I only watched the first few minutes and thought no way touch is that simple! It isn’t. Because most of these utilities have a lot of scar tissue handling years of various edge cases and conditions. Something to be said for software engineering. If you try to build for all the edge conditions then getting a somewhat working example out would take too long
I did appreciate the ways he used the man pages to build out the tools.
if it only needs to work on his PC then its not tok bad. these utils are often full of crap because they need to work on everyones pc with wacky locales and filesystems and all sortsof crap this guy wont run into.
its always the case if ud implmenet everything just for your case you can skip most edge cases and software can be much more lean.
granted his claim is maybe a bit broadly worded regardin this point ofcourse.
This could just be from a classic issue that comes up with organic intelligences. Where do people most commonly see ‘touch’ used, after all? It's very common to see it in build processes, used only on empty files, whereas using it to update the timestamp of a file with contents is more obscure (I don't know if I've ever seen it in the wild, personally). If you asked random developers “what does ‘touch’ do?”, and they weren't primed to think unusually carefully about it, you might well get something around creating empty files as the dominant answer, via extrapolating from a biased pile of examples rather than going to an authoritative source.
In natural language, this leads to semantic drift. In programming languages, semantic drift leads to bugs.
Well, if the point is to have something that works well and follows the standard, he shouldn't have done anything at all and just use what was already made.
Why would one would like to truncate an existing file? To overwrite it. It's a very common use case.
What I have never found any use for personally is opening a file in read-write mode, while also truncating it, which is the full extent of w+ behaviour. I cannot think of many reasons to have an output stream to also function as an input.
> Why would one would like to truncate an existing file? To overwrite it. It's a very common use case.
Yeah, but we have plain "w" for that already.
> What I have never found any use for personally is opening a file in read-write mode, while also truncating it, which is the full extent of w+ behaviour. I cannot think of many reasons to have an output stream to also function as an input.
Exactly my point. If I am about to fully overwrite the file content, why would I want to re-read them without even an intermediate fclose()? I know that content already, I've just wrote it out! The only fringe use cases I can think of are really the edge cases of some general scenarios that would normally use "r+" (e.g. databases) but could occasionally go "okay, let's drop everything and start ab initio" — but those can just call ftrunc() in those circumstances. Or just open the file twice with different modes, the filesystem being what it is (global shared mutable state), you'll see your changes through.
well, if we go by the title then I would expect that he's writing the code on a piece of paper and not typing into a computer - heh
But, okay, that could make a valid excuse.
But, what you cannot excuse yourself for, is that you don't really add any functionality. In fact, you don't reproduce all the functionality let alone adding more.
I really hate how the term "slop" has been misused so much that it lost its original meaning (just like a particular group of ppl calling anything they dislike woke)
That's my personal suspicion, why there's no real coding tutorials like that on YouTube.
Having so many eyes to basically rip to shreds your output is probably the reason.
Except now that no one cares about handwritten code, maybe instead people will celebrate the diversity of ways one can write code to do things. Might see a resurgence in little artisanal videos like these
I kind of expected this to happen: Coding will become an "artisanal craft" - much like woodworking for example. Sure for most of your furniture you can just go to ikea but every now and then you might get joy from making something with your own hands.
I'm kinda bored to read these pro AI arguments, which seem to come by waves. Suddenly everyone seems to say the same thing at the same time and it's been like this for months. Previously it was "it's there to stay", and then "you will be left behind if you don't use it", and this week it's "coding by yourself is for hobbies" or "is like artisanal work [which is cute but not really productive]".
Is it the $3B anthropic pretends to spend on marketing doing its job?
I tend to agree with the fact that there seem to be a lot of what are essentially talking points, that recur.
I think this comes from any widespread difference of opinion. The people generally do believe what they are saying, but when it gets to the point where arguments and counterarguments go back and forth unchanged it seems more like declaring their position than trying to convince others.
People can legitimately share a common belief and talk about it. So similar thinga can turn up in good faith. We just have to remember not to talk past each other.
Yup. As a principle SWE with over 30 years writing code, I can say that the business case for hand-written code is now almost completely gone. If you are still writing code entirely by hand and expect to profit from that labor, you are doing yourself or your employer a gross disservice.
As an engineer with 20 years experience, I can say this is bogus. LLMs still - right now, today, with the latest model - make errors that humans would never make. They cannot be trusted. So you have to carefully review everything they produce, which negates the productivity gains. You may as well just write your own code, it will be similarly fast and better quality.
I think both of you could be correct based on a whole host of qualifiers like current SDLC landscape, specific business domain, current delivery pipelines, etc.
You will at least need to use LLM's to check for your manually introduced inconsistencies and security problems, like we see right now in kernels and libraries we considered peak of human achievement.
If an LLM can spot those better than a human, I'd wager it can write code without introducing them in the first place.
Eh, I dunno -- I have an LLM write a feature, then create a PR. Subsequent bot reviews list tons of "problems" (some relevant, others not so much). Why couldn't the first LLM get it right the first time?
Ask your LLM to review the PR before submitting it. Then make it fix the problems. Then repeat the process until there are no more problems. At each step, document where it went wrong in a guide; next time, it can use that guide and will make less errors.
In other words, treat it like a junior that has infinite enthusiasm and talent but zero existing habits or wisdom. If you are not doing something like this, then you are going to have a bad time.
I mean you can be bored all you want, but the technology works. That's the difference. This isn't cryptocurrency where "you gotta get in to make it big (because I need to sell it to you)". It works: the results are real, the code runs, you can read it - the systems it provides are real.
Even the security argument has evaporated: AIs now hack better then people. They're better bug-finders then developers. The tool which exploits your app itself would be another AI system. On a good day you might be smarter then it...but you're not faster, and you don't have the perseverance.
My challenge would simply be: buy a low-tier Claude license. Or some other system - I use Claude so that's the example. Point it at something you wrote. And realize in the span of about an hour how obsolete the human developer writing code truly is.
Yeah, we know. The technology is definitely impressive, and kinda fascinating actually, I must admit. If you ignore the huge externalities and all the problems it creates, that is.
Can you just guys go vibecoding and stop trying to convince us or yourselves?
I don’t know, I feel like he is using a lot more convincing arguments than just being dismissive without real substance like you are. When you say huge externalities are you referring to something that hasn’t been seen in other major industrial paradigm shifts?
Whats the arguments?? They're making sensationalist claims based of headlines and twitter discourse. This is all zeitgeist nonsense, anyone using AI on the day to day for hard problems knows its capabilities are far less than what you'll see claimed on this site, x the everything app, and LinkedIn.
There's genuinely just a ton of people who work on 40th percential problems and are super impressed. There's people who work on 90th percentile problems who are impressed too, but if you put those 40th percentile problem solvers on the 90th percentile problems they would not have the same outcomes. The agents by themselves are not as impressive as they're made out to be.
its like owning a personal computer in 1986. the tv talking how its going to revolutionize everything and yet their too expensive for most people to own or and 40% problem people were getting bogged down with just owning a pc, writing dear john letters to their VCRs for ever calling them difficult to use.
And all the opinions were conflicting, and all the opinions were correct.
Dan Luu's recent blog post [0] reminded me of Gas Town. Gas Town was all the hype when it debuted. Us, mere mortals, had a few days left in the business of software. Now, only the crickets remain in the Gas Town. Not even its creator is there [1].
Gas Town was never "the hype", it was always a bit of a joke, in fact the original post was hard to separate from satire. Linking to the HN comments here would show that.
I went down a rabbit hole of copyright and one of the things that clicked along the way is that a vibe coded app is not copyrightable. If someone says their app is completely Built with AI, they’re also saying: here is a public domain program that you can repackage as you wish.
It’s not quite that simple - there may be minor things here and there that are copyrightable like the name. But anything purely prompt generated is fair game to be copied and reused.
It will be interesting to see this tested, but companies telling people AI is writing their software may be a liability in the not-to distant future.
(This is based on USCO opinions and congressional testimony: https://www.copyright.gov/ai/ . EULAs are different things, and may restrict rights otherwise granted by the law, etc., etc.)
> But anything purely prompt generated is fair game to be copied and reused.
This isn't true.
While the "copyrightability" report[1] mostly deals with images (it was written before coding agents) it's quite clear how code generated by an AI driven by humans can be copyrightable:
> The Office finds that the compilation of these images and text throughout the
Work contains sufficient creativity under Feist to be protected by copyright.
Specifically, the Office finds the Work is the product of creative choices with
respect to the selection of the images that make up the Work and the placement
and arrangement of the images and text on each of the Work’s pages.
and
> Unlike prompts alone, these tools can enable the user to control the selection and placement of individual creative elements. Whether such modifications rise to the minimum standard of originality required under Feist will depend on a case-by-case determination. In those cases where they do, the output should be copyrightable
It depends how that selection is done. If an AI prompts a human with 10 layouts it generates and the human picks ones, it is not copyrightable.
If the human takes some number of individually generated components and composes them in some creative way, the composition is copyrightable but the individual components are not.
One existing precedent was the layout of a graphic novel consisting AI generated panels. The layout was done by a human and protected by copyright. The content of each individual panel was generated by AI and was not. https://www.copyright.gov/docs/zarya-of-the-dawn.pdf
Code is no different. If AI generates a library and GUI, neither are copyrightable. If a dev glues them together, the expressiveness of that work may be, but it also might be the logical application of attaching interfaces and not creative enough to meet the threshold.
a professional software developer can simultaneously become completely dependent on a remote account each hour of each day, and lose the ability to copyright the work that the professional software developer makes, or attach a LICENSE.
Welcome to PCB design only imagine spending an order of magnitude more on tooling.
This is actually my fear for SWEs, there is SO MUCH PRECEDENT for making the entire career path completely dependent on a single company selling a subscription at absolutely deranged prices.
I’m tired of the “omg huge externalities!” argument.
The externalities can literally be solved with just more AI. Especially even better AI as it emerges every year. Problems that primitive agents created back in 2024 have now long been solved by the advanced agents of 2026. And in 2028, agents will solve problems the 2026 agents leave behind.
There is never going to be a moment where a massive problem created by AI finally arrives and everyone throws their hands up and says “oh no we should have just coded by hand all along!” Most software doesn’t really even live that long before it has to be changed due to new requirements.
Every problem will be solved by AI and if not then everything just gets trivially rewritten.
Mind you, we have not even entered the insane era yet where local LLMs etched in silicon chips will be putting out code at an insane 16k/tokens per second for just the cost of electricity. Let that sink in.
It’s over for handwritten code (in professional contexts).
> There is never going to be a moment where a massive problem created by AI finally arrives and everyone throws their hands up and says “oh no we should have just coded by hand all along!”
This is a shortcoming of this way of thinking, IMHO. It's incredibly unlikely that we will never, ever face a massive problem caused by AI.
The problem won’t be because of AI code, it will be because the people controlling the AI have made incorrect fundamental decisions, which would lead to the same problem even if humans were the ones coding.
So I would completely agree with you concerning things like my dayjob. But maybe there are some details and edge cases being overlooked.
Like I would love to make an indie game as good as like Shovel Knight, Celeste, or Into the Breach. I know these are extremely ambitious goals but AI should make my 1000x more productive, right? ... right?
But what Imve realized is that for someone not in the industry and without the experience I'll provably never make it.
You could blame this on the fact that games are asset heavy and involve interaction, but even then, it seems like just pulling a slot machine after your project gets to a certain size.
I'll accept that AI will most certainly be used to great effect in this industry but I don't believe some factory paradigm will arrive leveling an ignoramus like me and an "artisan" to be equal.
Programming has arguably never been the most important part of making video games though, except when people stuff it up and a game doesn't run well on the hardware people want to play it on.
Story, design, game mechanics...these all have something in common, and it's that at a moderate sized studio they're not even slightly the programmers job.
Arguably if you wanted to make games then now is the best time to do it because a lot of the boilerplate won't be in your way.
I agree with this in general and that a game in theory is so much more than code. The the one caveat is that code is the ultimate bottleneck of games. Assets, story, mechanics, design are all mediated my code and that code complexity is what limits scope.
Also, a lot of these aspects are touted as stuff that AI is so good good at: AI can write better fiction, AI can design better, AI can produce better art, AI can infinitely churn out 3D models, etc. There are people 100% sincerely drumming the same beat that we see here concerning code for all these areas... and yet something really isn't there yet.
> Even the security argument has evaporated: AIs now hack better then people.
They hack better (faster, at least) than humans if you have thousands of dollars to spend.
Can you afford to spend that money regularly pentesting all the code you have your AI to write for you, though?
And if you can't, then what link is there between being good at hacking, and being good at writing secure code?
(Those aren't rhetorical questions.)
Another one: When will Bun (at this point a mostly Claude-written project, managed by Anthropic themselves) become at least as secure or reliable as Node.js (a mostly human-written project)? If you're right, it should be any day now. And maybe you are right - maybe by tomorrow Claude will have written a formal proof that Bun is memory-safe and it'll never segfault again.
They arent better bug finders, I can assure you. They definately are great fuzzing tools, but if you really think AI is a better bug hunter than humans there's millions in bounties waiting for you and your claude code session to pick up off the ground. I can assure you its not what you think it is.
I can't personally vouch for whether top AIs have broadly exceeded (median) human ability here, but as far as bounties go (or even just fixing things in existing projects), what I usually hear about involves a lot of maintainers auto-rejecting AI-assisted contributions because they are AI-assisted, with the charitable reason usually being that it's necessary to have some kind of attention gate or else they get flooded out. (To be clear, this seems like a reasonable response to me; the existing structure often wasn't made to handle that kind of influx, and the spectrum of current-day AI includes a ton of lower-quality models and harnesses that will be cheerfully overconfident. The meta signal can be ruined even if there is also high-quality stuff that people who know how to get it can get!) This plus popular anger over AI developments seems to mean that people who do successfully use AI to find bugs have a lot of incentive to not publish anything or try to get it upstreamed because they'll immediately come under fire. Highly reputable people and organizations sometimes use their existing clout to vouch for AI-assisted bug reports for prominent projects, but a lot of others just quietly adopt the AI fix in their own tree and give up on sharing it.
I'm not sure that societal trends can be extrapolated from YouTube videos, certainly not from one YouTube video, or channel. There's a lot of oddball stuff on YouTube that, if taken to be definitive would have you, for example, believing that the human race is becoming MMORPG players from New Zealand.
What it might illustrate is that "No AI" and long-format are currently seen as ways of attracting attention.
The difference is that Ikea is made out of cardboard, while that expensive woodworking table is still made out of solid wood, and the whole construction is quite different too. The manual process still has many benefits over the industrial one, even if its more expensive and labor intensive.
Code on the other side is just text files, if AI writes them or a human makes no difference. You still end up with the same old files. At least for me, AI has long crossed the point where I feel like I can contribute anything useful. The AI understands my own programs better then myself and is far faster and more flexible than I could ever hope to get anywhere near.
I am sorry, but this kind of comment has become too common on this site. It underestimates the kind of code a highly efficient human being coder is capable of writing and managing over a long period of time ( which is what is required in a real job, btw ).
The only thing that I think coding agents are better than good human coders is the ability to generate a lot of usable code in a shorter period of time.
But can they maintain it over a long period of time? I highly doubt that, given the evidence that I have with me. Coding agents under-perform quite often in my line of work. They overproduce and yet are not able to cover the given requirements.
Sometimes they do 100% of the given task, but miss a slight detail ( given clearly in the prompt ), that renders their product unusable. This quite often results in a complete rewrite - consuming more, guess what, tokens!!!
So anyone who thinks that right now, coding agents can replace human coders completely, is really not working with these agents long enough or hard enough.
Once you start demanding from agents the same quality of code and the added responsibility/liability that you would associate with a human code, you will sure see the difference and I can tell you that its a lot.
I work with Opus 5.5 and Codex Astra 6 on a wide range of projects and they are found completely wanting in certain situations where a human would be more trustworthy and produce better results.
While I like the analogy, I'm not sure it holds how I think of AI. I'd do woodworking because I want better quality for less money, and that's why I use AI now. "Hand crafting" code might be fun but it doesn't get you it cheaper or better anymore IMHO
Better quality for less money? No. You do woodworking (if in fact you do) because you enjoy it. The time spent making a decent piece of furniture you could work at McDonald’s and buy a nicer piece in less time it takes to master woodworking.
So, don’t kid yourself! You enjoy it, and if you’ve mastered the skill then naturally you can save money over time. But I’d wager you spend more in hours mastering the skill than if you were to buy a couple pieces of furniture. ;)
It most certainly gets you better. Arguably not cheaper, though that depends on whether you're writing the code yourself (in which case it's cheaper) or hiring it out (in which case it's more expensive). But it's definitely better, because LLMs still are worse than humans at programming.
We are not sure yet.
As every tech we show the benefits first then we see the cracks.
I think we might see more soon.
I think the fact the are hiding the harness and the "real" model is a signal they are not confident they will keep improving.
openai reported 95%+ on some benchmarks with an unpublished harness but when running with public harness and public model it is not a 30%+ improvement it is a 6% if that much...
Having purchased furniture recently, it is not easy to get what you exactly have in mind or get matching things for a room. So there is value in building furniture by hand.
Not sure about coding though. There could be some thing which humans will do better. Maybe humans will always be better at understanding intent of other humans as they share same biology.
Just like how IKEA furniture is plasterboard trash compared to the quality that we used to have with furniture (and can still get if we pay large sums), so has ai reduced code quality to this messy homogeneous pulp.
maybe code will go the way of preciesely specified requirements. I guess that kind of exists already with formal proving, but maybe new languages will along that merge the two even closer. For example, instead of specifying a map data structure, you will specify a way two data elements associate, and AI will try various DS until it hones in on the "best" (highly dubious word there) one.
Or things will settle down enough that people start looking at the numbers and doing the math. LLM tooling can generate code and often that code is correct. In my experience it takes about as long for the tooling to generate the code as it would for me to write the code myself and it also costs more money (my time plus the token cost). The more complicated the task the more likely I have to go back and get the model to re-work the majority of it as it made some assumptions that, after a little thought, turn out to be problematic.
IMHO, at the end of the day it's been about the same amount of time, the cost is a little higher and I don't feel as confident in the end product. I suspect that I am not the only one that feels this way!
I can, but I can also find problems with all the analogies offered by AI apologists.
To me it's not the human in the loop that's weak, it's programming tools that fail to create and utilize bandwidth in the human/computer interaction cycle. Look at how much raw UX bandwidth people can show off when they're pushed to the limits! https://www.youtube.com/watch?v=Ki9EeaF4MLI
I just think the people selling the idea that AI will outstrip people's abilities to express themselves haven't actually considered what the practical constrictions on self-expression are and whether or not the limits we experience in practice are really the limits of human expressive capability.
If we're at the limits of expressive capability, the only improvements will come from replacing humans. If not, we may be able to win quite a lot of ground back just by building better instruments. ...that's why I'm working on replacing git
This is me most days at work, writing code that will be shipped in a mass market product. So long as I'm getting my work done nobody gets to tell me how to do my job.
We mostly saw the guy typing. The clicking sound is very distracting. Using vim didn't do him any favour. I would be much more impressed if we could get more insight into to the thinking and problem solving process. Also smoking is not commonly accepted productivity booster. He could have watched porn instead.
I'm not as certain as you seem to be, but I agree it's at least suspicious. The keypress sounds are admitted to be fake in the comments. The way the camera shakes (or doesn't) in response to keypresses seems odd and physically inconsistent over time; same goes for his apparent manual adjustments of its position/angle. Also the way his eyes track when he looks at his notes doesn't seem to be consistent with the screen layout as shown - he looks noticeably far downward when the notes pane on the screen is horizontally level with vim.
IYKYK. Zero rice. Zero rizz. Sort of person who does things like this would have at least a little rice. Look at Tsoding’s i3status to see what I’m talking about.
I really don't. I've realized I never liked coding. I liked designing and watching manual tasks become autonomous.
So AI software development is basically giving me everything I want: see a problem, iterate on the design I think will fix it, try it out. And then it works.
I have personal Android apps which "I developed" by having Claude update the react native hot reload view as I used them for the main application, and the biggest change in my home server infrastructure has been retooling it to support the fact that almost all the apps I use going forwards will be built and hosted locally.
Using AI more efficiently on your coding tasks? The software engineering aspect of human review is still an indispensable skill on the age of AI.
Will just relying on AI to code everything for you, while you just review and make decisions, keep your coding skills sharp? (At worst, atrophy your manual coding skills)
It's when I wrote some code at work. I couldn't easily complete the task. Each day I go to work and write even more code. The task is far from complete. I have written hundreds of thousands of code so far.
I seem to be closing in the completing the task nowadays, though.
A lot of people have recently started smoking with the expectation that cancer will soon be a nonissue. He just promised not to use AI for the coding itself, but for fixing problems created along the way... ;)
I only watched the first few minutes and thought no way touch is that simple! It isn’t. Because most of these utilities have a lot of scar tissue handling years of various edge cases and conditions. Something to be said for software engineering. If you try to build for all the edge conditions then getting a somewhat working example out would take too long
I did appreciate the ways he used the man pages to build out the tools.
Actual touch.c source: https://github.com/coreutils/coreutils/blob/master/src/touch...
GNU utils are very maximalist though. Busybox touch is much simpler, especially if you cut out the optional ifdeffed susv3 features: https://github.com/mirror/busybox/blob/master/coreutils/touc...
plan9 touch is also pretty clean: https://9p.io/sources/plan9/sys/src/cmd/touch.c
if it only needs to work on his PC then its not tok bad. these utils are often full of crap because they need to work on everyones pc with wacky locales and filesystems and all sortsof crap this guy wont run into.
its always the case if ud implmenet everything just for your case you can skip most edge cases and software can be much more lean.
granted his claim is maybe a bit broadly worded regardin this point ofcourse.
I just looked at the first example / utility he implemented; which was 'touch'
Will this actually update the 'last modified' date for an already existing file?
Also, there's a standard way of processing command line arguments on Linux/Unix, which is not being followed here.
No it will *replace* or create the files "touched" by an empty file.
Major bug in the first command.
Technically, yes, it updates the last modified date, by actually emptying the file. He should have use "w+" or "a" instead of "w".
PS: I like this guy.
This could just be from a classic issue that comes up with organic intelligences. Where do people most commonly see ‘touch’ used, after all? It's very common to see it in build processes, used only on empty files, whereas using it to update the timestamp of a file with contents is more obscure (I don't know if I've ever seen it in the wild, personally). If you asked random developers “what does ‘touch’ do?”, and they weren't primed to think unusually carefully about it, you might well get something around creating empty files as the dominant answer, via extrapolating from a biased pile of examples rather than going to an authoritative source.
In natural language, this leads to semantic drift. In programming languages, semantic drift leads to bugs.
I had forgotten that touch creates files, not just updates the timestamp. I was watching the video and was confused until he demoed its usage.
interesting, i for one didn’t know about updating the timestamps haha, i’ve only ever used touch to create an empty file
Should have used AI
Well, if the point is to have something that works well and follows the standard, he shouldn't have done anything at all and just use what was already made.
It seems 50/50 to me that an LLM would come up with a similar solution.
That's very GitHub Copilot + GPT 4o energy.
its not neccesarily a bug it could be a feature. :D
w+ truncates
Such a baffling option, to be fair, I don't think I've ever used it. Like, what's even the point?
Why would one would like to truncate an existing file? To overwrite it. It's a very common use case.
What I have never found any use for personally is opening a file in read-write mode, while also truncating it, which is the full extent of w+ behaviour. I cannot think of many reasons to have an output stream to also function as an input.
> Why would one would like to truncate an existing file? To overwrite it. It's a very common use case.
Yeah, but we have plain "w" for that already.
> What I have never found any use for personally is opening a file in read-write mode, while also truncating it, which is the full extent of w+ behaviour. I cannot think of many reasons to have an output stream to also function as an input.
Exactly my point. If I am about to fully overwrite the file content, why would I want to re-read them without even an intermediate fclose()? I know that content already, I've just wrote it out! The only fringe use cases I can think of are really the edge cases of some general scenarios that would normally use "r+" (e.g. databases) but could occasionally go "okay, let's drop everything and start ab initio" — but those can just call ftrunc() in those circumstances. Or just open the file twice with different modes, the filesystem being what it is (global shared mutable state), you'll see your changes through.
to be fair the title didn't promise it would be done well
well, if we go by the title then I would expect that he's writing the code on a piece of paper and not typing into a computer - heh
But, okay, that could make a valid excuse.
But, what you cannot excuse yourself for, is that you don't really add any functionality. In fact, you don't reproduce all the functionality let alone adding more.
But it's no AI, it's organic, hand woven, the toil and pain of really putting it down by hand... or something of the sort.
Primitive technology, modern look
True slop ?
The true ultimate expression of the human condition.
I really hate how the term "slop" has been misused so much that it lost its original meaning (just like a particular group of ppl calling anything they dislike woke)
woke slop
[dead]
[dead]
That's my personal suspicion, why there's no real coding tutorials like that on YouTube. Having so many eyes to basically rip to shreds your output is probably the reason.
Except now that no one cares about handwritten code, maybe instead people will celebrate the diversity of ways one can write code to do things. Might see a resurgence in little artisanal videos like these
I kind of expected this to happen: Coding will become an "artisanal craft" - much like woodworking for example. Sure for most of your furniture you can just go to ikea but every now and then you might get joy from making something with your own hands.
I'm kinda bored to read these pro AI arguments, which seem to come by waves. Suddenly everyone seems to say the same thing at the same time and it's been like this for months. Previously it was "it's there to stay", and then "you will be left behind if you don't use it", and this week it's "coding by yourself is for hobbies" or "is like artisanal work [which is cute but not really productive]".
Is it the $3B anthropic pretends to spend on marketing doing its job?
One word: bots.
Talk to other Devs face to face around you. "AI" is suddenly not that great, but a good tool nonetheless.
That's my experience anyway.
We have Linus Torvalds himself using and promoting AI. This kind of rhetoric is ridiculous. It feels like coping
"but a great tool nonetheless".
I run self hosted models, don’t go anywhere near Anthropic/OAI, and I share this view.
Maybe it’s also folks adapting to the new landscape ahead of them, not desperately holding on to the old norm.
I tend to agree with the fact that there seem to be a lot of what are essentially talking points, that recur.
I think this comes from any widespread difference of opinion. The people generally do believe what they are saying, but when it gets to the point where arguments and counterarguments go back and forth unchanged it seems more like declaring their position than trying to convince others.
People can legitimately share a common belief and talk about it. So similar thinga can turn up in good faith. We just have to remember not to talk past each other.
Being realistic about the future doesn't necessarily mean you're in favour of it
It’s the most obvious conclusion, once you start using LLMs. There’s no grand conspiracy here
Yup. As a principle SWE with over 30 years writing code, I can say that the business case for hand-written code is now almost completely gone. If you are still writing code entirely by hand and expect to profit from that labor, you are doing yourself or your employer a gross disservice.
As an engineer with 20 years experience, I can say this is bogus. LLMs still - right now, today, with the latest model - make errors that humans would never make. They cannot be trusted. So you have to carefully review everything they produce, which negates the productivity gains. You may as well just write your own code, it will be similarly fast and better quality.
I think both of you could be correct based on a whole host of qualifiers like current SDLC landscape, specific business domain, current delivery pipelines, etc.
You will at least need to use LLM's to check for your manually introduced inconsistencies and security problems, like we see right now in kernels and libraries we considered peak of human achievement.
If an LLM can spot those better than a human, I'd wager it can write code without introducing them in the first place.
Eh, I dunno -- I have an LLM write a feature, then create a PR. Subsequent bot reviews list tons of "problems" (some relevant, others not so much). Why couldn't the first LLM get it right the first time?
Why don’t humans get it right the first time?
Ask your LLM to review the PR before submitting it. Then make it fix the problems. Then repeat the process until there are no more problems. At each step, document where it went wrong in a guide; next time, it can use that guide and will make less errors.
In other words, treat it like a junior that has infinite enthusiasm and talent but zero existing habits or wisdom. If you are not doing something like this, then you are going to have a bad time.
It doesn't really if you're already reviewing you team's code anyway. In that case it still saves the time writing the software
I mean you can be bored all you want, but the technology works. That's the difference. This isn't cryptocurrency where "you gotta get in to make it big (because I need to sell it to you)". It works: the results are real, the code runs, you can read it - the systems it provides are real.
Even the security argument has evaporated: AIs now hack better then people. They're better bug-finders then developers. The tool which exploits your app itself would be another AI system. On a good day you might be smarter then it...but you're not faster, and you don't have the perseverance.
My challenge would simply be: buy a low-tier Claude license. Or some other system - I use Claude so that's the example. Point it at something you wrote. And realize in the span of about an hour how obsolete the human developer writing code truly is.
Yeah, we know. The technology is definitely impressive, and kinda fascinating actually, I must admit. If you ignore the huge externalities and all the problems it creates, that is.
Can you just guys go vibecoding and stop trying to convince us or yourselves?
I don’t know, I feel like he is using a lot more convincing arguments than just being dismissive without real substance like you are. When you say huge externalities are you referring to something that hasn’t been seen in other major industrial paradigm shifts?
Whats the arguments?? They're making sensationalist claims based of headlines and twitter discourse. This is all zeitgeist nonsense, anyone using AI on the day to day for hard problems knows its capabilities are far less than what you'll see claimed on this site, x the everything app, and LinkedIn.
There's genuinely just a ton of people who work on 40th percential problems and are super impressed. There's people who work on 90th percentile problems who are impressed too, but if you put those 40th percentile problem solvers on the 90th percentile problems they would not have the same outcomes. The agents by themselves are not as impressive as they're made out to be.
its like owning a personal computer in 1986. the tv talking how its going to revolutionize everything and yet their too expensive for most people to own or and 40% problem people were getting bogged down with just owning a pc, writing dear john letters to their VCRs for ever calling them difficult to use.
And all the opinions were conflicting, and all the opinions were correct.
Dan Luu's recent blog post [0] reminded me of Gas Town. Gas Town was all the hype when it debuted. Us, mere mortals, had a few days left in the business of software. Now, only the crickets remain in the Gas Town. Not even its creator is there [1].
[0] We should be more ambitious with the software we build. https://news.ycombinator.com/item?id=50225695
[1] https://x.com/danluu/status/2099800154639581353
Gas Town was never "the hype", it was always a bit of a joke, in fact the original post was hard to separate from satire. Linking to the HN comments here would show that.
I went down a rabbit hole of copyright and one of the things that clicked along the way is that a vibe coded app is not copyrightable. If someone says their app is completely Built with AI, they’re also saying: here is a public domain program that you can repackage as you wish.
It’s not quite that simple - there may be minor things here and there that are copyrightable like the name. But anything purely prompt generated is fair game to be copied and reused.
It will be interesting to see this tested, but companies telling people AI is writing their software may be a liability in the not-to distant future.
(This is based on USCO opinions and congressional testimony: https://www.copyright.gov/ai/ . EULAs are different things, and may restrict rights otherwise granted by the law, etc., etc.)
> But anything purely prompt generated is fair game to be copied and reused.
This isn't true.
While the "copyrightability" report[1] mostly deals with images (it was written before coding agents) it's quite clear how code generated by an AI driven by humans can be copyrightable:
> The Office finds that the compilation of these images and text throughout the Work contains sufficient creativity under Feist to be protected by copyright. Specifically, the Office finds the Work is the product of creative choices with respect to the selection of the images that make up the Work and the placement and arrangement of the images and text on each of the Work’s pages.
and
> Unlike prompts alone, these tools can enable the user to control the selection and placement of individual creative elements. Whether such modifications rise to the minimum standard of originality required under Feist will depend on a case-by-case determination. In those cases where they do, the output should be copyrightable
[1] https://www.copyright.gov/ai/Copyright-and-Artificial-Intell...
It depends how that selection is done. If an AI prompts a human with 10 layouts it generates and the human picks ones, it is not copyrightable.
If the human takes some number of individually generated components and composes them in some creative way, the composition is copyrightable but the individual components are not.
One existing precedent was the layout of a graphic novel consisting AI generated panels. The layout was done by a human and protected by copyright. The content of each individual panel was generated by AI and was not. https://www.copyright.gov/docs/zarya-of-the-dawn.pdf
Code is no different. If AI generates a library and GUI, neither are copyrightable. If a dev glues them together, the expressiveness of that work may be, but it also might be the logical application of attaching interfaces and not creative enough to meet the threshold.
The 2024 opinion reinforces that determination.
a professional software developer can simultaneously become completely dependent on a remote account each hour of each day, and lose the ability to copyright the work that the professional software developer makes, or attach a LICENSE.
So, what's not to like?
Welcome to PCB design only imagine spending an order of magnitude more on tooling.
This is actually my fear for SWEs, there is SO MUCH PRECEDENT for making the entire career path completely dependent on a single company selling a subscription at absolutely deranged prices.
I’m tired of the “omg huge externalities!” argument.
The externalities can literally be solved with just more AI. Especially even better AI as it emerges every year. Problems that primitive agents created back in 2024 have now long been solved by the advanced agents of 2026. And in 2028, agents will solve problems the 2026 agents leave behind.
There is never going to be a moment where a massive problem created by AI finally arrives and everyone throws their hands up and says “oh no we should have just coded by hand all along!” Most software doesn’t really even live that long before it has to be changed due to new requirements.
Every problem will be solved by AI and if not then everything just gets trivially rewritten.
Mind you, we have not even entered the insane era yet where local LLMs etched in silicon chips will be putting out code at an insane 16k/tokens per second for just the cost of electricity. Let that sink in.
It’s over for handwritten code (in professional contexts).
> There is never going to be a moment where a massive problem created by AI finally arrives and everyone throws their hands up and says “oh no we should have just coded by hand all along!”
This is a shortcoming of this way of thinking, IMHO. It's incredibly unlikely that we will never, ever face a massive problem caused by AI.
The problem won’t be because of AI code, it will be because the people controlling the AI have made incorrect fundamental decisions, which would lead to the same problem even if humans were the ones coding.
So I would completely agree with you concerning things like my dayjob. But maybe there are some details and edge cases being overlooked.
Like I would love to make an indie game as good as like Shovel Knight, Celeste, or Into the Breach. I know these are extremely ambitious goals but AI should make my 1000x more productive, right? ... right?
But what Imve realized is that for someone not in the industry and without the experience I'll provably never make it.
You could blame this on the fact that games are asset heavy and involve interaction, but even then, it seems like just pulling a slot machine after your project gets to a certain size.
I'll accept that AI will most certainly be used to great effect in this industry but I don't believe some factory paradigm will arrive leveling an ignoramus like me and an "artisan" to be equal.
Programming has arguably never been the most important part of making video games though, except when people stuff it up and a game doesn't run well on the hardware people want to play it on.
Story, design, game mechanics...these all have something in common, and it's that at a moderate sized studio they're not even slightly the programmers job.
Arguably if you wanted to make games then now is the best time to do it because a lot of the boilerplate won't be in your way.
I agree with this in general and that a game in theory is so much more than code. The the one caveat is that code is the ultimate bottleneck of games. Assets, story, mechanics, design are all mediated my code and that code complexity is what limits scope.
Also, a lot of these aspects are touted as stuff that AI is so good good at: AI can write better fiction, AI can design better, AI can produce better art, AI can infinitely churn out 3D models, etc. There are people 100% sincerely drumming the same beat that we see here concerning code for all these areas... and yet something really isn't there yet.
> Even the security argument has evaporated: AIs now hack better then people.
They hack better (faster, at least) than humans if you have thousands of dollars to spend.
Can you afford to spend that money regularly pentesting all the code you have your AI to write for you, though?
And if you can't, then what link is there between being good at hacking, and being good at writing secure code?
(Those aren't rhetorical questions.)
Another one: When will Bun (at this point a mostly Claude-written project, managed by Anthropic themselves) become at least as secure or reliable as Node.js (a mostly human-written project)? If you're right, it should be any day now. And maybe you are right - maybe by tomorrow Claude will have written a formal proof that Bun is memory-safe and it'll never segfault again.
They arent better bug finders, I can assure you. They definately are great fuzzing tools, but if you really think AI is a better bug hunter than humans there's millions in bounties waiting for you and your claude code session to pick up off the ground. I can assure you its not what you think it is.
I can't personally vouch for whether top AIs have broadly exceeded (median) human ability here, but as far as bounties go (or even just fixing things in existing projects), what I usually hear about involves a lot of maintainers auto-rejecting AI-assisted contributions because they are AI-assisted, with the charitable reason usually being that it's necessary to have some kind of attention gate or else they get flooded out. (To be clear, this seems like a reasonable response to me; the existing structure often wasn't made to handle that kind of influx, and the spectrum of current-day AI includes a ton of lower-quality models and harnesses that will be cheerfully overconfident. The meta signal can be ruined even if there is also high-quality stuff that people who know how to get it can get!) This plus popular anger over AI developments seems to mean that people who do successfully use AI to find bugs have a lot of incentive to not publish anything or try to get it upstreamed because they'll immediately come under fire. Highly reputable people and organizations sometimes use their existing clout to vouch for AI-assisted bug reports for prominent projects, but a lot of others just quietly adopt the AI fix in their own tree and give up on sharing it.
What's your opinion?
I'm not sure that societal trends can be extrapolated from YouTube videos, certainly not from one YouTube video, or channel. There's a lot of oddball stuff on YouTube that, if taken to be definitive would have you, for example, believing that the human race is becoming MMORPG players from New Zealand.
What it might illustrate is that "No AI" and long-format are currently seen as ways of attracting attention.
There has been a "handmade" software subculture for years. Think of handmade hero or treating he better software conference.
>for example, believing that the human race is becoming MMORPG players from New Zealand.
To be fair this is just about everyone I meet these days. Nzgdc was just full of them.
The difference is that Ikea is made out of cardboard, while that expensive woodworking table is still made out of solid wood, and the whole construction is quite different too. The manual process still has many benefits over the industrial one, even if its more expensive and labor intensive.
Code on the other side is just text files, if AI writes them or a human makes no difference. You still end up with the same old files. At least for me, AI has long crossed the point where I feel like I can contribute anything useful. The AI understands my own programs better then myself and is far faster and more flexible than I could ever hope to get anywhere near.
I am sorry, but this kind of comment has become too common on this site. It underestimates the kind of code a highly efficient human being coder is capable of writing and managing over a long period of time ( which is what is required in a real job, btw ).
The only thing that I think coding agents are better than good human coders is the ability to generate a lot of usable code in a shorter period of time.
But can they maintain it over a long period of time? I highly doubt that, given the evidence that I have with me. Coding agents under-perform quite often in my line of work. They overproduce and yet are not able to cover the given requirements.
Sometimes they do 100% of the given task, but miss a slight detail ( given clearly in the prompt ), that renders their product unusable. This quite often results in a complete rewrite - consuming more, guess what, tokens!!!
So anyone who thinks that right now, coding agents can replace human coders completely, is really not working with these agents long enough or hard enough.
Once you start demanding from agents the same quality of code and the added responsibility/liability that you would associate with a human code, you will sure see the difference and I can tell you that its a lot.
I work with Opus 5.5 and Codex Astra 6 on a wide range of projects and they are found completely wanting in certain situations where a human would be more trustworthy and produce better results.
I can't believe that in the space of only a few short years the notion of coding by hand is, by many people, seen as over. It's not.
While I like the analogy, I'm not sure it holds how I think of AI. I'd do woodworking because I want better quality for less money, and that's why I use AI now. "Hand crafting" code might be fun but it doesn't get you it cheaper or better anymore IMHO
Better quality for less money? No. You do woodworking (if in fact you do) because you enjoy it. The time spent making a decent piece of furniture you could work at McDonald’s and buy a nicer piece in less time it takes to master woodworking.
So, don’t kid yourself! You enjoy it, and if you’ve mastered the skill then naturally you can save money over time. But I’d wager you spend more in hours mastering the skill than if you were to buy a couple pieces of furniture. ;)
It most certainly gets you better. Arguably not cheaper, though that depends on whether you're writing the code yourself (in which case it's cheaper) or hiring it out (in which case it's more expensive). But it's definitely better, because LLMs still are worse than humans at programming.
We are not sure yet. As every tech we show the benefits first then we see the cracks.
I think we might see more soon. I think the fact the are hiding the harness and the "real" model is a signal they are not confident they will keep improving. openai reported 95%+ on some benchmarks with an unpublished harness but when running with public harness and public model it is not a 30%+ improvement it is a 6% if that much...
I really wonder why they need to do that
The furniture you buy perhaps, but actual construction projects still use actual carpenters, not ikea.
I think theres a good chance coding will go down the artisan route, but thats far less niche commercially than you think.
just like artisanal pasta (read: a human willingly spend time to cut the pasta from the pasta machine), we'll charge double
Having purchased furniture recently, it is not easy to get what you exactly have in mind or get matching things for a room. So there is value in building furniture by hand.
Not sure about coding though. There could be some thing which humans will do better. Maybe humans will always be better at understanding intent of other humans as they share same biology.
Just like how IKEA furniture is plasterboard trash compared to the quality that we used to have with furniture (and can still get if we pay large sums), so has ai reduced code quality to this messy homogeneous pulp.
IKEA has a variety of furniture, ranging from fiberboard trash (which isn’t trash if used correctly) to solid hardwood with thick steel legs.
You get what you pay for. With Ikea, it’s at least obvious up front what quality band it’s in.
maybe code will go the way of preciesely specified requirements. I guess that kind of exists already with formal proving, but maybe new languages will along that merge the two even closer. For example, instead of specifying a map data structure, you will specify a way two data elements associate, and AI will try various DS until it hones in on the "best" (highly dubious word there) one.
Or things will settle down enough that people start looking at the numbers and doing the math. LLM tooling can generate code and often that code is correct. In my experience it takes about as long for the tooling to generate the code as it would for me to write the code myself and it also costs more money (my time plus the token cost). The more complicated the task the more likely I have to go back and get the model to re-work the majority of it as it made some assumptions that, after a little thought, turn out to be problematic.
IMHO, at the end of the day it's been about the same amount of time, the cost is a little higher and I don't feel as confident in the end product. I suspect that I am not the only one that feels this way!
Uh huh.
Like writing! Everyone knows that after the printing press, nobody needed to know how to read or write anymore
I’m with you in spirit but surely you can find problems with this analogy.
I can, but I can also find problems with all the analogies offered by AI apologists.
To me it's not the human in the loop that's weak, it's programming tools that fail to create and utilize bandwidth in the human/computer interaction cycle. Look at how much raw UX bandwidth people can show off when they're pushed to the limits! https://www.youtube.com/watch?v=Ki9EeaF4MLI
I just think the people selling the idea that AI will outstrip people's abilities to express themselves haven't actually considered what the practical constrictions on self-expression are and whether or not the limits we experience in practice are really the limits of human expressive capability.
If we're at the limits of expressive capability, the only improvements will come from replacing humans. If not, we may be able to win quite a lot of ground back just by building better instruments. ...that's why I'm working on replacing git
This is me most days at work, writing code that will be shipped in a mass market product. So long as I'm getting my work done nobody gets to tell me how to do my job.
[dead]
I discovered his videos a few days ago. I imagine the "algorithm" pushed those into a lot of HN users' feeds.
Just this morning, I was wondering if it would make sense to share it on HN. And here it it on the front page. What a small world.
Finally a video where creator's face overlaid on top of the content actually contributes value.
We mostly saw the guy typing. The clicking sound is very distracting. Using vim didn't do him any favour. I would be much more impressed if we could get more insight into to the thinking and problem solving process. Also smoking is not commonly accepted productivity booster. He could have watched porn instead.
No splits timer. No frame perfect tricks. Uses "exploits" (touch truncates file). Terrible strats.
Swapping to printf("Usage: %s ...\n",argv[0]); could easily save 60 key presses.
Let us know when somebody creates a No-ABE category.
Most (if not all) of this video is AI.
I'm not as certain as you seem to be, but I agree it's at least suspicious. The keypress sounds are admitted to be fake in the comments. The way the camera shakes (or doesn't) in response to keypresses seems odd and physically inconsistent over time; same goes for his apparent manual adjustments of its position/angle. Also the way his eyes track when he looks at his notes doesn't seem to be consistent with the screen layout as shown - he looks noticeably far downward when the notes pane on the screen is horizontally level with vim.
How so?
IYKYK. Zero rice. Zero rizz. Sort of person who does things like this would have at least a little rice. Look at Tsoding’s i3status to see what I’m talking about.
Just a guy using libc in fact...
Should have done it without a mouse
This reminds of the CoderPad-style interview. Kinda triggering, but might be a good way to practice the muscle! Anyone else agree?
Would expect this on hackertube.com.
https://news.ycombinator.com/item?id=50319510
- waiting for no AI, no game engine, pure c++ modern warfare 4 game from scratch
What did the LLM bug check say?
[flagged]
Thanks to AI, yes.
:(
This cursor sound is almost asmr-ish, damn.
sometimes, I really miss coding by hand
I really don't. I've realized I never liked coding. I liked designing and watching manual tasks become autonomous.
So AI software development is basically giving me everything I want: see a problem, iterate on the design I think will fix it, try it out. And then it works.
I have personal Android apps which "I developed" by having Claude update the react native hot reload view as I used them for the main application, and the biggest change in my home server infrastructure has been retooling it to support the fact that almost all the apps I use going forwards will be built and hosted locally.
What is stopping you?
Probably "I miss coding by hand and being paid for that".
Might be a 9-5 coder.
Personally I let Claude handle most code while I do some coding challenges by hand on the side to keep the brain sharp.
This is the way.
It would be better if someone knew or thought about where the way leads...
Using AI more efficiently on your coding tasks? The software engineering aspect of human review is still an indispensable skill on the age of AI.
Will just relying on AI to code everything for you, while you just review and make decisions, keep your coding skills sharp? (At worst, atrophy your manual coding skills)
This is the way.
What's wrong with being a 9-5 coder?
It's when I wrote some code at work. I couldn't easily complete the task. Each day I go to work and write even more code. The task is far from complete. I have written hundreds of thousands of code so far.
I seem to be closing in the completing the task nowadays, though.
It has been 5 years so far.
Nothing wrong with a 9-5 coder IMHO by itself.
Thing is if you do things by hand or do not, It's still for company. You'd be penalized in some cases for not using.
For personal software yes it's a good exercise.
Now, for contrast, I'd love to see what Opus 5.5 and/or Astra can do when set to do the same thing on a 52m wall clock limit.
The thing with AI is that nobody really interested in what it can do (to some degree).
We all know it's super-duper-productive and can write a lot of code in a short period of time. But I really doubt people would watch a stream on that.
*while smoking cigs
A lot of people have recently started smoking with the expectation that cancer will soon be a nonissue. He just promised not to use AI for the coding itself, but for fixing problems created along the way... ;)
Dude is a bit unhinged, but I like him. I'll keep an eye on him.