is too simplistic of an analogy. It's more like working with a savant.
The type of thing you can tell AI to do is like this: You tell it to code a website... it does it, but you don't like the pattern.
Say, "use functional programming", "use camel-case" don't use this pattern, don't use that. And then it does it. You can leave it in the agent file and those instructions become burned into it forever.
That's all to say the learning curve with LLMs is how to say things a specific way to reliability get an outcome.
I recently inherited an over decade old web project full of EOL'd libraries and OS packages that desperately needed to be modernized.
Within 3 hours I had a working test suite with 80% code coverage on core business functionality (~300 tests). Now - maybe the tests aren't the best designs given there is no way I could review that many tests in 3 hours, but I know empirically that they cover a majority of the code of the core logic. We can now incrementally upgrade the project and have at least some kind of basic check along the way.
There's no way I could have pieced together as large of a working test suite using tech of that era in even double that time.
If you haven't reviewed and signed off then you have to assume that the stuff is garbage.
This is the crux of using AI to create anything and it has been a core rule of development for many years that you don't use wizards unless you understand what they are doing.
I'd say for what I'm trying to do - which is upgrade a very old version of PHP to something that is supported, this is completely acceptable. These are basically acting as smoke tests.
For God's sake that's completely slop.
There is obvious division of ideas here. But calling one side stupid or referring to them as charlatans is outright wrong and biased.
There is a reason why they struggle selling them and executives are force feeding them to their workers.
Charlatan is the perfect term for those that stand to make money selling half baked goods and forcing more mass misery upon society.
I think uncritical AI enthusiasts are just essentially making the bet that the rising mountains of tech debt they are leaving in their wake can be paid off later on with yet more AI. And you know, that might even work out. Until such a time, though, and as things currently stand, I struggle to understand how one can view raw LLM code and find it acceptable by any professional standard.
The same coworker asked to update a service to Spring Boot 4. She made a blog post about. She used LLM for it. So far every point which I read was a lie, and her workarounds make, for example tests, unnecessarily less readable.
So yeah, “it works”, until it doesn’t, and when it hits you, that you need to work more in sum at the end, because there are more obscure bugs, and fixing those are more difficult because of terrible readability.
There are many ways to skin a cat, and in programming the happens-in-a-digital-space aspect removes seemingly all boundaries, leading to fractal ways to "skin a cat".
A lot of programmers have hard heads and know the right way to do something. These are the same guys who criticized every other senior dev as being a bad/weak coder long before LLMs were around.
Your own profile says you are a PM whose software skills amount to "Script kiddie at best but love hacking things together."
It seems like the "separate worlds" you are describing is the impression of reviewing the code base from a seasoned engineer vs an amateur. It shouldn't be even a little surprising that your impression of the result is that the code is much better looking than the impression of a more experienced developer.
At least in my experience, learning to quickly read a code base is one of the later skills a software engineer develops. Generally only very experienced engineers can dive into an open source code base to answer questions about how the library works and is used (typically, most engineers need documentation to aid them in this process).
I mean, I've dabbled in home plumbing quite a bit, but if AI instructed me to repair my pipes and I thought it "looked great!" but an experienced plumber's response was "ugh, this doesn't look good to me, lots of issues here" I wouldn't argue there are "two separate worlds".
This really is it: AI produces bad to mediocre code. To someone who produces terrible code mediocre is an upgrade, but to someone who produces good to excellent code, mediocre is a downgrade.
You think it'll rapidly get smarter, but it just recreates things from all the terrible code it was fed. Code and how it is written also rapidly changes these days and LLMs have some trouble drawing lines between versions of things and the changes within them.
Sure, they can compile and test things now, which might make the code work and able to run. The quality of it will be hard to increase without manually controlling and limiting the type of code it 'learns' from.
Still bad code though
And by bad I'm not making a stylistic judgement. I mean it'll be hell to work with, easy to add bugs, and slow to change
Rather, to me it looks like all we're getting with additional time is marginal returns. What'll it be in 1 year? Marginally better than today, just like today is marginally better compared to a year ago. The exponential gains in performance are already over. What we're looking at now is exponentially more work for linear gains in performance.
The problem is the 0.05X developers thought they were 0.5X and now they think they're 20X.
Plenty of respect to the craft of code but the AI of today is the worst is is ever going to be.
That's all before you even get to all of the other quirks with LLMs.
Getting code to do exactly what, based on using and prompting Opus in what way?
Of course it works well for some things.
I have ones that describe what kinds of functions get unit vs integration tests, how to structure them, and the general kinds of test cases to check for (they love writing way too many tests IME). It has reduced the back and forth I have with the LLM telling it to correct something.
Usually the first time it does something I don't like, I have it correct it. Once it's in a satisfactory state, I tell it to write a Cursor rule describing the situation BRIEFLY (it gets way to verbose by default) and how to structure things.
That has made writing LLM code so much more enjoyable for me.
Because of Beads I can have Claude do a code review for serious bugs and issues and sure enough it finds some interesting things I overlooked.
I have also seen my peers in the reverse engineering field make breakthroughs emulating runtimes that have no or limited existing runtimes, all from the ground up mind you.
I think the key is thinking of yourself as an architect / mentor for a capable and promising Junior developer.
For example, someone may ask an LLM to write a simple http web server, and it can do that fine, and they consider that complex, when in reality its really not.
This is an extremely false statement.
There also seem to be people hearing big names like Karpathy and Linus Torvalds say they are vibe coding on their hobby projects, meaning who knows what, and misunderstanding this as being an endorsement of "magic genie" creation of professional quality software.
Results of course also vary according to how well what you are asking the AI to do matches what it was trained on. Despite sometimes feeling like it, it is not a magic genie - it is a predictor that is essentially trying to best match your input prompt (maybe a program specification) to pieces of what it was trained on. If there is no good match, then it'll have a go anyway, and this is where things tend to fall apart.
It seems clear that Karpathy himself is well aware of the difference between "vibe coding" as he defined it (which he explicitly said was for playing with on hobby projects), and more controlled productive use of AI for coding, which has either eluded him, or maybe his expectations are too high and (although it would be surprising) he has not realized the difference between the types of application where people are finding it useful, and use cases like his own that do not play to its strength.
You have to pick people with nothing to gain. https://x.com/rough__sea/status/2013280952370573666
You don't have to be bad at coding to use LLMs. The argument was specifically about thinking that LLMS can be great at accomplishing complex tasks (which they are not)
The simple case is that if you ask an agent to do a whole bunch of modifications across a large number of files, it often loses context due to context windows.
Now, you can make your own agents with custom mcp servers to basically improve its ability to do tasks, but then you are basically just building automation tools in the first place.
Great programmers wouldn't support or back AI if it couldn't handle complex tasks. AI can handle complex tasks inconsistently when operating on it's own. They can handle complex tasks consistently when pair programming with a human operator.
I hold a result of AI in front of your face and they still proclaim it’s garbage and everything else is fraudulent.
Let’s be clear. You’re arguing against a fantasy. Nobody even proponents of AI claims that AI is as good as humans. Nowhere near it. But they are good enough for pair programming. That is indisputable. Yet we have tons of people like you who stare at reality and deny it and call it fraudulent.
Examine the lay of the land if that many people are so divided it really means both perspectives are correct in a way.
Nobody smart is going to disagree that LLMs are a huge net positive. The finer argument is whether or not at this point you can just hand off coding to an LLM. People who say yes simply just haven't had enough experience with using LLMs to a large extent. The amount of time you have to spend prompt engineering the correct response is often the same amount of time it takes for you to write the correct code yourself.
And yes, you can put together AGENT.md files, mcp servers, and so on, but then it becomes a game of this. https://xkcd.com/1205/
Because the dirty secret is a lot of successful people aren't actually smart or talented, they just got lucky. Or they aren't successful at all, they're just good at pretending they are, either through taking credit for other people's work or flat out lying.
I've run into more than a few startups that are just flat out lying about their capabilities and several that were outright fraud. (See DoNotPay for a recent fraud example lol)
Pointing to anyone and going "well THEY do it, it MUST work" is frankly engineering malpractice. It might work. But unless you have the chops to verify it for yourself, you're just asking to be conned.
Steve Yegge (Veteran engineer, formerly Google and Amazon): A leading technical voice who describes vibe coding as acting as an orchestrator. He maintains that engineers who do not master "agentic engineering" and AI-driven workflows will be left behind as the industry moves toward "hyperproductivity".
Patrick Debois (Founder of DevOps): Often called the "godfather of DevOps," Debois now advocates for the "AI native developer". He views vibe coding as a high-level abstraction where the engineer's role shifts from a "producer" of lines of code to a "supervisor" of complex automated systems.
Simon Willison (Co-creator of Django): Recognized for his highly technical workflows that use AI to handle mechanical implementation while he focuses on rigorous documentation, tool coverage, and validation—a process often cited as the professional gold standard for vibe coding.
Stephen Blum (Founder/CTO of PubNub): A technical leader who has integrated generative coding into production-scale architecture. He characterizes the 2026 developer's role as directing agents for everything from database migrations to security audits rather than manually performing these tasks.
Gene Kim (Renowned DevOps researcher and author): Co-author of The Phoenix Project, Kim has publicly championed vibe coding as one of the most enjoyable technical experiences of his career, citing how it allows him to build sophisticated prototypes in minutes rather than days.
Patrick Debois (Founder of DevOps): Often referred to as the "godfather of DevOps," Debois advocates for "AI-native engineering". He views vibe coding as a mature abstraction layer where elite engineers focus on "system orchestration" rather than producing individual lines of code.
Geoffrey Huntley (Founder of the "Vibe Coding Academy"): A highly technical engineer known for pushing the boundaries of AI-driven development. He is a primary source for experimental techniques that use agents for everything from infrastructure to core logic.
Boris Cherny (Author of Programming TypeScript): An authority on type systems and engineering rigor, Cherny now provides deep technical guidance on how to integrate high-level intent with reliable, production-ready source code using tools like Claude Code.
Stephen Webb (UK CTO at Capgemini): A key industry figure declaring 2026 as the year "AI-native engineering goes mainstream". He supports vibe coding as a legitimate method for rewriting legacy systems and refactoring entire modules autonomously.
Linus Torvalds (Creator of Linux and Git): In a significant endorsement for the paradigm, Torvalds reported in early 2026 that he used Google Antigravity to vibe code a Python visualizer for his AudioNoise project. He noted in the project's documentation that the tool was "basically written by vibe-coding".
Theo Browne (Founder of Ping.gg, T3.gg): Known for his deep technical influence on the web development community, Browne is a primary educator for tools like Claude Code. He advocates for vibe coding as a way to bypass the "boring parts" of development, allowing engineers to focus on higher-level architecture and product logic.
McKay Wrigley (Developer and AI educator): A leading technical figure focused on structured tutorials and advanced workflows for agentic programming. He is widely followed by senior engineers seeking to move beyond simple chat interfaces into full-scale autonomous software generation.
Charlie Holtz (Software engineer and infrastructure specialist): Known for building advanced infrastructure tools, Holtz is recognized as an engineer "pushing the boundaries" of what can be built using vibe coding for complex, back-end systems.
Cian Clarke (Principal Engineer at NearForm): A veteran in the Node.js ecosystem who has transitioned toward spec-driven development. He advocates for "AI native engineering" where specialized agentic roles (such as security or performance agents) are orchestrated to build and refactor large-scale enterprise systems.
IndyDevDan (Senior developer and educator): A highly technical voice advocating for "deep mastery" of AI-assisted engineering. He focuses on teaching developers how to maintain rigorous engineering standards while leveraging the speed of vibe coding.
Mitchell Hashimoto (Founder of HashiCorp, Creator of Terraform): Now focused on his terminal project Ghostty, Hashimoto has become a leading voice on "pragmatic AI coding." In 2026, he detailed his workflow of using reasoning models (like o3) to generate comprehensive architecture plans before writing a single line of code. He argues this "learning accelerator" approach allows him to build outside his primary expertise (e.g., frontend) while maintaining strict engineering rigor by reviewing the output line-by-line.
Kent C. Dodds (Renowned Web Development Educator & Engineer): A highly influential figure in the React community, Dodds has fully embraced the paradigm, stating in 2026 that he has "never had so much fun developing software." He advocates for a "problem elimination" mindset where AI handles the implementation details, allowing senior engineers to focus entirely on user experience and application architecture.
Guillermo Rauch (CEO of Vercel, Creator of Next.js/Socket.io): Rauch has been a vocal proponent of vibe coding as the bridge between business logic and shipping software. He argues that vibe coding solves the "execution gap," enabling technical founders and engineers to ship complex products without getting bogged down in boilerplate, effectively treating the AI as a "junior engineer with infinite stamina" that requires high-level direction.
DHH (David Heinemeier Hansson) (Creator of Ruby on Rails & CTO at 37signals): Historically a skeptic of industry hype, DHH has acknowledged the "tipping point" in 2026, noting that agentic coding has become a viable tool for experienced developers to deliver on specs rapidly. His shift represents a major endorsement from the "craftsman" sector of the industry, validating that AI tools can coexist with high standards for code quality.
Rich Harris (Creator of Svelte): Harris has spoken about how AI-driven workflows liberate developers from "code preferences" and syntax debates. He views the 2026 landscape as one where the engineer's job is to focus on the "what" and "why" of a product, while AI increasingly handles the "how," allowing for a renaissance in creativity and shipping speed.
Addy Osmani (Engineering Manager for Chrome Web Platform): While deeply embedded in the browser ecosystem, Osmani has published extensively on his "AI-augmented" workflow in 2026. He characterizes the modern senior engineer not as a typist but as a "Director," whose primary skill is effectively guiding AI agents to execute complex engineering tasks while maintaining architectural integrity.
The above is just a smattering of individuals. I can keep going.
This is an orthogonal off topic point. My or anyones skills don't have to do with the topic at hand. The topic at hand is AI.
>Because the dirty secret is a lot of successful people aren't actually smart or talented, they just got lucky. Or they aren't successful at all, they're just good at pretending they are, either through taking credit for other people's work or flat out lying.
Again orthoganol to the point. But I'll entertain it. There's another class of people who are delusional. They think they're good, but they're not good at all. I've seen plenty of that in the industry. More so then people who lie, it's people who lie to themselves and believe it. Engineers so confident in their skills, but when I look at them I think they're raw dog shit.
>I've run into more than a few startups that are just flat out lying about their capabilities and several that were outright fraud. (See DoNotPay for a recent fraud example lol)
Again so?
>Pointing to anyone and going "well THEY do it, it MUST work" is frankly engineering malpractice. It might work. But unless you have the chops to verify it for yourself, you're just asking to be conned.
Of course. But it's idiotic when there is a huge population of people who are smarter than you better than you and proven to be more capable than you saying they can do it. I need to emphasize it's not just one person saying it. Tons and tons and tons of people are saying it.
Fraud happens in the margins of society it rarely ever happens at a macro level, and if it does happen at a macro level the trend doesn't last long and will mostly die within a year at most.
So when multitudes of highly reputed people are saying one thing, and your on the ground self verification of that thing is directly opposite of what they are saying. Then you need to re-evaluate your OWN verification. You need to investigate WHY there is a discrepancy, because it is utter stupidity to deny what others have seen as fraud and believe that your own judgements and verifications are flawless.
No offense, my dude, but your philosophy on this topic embodies the delusional stupidity I am talking about. People lie to themselves. That is the key metric here.
I don't need to explain ANY of this to you. You know it, because every explanation I just gave is an OBVIOUS facet of life in general. It needs to be explained to someone like you despite it's obviousness because of self delusion.
If we're being honest with ourselves, Opus 4.5 / GPT 5.2 etc are maybe 10-20% better than GPT 3.5 at most. It's a total and absolute catastrophic failure that will go down in history as one of humanity's biggest mistakes.
His tweets were getting ~40k views average. He made his big proclamation about AI and boom viral 7 million
This is happening over, and over, and over again
I'm not saying he's making shit up but you're naive if you don't think they're slightly tempted by the clear reaction this content gets
A complete exercise in frustration that has turned me off of all agentic code bullshit. The only reason I still have Claude Code installed is because I like the `/multi-commit` skill I made.
These cases are common enough to where it's more systemic than isolated.
I read these comments and articles and feel like I am completely disconnected from most people here. Why not use GenAI the way it actually works best: like autocomplete on steroids. You stay the architect, and you have it write code function by function. Don't show up in Claude Code or Codex asking it to "please write me GTA 6 with no mistakes or you go to jail, please."
It feels like a lot of people are using GenAI wrong.
That argument doesn’t fly when the sellers of the technology literally sing at you “there’s no wrong way to prompt”.
That's exactly the point. Modern coding agents aren't smart software engineers per se; they're very very good goal-seekers whose unit of work is code. They need automatable feedback loops.
It would correctly modify a single method. I would ask it to repeat for next and it would fail.
The code that our contractors are submitting is trash and very high loc. When you inspect it you can see that unit tests are testing nothing of value.
when(mock.method(foo)).thenReturn(bar)
assert(bar == bar)
its all fake coverage, for fake tests, for fake OKRs
what are people actually getting done? I've sat next to our top evangelist for 30 minutes pair programming and he just fought the tool saying something was wrong with the db while showing off some UI I dont care about.
like that seems to be the real issue to me. i never bother wasting time with UI and just write a tool to get something done. but people seem impressed that AI did some shitty data binding to a data model that cant do anything, but its pretty.
it feels weird being an avowed singularitarian but adamant that these tools suck now.
You try a gamut of sample inputs and observe where its going awry? Describe the error to it and see what it does
A trivial example is your happy path git workflow. I want:
- pull main
- make new branch in user/feature format
- Commit, always sign with my ssh key
- push
- open pr
but it always will
- not sign commits
- not pull main
- not know to rebase if changes are in flight
- make a million unnecessary commits
- not squash when making a million unnecessary commits
- have no guardrails when pushing to main (oops!)
- add too many comments
- commit message too long
- spam the pr comment with hallucinated test plans
- incorrectly attribute itself as coauthor in some gorilla marketing effort (fixable with config, but whyyyyyy -- also this isn't just annoying, it breaks compliance in alot of places and fundamentally misunderstands the whole point of authorship, which is copyright --- and AIs can't own copyright )
- not make DCO compliant commits ...
Commit spam is particularly bad for bisect bug hunting and ref performance issues at scale. Sure I can enforce Squash and Merge on my repo but why am I relying on that if the AI is so smart?
All of these things are fixed with aliases / magit / cli usage, using the thing the way we have always done it.
Because it's not? I use these things very extensively to great effect, and the idea that you'd think of it as "smart" is alien to me, and seems like it would hurt your ability to get much out of them.
Like, they're superhuman at breadth and speed and some other properties, but they don't make good decisions.
Yet. Most of my criticism is not after running the code, but after _reading_ the code. It wrote code. I read it. And I am not happy with it. No even need to run it, it's shit at glance.
"SELECT * FROM table WHERE id=1;"
$result = $db->query("SELECT * FROM table;");
for ($row in $result)
if ($["id"] == 1)
return $row;
It sounds like you know what the problem with your AI workflow is? Have you tried using an agent? (sorry somewhat snarky but… come on)
The thing that differentiates LLM's from my stupid but cute vacuum cleaner, is that the (at least OpenAI's) AI model is cocksure and wrong, which is infinitely more infuriating than being a bit clueless and wrong.
ETA: I've probably gotten 10k worth of junior dev time out of it this month.
Im not crazy about signing up for a subscription service, it depends on you remembering to cancel and not have a headache when you do cancel.
My theory is that the people who are impressed are trying to build CRUD apps or something like that.
"It Is Difficult to Get a Man to Understand Something When His Salary Depends Upon His Not Understanding It"
...might be my issue indeed. Trying to balance it by not being too stubborn though. I'm not doing AI just to be able to dump on them, you know.
They're certainly not perfect, but many of the issues that people post about as though they're show-stoppers are easily resolved with the right tools and prompting.
A lot of the failures people talk about seem to involve expecting the models to one-shot fairly complex requirements.
It is hands down good for code which is laborious or tedious to write, but once done, obviously correct or incorrect (with low effort inspection). Tests help but only if the code comes out nicely structured.
I made plenty of tools like this, a replacement REPL for MS-SQL, a caching tool in Python, a matplotlib helper. Things that I know 90% how to write anyway but don't have the time, but once in front of me, obviously correct or incorrect. NP code I suppose.
But business critical stuff is rarely like this, for me anyway. It is complex, has to deal with various subtle edge cases, be written defensively (so it fails predictably and gracefully), well structured etc. and try as I might, I can't get Claude to write stuff that's up to scratch in this department.
I'll give it instructions on how to write some specific function, it will write this code but not use it, and use something else instead. It will pepper the code with rookie mistakes like writing the same logic N times in different places instead of factoring it out. It will miss key parts of the spec and insist it did it, or tell me "Yea you are right! Let me rewrite it" and not actually fix the issue.
I also have a sense that it got a lot dumber over time. My expectations may have changed of course too, but still. I suspect even within a model, there is some variability of how much compute is used (eg how deep the beam search is) and supply/demand means this knob is continuously tuned down.
I still try to use Claude for tasks like this, but increasingly find my hit rate so low that the whole "don't write any code yet, let's build a spec" exercise is a waste of time.
I still find Claude good as a rubber duck or to discuss design or errors - a better Stack Exchange.
But you can't split your software spec into a set of SE questions then paste the code from top answers.
> It is hands down good for code which is laborious or tedious to write, but once done, obviously correct or incorrect (with low effort inspection).
The problem here is, that it fills in gaps that shouldn't be there in the first place. Good code isn't laborious. Good code is small. We learn to avoid unnecessary abstractions. We learn to minimize "plumbing" such that the resulting code contains little more than clear and readable instructions of what you intend for the computer to do.
The perfect code is just as clear as the design document in describing the intentions, only using a computer language.
If someone is gaining super speeds by providing AI clear design documents compared to coding themselves, maybe they aren't coding the way they should.
My biggest LLM success resulted in something operationally correct but was something that I would never want to try to modify. The LLM also had an increasingly difficult time adding features.
Meanwhile my biggest 'manual' successes have resulted in something that was operationally correct, quick to modify, and refuses to compile if you mess anything up.
The only thing I think I learned from some of those exchanges was that xslt adherents are approximately as vocal as lisp adherents.
I still use it from time to time for config files that a developer has to write. I find it easier to read that JSON, and it supports comments. Also, the distinction between attributes and children is often really nice to have. You can shoehorn that into JSON of course, but native XML does it better.
Obviously, I would never use it for data interchange (e.g. SOAP) anymore.
Well, those comments were arguing about how it is the absolute best for data interchange.
> I still use it from time to time for config files that a developer has to write.
Even back when XML was still relatively hot, I recalled thinking that it solved a problem that a lot of developers didn't have.
Because if, for example, you're writing Python or Javascript or Perl, it is dead easy to have Python or Javascript or Perl also be your configuration file language.
I don't know what language you use, but 20 years ago, I viewed XML as a Java developer's band-aid.
Sure. Like C header files. It's the easiest option - no arguments there.
But there are considerations beyond being easy. I think there's a case to be made that a config file should be data, not code.
If people are really technical, then a language subset is fine.
If they're not really technical, then you might need a separate utility to manipulate the config file, and XML is OK if you need a separate utility. There are readers/writers available in every language, and it's human readable enough for debugging, but if a non-technical human mistakenly edits it, it might take some repair to make it usable again.
Even if you've decided on a separate config language, there are a lot of reasons why you might want to use something other than XML. The header/key/value system (e.g. the one that .gitconfig and a lot of /etc files use) remains popular.
I could be wrong, but it always seemed to me that XML was pushed as a doc/interchange format, and its use in config files was driven by "I already have this hammer and I know how to use it."
I don't how much scope realistically there is for writing these kinds of code nicely.
You can't dispense with yourself in those scenarios. You have to read, think, investigate, break things down into smaller problems. But I employ LLM's to help with that all the time.
Granted, that's not vibe coding at all. So I guess we are pretty much in agreement up to this point. Except I still think LLMs speed up this process significantly, and the models and tools are only going to get better.
Also, there are a lot of developers that are just handed the implementation plan.
That's your job.
The great thing about coding agents is that you can tell them "change of design: all API interactions need to go through a new single class that does authentication and retries and rate-limit throttling" and... they'll track down dozens or even hundreds of places that need updating and fix them all.
(And the automated test suite will help them confirm that the refactoring worked properly, because naturally you had them construct an automated test suite when they built those original features, right?)
Going back to typing all of the code yourself (my interpretation of "writing by hand") because you don't have the agent-managerial skills to tell the coding agents how to clean up the mess they made feels short-sighted to me.
I dunno, maybe I have high standards but I generally find that the test suites generated by LLMs are both over and under determined. Over-determined in the sense that some of the tests are focused on implementation details, and under-determined in the sense that they don't test the conceptual things that a human might.
That being said, I've come across loads of human written tests that are very similar, so I can see where the agents are coming from.
You often mention that this is why you are getting good results from LLMs so it would be great if you could expand on how you do this at some point in the future.
Or I can say "use pytest-httpx to mock the endpoints" and Claude knows what I mean.
Keeping an eye on the tests is important. The most common anti-pattern I see is large amounts of duplicated test setup code - which isn't a huge deal, I'm much more more tolerant of duplicated logic in tests than I am in implementation, but it's still worth pushing back on.
"Refactor those tests to use pytest.mark.parametrize" and "extract the common setup into a pytest fixture" work really well there.
Generally though the best way to get good tests out of a coding agent is to make sure it's working in a project with an existing test suite that uses good patterns. Coding agents pick the existing patterns up without needing any extra prompting at all.
I find that once a project has clean basic tests the new tests added by the agents tend to match them in quality. It's similar to how working on large projects with a team of other developers work - keeping the code clean means when people look for examples of how to write a test they'll be pointed in the right direction.
One last tip I use a lot is this:
Clone datasette/datasette-enrichments
from GitHub to /tmp and imitate the
testing patterns it uses
Yeah, this is where I too have seen better results. The worse ones have been in places where it was greenfield and I didn't have an amazing idea of how to write tests (a data person working on a django app).
Thanks for the information, that's super helpful!
I am not sure why, but it kept trying to do that, although I made several attempts.
Ended up writing it on my own, very odd. This was in Cursor, however.
If you start with an example file of tests that follow a pattern you like, along with the code the tests are for, it's pretty good at following along. Even adding a sentence to the prompt about avoiding tautological tests and focusing on the seams of functions/objects/whatever (integration tests) can get you pretty far to a solid test suite.
Another agent reviews the tests, finds duplicate code, finds poor testing patterns, looks for tests that are only following the "happy path", ensures logic is actually tested and that you're not wasting time testing things like getters and setters. That agent writes up a report.
Give that report back to the agent that wrote the test or spin up a new agent and feed the report to it.
Don't do all of this blindly, actually read the report to make sure the llm is on the right path. Repeat that one or two times.
Just writing one line in CLAUDE.md or similar saying "don't test library code; assume it is covered" works.
Half the battle with this stuff is realizing that these agents are VERY literal. The other half is paring down your spec/token usage without sacrificing clarity.
Just like anything else in software, you have to iterate. The first pass is just to thread the needle.
I don't get it. I have insanely high standards so I don't let the LLM get away with not meeting my standards. Simple.
Incidentally, I wonder if anyone has used LLMs to generate complex test scenarios described in prose, e.g. “write a test where thread 1 calls foo, then before hitting block X, thread 2 calls bar, then foo returns, then bar returns” or "write a test where the first network call Framework.foo makes returns response X, but the second call returns error Y, and ensure the daemon runs the appropriate mitigation code and clears/updates database state." How would they perform in this scenario? Would they add the appropriate shims, semaphores, test injection points, etc.?
> [Agents write] units of changes that look good in isolation.
I have only been using agents for coding end-to-end for a few months now, but I think I've started to realise why the output doesn't feel that great to me.
Like you said; "it's my job" to create a well designed code base.
Without writing the code myself however, without feeling the rough edges of the abstractions I've written, without getting a sense of how things should change to make the code better architected, I just don't know how to make it better.
I've always worked in smaller increments, creating the small piece I know I need and then building on top of that. That process highlights the rough edges, the inconsistent abstractions, and that leads to a better codebase.
AI (it seems) decides on a direction and then writes 100s of LOC at one. It doesn't need to build abstractions because it can write the same piece of code a thousand times without caring.
I write one function at a time, and as soon I try to use it in a different context I realise a better abstraction. The AI just writes another function with 90% similar code.
We expect the spec writing and prompt management to cover the "work smarter" bases, but part of the work smarter "loop" is hitting those points where "work harder" is about to happen, where you know you could solve a problem with 100s or 1000s of lines of code, pausing for a bit, and finding the smarter path/the shortcut/the better abstraction.
I've yet to see an "agentic loop" that works half as well as my well trained "work smarter loop" and my very human reaction to those points in time of "yeah, I simply don't want to work harder here and I don't think I need hundreds more lines of code to handle this thing, there has to be something smarter I can do".
In my opinion, the "best" PRs delete as much or more code than they add. In the cleanest LLM created PRs I've never seen an LLM propose a true removal that wasn't just "this code wasn't working according to the tests so I deleted the tests and the code" level mistakes.
I increasingly feel a sort of "guilt" when going back and forth between agent-coding and writing it myself. When the agent didn't structure the code the way I wanted, or it just needs overall cleanup, my frustration will get the best of me and I will spend too much time writing code manually or refactoring using traditional tools (IntelliJ). It's clear to me that with current tooling some of this type of work is still necessary, but I'm trying to check myself about whether a certain task really requires my manual intervention, or whether the agent could manage it faster.
Knowing how to manage this back and forth reinforces a view I've seen you espouse: we have to practice and really understand agentic coding tools to get good at working with them, and it's a complete error to just complain and wait until they get "good enough" - they're already really good right now if you know how to manage them.
> So I’m back to writing by hand for most things. Amazingly, I’m faster, more accurate, more creative, more productive, and more efficient than AI, when you price everything in, and not just code tokens per hour
At least he said "most things". I also did "most things" by hand, until Opus 4.5 came out. Now it's doing things in hours I would have worked an entire week on. But it's not a prompt-and-forget kind of thing, it needs hand holding.
Also, I have no idea _what_ agent he was using. OpenAI, Gemini, Claude, something local? And with a subscription, or paying by the token?
Because the way I'm using it, this only pays off because it's the 200$ Claude Max subscription. If I had to pay for the token (which once again: are hugely marked up), I would have been bankrupt.
"vibe coding" didn't really become real until 2025, so how were they vibe coding for 2 years? 2 years ago I couldn't count on an llm to output JSON consistently.
Overall the article/video are SUPER ambiguous and frankly worthless.
I remember being amazed and at the time thinking the game had changed. But I've never been able to replicate it since. Even the latest and greatest models seem to always go off and do something stupid that it can't figure out how to recover from without some serious handholding and critique.
LLMs are basically slot machines, though, so I suppose there has always been a chance of hitting the jackpot.
No, that isn't. To quote your own blog, his job is to "deliver code [he's] proven to work", not to manage AI agents. The author has determined that managing AI agents is not an effective way to deliver code in the long term.
> you don't have the agent-managerial skills to tell the coding agents how to clean up the mess they made
The author has years of experience with AI assisted coding. Is there any way we can check to see if someone is actually skilled at using these tools besides whether they report/studies measure that they do better with them than without?
Or those skills are a temporary side effect of the current SOTA and will be useless in the future, so honing them is pointless right now.
Agents shouldn't make messes, if they did what they say on the tin at least, and if folks are wasting considerable time cleaning them up, they should've just written the code themselves.
Exactly.
AI assisted development isn't all or nothing.
We as a group and as individuals need to figure out the right blend of AI and human.
Vibe coding is the extreme end of using AI, while handwriting is the extreme end of not using AI. The optimal spot is somewhere in the middle. Where exactly that spot is, I think is still up for debate. But the debate is not progressed in any way by latching on to the extremes and assuming that they are the only options.
Because when I see people that are downplaying LLMs or the people describing their poor experiences it feels like they're trying to "vibe code" but they expect the LLM to automatically do EVERYTHING. They take it as a failure that they have to tell the LLM explicitly to do something a couple times. Or they take it as a problem that the LLM didn't "one shot" something.
But I'm thankful for you devs that are giving me job security.
But that said, there are still plenty of ops-y situations where AI can be very helpful. Even just "here's a 125k lines of prod logs. Can you tell me what is going wrong?" has saved me lots of time in the past, especially for apps that I'm not super familiar with. It's (sometimes) pretty good at finding the needle in the haystack. The most common workflow I have now is to point an agent at it and while it's griding on it I'll do some hand greps and things. I've gotten to the bottom of some really tricky things much faster because of it. Sometimes it points me in the wrong direction (for example, one time it noticed that we were being rate-limited by the Cloudflare API, and instead of adding a single flag to the library calls it wrote it's own very convoluted queue system. But it was still helpful because at least it pinpointed the problem).
The other "small pieces of software" I find it very helpful for are bash functions or small scripts to do things. The handwritten solution is usually quick, but rarely as resilient/informative as it could be because writing a bunch of error handling can 5x or 10x the handwritten time. I will usually write the quick version, then point AI at it and have it add arg passing/handling, error handling, and usage info/documentation. It's been great for that.
For example, most political flamefests
> AI assisted development isn't all or nothing.
> We as a group and as individuals need to figure out the right blend of AI and human.
"We as a group need to figure out the right blend of strong static and weak dynamic typing."
One can look around and see where that old discussion brought us. In my opinion, nowhere, things are same as they were.
So, where will LLM-assisted coding bring us? By rhyming it with the static types, I see no other variants than "nowhere."
For small projects, I don’t think it makes a huge difference.
But for large projects, I’d guess that most die-hard dynamic people who have tried typescript have now seen the light and find lots of benefits to static typing.
My own experience suggest that if you need to develop heavily multithreaded application, you should use Haskell and you need some MVars if you are working alone and you need software transactional memory (STM) if you are working as part of a team, two and more people.
STM makes stitching different parts of the parallel program together as easy as just writing sequential program - sequential coordination is delegated to STM. But, STM needs control of side effects, one should not write a file inside STM transaction, only before transaction is started or after transaction is finished.
Because of this, C#, F#, C++, C, Rust, Java and most of programming languages do not have a proper STM implementation.
For controlling (and combining) (side) effects one needs higher order types and partially instantiated types. These were already available in Haskell (ghc 6.4, 2005) at the time Rust was conceived (2009), for four years.
Did Rust do anything to have these? No. The authors were a little bit too concerned to reimplement what Henry Baker did at the beginning of 1990-s, if not before that.
Do Rust authors have plans to implement these? No, they have other things to do urgently to serve community better. As if making complex coordination of heavily parallel programs is not a priority at all.
This is where I get my "rhyme" from.
Like people in here complaining about how poor the tests are... but did they start another agent to review the tests? Did they take that and iterate on the tests with multiple agents?
I can attest that the first pass of testing can often be shit. That's why you iterate.
So far, by the time I’m done iterating, I could have just written it myself. Typing takes like no time at all in aggregate. Especially with AI assisted autocomplete. I spend far more time reading and thinking (which I have to do to write a good spec for the AI anyways).
There's been such a massive leap in capabilities since claude code came out, which was middle/end of 2025.
2 years ago I MAYBE used an LLM to take unstructured data and give me a json object of a specific structure. Only about 1 year ago did I start using llms for ANY type of coding and I would generally use snippets, not whole codebases. It wasn't until September when I started really leveraging the LLM for coding.
I shipped a small game that way (https://love-15.com/) -- one that I've wished to make for a long time but wouldn't have been worth building other wise. It's tiny, really, but very niche -- despite being tiny, I hit brick walls multiple times vibing it, and had to take a few brief breaks from vibing to get it unstuck.
Claude Code was a step change after that, along with model upgrades, about 9 months ago. That size project has been doable as a vibe coded project since then without hitting brick walls.
All this to say I really doubt most claims about having been vibe coding for more than 9-15 months.
Now I expect to start seeing job postings asking for "3 years of experience vibe coding"
It's used more broadly now, but still to refer to the opposite end of the spectrum of AI-assisted coding to what you described.
Best case is still operationally correct but nightmare fuel on the inside. So maybe good for one off tools where you control inputs and can vibe check outputs without diaster if you forget to carry the one.
Well yea, but you can guard against this in several ways. My way is to understand my own codebase and look at the output of the LLM.
LLMs allow me to write code faster and it also gives a lot of discoverability of programming concepts I didn't know much about. For example, it plugged in a lot of Tailwind CSS, which I've never used before. With that said, it does not absolve me from not knowing my own codebase, unless I'm (temporarily) fine with my codebase being fractured conceptually in wonky ways.
I think vibecoding is amazing for creating quick high fidelity prototypes for a green field project. You create it, you vibe code it all the way until your app is just how you want it to feel. Then you refactor it and scale it.
I'm currently looking at 4009 lines of JS/JSX combined. I'm still vibecoding my prototype. I recently looked at the codebase and saw some ready made improvements so I did them. But I think I'll start to need to actually engineer anything once I reach the 10K line mark.
Then you are not vibe coding. The core, almost exclusive requirement for "vibe coding" is that you DON'T look at the code. Only the product outcome.
You don't even look at the diffs. You just yolo the code.
> It’s not until I opened up the full codebase and read its latest state cover to cover that I began to see what we theorized and hoped was only a diminishing artifact of earlier models: slop.
This is true vibe coding, they exclusively interacted with the project through the LLM, and only looked at its proposed diffs in a vacuum.
If they had been monitoring the code in aggregate the entire time they likely would have seen this duplicative property immediately.
> What’s worse is code that agents write looks plausible and impressive while it’s being written and presented to you. It even looks good in pull requests (as both you and the agent are well trained in what a “good” pull request looks like).
Which made me think that they were indeed reading at least some of the code - classic vibe coding doesn't involve pull requests! - but weren't paying attention to the bigger picture / architecture until later on.
Is it a skill for the layman?
Or does it only work if you have the understanding you would need to manage a team of junior devs to build a project.
I feel like we need a different term for those two things.
Programming together with AI however, is a skill, mostly based on how well you can communicate (with machines or other humans) and how well your high-level software engineering skills are. You need to learn what it can and cannot do, before you can be effective with it.
I call the act of using AI to help write code that you review, or managing a team of coding agents "AI-assisted programming", but that's not a snappy name at all. I've also skirted around the idea of calling it "vibe engineering" but I can't quite bring myself to commit to that: https://simonwillison.net/2025/Oct/7/vibe-engineering/
I think we need another term for using an LLM to write code but absolutely not forgetting the code exists.
I often use LLMs to do refactoring and, by definition, refactoring cannot be vibe-coding because that's caring about the code.
That is now what software engineering is.
Normally I'd know 100% of my codebase, now I understand 5% of it truly. The other 95% I'd need to read it more carefully before I daresay I understand it.
I agree there is a spectrum, and all the way to the left you have "vibe coding" and all the way to the right you have "manual programming without AI", of course it's fine to be somewhere in the middle, but you're not doing "vibe coding" in the way Karpathy first meant it.
This is the bit I think enthusiasts need to argue doesn't apply.
Have you ever read a 200 page vibewritten novel and found it satisfying?
So why do you think a 10 kLoC vibecoded codebase will be any good engineering-wise?
I've been coding a side-project for a year with full LLM assistance (the project is quite a bit older than that).
Basically I spent over a decade developing CAD software at Trimble and now have pivoted to a different role and different company. So like an addict, I of course wanted to continue developing CAD technology.
I pretty much know how CAD software is supposed to work. But it's _a lot of work_ to put together. With LLMs I can basically speedrun through my requirements that require tons of boilerplate.
The velocity is incredible compared to if I would be doing this by hand.
Sometimes the LLM outputs total garbage. Then you don't accept the output, and start again.
The hardest parts are never coding but design. The engineer does the design. Sometimes I pain weeks or months over a difficult detail (it's a sideproject, I have a family etc). Once the design is crystal clear, it's fairly obvious if the LLM output is aligned with the design or not. Once I have good design, I can just start the feature / boilerplate speedrun.
If you have a Windows box you can try my current public alpha. The bugs are on me, not on the LLM:
https://github.com/AdaShape/adashape-open-testing/releases/t...
I shared the app because it’s not confidential and it’s concrete - I can’t really discuss work stuff without stressing out what I can share and what not.
At least in my workplace everyone I know is using Claude Code or Cursor.
Now, I don’t know why some people are productive with tools and some aren’t.
But the code generation capabilities are for real.
About the project itself, do you plan to open source if eventually? LLM discussion aside, I've long been frustrated by the lack of a good free desktop 3D CAD software.
I would love to build this eventually to a real product so am not currently considering open sourcing it.
I can give you a free foreverlicense if you would like to be an alpha tester though :) - but am considering in any case for the eventual non-commercial licenses to be affordable&forever.
IMHO what the world needs is a good textbook on how to build CAD software. Mäntylä’s ”Solid modeling” is almost 40 years old. CAD itself is pushing 60-70 years.
The highly non-trivial parts in my app are open source software anyways (you can check the attribution file) and what this contributes is just a specific, opinionated way of how a program like this should work in 2020’s.
What I _would_ like to eventually contribute is a textbook in how to build something like this - and after that re-implementation would be a matter of some investment to LLM inference, testing, and end-user empathy. But that would have to wait either for my financial independence, AI-communism or my retirement :)
Thank you!
—
I would never use, let alone pay for, a fully vibe-coded app whose implementation no human understands.
Whether you’re reading a book or using an app, you’re communicating with the author by way of your shared humanity in how they anticipate what you’re thinking as you explore the work. The author incorporates and plans for those predicted reactions and thoughts where it makes sense. Ultimately the author is conveying an implicit mental model (or even evoking emotional states or sensations) to the reader.
The first problem is that many of these pathways and edge cases aren’t apparent until the actual implementation, and sometimes in the process the author realizes that the overall product would work better if it were re-specified from the start. This opportunity is lost without a hands on approach.
The second problem is that, the less human touch is there, the less consistent the mental model conveyed to the user is going to be, because a specification and collection of prompts does not constitute a mental model. This can create subconscious confusion and cognitive friction when interacting with the work.
We’re moving into a world where suboptimal code doesn’t matter that much because it’s so cheap to produce.
Have you ever read a 200 page vibewritten novel and found it satisfying?
(The model was asked to generate a chapter at a time. At each step, it was given the full outline of the novel, the characters, and a summary of each chapter so far.)
Did the model also come up with the idea for the novel, the characters, the outline?
For the other, my son wrote ~200 words total describing the story idea and the characters.
In each case, the model created the detailed outline and did all the writing.
I suspect part of the reason we see such a wide range of testimonies about vibe-coding is some people are actually better at it, and it would be useful to have some way of measuring that effectiveness.
If you’re writing novel algorithms all day, then I get your point. But are you? Or have you ever delegated work? If you find the AI losing its train of thought all it takes is to try again with better high level instructions.
It wasn't fully autonomous (the reliability was a bit low -- e.g. had to get the code out of code fences programmatically), and it wasn't fully original (I stole most of it from Auto-GPT, except that I was operating on the AST directly due to the token limitations).
My key insight here was that I allowed GPT to design the apis that itself was going to use. This makes perfect sense to me based on how LLMs work. You tell it to reach for a function that doesn't exist, and then you ask it to make it exist based on how it reached for it. Then the design matches its expectations perfectly.
GPT-4 now considers self modifying AI code to be extremely dangerous and doesn't like talking about it. Claude's safety filters began shutting down similar conversations a few months ago, suggesting the user switch to a dumber model.
It seems the last generation or two of models passed some threshold regarding self replication (which is a distinct but highly related concept), and the labs got spooked. I haven't heard anything about this in public though.
Edit: It occurs to me now that "self modification and replication" is a much more meaningful (and measurable) benchmark for artificial life than consciousness is...
BTW for reference the thing that spooked Claude's safety trigger was "Did PKD know about living information systems?"
I speculate that this has more to do with recent high-profile cases of self harm related to "AI psychosis" than any AGI-adjacent danger. I've read a few of the chat transcripts that have been made public in related lawsuits, and there seems to be a recurring theme of recursive or self-modifying enlightenment role-played by the LLM. Discouraging exploration of these themes would be a logical change by the vendors.
When some people say vibe coding, they mean they're copy-pasting snippets of code from ChatGPT.
When some people say vibe coding, they give a one sentence prompt to their cluster of Claude Code instances and leave for a road trip!
You don't need a "fully agentic" tool like Claude Code to write code. Any of the AI chatbots can write code too, obviously doing so better since the advent of "thinking" models, and RL post-training for coding. They also all have had built-in "code interpreter" functionality for about 2 years where they can not only write code but also run and test it in a sandbox, at least for Python.
Recently at least, the quality of code generation (at least if you are asking for something smallish) is good enough that cut and pasting chatbot output (e.g. C++, not Python) to compile and run yourself is still a productivity boost, although this was always an option.
Just more FUD from devs that think they're artisans.
Karpathy absolutely did not invent coding with an LLM. So yes OP can easily have been doing it for 2 years.
Stop worshipping rich kids.
I actually invented what you all know as 4o and Gemini, that’s how that charlatan even started looking at me.
On a personal note, vibe coding leaves me with that same empty hollow sort of tiredness, as a day filled with meetings.
And as a added benefit: I feel accomplished and proud of the feature.
We need to find the Goldilocks optimal level of AI assistance that doesn't leave everyone hating their jobs, while still boosting productivity.
I also like to think that I'm utilising the training done on many millions of lines of code while still using my experience/opinions to arrive at something compared to just using my fallible thinking wherein I could have missed some interesting ideas. Its like me++. Sure, it does a lot of heavy lifting but I never leave the steering wheel. I guess I'm still at the pre-agentic stage and not ready to letting go fully.
I’m not sure if this counts as “vibe coding” per se, but I like that this mentality keeps my workday somewhat similar to how it was for decades. Finding/creating holes that the agent can fill with minimal adult supervision is a completely new routine throughout my day, but I think obsessing over maintainability will pay off, like it always has.
It's crazy to me nevertheless that some people can afford the luxury to completely renounce AI-assisted coding.
What's the luxury? The luxury is AI-assisted coding considering how expensive it is for tokens/$.
My habit now: always get a 2nd or 3rd opinion before assuming one LLM is correct.
All code written by an LLM is reviewed by an additional LLM. Then I verify that review and get one of the agents to iterate on everything.
Might be my skills but I can tell you right now I will not be as fast as the AI especially in new codebases or other languages or different environments even with all the debugging and hell that is AI pull request review.
I think the answer here is fast AI for things it can do on its own, and slow, composed, human in the loop AI for the bigger things to make sure it gets it right. (At least until it gets most things right through innovative orchestration and model improvement moving forward.)
I have AI build self-contained, smallish tasks and I check everything it does to keep the result consistent with global patterns and vision.
I stay in the loop and commit often.
Looks to me like the problem a lot of people are having is that they have AI do the whole thing.
If you ask it "refactor code to be more modern", it might guess what you mean and do it in a way you like it or not, but most likely it won't.
If you keep tasks small and clearly specced out it works just fine. A lot better than doing it by hand in many cases, specially for prototyping.
it'll be really interesting to see in the decades to come what happens when a whole industry gets used to releasing black boxes by vb coding the hell out of it
It's worth mentioning that even today, Copilot is an underwhelming-to-the-point-obstructing kind of product. Microsoft sent salespeople and instructors to my job, all for naught. Copilot is a great example of how product > everything, and if you don't have a good product... well...
While this is likely feasible, I imagine it is also an instant fireable offense at these sites if not already explicitly directed by management. Also not sure how Microsoft would react upon finding out (never seen the enterprise licensing agreement paperwork for these setups). Someone's account driving Claude Code via Github Copilot will also become a far outlier of token consumption by an order(s) of magnitude, making them easy to spot, compared to their coworkers who are limited to the conventional chat and code completion interfaces.
If someone has gotten the enterprise Github Copilot integration to work with something like Claude Code though (simply to gain access to the models Copilot makes available under the enterprise agreement, in a blessed golden path by the enterprise), then I'd really like to know how that was done on both the non-technical and technical angles, because when I briefly looked into it all I saw were very thorny, time-consuming issues to untangle.
Outside those environments, there are lots of options to consume Claude Code via Github Copilot like with Visual Studio Code extensions. So much smaller companies and individuals seem to be at the forefront of adoption for now. I'm sure this picture will improve, but the rapid rate of change in the field means those whose work environment is like those enterprise constrained ones I described but also who don't experiment on their own will be quite behind the industry leading edge by the time it is all sorted out in the enterprise context.
All under one subscription.
Does not support upload / reading of PDF files :(
Yes, definitely. I use it mostly in Agent mode, then switch to Ask mode to ask it questions.
> How's the autocomplete?
It works reasonably well, but I'm less interested in autocomplete.
As I have never tried Claude Code, I can't say how much better it is. But Copilot is definitely more then auto-complete. Like I already wrote, it can do Planning mode, edit mode, mcp, tool calling, web searches.
I don't "vibecode" though, if I don't understand what it's doing I don't use it. And of course, like all LLMs, sometimes it goes on a useless tangent and must be reigned in.
I tried minimalist example where it totally failed few years back, and still, ChatGPT 5 produced 2 examples for "Async counter in Rust" - using Atomics and another one using tokio::sync::Mutex. I learned it was wrong then the hard way, by trying to profile high latency. To my surprise, here's quote from Tokio Mutex documentation:
Contrary to popular belief, it is ok and often preferred to use the ordinary Mutex from the standard library in asynchronous code.
The feature that the async mutex offers over the blocking mutex is the ability to keep it locked across an .await point.
I have to go out of my way to get this out of llms. But with enough persuasion, they produce roughly what I would have written myself.
Otherwise they default to adding as much bloat and abstraction as possible. This appears to be the default mode of operation in the training set.
I also prefer to use it interactively. I divide the problem to chunks. I get it to write each chunk. The whole makes sense. Work with its strengths and weaknesses rather than against them.
For interactive use I have found smaller models to be better than bigger models. First of all because they are much faster. And second because, my philosophy now is to use the smallest model that does the job. Everything else by definition is unnecessarily slow and expensive!
But there is a qualitative difference at a certain level of speed, where something goes from not interactive to interactive. Then you can actually stay in flow, and then you can actually stay consciously engaged.
And I also might "vibe code" when I need to add another endpoint on a deadline to earn a living. To be fair - I review and test the code so not sure it's really vibe coding.
For me it's not that binary.
- No ai engineers - Minimal AI autocomplete engineers - Simple agentic developers - Vibe coders who review code they get - Complete YOLO vibe coders who have no clue how their "apps" work
And that spectrum will also correlate to the skill level in engineering: from people who understand what they are doing and what their code is doing - to people who have lost (or never even had) software engineering skills and who only know how to count lines of code and write .md files.
We can identify 3 levels of "vibe coding":
1. GenAI Autocomplete
2. Hyperlocal prompting about a specific function. (Copilot's orginal pitch)
3. Developing the app without looking at code.
Level 3 is hardly considered "vibe" coding, and Level 2 is iffy.
"90% of code written by AI" in some non-trivial contexts only very recently reached level 3.
I don't think it ever reached Level 2, because that's just a painfully tedious way of writing code.
It came very close to success, but there were 2 or 3 big show-stopping bugs such as it forgetting to update the spatial partitioning when the entities moved, so it would work at the start but then degrade over time.
It believed and got stuck on thinking that it must be the algorithm itself that was the problem, so at some point it just stuck a generic boids solution into the middle of the rest. To make it worse, it didn't even bother to use the spatial partitioning and they were just brute force looking at their neighbours.
Had this been a real system it might have made its way into production, which makes one think about the value of the AI code out there. As it was I pointed out that bit and asked about it, at which point it admitted that it was definitely a mistake and then it removed it.
I had previously implement my own version of the algorithm and it took me quite a bit of time, but during that I built up the mental code model and understood both the problem and solution by the end. In comparison it easily implemented it 10-30x faster than I did but would never have managed to complete the project on its own. Also if I hadn't previously implemented it myself and had just tried to have it do the heavy lifting then I wouldn't have understood enough of what it was doing to overcome its issues and get the code working properly.
Once I mastered the finite number of operations and behaviors, I knew how to tell "it" what to do and it would work. The only thing different about vibe coding is the scale of operations and behaviors. It is doing exactly what you're telling it to do. And also expectations need to be aligned. Don't think you can hand over architecture and design to the LLM; that's still your job. The gain is, the LLM will deal with the proper syntax, api calls, etc. and work as a reserach tool on steroids if you also (from another mentor later in life) ask good questions.
If what you're doing is proprietary, or even a little bit novel. There is a really good chance that AI will screw it up. After all, how can it possibly know how to solve a problem it has never seen before?
I am writing a game in Monogame, I am not primarily a game dev or a c sharp dev. I find AI is fantastic here for "Set up a configuration class for this project that maps key bindings" and have it handle the boiler plate and smaller configuration. Its great at give me an A start implementation for this graph. But when it becomes x -> y -> z without larger contexts and evolutions it falls flat. I still need creativity. I just don't worry too much about boiler plate, utility methods, and figuring out specifics of wiring a framework together.
I will have a conversation with the agent. I will present it with a context, an observed behavior, and a question... often tinged with frustration.
What I get out of this interaction at the end of it is usually a revised context that leads me figure out a better outcome. The AI doesn't give me the outcome. It gives me alternative contexts.
On the other hand, when I just have AI write code for me, I lose my mental model of the project and ultimately just feel like I'm delaying some kind of execution.
As a PRODUCT person, it writes code 100x faster than I can, and I treat anything it writes as a "throwaway" prototype. I've never been able to treat my own code as throwaway, because I can't just throw away multiple weeks of work.
It doesn't aid in my learning to code, but it does aid in me putting out much better, much more polished work that I'm excited to use.
Nobody forces you to completely let go of the code and do pure vibe coding. You can also do small iterations.
Option 1: The cost/benefit delta of agentic engineering never improves past net-zero, and bespoke hand-written code stays as valuable as ever.
Option 2: The cost/benefit becomes net positive, and economics of scale forever tie the cost of code production directly to the cost of inference tokens.
Given that many are saying option #2 is already upon us, I'm gonna keep challenging myself to engineer a way past the hurdles I run into with agent-oriented programming.
The deeper I get, the more articles like this feel like the modern equivalent of saying "internet connections are too slow to do real work" or "computers are too expensive to be useful for regular people".
What AI (LLMs) do is raises the level of abstraction to human language via translation. The problem is human language is imprecise in general. You can see this with legal or science writing. Legalese is almost illegible to laypeople because there are precise things you need to specify and you need be precise in how you specify it. Unfortunately the tech community is misleading the public and telling laypeople they can just sit back and casually tell AI what you want and it is going to give you exactly what you wanted. Users are just lying to themself, because most-likely they did not take the time to think through what they wanted and they are rationalizing (after the fact) that the AI is giving them exactly what they wanted.
I work on game engines which do some pretty heavy lifting, and I'd be loath to let these agents write the code for me.
They'd simply screw too much of it up and create a mess that I'm going to have to go through by hand later anyway, not just to ensure correctness but also performance.
I want to know what the code is doing, I want control over the fine details, and I want to have as much of the codebase within my mental understanding as possible.
Not saying they're not useful - obviously they are - just that something smells fishy about the success stories.
Then, I can reason through the AI agent's responses and decide what if anything I need to do about them.
I just did this for one project so far, but got surprisingly useful results.
It turns out that the possible bugs identified by the AI tool were not bugs based on the larger context of the code as it exists right now. For example, it found a function that returns a pointer, and it may return NULL. Call sites were not checking for a NULL return value. The code in its current state could never in fact return a NULL value. However, future-proofing this code, it would be good practice to check for this case in the call sites.
Examples.
Thanks to Claude I've finally been able to disable the ssh subsystem of the GNOME keyring infrastructure that opens a modal window asking for ssh passhprases. What happened is that I always had to cancel the modal, look for the passhprase in my password manager, restart what made the modal open. What I have now is either a password prompt inside a terminal or a non modal dialog. Both ssh-add to a ssh agent.
However my new emacs windows still open in an about 100x100 px window on my new Debian 13 install, nothing suggested by Claude works. I'll have to dig into it but I'm not sure that's important enough. I usually don't create new windows after emacs starts with the saved desktop configuration.
It is quite scary that junior devs/college kids are more into vibe coding than putting in the effort to actually learn the fundamentals properly. This will create at least 2-3 generations of bad programmers down the line.
So while there’s no free lunch, if you are willing to pay - your lunch will be a delicious unlimited buffet for a fraction of the cost.
I think coding with an AI changes our role from code writer to code reviewer, and you have to treat it as a comprehensive review where you comment not just on code "correctness" but these other aspects the author mentions, how functions fits together, codebase patterns, architectural implications. While I feel like using AI might have made me a lazier coder, it's made me a me a significantly more active reviewer which I think at least helps to bridge the gap the author is referencing.
I admit I could be an outlier though.
In order to get high accuracy PRs with AI (small, tested commits that follow existing patterns efficiently), you need to spend time adding agents (claude.md, agents.md), skills, hooks, and tools specific to your setup.
This is why so much development is happening at the plugin layer right now, especially with Claude code.
The juice is worth the squeeze. Once accuracy gets high enough you don't need to edit and babysit what is generated, you can horizontally scale your output.
That's exactly why this whole (nowadays popular) notion of AI replacing senior devs who are capable of understanding large codebases is nonsense and will never become reality.
The opener is 100% true. Our current approach with AI code is "draft a design in 15mins" and have AI implement it. The contrasts with the thoughtful approach a human would take with other human engineers. Plan something, pitch the design, get some feedback, take some time thinking through pros and cons. Begin implementing, pivot, realizations, improvements, design morphs.
The current vibe coding methodology is so eager to fire and forget and is passing incomplete knowledge unto an AI model with limited context, awareness and 1% of your mental model and intent at the moment you wrote the quick spec.
This is clearly not a recipe for reliable and resilient long-lasting code or even efficient code. Spec-driven development doesn't work when the spec is frozen and the builder cannot renegotiate intent mid-flight..
The second point made clearer in the video is the kind of learned patterns that can delude a coder, who is effectively 'doing the hard part', into thinking that the AI is the smart one. Or into thinking that the AI is more capable than it actually is.
I say this as someone who uses Claude Code and Codex daily. The claims of the article (and video) aren't strawman.
Can we progress past them? Perhaps, if we find ways to have agents iteratively improve designs on the fly rather than sticking with the original spec that, let's be honest, wasn't given the rigor relative to what we've asked the LLMs to accomplish. If our workflows somehow make the spec a living artifact again -- then agents can continuously re-check assumptions, surface tradeoffs, and refactor toward coherence instead of clinging to the first draft.
Perhaps that is the distinction between reports of success with AI and reports of abject failure. Your description of "Our current approach" is nothing like how I have been working with AI.
When I was making some code to do a complex DMA chaining, the first step with the AI was to write an emulator function that produced the desired result from the parameters given in software. Then a suite of tests with memory to memory operations that would produce a verifiable output. Only then started building the version that wrote to the hardware registers ensuring that the hardware produced the same memory to memory results as the emulator. When discrepancies occurred, checking the test case, the emulator and the hardware with the stipulation that the hardware was the ground truth of behaviour and the test case should represent the desired result.
I occasionally ask LLMs to one shot full complex tasks, but when I do so it is more as a test to see how far it gets. I'm not looking to use the result, I'm just curious as to what it might be. The amount of progress it makes before getting lost is advancing at quite a rate.
It's like seeing an Atari 2600 and expecting it to be a Mac. People want to fly to the moon with Atari 2600 level hardware. You can use hardware at that level to fly to the moon, and flying to the moon is an impressive achievement enabled by the hardware, but to do so you have to wrangle a vast array of limitations.
They are no panacea, but they are not nothing. They have been, and will remain, somewhere between for some time. Nevertheless they are getting better and better.
"AI can be good -- very good -- at building parts. For now, it's very bad at the big picture."
I disagree though. There’s no good reason that careful use of this new form of tooling can’t fully respect the whole, respect structural integrity, and respect neighboring patterns.
As always, it’s not the tool.
That's a very bad way to look at these tools. They legit know nothing, they hallucinate APIs all the time.
The only value they have at least in my book is they type super fast.
this is such an individualized technology that two people at the same starting point two years ago, could've developed wildly different workflows.
It's just a tool with a high level of automation. That becomes clear when you have to guide it to use more sane practices, simple things like don't overuse HTTP headers when you don't need them.
Good take though.
You should never just let AI "figure it out." It's the assistant, not the driver.
his points about why he stopped using AI: these are the things us reluctant AI adopters have been saying since this all started.
I just bootstrapped a 500k loc MVP with AI Generator, Community and Zapier integration.
www.clases.community
And is my 3rd project that size, fully vibe coded
I have been tolerably successful. However, I have almost 30 years of coding experience, and have the judgement on how big a component should be - when I push that myself _or_ with AI, things go hairy.
ymmv.
There are many instances where I get to the final part of the feature and realize I spent far more time coercing AI to do the right thing than it would have taken me to do it myself.
It is also sometimes really enjoyable and sometimes a horrible experience. Programming prior to it could also be frustrating at times, but not in the same way. Maybe it is the expectation of increased efficiency that is now demanded in the face of AI tools.
I do think AI tools are consistently great for small POCs or where very standard simple patterns are used. Outside of that, it is a crapshoot or slot machine.
Have people always been this easy to market to?
You gotta have a better argument than "AI Labs are eating their own dogfood". Are there any other big software companies doing that successfully? I bet yes, and think those stories carry more weight.
I think the most I can say I've dove in was in the last week. I wrangled some resources to build myself a setup with a completely self-hosted and agentic workflow and used several open-weight models that people around me had specifically recommended, and I had a work project that was self-contained and small enough to work from scratch. There were a few moving pieces but the models gave me what looked like a working solution within a few iterations, and I was duly impressed until I realized that it wasn't quite working as expected.
As I reviewed and iterated on it more with the agents, eventually this rube-goldberg machine started filling in gaps with print statements designed to trick me and sneaky block comments that mentioned that it was placeholder code not meant for production in oblique terms three lines into a boring description of what the output was supposed to be. This should have been obvious, but even at this point four days in I was finding myself missing more things, not understanding the code because I wasn't writing it. This is basically the automation blindness I feared from proprietary workflows that could be changed or taken away at any time, but much faster than I had assumed, and the promise of being able to work through it at this higher level, this new way of working, seemed less and less plausible the more I iterated, even starting over with chunks of the problem in new contexts as many suggest didn't really help.
I had deadlines, so I gave up and spent about half of my weekend fixing this by hand, and found it incredibly satisfying when it worked, but all-in this took more time and effort and perhaps more importantly caused more stress than just writing it in the first place probably would have
My background is in ML research, and this makes it perhaps easier to predict the failure modes of these things (though surprisingly many don't seem to), but also makes me want to be optimistic, to believe this can work, but I also have done a lot of work as a software engineer and I think my intuition remains that doing precision knowledge work of any kind at scale with a generative model remains A Very Suspect Idea that comes more from the dreams of the wealthy executive class than a real grounding in what generative models are capable of and how they're best employed.
I do remain optimistic that LLMs will continue to find use cases that better fit a niche of state-of-the-art natural language processing that is nonetheless probabilistic in nature. Many such use cases exist. Taking human job descriptions and trying to pretend they can do them entirely seems like a poorly-thought-out one, and we've to my mind poured enough money and effort into it that I think we can say it at the very least needs radically new breakthroughs to stand a chance of working as (optimistically) advertised
I chuckled at this. This describes pretty much every large piece of software I've ever worked on. You don't need an LLM to create a giant piece of slop. To avoid it takes tons of planning, refinement, and diligence whether it's LLM's or humans writing it.
Homelab is my hobby where I run Proxmox, Debian VM, DNS, K8s, etc, all managed via Ansible.
For what it is worth, I hate docker :)
I wanted to setup a private tracker torrent that should include:
1) Jackett: For the authentication
2) Radarr: The inhouse browser
3) qBitorrent: which receives the torrent files automatically from Radarr
4) Jellyfin: Of course :)
I used ChatGPT to assist me into getting the above done as simple as possible and all done via Ansible:
1) Ansible playbook to setup a Debian LXC Proxmox container
2) Jackett + Radarr + qBitorrent all in one for simplicity
3) Wireguard VPN + Proton VPN: If the VPN ever go down, the entire container network must stop (IPTables) so my home IP isn't leaked.
After 3 nights I got everything working and running 24/7, but it required a lot of review so it can be managed 10 years down the road instead of WTF is this???
There were silly mistakes that make you question "Why am I even using this tool??" but then I remember, Google and search engines are dead. It would have taken me weeks to get this done otherwise, AI tools speed that process by fetching the info I need so I can put them together.
I use AI purely to replace the broken state of search engines, even Brave and DuckDuckGo, I know what I am asking it, not just copy/paste and hope it works.
I have colleagues also into IT field whose the company where they work are fully AI, full access to their environment, they no longer do the thinking, they just press the button. These people are cooked, not just because of the state of AI, if they ever go look for another job, all they did for years was press a button!!
For the record, I use AI to generate code but not for "vibecoding". I don't believe when people tell me "you just prompt it badly". I saw enough to lose faith.
"Amazingly, I’m faster, more accurate, more creative, more productive, and more efficient than AI, when you price everything in, and not just code tokens per hour."
For 99.99% of developers this just won't be true.
AI is far from perfect, but the same is true about any work you may have to entrust to another person. Shipping slop because someone never checked the code was literally something that happened several times at startups I have worked at - no AI necessary!
Vibecoding is an interesting dynamic for a lot of coders specifically because you can be good or bad at vibecoding - but the skill to determine your success isn't necessarily your coding knowledge but your management and delegation soft skills.
I also keep seeing that writing more detailed specs is the answer and retorts from those saying we’re back to waterfall.
That isn’t true. I think more of the iteration has moved to the spec. Writing the code is so quick now so can make spec changes you wouldn’t dare before.
You also need gates like tests and you need very regular commits.
I’m gradually moving towards more detailed specs in the form of use cases and scenarios along with solid tests and a constantly tuned agent file + guidelines.
Through this I’m slowly moving back to letting Claude lose on implementation knowing I can do scan of the git diffs versus dealing with a thousand ask before edits and slowing things down.
When this works you start to see the magic.
This is no different. And I'm not talking about vibe coding. I just mean having an llm browser window open.
When you're losing your abilities, it's easy to think you're getting smarter. You feel pretty smart when you're pasting that code
But you'll know when you start asking "do me that thingy again". You'll know from your own prompts. You'll know when you look at older code you wrote with fear and awe. That "coding" has shifted from an activity like weaving cloth to one more like watching YouTube.
Active coding vs passive coding
Relevant xkcd: https://xkcd.com/568/
Even if we reach the point where it's as good as a good senior dev. We will still have to explain what we want it to do.
That's how I find it most helpful too. I give it a task and work out the spec based on the bad assumptions it makes and manually fix it.
The result stunned everyone I work with. I would never in a million years put this code on Github for others. It's terrible code for a myriad reasons.
My lived experience was... the task was accomplished but not in a sustainable way over the course of perhaps 80 individual sessions with the longest being multiple solid 45 minute refactors...(codex-max)
About those. One of things I spotted fairly quickly was the tendency of models to duplicate effort or take convoluted approaches to patch in behaviors. To get around this, I would every so often take the entire codebase, send it to Gemini-3 Pro and ask it for improvements. Comically, every time, Gemini-3-Pro responds with "well this code is hot garbage, you need to refactor these 20 things". Meanwhile, I'm side-eying like.. dude you wrote this. Never fails to amuse me.
So, in the end, the project was delivered, was pretty cool, had 5x more features than I would have implemented myself and once I got into a groove -- I was able to reduce the garbage through constant refactors from large code reviews. Net Positive experience on a project that had zero commercial value and zero risk to customers.
But on the other hand...
I spend a week troubleshooting a subtle resource leak (C#) on a commercial project that was introduced during a vibe-coding session where a new animation system was added and somehow added a bug that caused a hard crash on re-entering a planet scene.
The bug caused an all-stop and a week of lost effort. Countless AI Agent sessions circularly trying to review and resolve it. Countless human hours of testing and banging heads against monitors.
In the end, on the maybe random 10th pass using Gemini-3-Pro it provided a hint that was enough to find the issue.
This was a monumental fail and if game studios are using LLMs, good god, the future of buggy mess releases is only going to get worse.
I would summarize this experience as lots of amazement and new feature velocity. A little too loose with commits (too much entanglement to easily unwind later) and ultimately a negative experience.
A classic Agentic AI experience. 50% Amazing, 50% WTF.
2026: "If I can just write the specs so that the machine understands them it will write me code that works."
Like it not, as a friend observed, we are N months away a world where most engineers never looks at source code; and the spectrum of reasons one would want to will inexorably narrow.
It will never be zero.
But people who haven't yet typed a word of code never will.
Or how I would start spamming SQL scripts and randomly at some point nuke all my work (happened more than once)... luckily at least I had backups regularly but... yeah.
I'm sorry but no, LLMs can't replace software engineers.
It requires refactoring at scale, but GenAI is fast so hitting the same code 25 times isn’t a dealbreaker.
Eventually the refactoring is targeted at smaller and smaller bits until the entire project is in excellent shape.
I’m still working on Sharpee, an interactive fiction authoring platform, but it’s fairly well-baked at this point and 99% coded by Claude and 100% managed by me.
Sharpee is a complex system and a lot of the inner-workings (stdlib) were like coats of paint. It didn’t shine until it was refactored at least a dozen times.
It has over a thousand unit tests, which I’ve read through and refactored by hand in some cases.
The results speak for themselves.
https://sharpee.net/ https://github.com/chicagodave/sharpee/
It’s still in beta, but not far from release status.
Sharpee’s success is rooted in this and its recorded:
https://github.com/ChicagoDave/sharpee/tree/main/docs/archit...