But a big yes for still typing code by hand, and not leaving it to the llm. Except it has to be the code generated by your brain.
That is what will create new neurons and new connections, which is what will keep away the cognitive decline.
And the constraint of not having to use llms will enhance creativity.
Actually, the constraints llms add to your code are more in number than the former. llms code in only the specific ways they've been trained on. So you won't ever come across of other ways.
Off the top of my head.. here's RubyQuiz.com [0] which I came across when I was learning ruby more than a decade ago. Looking at the many user-submitted solutions (you have to download the zip file!) you'll see completely different ways the problems were solved.
Sure, many won't be deemed efficient or standard by today's llm or rubocop checks, but looking at their code.. and retyping them and seeing them work.. was crucial in how I was able to think in Ruby for solving coding problems.
I did the same with Go too, with the "learn go with tests" guide [1].
[0] - http://rubyquiz.com/
What ended up happening is that I was typing, I'd notice little weird bits here and there, and I'd ask questions about it, explore alternatives, etc. So the end result was partially generated by ChatGPT, but also partially influenced by me.
But the biggest win was that within 3-4 hours, I was comfortable enough with the syntax that I was writing it entirely by hand, without AI assistance.
Now, if all you're doing is literally typing the characters and not thinking at all about what you're typing, and if you accept the LLMs output verbatim, then yes I'd say that you aren't really learning anything. But this is no different than just copying things from Stack Overflow. That behavior is nothing new.
The core distinction, as always, is whether you are exercising your brain.
> That is what will create new neurons and new connections, which is what will keep away the cognitive decline.
100% agree with this.
The problem is, your employer doesn’t care whether your brain is creating new neurons and connections. They care about productivity and profit.
I feel like the folks that believe we can continue to write code by hand are either in denial that LLMs will eventually (if not already) outperform handwritten code, or are in denial that employers will be ok with lower levels of productivity.
If you have the luxury of working for an employer that either doesn’t care or is oblivious to this, then 100% continue writing code by hand.
Or if you do work on the side (whether for fun or for profit) and are ok with lower productivity, then yes, stick to handwritten code.
But just because handwriting code is better for your brain and cognitive development, doesn’t mean that the industry will be in support of it.
At least, I'm not aware of any actual reasons, backed by a proper theory.
On the other hand. I've noticed some persistent issues with code generated by agents, especially poorly supervised agents. If engineers become less vigilant, agents never get to the point of not needing supervision and juniors never pick up required skills, this could lead to real trouble.
Did business ever cared about this? Like ever?.. I've seen too many code bases with awful code by humans, now AI simply exacerbated the issue but this is nothing new in our field.
I just don't see how this is going to be a battle engineering departments can win. As long as AI is brining money to the business, they are not going to listen.
It seems they run into problems with larger concepts and general organization and need guidance at the moment but for a single code file they often do better then I would have.
This is trolling right?
I’m simply playing devil’s advocate, because engineers can believe one thing, but until employers believe it, it doesn’t matter much.
I’m also not convinced that the reality of LLMs will never catch up with what employers think they can do.
It may never happen, but it very well could.
Either way, I feel the days of employers being ok with handwriting code are limited.
In theory (never happens) the software will become feature complete and you only pay the LLM for security updates.
A rough proxy for this would be if they say contribute to the C++ standards process, which Google Microsoft nVidia do, and basically zero "normal" companies (say SaaS) do. It means they are investing in multi-decade maintenance of their codebase
It was/is a struggle to get them to recognize and balance tech dept, we must now push to balance cognitive dept as well.
Some employers will eventually get it as they have with tech dept. And those will be the ones we see succeed.
I feel like this is an odd thing to be reticent about. You're working for someone who doesn't care about you and, in this changing landscape, you're compartmentalizing that in order to determine how best to care for yourself. Logically, to care for yourself you'd need to find an environment in which your needs aren't being neglected.
More like sweeping the dust under the rug for later. The initial productivity boost is massive but at the cost of massive upfront tech debt. Given that LLMs aren't as good at removing code as they are at generating it, this is a problem.
How about moving up the stack?
When the bill for the latter lands with a heavy thud, moderation and common sense start to look like a pretty good idea.
Sure, but again, this assumes both that handwriting code sometimes outperforms LLMs, and also that your employer agrees with this.
I think it’s only a matter of time (again, if we’re not already there) before LLMs outperform handwritten code nearly all of the time.
And, even if that’s not the case, I’m pretty convinced nearly all employers believe this is true, whether it is or not.
So your employer only sees a “middle ground” as room for higher productivity.
Because to be honest, more and more it seems many businesses have no real purpose other than to act as a sort of adult daycare for otherwise useless people.
Which is why businesses, tech and otherwise, are falling apart constantly unless they achieve gigafuck scale status, at which point they're unkillable no matter how much of this cancer infests their management.
I don't know how we arrived at a social norm where it's just completely fine that leadership in massive companies is just absolutely useless at their jobs, but the firms that last another hundred years are going to be the ones that figure out how to fire them all.
Probably what you want to do is use it a lot when you first start a job to establish some productivity credence, then as you become ready to switch jobs, use it less and less until you're writing everything by hand and taking interviews.
They certainly don't today. Perhaps they will in the future, but based on the lack of improvement thus far it seems unlikely that they will get to this point. But whether or not they do improve to that point, the reality is that today, LLMs do not give you a productivity boost unless you give up on quality and just YOLO stuff the LLM gives you without actually checking it. And nobody should be willing to do the latter, because that is how you get software/infrastructure that doesn't actually work.
Cooking a steak and then flushing it straight down the toilet outperforms all the slow and tedious business of eating and digesting it, too.
Why should I use an LLM to write code? It cannot do the job I want it to. It cannot type the code I've already written and it cannot generate correct code.
That would be more equivalent to writing the code and immediately deleting it.
Except this code is being deployed, and it is still (mostly) functional, at least functional enough to satisfy your employer (their “hunger” in your analogy).
> Why should I use an LLM to write code? It cannot do the job I want it to.
Because your employer is convinced that it can do the job they want it to (whether it does the job you want it to do is irrelevant, unless you're the employer, in which case 100% handwrite your code).
Bzzzt. Wrong. I don't give a fuck who's satisfied by my code. I only care about writing it.
> Because your employer is convinced that it can do the job they want it to
Bzzzt. Wrong. The code is incorrect because it is not precisely the code I would have written.
LLMs are useless to me because they don't solve the problem of actually typing the code I've written into a computer.
All that matters is whether they agree with you.
The only control you have over this is to either a) work for yourself, or b) keep looking until you find an employer that agrees with you.
My argument is that the number of employers that fall under (b) is shrinking.
That's great. Sucks to be them. I've doubled my prices and doubled my time estimates and I'm still turning work away.
The first thing I do when I'm asked to clean up the LLM slop is "rm -rf ." and start from scratch.
You will pay a fortune for this.
I mean in your case, you are not able to use LLMs effectively to build features, so you should type the code by hand.
However, upper management is excited about AI because it can push features quickly to production. Granted, they will be giant balls of slop but they don't care. As long as it works in the UI, they are happy.
If no one could get anything done with AI, there wouldn't be billions invested i n it.
I never write code sitting at a computer. Mostly I write it in the car, when I have no distractions. I just type it in when I'm sitting at the computer.
I've been using Claude as an experiment for some "fun" projects that I'm unlikely to release, just to see how it is.
The thing is, at least for the stuff where it's getting deep into the far edges of synthesizer firmware, disk image decoding, audio generation and the like, Claude is mostly just quoting my own github repositories back to me.
So, I guess I'm doing something right then, or maybe I'm doing it wrong.
For example:
Sometimes I just need to blast something out which works well enough so I can either consider the output or play with the tool it generated to see if I'm on the track for solving my bigger problem. In these instances, I just say, do it, and don't focus too much on the code it produces.
Sometimes I am using languages or tools I am not an expert in. In many of these cases, I do manually type out the LLM's suggestions so that I am more intimately familiar with the syntax, libraries and idioms being used. Frequently pausing to look at actual documentation or find examples of sample code. It's a more interactive version of what I might have been doing a couple years ago googling solutions on stack overflow.
Then there are sluggish corporate jobs. Here, even 10x speed up of development won't change delivery dates significantly or at all. I am in one such job, team leader has claude code higher tier and basically uses it for some more complex bash scripts and thats mostly it. Given this topic, I dont complain much, I value my long term senior skills way more than those new agentic ones.
The middle is... well somewhere middle.
Personally, I don't think we will ever be able to reconcile using LLMs and cognitive debt. Even before LLMs we were aware if it: we knew people moving to managerial/PM roles eventually get their coding skills rusted. Well, now we are all in those managerial roles...
I still think that you can build this model in your head even with LLM's but I'm not sure neither one way or the other.
> Even before LLMs we were aware if it: we knew people moving to managerial/PM roles eventually get their coding skills rusted. Well, now we are all in those managerial roles...
Retyping things is inefficient for learning. It's like trying to retype calculus solutions — you don't learn from it. Even if there is an explanation of why the code is written in such a way, you did not come up with it, and you don't know alternative solutions. It is a practice for memorizing, not for building your intuition.
A better option is to write it yourself first and ask LLMs for better options. They are pretty good at it, especially when you need to optimize hot loops.
14 years later, as a software engineer, I still think about that book and the way I can trace back a lot of my initial improvements to its requirement that code got written line-by-line. I still maintain the habit of rote text copying as a way to pick up new tools and commands (i.e. copying documentation examples exactly into my editor), and also when people make an assertion like:
> It's like trying to retype calculus solutions — you don't learn from it.
Maybe not everyone learns that way, but I sure did!
[0] - https://www.amazon.co.uk/iPhone-iPad-Apps-Absolute-Beginners...
I can definitely say it helped me learn a lot more than just blindly copy/pasting everything over.
It seems like an exercise to force you to attend to the details. I had a similar experience with a "certified X programmer" exam my employer bought for me. I wouldn't say passing such an exam makes anyone a good programmer, but it was an unexpectedly valuable experience because it forced me to attend to a broad set of details (instead of getting a spotty command of them, due to path-dependent experience).
Similar thing happens with notes: I rarely reference them but taking them keeps my mind from wandering.
Except that you do. Otherwise you could just sit in school or university and just listen and do nothing, but that way you just learn A LOT less, instead of taking down the lecture/lesson.
Having to spend 95% of the time taking notes of things that could be distributed in a PDF has only ever hindered me academically. It's mechanical and boring, my mind tunes out and I get nothing out of it.
Guess what he would ask about in the exams?
So transcription can definitely be helpful when part of a broader, intentional process, while also being insufficient to do much on its own.
And the next logical question to ask is whether there's a better form of intentional behavior that might be more effective.
The solution in TFA feels like an on-ramp to cargo culting somehow; observing that manual typing and good results often go together, but then thinking that it's the typing that directly causes the good results rather than the thought process that accompanied the typing.
There's a much better article hiding inside the current one that's titled "Prevent cognitive debt by understanding deeply the code that your LLM spits out", but that sounds like hard work and would probably not be very popular. It's much better for audience engagement to provide a simple solution that anyone can do and does not require a lot of deep thought like "manually retype everything the LLM generated", even if it doesn't actually work.
> If you absolutely cannot do it then go home and think but for heavens sakes don't look it up in a book till you give up. Looking it up in a book is giving up. > > Paul Halmos (https://www.robots.ox.ac.uk/~adutta/blog/quotations-powerful...)
I strongly disagree. I used this strategy for learning how to reverse engineer and hook functions in a game with C++ and learned a ton.
I also used this strategy to learn Imgui and it worked great. Before LLMs I did this when learning from books too.
A lot of calculus is rewriting blocks of solutions and applying them to problems. There is a huge chasm between how calculus and real analysis are taught. By your logic calculus should be of very low value. Yet somehow it opens a lot of doors for people to learn other things based on it.
Trying it first sounds slower, but definitely better for cognitive training :)
Having Codex/Claude write all of it won't really benefit you imho.
Also essays too and other texts non code from llm or books, it helps.
Says who? You're saying this unequivocally like you have research that supports this.
I used to re-write the notes I took for studying and it was like night and day for how well I did on tests. IT also gave me a chance to tighten the information I was receiving. And it's exactly what's happening here.
I don't remember if I read this advice or just intuited it myself (perhaps after some hard lessons), but it's a programming habit I've kept for as long as I can remember (I started coding in the 90s). If I feel rushed, e.g. someone looking over my shoulder, and I copy+paste something, it always leaves me with a sense of unease. It creates a memory & comprehension hole that sticks out like a sore thumb, even for seemingly simple snippets. You can't really be sure it's simple without stepping through it carefully, and simple can be deceptive because it's usually the interactions and assumptions wrt surrounding code that lead to surprises. Typing out code manually gives you time and space to consider the broader picture.
For me I feel that LLMs have exploded (in a good way) my cognitive capabilities. I'm now the general of an army, rather than playing the role of a soldier. Of course that means that I lose the experience of being the lonely soldier, but it is a no-brainer tradeoff for me.
Anyway, I have to go now so I can push my car to the grocery store (so I don't forget how to walk), me and my giant calves will be back in a few hours.
All meant in good spirit. Keep doing what you're doing, thanks for sharing, and hope people are kind and only give good natured ribbings.
If you're going to walk to the store, don't push you car there! Just walk. So if you want to write code, don't retype what the LLM said, create your code from your own brain.
For example, I don’t handwrite the code that the compiler produces, but I know how to troubleshoot the IR/assembly and fix optimization issues that arise occasionally. That makes me better at my job of directing the compiler.
Unfortunately this feels less easy to do with modern tools. For example, Claude Code expects to edit your actual source files, and the Claude chat is much worse at writing code.
I was pretty good at it - mostly remembered to change the name at the top of the paper too.
Struggled at moderated exams; think it must have been the time pressure or something.
When you say "long known" it sounds like this is established science. Is there a link you can share?
For example, when I am reading a book, I often take notes of my favorite quotes/sentences/things that stick out or perk up my ears. Usually if I own the book or it's a used copy of mine I make a note of the page number on the very last blank page(s) and write out the full sentence/quote. Then when I go back months or years later I just open to the last page and can find all my favorite/important quotes. Most 200 page books typically have ~20 useful sentences for me to reference [going forward] - it's basically like making your own index or table of contents.
That long aside, just to say, if I'm not "very aware" or "consciously mindful" of what I'm copying down or writing, I'm able to do it very well on autopilot without etching anything into my memory -- as if no recallable moment gets etched into the wet cement of the brain unless I'm actively "hip to it." I wonder if others have this "autopilot" problem.
This is just a miserable career of "paint-by-number" because people can't be bothered to have a creative thought about their professional work or programming hobbies.
Software developers think that they are being clever with these kinds of strategies to "keep their skills sharp", but unfortunately the entire industry knows about this, and especially the upper management who are already eliminating these assembly line, JIRA-ticket-taker software jobs en masse.
With AI, our role also shifts. It's mainly to know what to spend effort on, to set priorities and, to be able to verbalize requirements, missing social context and unwritten rules, to anticipate what additional documents the agent needs, to prioritize deadlines, feature necessity, and other judgment calls.
We are right at the stage where our coding ability and review ability is still needed though, but this stage won't last long. Soon there will be as little point to a human diving into the code as to trying to beat a chess engine, or humans constructing buildings by hand. Of course the discussion and prioritization may involve looking at the code itself, to get a better idea of why the agent says that a certain feature would be tedious to implement in the current architecture, but then most people will just learn to take its word for it, just as you may want to understand a chess engine's step, but you typically wouldn't want to override it.
Did you not do that before AI? It’s so strange to me when people are calling out these kind of tasks like they were not already a requirement for the job. What were you doing before?
> Soon there will be as little point to a human diving into the code as to trying to beat a chess engine, or humans constructing buildings by hand
Chess is way less complex than coding. The rules are like a few pages. While the specs for an 8 bit chip like the AVR is in the hundreds of pages. Books like “The Linux Programming Interface” are thousand pages long.
Also humans are using tools for building. Tools that do exactly what you control them to do. When you use a drill for a hole, you don’t have to worry that pressing the trigger have a good chance of sending the bit in your guts.
No, you didn't have to explicitly say it in words. My mind doesn't run on internal monologue. Many people can just do their work without ever having reflected on it in words. Tacit knowledge, routines, shared assumptions and culture in a team, common knowledge etc. People have a hard time using AI because they are bad at modeling the knowledge state / information context from the AI POV. You need good theory of mind for this, and being a good programmer is distinct from that soft skill. Yes yes blabla soft skills are more important than hard skills blabla, I don't buy it. It used to be valuable to be great at the hard skills even with mid-tier soft skills. You can have a ton of smooth talkers who are attuned to feel each others emotion and desires super well, but the thing has to actually work too.
> What were you doing before?
Wrote code. Yes, you have to explain the outcome to your boss or your team at some point, but people generally have better developed theories of mind for people than for AI.
> Also humans are using tools for building. Tools that do exactly what you control them to do. When you use a drill for a hole, you don’t have to worry that pressing the trigger have a good chance of sending the bit in your guts.
Right. I'm not sure how to reconcile the two though. A tool whose job is to do some of the thinking part seems to be a contradiction to me. If I so much know what there is to do that it's pure execution and can reliably be executed in a way that basically ensures no potential surprises to me, then I wouldn't need more thinking. But I agree, it would be better to somehow find a hybrid that is both doing thinking and feels more like a tool also while using it.
Were you a solo developer? I think the amount of reports and explanations (either written or verbal) dwarfs by large the amount of code I’ve written. From training juniors to drafting a design specs for a feature. That is why people say code is the trivial part of the job because it’s easy. I’m pretty sure the mailing list of the Linux kernel is bigger than the code itself.
And here’s another fallacy that is always thrown around. Always underestimating the other’s tasks when you are not an expert in it and don’t know intimately what’s involved.
I don't think we should be making sweeping claims. But everyone generalizes. The person on top claimed that productivity doesn't increase for competent people and that coding is anyway very easy. That may be true in some contexts but not in others. Some software jobs require a lot of face-to-face, others less. Some managers micromanage, others just care about results on a longer timeframe. Some people build technically difficult, computationally heavy code, deployed on special hardware with efficiency requirements, others deploy website-like iPhone apps and CRUD webapps all day. There's no single shape for a software-related job.
So comes the programmer as a translator. Why the translation itself takes some effort, but most of the intellectual works is understand the intent of the change request or the specs (nothing to do with the code) and grasping the semantics of some features of the platform (library docs,…). Once that’s done coding is fairly easy as long as you follow the consistency of the code (architecture and other design patterns).
So on one side you get the formal world of computing and buggy libraries (backed by electronics which is a bit more chaotic) and on the other side you got natural languages, the various jargons, and other formal notations. The programmer translates one to another and overtime become proficient in both.
You then add AI into the mix and the same ickiness appears. While non speakers are enjoyed that they can write code, all programmers are saying that the code is bad. Not because of syntax, but because it lacks any consistency with the program which is the formal representation of the specs/solution. Thus the rise of harness and agents. But the machine translation still does have a lot of issues that professional programmers have to clean up.
> Some software jobs require a lot of face-to-face, others less.[…] There's no single shape for a software-related job.
But there’s one fact. Code is a formal notation for an algorithm, and that notation can be executed by a computer, thus making the process have a tangible effect in the real world. That effect is what is useful and making it controllable, thus modeling it and designing an interface is a worthwhile effort. Once that’s done, you can formally write it. But we have found that it is hard to do it in one go, so incremental approach is best.
So everything about software development is about modeling and having a suitable interface. The code itself is only a tool. Just like a ruler is only a tool to get a straight line. The straight line is what’s valuable. Not the ruler.
Yes, of course, but a lot of it is boilerplate and just figuring out how some function is called or how exactly it is exposed in the library, dealing with some frustrating but trivially solved inconsistencies or incompatibilities or oh you first have to initialize this or that, or actually this feature is missing from version X but can be emulated by setting an environment variable to this other value and blablabla.
> all programmers are saying that the code is bad.
Ridiculous. When did you last use AI for actual work where you wanted it to actually work? I know it may sound like saying James Randi's negative attitude blocks my telepathy or dowsing skill, but I get this impression of many people. They tried agentic coding a year ago, decided that it doesn't work and if they tried it ever since, they didn't provide context, didn't set it up for a possibility of success then immediately closed the program and sighed a sigh or relief that they are safe. It's not like that. The current frontier is very capable on higher level design tasks as well. If your problem is spiritual or social or economic, then let's skip this "not productive enough" alibi talk.
Machine translation today may not be up to the levels of the best professional translators who work on something with full focus and full effort, given all the time they want to polish it, but it's better than an ordinary (not particularly the top of the top) translator working multiple project with tight deadlines. You have to compare like for like and take cost into account too. There is an immense amount of mistranslations by "professional" translators, not sure if you've ever seen movie subtitles or TV series dubs. Lots of paid professionals hired by production studios make huge mistakes, clearly miss cultural references even sometimes grammatical structures, which modern machine translation aces and even chat models are great at it and know basically all obscure pop culture references that e.g. a regular American audience member is expected to know but a translator working in Estonia may not get. Yes you should have a guy checking the output and doing some light touchups, given the draft and explanations and references provided by the AI, but the time of doing it fully manually and being competitive on the real world translator market has ended.
My problem is not about the productivity, it's about how this productivity is measured.
> Yes you should have a guy checking the output and doing some light touchups, given the draft and explanations and references provided by the AI, but the time of doing it fully manually and being competitive on the real world translator market has ended.
Have this been measured that this is actually faster while ensuring a quality baseline? I believe it's still the same "LGTM". And a lot of complaints about low quality text. And bugs in the case of software.
So you got a lot of messaging about how AI is supposed to make you faster. And now everyone is trying to show how fast they are with AI while the products are getting buggier.
Is there any proof of this?
https://thefederal.com/category/news/h1b-visa-indian-tech-wo...
> According to the discussion, foreign hiring at Google has fallen by more than half, while approvals at Amazon have dropped by nearly a third.
> According to Xfino's Active Tech Jobs Outlook, active technology job openings fell to 93,000 in June, down 14 per cent from 108,000 a month earlier.
That approach works well for me, so why would I retype AI-generated code? I wouldn't do that with code written by another developer, and their code can be just as buggy or weird in theory. You could argue that at least the developer knows how it works. That's true - for a week, a month, or a few years if you're lucky. Until he forgets or leaves.
Given that those things are checked off, I am having a tough time believing a frontier model, in conjunction with a new dev on a project, can't be immediately valuable to determining what is going on in a repo.
I've been trying to address this by telling LLMs primarily how the code should be structured, rather than only what it should do. Still, any design I hand to an LLM will be underspecified in one way or another (if it were fully specified, it would just be code), and the LLM fills those gaps somehow - which adds to the cognitive debt, slowly but surely.
Retyping LLM-generated code is an interesting solution. You'd certainly end up understanding the generated code better than if you merely reviewed it, but I doubt it produces a mental model as reliable as the one you'd build writing the code yourself. The longer you think, the better your mental model gets - and outsourcing the thinking to an LLM means thinking less.
That said, I've started to wonder whether I'm solving the wrong problem. Should I really insist on an accurate mental model of the code I own? We'd find it strange for a non-engineering manager to try to fully understand every piece of code their reports produce. If that's the right analogy, then as LLMs' agentic capabilities improve, maybe we should stop treating LLMs as tools that boost our own productivity as a software engineer and start treating them as independent agents we manage and steer.
Exactly as we don't write machine code letting the compiler do that, now and in the near future we won't be writing high level programming language code. We are moving towards working on a higher level of abstraction.
When I ask a frontier model to write a loop 10 different ways in Python and TypeScript and test the performance of each using a 1,000,000 iterations, it isn't creating cognitive debt. For the time being, I'm still racking my brain asking the question, how does garbage collection affect the performance.
Knowledge abstractions have historically been built by people with a detailed understanding of a problem domain and with a period of vetting the abstraction by many people. That is simply not true with LLM-generated code. We do not really understand what techniques LLMs are using to understand directions, value aesthetic/legibility characteristics, assess tradeoffs, retrieve contextual information, generate output, etc. Embedded in everything LLMs do are small decisions.
If you reach for formulas that you don't understand the applicability of, you will similarly get bad outcomes! But a calculator has almost no resemblance to what happens when you give general instructions to an LLM and get a generated code back, because in between that prompt and the result are many layers of decisions you the developer did not make and likely will not understand without carefully assessing the output!
If you have no discomfort with being culpable for something and doing none of the work to understand it, by all means, don't do anything to understand the output! Don't read the code, don't re-type the result, etc. I have concerns that you'll have a job in a year, but that's your decision to make!
> I have concerns that you'll have a job in a year
Me too because I might be the best TypeScript coder on Earth and there is no demand for those skills. I'm doing a pivot.
As good as they are, they are still fallible. They still write bugs, they still misunderstand instructions. They even get defensive about bad choices! And even in a world where that appears to no longer be true, we will be making society extremely fragile if we give an alien intelligence total autonomy to do what they think is right.
If you're not going to be writing code, you better be a vigilant babysitter. Otherwise I firmly believe that more than your job is at stake.
Creating while writing allowed us to build a mental context but in a unproductive way, it never scalled.
I believe we need to move onto a new way of reading codebases that go beyond reading line by line.
I know people have explored representing code in 3D spaces. I don't know the solution. But I believe that is the problem.
By default, if you ask an AI to "generate documentation for this code" it generates the same broken documentation all the humans do too; an enumeration of all the modules in the code and what their API is. I'm not surprised, the training data is biased probably at least 25:1 in favor of this rather than the useful data flow documentation. Fred Brooks was complaining about this over 50 years ago and the discipline as a whole still gets this wrong.
I'm not saying this is a future solution to all problems, but it is a now solution to some problems.
3D doesn't help. We live in a 3D world but our vision is 2D with a bit of augmentation from a second view point just a bit away. We derive some depth information from that, but we don't really "see in 3D". To do that we'd need to be 4D beings. There's a lot less juice in the 3D squeeze than meets the eye.
You can ask Claude to "sdoc me an annotated code walkthrough of this MR" and it will produce something like this:
https://smalldocs.org/s/Ju9GOmWZ0JXTtzqCVfgt1q#k=6HBrpcCjIu7...
I use this a lot to stay in touch with the code the LLM is producing.
[1] https://smalldocs.org [2] https://github.com/espressoplease/smalldocs
This is part of how I learned how to code! Well, sans LLMs. Instead of copypasta-ing code blocks from books, Expert Sexchange and Stack Overflow, I manually retyped everything, looked up what I didn't understand and changed what I could.
I also tried where I would print out the code (with color) and then go read it with a red pen. Faster and similar results of forcing me to be able to read it enough to be "an editor" to the code.
And I have done the same for LLM-generated code and text, back in '23 this meant more taking their so-so output and then turning it into my own, but between now and then come up with a number of techniques to improve the AI output to more acceptable out of the gate so more learning than being inspired.
These are good techniques as it causes your own brain to rapidly learn the material, but no matter how good it is I have never met anyone else that does it so the real ponderable is assuming that everyone else doesn't do this and what does that mean?
The goal is to maintain the taste, for lack of a better word, that I've developed over decades of programming.
Claude put me on to the concept "Étude", so I've taken to calling it my Étude project.
You could try doing the high level design yourself at least. Ask for its review and iterate without asking it to do it all.
Once it has generated some implementation, critique it and ensure you understand its approach And you agree with it, steer it otherwise.
If there’s anything unclear to you say so and have it rewrite it in an easier way to understand.
For example I had a working linear decay velocity boost function added to a godot scene and it's suggested refactoring actually reduced lines of code just today. So kind of happy when that works out, but yes I also applied the changes manually so I would grok them and hopefully remember it better.
> For your next ticket or feature, engage with your LLM as you normally would to produce a design and implementation plan, but with that plan in hand, make the mechanical edits yourself...You can expect this to improve your knowledge retention compared to merely reading a diff. You want to go slower now so that later you can go at all.
"professional malpractice" LOL. This is so dumb. I can't believe this shit is getting upvoted.
Managers never got to write and internalize every line of code anyway. Arguably their essential skill is producing useful stuff without needing to do so. So is that 'cognitive debt'? When I use an agent to code, I do the same things: I validate the direction, organization, and core decisions, but delegate the details. It's an intentional calculation of risk, managing the scope of future problems if there's a failure of implementation somewhere.
It's a question of where you put the value of cognitive focus. I don't understand 'line by line' how my car's ICE works, even though it will likely impact me at some point, because while I might be able to fix it myself if it breaks, the likely applied solution will be to delegate that work anyway.
You may be correct now and it doesn’t matter one bit.
What I do see is that the frontier models can get easy stuff mostly right, but they can’t maintain a small codebase such that it behaves correctly across several mutations. When the going requires a lot of context, tests get rewritten to not do anything and features that manipulate similar data get divergent implementations. That’s fine for goofing around personally, but I see vibe coders getting their asses handed to them when they mistake that for a well maintained codebase and try to sell a product.
Please, don’t mistake an impressive technology for a miracle; there will still be people needed to make codebases healthy for a long time to come.
This reads like it can write a better javascript function that executes a SQL statement and maps the parameters.
> They’ll get to 95% of most niche coding domains by December and likely all coding tasks sometime next year; 99% better at all tasks by December 2028.
Basing this prediction on LLMs writing what something like an ORM would generate for you doesn't connect for me.
I'm not hanging my hat yet and can still see where LLMs struggle. I think it's the best bet you can make: keep working for the future where you'll still be needed, because you can't prepare for the future where you're not.
This is specifically what people pumping these ideas are doing as you can see like Sam Altman or Dario.
Absolutely delusional if you think the two aren't related
I also think I write better code than an LLM in certain areas, but in most programming domains, the LLM knows more than I do across many dimensions. As prompts get deeper, LLMs are already producing PhD-level code—and that's been shown in research. The vast majority of people don't have that level of education. Of course, having a PhD doesn't guarantee good coding, but at least it's clear that LLMs can handle that level of capability.
People might disagree, but my view is firm on this.
Remember that the quality of the LLMs code in the areas you don’t know is as good as the quality as the area you do know. You’re only able to gauge the quality of what you do know.
Executable specifications
Unit, integration, and property tests
Official API documentation
All of these provide ways to judge quality. There are so many metrics that the problem is actually choosing which ones to use.
1.Does it produce output A for input B? 2.Can it process 100,000 records within 5ms? 3.Is memory usage within the defined limit? 4.Does the protocol handle error conditions properly?
You don't need to be an expert to test these. People forget that programming is one of the few fields where judgment criteria can be easily translated into machine executable tests.
The biggest problem with epistemological objections is that they often assume a binary: 'experts can judge, non-experts cannot.' But in reality, it's a continuous process.
You run the code, notice something odd, look up the terminology, compare it with official documentation and reference implementations, add tests, and gradually build up judgment in that area. Rewriting LLM generated code compresses that entire learning curve.
In the past, entering an unfamiliar domain meant starting from a blank file and reading documentation. Now, you start with a working hypothesis and modify it. That's far cheaper than starting from scratch.
If you treat LLM generated code as executable teaching material and a falsifiable hypothesis, it's a very powerful resource.
I feel uncomfortable when people reduce programming to aesthetic quality alone. We were all trained to measure things.
Also, error handling.
Tests, specs, and docs are are all downstream of that.
So far I've found that AI does an adequate-to-very-good job up to a certain amount of code, then things tend to fall apart. The solution is modularity and clean interfaces - as it always was.
That's not true. You can follow the logic step by step, search for similar domains, and form a comparison group.
Realistically, if we take this logic to its conclusion, it's like asking: 'You're just a spectator, so can you really judge whether a football player is good at football?
Realistically, LLMs write code well. I'm a C# programmer, and as you know, the current trend in modern programming is to refine error handling—using try/catch at higher layers, working with monads, preserving computable contexts, and branching accordingly. That kind of code typically takes 1 to 2 hours of modeling before you even start writing.
But an LLM can do it in just 5 minutes.
So I think the opposite way: if LLMs can do this well in a domain I'm familiar with, they're probably just as capable in other domains too. And that makes them feel trustworthy.
And more importantly, no matter how great a human programmer is, once you step outside your domain, you're in unknown territory. In those areas, LLMs are surprisingly helpful.
Epistemologically, it might be true that you can't evaluate code in a domain you don't know. But people keep forgetting the fact that you can learn through that code. This very post is itself an example of learning through LLM-typed code.
And programming actually provides clearer results than many other domains. It's not based on subjective UX like UI design. You have:
1.Executable specifications
2.Performance measurements
3.Comparisons with official documentation
These are clear engineering outcomes that allow for objective comparison. For example, you can verify whether a program produces results within 5ms when executed.
Many epistemological arguments tend to overlook this.
A lot of code is open source. You can find mature implementations anywhere. Finding a repo from someone claiming 10 or 20 years of experience and running a differential comparison is not difficult at all."
This gave me a chuckle, "PhD-level code" is gross actually. Have you ever looked at the code of research papers?
When I talk about PhD level code, I'm not referring to research code quality. I'm talking about research level algorithms.
Production code is built around reusability, policies, monadic chaining, and various code hygiene practices. Research code isn't. So what I'm saying is: 'Yes, I've collaborated with professors from top Korean universities.'
That said, I'm not writing PhD level code myself. The reason is that I take existing algorithms and apply code hygiene to them. I'm not actually developing new algorithms at a PhD level of understanding
So from what you're saying, it sounds like:
'Are you talking about those horrible research paper codes?'
But what exists in those research papers is: 'The domain modeling itself is difficult.'
I learned about code hygiene in production code—things like using functional approaches, and so on. Research code often lacks those things, so it might look bad. But that means the code quality is bad—not that the modeling itself is bad. And in fact, the core value is in being able to produce modeling at that level of difficulty.
In fact, most programmers, if trained like me, can do these things well enough. But domain modeling is different.
This touches on the distinction Frederick Brooks made in The Mythical Man-Month between 'Essential Complexity' and 'Accidental Complexity.' Code hygiene, functional chaining, monadic structures—these are techniques for reducing accidental complexity. With enough training, you can learn them. But domain modeling deals with essential complexity. It's about how difficult the problem itself is, and it's not easily solved through training alone. That's why the depth of modeling that researchers produce should be evaluated on a different axis from code quality.
This is what we call 'Essential Complexity.' From that perspective, saying code is 'good' usually means that the essential complexity is handled well on average. What I'm calling 'PhD-level' usually refers to modeling problems that are commonly difficult to solve.
So I can see why it might be considered low-quality. They don't use things like Result or policy types like we do in production code. But they're modifying or creating new algorithms, right?
The thing is, they don't need to do those things, so it looks lower-level—but the algorithm implementations themselves are actually pretty good, aren't they?
PhD thesis code often looks low-quality simply because there's no reason to reuse it like production code. The tensor computations inside are things that typical programmers can't easily do.
The point I'm making is that the core algorithmic work is something most programmers can't handle. And that's natural—highly skilled programmers have already built deep libraries for that.
When I talk about PhD-level code, I'm not referring to overall code quality. I'm talking about the level of the algorithms themselves.
This strikes me as a self-fulfilling prophecy. There are probably many people who would like to use the code published with PhD level research, but they don't because the code is not easy to adapt (or sometimes even to get running).
Then a decade later someone implements that same algorithm in a library with a well-designed interface and it suddenly becomes a useful tool for others. So ultimately there was a reason to reuse it, it was the quality of some aspects of the code that held it back.
This was true when I first learned to program, and is still true today. And I do find myself manually typing in really critical code. In those cases even if I do have an LLM alongside these days, LLM suggestions also then get manually typed.
The thing is that, probably like many of you, I love going deep in a side project, even if it lands nowhere. With so little time, working on them has become a very frustrating activity.
This is where I found my trap... do more with very little time by delegating to an LLM. You get dopamine shots, the feeling of achieving something but the cognitive dept is just crazy. So much that the activity becomes almost meaning less. After couple of months doing this, I'm not even sure it's a good use of this time. I get very little satisfaction on the long run.
I don't have a solution to this problem, not even sure there one. I think I have to accept that this is an activity that takes time, and only time gives the real gratification.
I think it is worth noting that not all code is equal... One could argue that adding a library is in some ways similar to copy/pasting code in as much as, one doesn't know what the code is doing, and yet that doesn't leave me with a sense of unease!
So it might be that as I'm working with an LLM there are parts (boring, as the author calls them) that are not worth "knowing" how they work, something uninteresting or that a correct output is all that us needed, I'm totally fine having the agent write that code, but the sections I need to know how they work, I think it might make sense to write those by hand!
It does, to me. Been burned enough times that I now, at minimum, audit the source code of third-party libraries before I use them.
NPM decencies is a bit of dumpster fire, supply chain attacks galore! A well known python (think numpy) certainly is less discomforting, and a "industry standard" C library like GSL or SDL3, I just use it and move on
I do think there is absolutely no way a learner should be using ai for code generation. I think code analysis is the only acceptable usecase.
At the very beginning stages you might be tempted to trace over an image but this builds a very shallow understanding. Instead you should quickly move onto replicating work you admire by sight. Consider the form, volume and values, conceptualise them in your mind and try to apply them in whatever medium you are using. This helps form your own mental model of the process. Eventually you can start using those techniques on original works.
I'm currently rethinking language learning too in a similar manner. Comprehensible input, shadowing. That sort of thing.
I've learned things I wouldn't otherwise have learned because I hadn't considered using the tools the LLM recommends. It's also a way to eliminate some hallucinating, given that critical questions are posed as unbiased as possible. For that, I also like to open a new chat with a different model and asking open-ended questions about a recommended tool I don't know much about, to double-check that the original LLM was likely correct in its recommendation in the first place.
File next to that 1970's era classic, "Prevent cognitive debt by manually retyping compiler-generated assembler"
It's important to retain the what why and how as a team to the degree that development can be efficient, extensible, and maintainable. (And ideally a good devx)
Needing to manually touch every line of code is not scalable.
This was true 10 years ago. It's still true today.
Don't let your skills atrophy.
I don't think this is a practice which will be sustainable for very long
If you write, you should write in your own words, to demonstrate your own understanding - the so-called Feynman technique. Never verbatim. That's as true for coding as it is for study notes.
In math classes back at school, it didn't matter how much the math professor explained how the formula works. What mattered is me putting in the effort to understand it. The implication to your example is, I should already be familiar enough to understand the generated code to the point where all the explanation that it's doing is effectively a "Quality of Life feature".
I ask it to plan the feature in a separate worktree.
In parallel I start coding without being biased by AI and vice versa.
At some point I read its plan and iterate on it all the while I am in implementation mode. This helps me improve my own vision.
Finally I ask the AI to review my implementation. It flags off bugs and gaps which are usually straightforward for it to fix.
The article is not wrong though that it pays off to have some imagination on how to use the models. For example, I want to use SIMD instructions in an ESP32-P4 CPU. Those instructions are undocumented for the most part, with just a couple of handwavey blog posts and some infuriatingly vague marketing material. So I just asked an LLM to create a `SIMD_P4.md` document with all the details. Lo and behold, it practically reverse-engineered the ISA. Now I can program in assembler by hand all I want and build that skill in my own brain, and whenever I find a slightly unclear op in the document, I ask the LLM to refine the documentation in that op.
Of course agentic coding tools are not trying to peer code or teach/inform you what they are doing, so being present in the moment doesn't help, but I suspect that copying it all down later doesn't help much either.
When you are/were developing software without AI, even for pretty large projects you do end up internalizing (memorizing, but not deliberately so) a lot of detail, but from my own experience I'd say it's more the design than the code. The design is what you put effort into, thought about, etc, so is both what you naturally end up memorizing, and is what you need to know to have a mental map of the project and therefore understand how best to modify it. The code itself was naturally always the last thing you did, and followed automatically from the design and module/component interfaces - not something you typically think much about other than while in the flow of just "coding it up".
By retyping LLM-generated code, it seems you are mostly going to be gaining familiarity with the wrong thing - the code and not the design. Memorizing the code is not going to help much in grokking the design.
Typing itself is irrelevant, it is the timing spent, even if only seconds, pondering at what each word or syntactic element is and why use it.
Being slower does not automatically make you learn better, focus on the learning is what makes the difference.
If you don't have the opportunity to learn, the time to actually think, then a faster tool is not helping.
TL;DR: what matters is why you are doing something, is it solely to get the task done or is it primary to learn, or both?
Since writing is thinking, coding is thinking since coding is writing. That means any time I am not certain how I’m gonna implement some feature or bug fix, I have to code it myself because that’s the only way I can force myself to think through it. Only when I get to a point where I’m line “ok I know exactly what to do now- all that’s left to do is type it out” that’s when AI can be employed - essentially as a autocomplete.
This is only for projects where I will be held responsible for outcomes and must understand how it works. For hackathon / personal projects, I vibe away.
I also use AI to brainstorm at the outset of the task when I don’t know where to start at all.
Is this inefficient? My take: no. It’s maximally efficient. Over the long term it gives me an edge over any teammates who just vibe code everything because I actually, you know, understand how stuff works. I become the guy who can save the day at 3am when the team’s business critical app goes down. I become the guy that gets pulled into meetings so the suits can ask “is this possible?”. I see opportunities and problems before my teammates because I have a relationship with our code and system that they never took the time to develop or think about.
Having LLMs write out their design and reviewing it seems more efficient. Have LLMs, maybe with a different model, check that the implementation meets the design.
Are people really out there just mass copy pasting llm code without even trying to understand it! lol
Whenever I encounter an especially preposterous proposal like this one, I like to imagine a USMC Drill Instructor wandering into the open plan office and having an interaction something like this:
USMC Drill Instructor: "What the actual fuck are you doing?!"
Smelly Recruit: "Sir, I'm hand typing the LLM output. Sir!"
USMC Drill Instructor: "Are you fucking with me recruit?! I said I wanted a SASS App, not a typing tutor! Drop and give me 20!"
Smelly Recruit: "SIR, YES SIR!"
And if so, in what work, and have you tried debugging issues with SOTA models?
From my experience it is certainly not the case that you cannot make progress after the first month.
I work on native mobile applications.
Without an active harness (eg. Appium) that can end-to-end deterministically verify the changes you make continue to work correctly it is almost impossible to continue to keep the same pace on the app.
Unsupervised LLMs (even fabel) are categorically incapable of running parallel unsupervised mobile app feature development.
That is my personal, first hand experience working in a team in this space.
What you are (I guess?) experiencing is user-in-the-loop light touch LLM development where you can 80% most tasks quite quickly (much faster than without assistance!) with a small number of human developers working on largely unrelated features and manually verifying they are correct and manually fixing the platform specific issues you encounter.
Maintaining a strong appium end-to-end test suite is still extremely challenging with notifications and maps.
Honestly, it blows my mind you could even being to claim that of all things, native apps using obscure languages like swift are suitable for this, compared to the much much easier path of web + react.
You might say “yeah yeah, but one month? Come on!”
…but have you actually seen how much code fabel can write in a month?
Its a lot.
So sure, you say, work at a slower pace. Don't just endlessly run a frontier model in unsupervised feature development mode.
Yes… you see, thats the point. Thats what the op is saying.
Move more slowly, and you can avoid building a spaghetti castle (ok sure! If you dont wanna, maybe don't retype every character by hand, but the point of that practice is not upping your wpm typing speed. :p It is to take the time to think, design and collaborate, not rush rush rush)
I doubt it brings any advantage over XCUITest here.
What is supposed to be the problem with parallel development? I use worktrees, and it works just fine with five agents in parallel.
Automated end-to-end testing on mobile is notoriously flaky, but that's nothing new, and I think it is now much easier to deal with.
Anyway the author did address that
> Using LLMs this way allows me to work faster than not using LLMs at all, but I'm still slower than those who are willing to allow the machine to think for them. Instead of being 10x faster, I'm probably only 2x faster. But what I lose out on in terms of speed, I gain in terms of a deeper understanding of my code.
Most of us read the specs, then raced into the coding phase, hands to IBM mechanical keyboards. One guy took a different approach. He took a legal pad and pencil, and wrote his whole program on paper before he ever approached a terminal. He’d do his own bug checking and syntax checking, instead of having the compiler do it ( compiles took longer in those days, and required JCL ). He avoided the entire compile/wait/read-with-dismay/quickly-try-again loop.
He was one of the top students, of course. And a lot less stressed, as I recall.
my workflow:
- ask not only for a solution to a problem but also for specific code (= tell the agent about your mental model of the codebase)
- ask for small stacked 'PRs/branches' and review/refactor heavily also using the agent (= refine your mental model of the codebase)
Mindlessly typing something is not much better than copy and pasting?
I could maybe see it if you asked it to spit out pseudocode you had to rewrite. At least there’s some translation there…
But this is bizarre. Write it yourself at that point. Is it any faster (or faster at all frankly) to prompt what you want, manually write it out, and maybe even make adjustments as you go? I’d argue not.
The way I wrote code in the past was to just first comment out what I wanted to do, and then underneath write the syntax for it. You could maybe do this too? Take the LLM code, and go through commenting what each section does to be able to effectively break it up? It still seems dumb.
Buy if you code for a job, good luck justifying this to management. “Yeah Claude already gave me the solution, I’ll take the rest of the week to type it out”
Isn't a working program itself the best textbook? It's just a difference in learning methods. Depending on Stack Overflow is also a dependency, and searching for code on GitHub is also a dependency. How much dependency you allow is purely a personal difference, and it varies depending on your own study habits and learning style. Whether your learning method is superior or not likely depends on how your brain works.
People tend to think that the more painful something is, the better it is.
I don't deny that there are talented people who can read the manual and build everything from scratch. But I think that analyzing and rebuilding a working template step by step is also valuable.
I agree with the view that LLMs may cause cognitive decline. But if you go down that path, Socrates already criticized writing for weakening human memory. And how did that turn out? Books became a universal medium for knowledge. Then the internet came along. When Stack Overflow appeared, there was opposition, but it also had explosive adoption. LLMs are just the next step in that sequence.
If there is cognitive decline, I think there's also compensation in other areas. Using LLMs clearly causes some cognitive decline. And I think there are areas that need to be reinforced to compensate.
But having a baseline to work from—modifying already-working code—is genuinely helpful. I don't see what's wrong with using that as a way to learn.
Realistically, LLMs write code much better than most people. In my domain, there are areas where I still write better code than an LLM, especially when it comes to physical constraints it might not understand, but there are far more domains where the LLM writes much better code than I do. In that sense, writing code with an LLM and keeping track of it feels more helpful than I expected.
Practicing solo coding for an hour a day often ends up being mechanical and not very useful. This might actually be more helpful.
1. HIGH-VALUE CODE:
I write it all myself. I will occasionally use AI for mostly mechanical changes, like cleaning up variable names or mass-changes when a function signature has changed. Either way, every line is read carefully. Sometimes this means isolating my high-value code as a library in a separate repo. Usually it's just a note in AGENTS.md, or even a well-written comment at the top of certain files. I'm not obsessive about it, though, as it can't hide from git. And learning what it's trying to change is sometimes a useful insight.
That doesn't stop me from using AI as a consultant. This is the one time I'll use a beast like Fable. Ask it to write a technical/security analysis on a section of code and damn it can pull out some impressive insights. It can't write new code particularly well, but it can inspect code like a boss. But that all stays in the chat window. (And despite being so infrequent, they ends up costing significantly more than all my other AI costs combined!)
2. BOILERPLATE/PROCEDURAL CODE:
I'll write the first draft, but once I've set the tone, I'll allow AI to build and maintain it. I keep on top of things like a senior manager, just to make sure it's not doing stupid things. Every few days I tell it to mow its own grass: AI is good at recognising its own stupidity, you just need to give it an opportunity to look.
3. TEST/HARNESS CODE:
Bring on the slop. If I get nothing else from the AI revolution, it's not having to write another stupid test unit. Nothing makes me happier than setting the AI to work writing every permutation of test I can think of. I will slop this code all day, and I won't read a single line of it. Why should I? If I ever doubt whether a particular test is correct, I'll test the test by breaking the code, not by reading the test. But I almost never catch it out. In my experience, AI is especially good at writing tests. Perhaps more than anything else.
Tests don't just take the form of a few mocks and props in a test harness. In one recent case, my project involved writing a library for the API of an obscure commercial microcontroller-powered device. I took the API documentation and made AI build me a complete simulator. I then made it write a full suite of tests using my client library within the test code. I then got it to run that test suite against real hardware and identify any inconsistencies. From there it could recursively modify the simulator until it became unreasonably good at mimicking the real hardware. I haven't read a single line of its code. But it's now core to the library's CI.
> Bring on the slop.
Bring on Volkswagen tests, right. Because reliably confirming that your code work is not a critical port of the project at all. /s
I can't reliably confirm that you read my post all the way to the end. I pointed out multiple ways where tests are proven. One is to verify the test by breaking the code under test. Another way is to build a fully independent, highly complex test rig that would never be (commercially) feasible without AI.