Given that some 80% of developers are now using AI in their regular work, blob-util is almost certainly the kind of thing that most developers would just happily have an LLM generate for them. Sure, you could use blob-util, but then you’d be taking on an extra dependency, with unknown performance, maintenance, and supply-chain risks.
Letting LLM write utility code is a sword that cuts both ways. You often create a throw-away code that is unproven and requires maintenance. It's not a guarantee that the blobutil or toString or whatever created by AI won't fail at some edge cases. That's why e.g. in Java there is Apache commons which is perceived as an industry standard nowadays.
Exactly. When you assume blob-util to be a utility library that has been in use for quite a while by many people in many different contexts, hasn't seen much changes and just "works", IMHO the risk of weird bugs is a lot larger with LLM-generated code. Code generated by LLM's often have the problem that the code seems logical, but then contain weird bugs that aren't immediately obvious.
This mostly sounds like a good thing to me from a utilitarian standpoint. Getting all your utility classes from somewhere like npm and creating dependencies on 20 different people and organizations who may or may not maintain their software has been a security nightmare with many highly public examples. If a LLM writes a utility class for me then my supply chain is smaller, meaning less surface area to attack plus I probably benefit from some form of security through obscurity for whatever non-trivial amount that's worth. "Downside" is I don't have some rando, probably unpaid labor out there updating a piece of my app for me...
Your supply chain is superficially fewer, but not smaller. The way you're counting the number of suppliers is heterogeneous: ChatGPT has a bigger surface area than 20 individuals.
If my unit tests run through, i don't have 'unproven' code. I have well working code which doesn't need to go through a dependency hell upgrade cycle just because one function in that lib, i don't use, has some CVE too high to be ignored.
It's not a new thing either, many years ago there was already the debate whether you should trust utility code copied from SO or use an NPM library. In fact, I'm 99% confident that the slew of single function NPM libraries became a thing because of that mindset.
Less incentive to write small libraries. Less incentive to write small tutorials on your own website. Unless you are a hacker or a spammer where your incentives have probably increased. We are entering the era of cheap spam of everything with little incentive for quality. All this for the best case outcome of most people being made unemployed and rolling the dice on society reorganising to that reality.
> or a spammer where your incentives have probably increased.
Slight pushback on this. The web has been spammed with subpar tutorials for ages now. The kind of medium "articles" that are nothing more than "getting started" steps + slop that got popular circa 2017-2019 is imo worse than the listy-boldy-emojy-filled articles that the LLMs come up with. So nothing gained, nothing lost imo. You still have to learn how to skim and get signals quickly.
I'd actually argue that now it's easier to winnow the slop. I can point my cc running in a devcontainer to a "tutorial" or lib / git repo and say something like "implement this as an example covering x and y, success condition is this and that, I want it to work like this, etc.", and come back and see if it works. It's like a litmus test of a tutorial/approach/repo. Can my cc understand it? Then it'll be worth my time looking into it. If it can't, well, find a different one.
I think we're seeing the "low hanging fruit" of slop right now, and there's an overcorrection of attitude against "AI". But I also see that I get more and more workflows working for me, more or less tailored, more or less adapted for me and my uses. That's cool. And it's powered by the same underlying tech.
The thing is, what is the actual point of this approach? Is it for leaning? I strongly believe there’s no learning without inmersion and practice. Is it for automation? The whole idea of automation is to not think about the thing again unless there’s a catastrophic error, it’s not about babysitting a machine. Is it about judgment? Judgment is something you hone by experiencing stuff then deciding whether it’s bad or not. It’s not something you delegate lightly.
The problem isn't that AI slop is doing something new. Phishing, blogspam, time wasting PRs, website scraping, etc have all existed before.
The problem is that AI makes all of that far, far easier.
Even using tooling to filter articles doesn't scale as slop grows to be a larger and larger percentage of content, and it means I'm going to have to consider prompt injections and running arbitrary code. All of this is a race to the bottom of suck.
The difference is that the cost of slop has decreased by orders of magnitude. What happens when only 1 in 10,000 of those tutorials you can find is any good, from someone actually qualified to write it?
I was searching a specific niche on Youtube today, and scrolled endlessly trying to find something that wasn't AI generated. Youtube is being completely spammed.
On that - naive proposition; shouldn't we establish say "humanTube" - service which would strictly prohibit AI content? With all this AI slop engulfing our web 2.0 - maybe this is the time and place to establish the new web for "nerds" i.e. people who care for the real thing? Just as our current web once was a place for scientists and engineers mostly, maybe we now need something as this? I feel the flaws in my own point, but maybe it's not all hopeless?
I see this as an absolute win. The state of micro dependencies of js was a nightmare that only happened because a lot of undereducated developers flooded the market to get that sweet faang money.
Now that both have dried up I hope we can close the vault door on js and have people learn how to code again.
Your first sentence cheers that we're moving from NPM micropackages to LLM-generated code, and then you say this will result in people having to learn to code again.
I don't see how the conclusion follows from this.
There will be many LLM-generated functions purporting to do the same thing, and a bug in one of them that gets fixed means only one project gets fixed instead of every project using an NPM package as a dependency.
There will be many LLM-generated functions doing the same thing, in the same project, unless the human pays attention.
I've been playing a lot recently with various models, lately with the expensive claude models (API, for the large context windows), and in every attempt things are really impressive at the beginning, and start going south once the codebase reaches about 10k to 15k lines of code. Even with tools split out into a separate library and separate documentation at that point it has a tendency to generate tool functions again in the module it's currently working on over taking the already defined one in the helper library.
Oh god, without tree shaking, lodash is such a blight.
I've seen so many tiny packages pull in lodash for some little utility method so many times. 400 bytes of source code becomes 70kb in an instant, all because someone doesn't know how to filter items in an array. And I've also seen plenty of projects which somehow include multiple copies of lodash in their dependency tree.
Its such a common junior move. Ugh.
Experienced engineers know how to pull in just what they need from lodash. But ... most experienced engineers I know & work with don't bother with it. Javascript includes almost everything you need these days anyway. And when it doesn't, the kind of helper functions lodash provides are usually about 4 lines of code to write yourself. Much better to do that manually rather than pull in some 70kb dependency.
I don't quite understand your argument. Wasn't the post about how users might replace transparent dependencies with transparent LLM drop-ins? I don't see how having an LLM to do the same job would enable someone to learn more. They're probably the kind of person who will ask the LLM to perform a refactor when problems arise, so they won't learn that much through osmosis.
Yeah, you're sort of swapping in one issue for another. I agree that micro dependencies needs to be tackled, but perhaps not using LLMs.
I could fear that rather than keep pushing for a Javascript standard library, which would encompass all these smaller function, we now just get more or less the same defective implementations generated by LLMs, but hidden in thousands of repos, where tools won't find security issues. At least with NPM we can pull in updated versions with NPM tells us that we're running an outdated version. Who is going to traverse your proprietary code base and let you know that the vibe coded left-pad Claude put in three years ago is buggy?
> the era of small, low-value libraries like blob-util is over.
Thankfully (not against blob-util specifically because I've never intentionally used it), I wouldn't completely blame llms either since languages like Go never had this dependency hell.
npm is a security nightmare not just because of npm the package manager, because the culture of the language rewards behavior such as "left-pad".
Instead of writing endless utilities for other project to re-use, write actual working things instead - that's where the value/fun is.
Yep something like blob util could be a blog post or a gist (or several stack overflow answers). And a lot of NPM library falls under that. They always have the anemic standard library of JavaScript forgetting that the C library is even smaller.
Another thing about Go, and more generally, the statically-typed languages, is that these utility functions are generally safer to get a once-over and a commit into the code base than in dynamically-typed languages. Something that joins an array of strings with a joiner in some non-trivial way is a lot easier to review when you don't have to worry about it getting an array of DOM nodes and a NaN for the join character, and all the other things that could come in. Well-tested little utilities for Javascript are valuable, and I don't 100% trust LLMs to not give me code based on common misconceptions rather than the actually correct code in that case. Or at least, correct for some definition of correct, which in the dynamic languages can itself be some work to figure out.
Bad take. In the first place, the appropriate time order of magnitude for how long static type checkers have existed for JS is "decades". In the second place, the function that "joins an array of strings with a joiner" is the wrong place to worry about whether that array is actually an array of DOM nodes or the joiner is NaN.
I don’t think open source is going anywhere. It’s posed to get significantly stronger — as the devs which care about it learn how to leverage AI tools to make things that corporate greasemonkeys never had the inspiration to. Low quality code spammers are just marketing themselves for jobs where they can be themselves: soulless and devoid of creative impulse.
That’s the thing: open source is the only place where the true value (or lack of value) of these tools can be established — the only place where one can test mettle against metal in a completely unconstrained way.
Did you ever want to build a compiler (or an equally complex artifact) but got stuck on various details? Try now. It’s going to stand up something half-baked, and as you refine it, you will learn those details — but you’ll also learn that you can productively use AI to reach past the limits of your knowledge, to make what’s beyond a little more palatable.
All the things people say about AI is true to some degree: my take is that some people are rolling the slots to win a CRUD app, and others are trying to use it to do things that they could only imagine before —- and open source tends to be the home of the latter group.
True innovation will come from open source for sure. As the developers don't have the same economic incentives to be "safe", "ethical" "profitable" or whatever. large corporations know this and fear this development. That's why i expect a significant lobbying to take hold in USA that will try and make local AI systems illegal. And I think they will be very convincing to the government. Because the government also fears the "peasants" and giving them any true semblance of real AGI like systems. I bet very soon we will start seeing various classifications that will define what is legal and what is not for a citizen to possess or use.
> That's why i expect a significant lobbying to take hold in USA that will try and make local AI systems illegal.
I think they're going to be using porn and terrorism (as usual) to do that, but also child suicide. I also think they're going to leverage this rhetoric to lock down OSes in general, by making them uninstallable on legally-available hardware unless approved, because approved OSes will only be able to run approved LLMs.
Meaning that I think LLMs/generative AI will be the lever to eliminate general-purpose computing. As mobile went, so will desktop.
I think this is inevitable. The real question for me is whether China will partner with the west on this, or whether we will be trading Chinese CPUs with each other like contraband in order to run what we want.
> any true semblance of real AGI like systems.
This is the only part I don't agree with. This isn't going to happen, but I'm not even sure it would be more useful than what we have. We have billions of full AGI machines walking around, and most of them aren't great. I'm talking about restrictions on something technically barely better than what we have now; maybe only a significant bit more compute-efficient. Training techniques will probably be where we get the most improvements.
It's really not. Every project of any significance is now fending off AI submissions from people who have not the slightest fucking clue about what is involved in working on long-running, difficult projects or how offensive it is to just slather some slop on a bug report and demand it is given scrutiny.
Even at the 10,000 feet view it has wasted people's time because they have to sit down and have a policy discussion about whether to accept AI submissions, which involves people reheating a lot of anecdotal claims about productivity.
Having learned a bit about how to write compilers I know enough to know that I can guarantee you that an AI cannot help you solve the difficult problems that compiler-building tools and existing libraries cannot solve.
It's the same as it is with any topic: the tools exist and they could be improved, but instead we have people shoehorning AI bollocks into everything.
This isn't an AI issue. It is a care issue. People shouldn't submit PRs to project where they don't care enough to understand the project they are submitting to or the code they are submitting. This has always been a problem, there is nothing new. The thing that is new is more people can get to a point where they can submit regardless of their care or understanding. A lot of people are trying to gild their resume by saying they contributed to a project. Blaming AI is blaming the wrong problem. AI is a a tool like a spreadsheet. Project owners should instead be working ways to filter out careless code more efficiently.
yeah we are getting lots of "I don't know how to do this and AI gave me this code that doesn't work, can you fix it" or "AI said it can do this" and the feature doesn't exist... some people will even argue and say "but AI said it doesn't take long, why won't you add it"
Sounds like a lot of FUD to me — if major projects balk at the emergence of new classes of tools, perhaps the management strategy wasn’t resilient in the first place?
Further: sitting down to discuss how your project will adapt to change is never a waste of time, I’m surprised you stated it like that.
In such a setting, you’re working within a trusted party — and for a major project, that likely means extremely competent maintainers and contributors.
I don’t think these people will have any difficulty adapting to the usage of these tools …
I am sure I am not the only one who thinks these micro-dependencies are worthless anyway. You'd be better off just listing the functions in a markdown file for people to copy over than ship an entire package for it.
This isn't "small" open source, "small" would be something you put together in a week or weekend. These are like "micro" projects, where more work goes into actually publishing and maintaining the repository than actually writing the library.
I like the approach C sometimes takes, with the "tiny header file" type of libraries. Though I guess that also stems from the lack of a central build system.
Vendoring dependencies is how I remember doing it for web projects pre-NPM. Find an open source, well tested library and copy the source into your project.
You have to manually update for any releases you care about, but that is also an incentive to keep dependency count low.
The security maintenance of the ten lines of code I have read and copied into my code is the same as the ten lines of code next to it, that I have written myself.
> I’m still trying to figure out what kinds of open source are worth writing in this new era
Is there any upside to opensourcing anything anymore? Anything published today becomes training data for the next model, with no attribution to the original work.
If the goal is to experiment, share ideas, or let others learn from the work, maybe the better default now is "source available", instead of FOSS in the classic sense. It gives people visibility while setting clearer boundaries on how the work can be used.
I learned most of what I know thanks to FOSS projects so I'm still on the fence on this.
I keep seeing this attitude and I don't really understand it at all; there's no upside to publishing open source work because it might be utilized by more people, is that correct?
Or is it the attribution? There are many many libraries I have used and continue to use and I don't know the author's internet handle or Christian name. Does that matter? Why?
I have written a lot of code that my name is no longer attached to. I don't care and I don't know why anyone does. If it were valuable I would have made more money off of it in the first place, and I don't have the ego to just care that people know it's my code either.
I want the things I do today to have an upside for people in the future. If that means I write code that gets incorporated into a model that people use to build things N number of years from now, that's great. That's awesome. Why the hell is that apparently so demotivating to some people?
> there's no upside to publishing open source work because it might be utilized by more people, is that correct?
I believe the perspective here is "I make code for fellow hackers to look into, critique, be educated on, or simply play with". If you see the hacker scene as a social one, LLM's are an awful black hole that sucks up everything around it and ruins this collaboration.
Not to mention that the hacker scene was traditionally thought to be a rejection of what we now call "Big Tech". Corporate was free to grab the code, but it didn't matter much as long as the scene was kept. Now even that invisible social contract is broken.
But I suppose if you're of a diehard FOSS mentality, "Free" means "Free". Free to be used to build, or destroy society at its whim. a hivemind to meld into and progress the overall understanding of science, for science's sake.
I'll admit the last few years have had me questioning what I truly want to do within the on these two mentalities.
> I want the things I do today to have an upside for people in the future.
I think most would agree with this, but the way things work today don't support it. As of now, AI gains are privatized while the losses are socialized. Until that one-sided imbalance is addressed, LLM's "use" of open source is unbounded and nonreciprocal.
Attribution is a big part of the human experience. Your response frames it as ego driven, but it's also what motivates people to maintain code that is not usually compensated, it's also what builds reputation, trust, communities, and even careers.
Until that’s figured out, we can still share, but maybe in ways that are closer to one another, or under distribution models that reflect the reality we’re in rather than the one we used to have.
People were publishing "open source" not because they gave a shit, but to pad their resumes to get hired at FAANG - like they're paid by the number of projects published on Github It was always human slop. I'm glad this practice is dying. Soon we might be able to reinstate the assumption that the remaining open source projects contain interesting ideas that are worth your eyeball time to look at.
This is kinda how I've felt for months. I don't have any interest in continuing existing open source projects and don't want to create any new ones.
What's the point?
All of my personal projects for the past few months have been entirely private, I don't even host them on Github anymore, I have a private Forgejo instance I use instead.
I also don't trust any new open source project I stumble upon anymore unless I know it was started at least a year ago.
You're not obligated to give away your mind for free. You're free to share, of course. But sharing implies reciprocity, a back and forth. The internet used to be like that, but if the environment changes, you adapt your behavior accordingly.
In the long run I think it's time to starve the system from input until it's attitude reverts to reciprocal. It's not what I'd want, but it seems necessary. People learn from consequences, not from words alone
Code is only useful if it's used. I could write a ton of code and be buried with it, or publish it for people (or AI software, or dolphins or aliens) to use. Who has the energy to have Anubis measure whether my code, or yours, is ethical enough? I'm going to die someday!
This author assumes that open sourcing a package only delivers value if is added as a dependency. Publicly sharing code with a permissive license is still useful and a radical idea.
Yep. Even just sharing code with any license is valuable. Much I have learned from reading an implementation of code I have never run even once. Solutions to tough problems are an under-recognized form of generosity.
This is a point where the lack of alignment between the free beer crowd and those they depend on is all too clear. The free beer enthusiast cannot imagine benefiting from anything other than a finished work. They are concerned about the efficient use of scarce development bandwidth without consciousness of why it is scarce or that it is not theirs to direct. They view solutions without a hot package cache as a form of waste, oblivious to how such solutions expedite the development of all other tools they depend on, commercial or free.
Yeah if I find some (small) unmaintained code I need, I will just copy it (then add in my metrics and logging standards :)
It shouldn't be a radical idea, it is how science overall works.
Also, as per the educational side, I find in modern software ecosystem, I don't want to learn everything. Excellent new things or dominantly popular new things, sure, but there are a lot of branching paths of what to learn next, and having Claude code whip up a good enough solution is fine and lets me focus on less, more deeply.
(Note: I tried leaving this comment on the blog but my phone keyboard never opened despite a lot of clicking, and on mastodon but hit the length limit).
A copyleft license is much better because it ensures that it will remain open and in most cases large companies won't use it, making sure they will have to shell out the money to hire someone to do it instead.
I do agree with this, but there are some caveats. At the end of the day it is time people invest into a project. And that is often unpaid time.
Now, that does not mean it has no value, but it is a trade-off. After about 14 years, for instance, I retired permanently from rubygems.org in 2024 due to the 100k download limit (and now I wouldn't use it after the shameful moves RubyCentral did, as well as the new shiny corporate rules with which I couldn't operate within anyway; it is now a private structure owned by Shopify. Good luck finding people who want to invest their own unpaid spare time into anything tainted by corporations here).
> Sure, you could use blob-util, but then you’d be taking on an extra dependency, with unknown performance, maintenance, and supply-chain risks.
Use of an AI to write your code is also a form of dependency. When the LLM spits out code and you just dump it in your project with limited vetting, that's not really that different from vendoring a dependency. It has a different set of risks, but it still has risks.
Part of the benefit over a dependency is that the code added will (hopefully) be narrowly tailored to your specific need, rather than the generic implementation from a library that likely has support for unused features.
Not including the unused features both makes the code you are adding easier to read and understand, but it also may be more efficient for your specific use case, since you don't have to take into account all the other possible use cases you don't care about.
But in a lot of cases you can't know all the dependencies, so you lean on the community trusting that a package solves the problem well enough that you can abstract it.
You can pin the dependency and review the changes for security reasons, but fully grasping the logic is non-trivial.
Smaller dependencies are fine to copy at first, but at some point the codebase becomes too big, so you abstract it and at that point it becomes a self-maintained dependency. Which is a fair decision, but it is all about tradeoffs and sometimes too costly.
> Part of the benefit over a dependency is that the code added will (hopefully) be narrowly tailored to your specific need, rather than the generic implementation from a library that likely has support for unused features.
In decent ecosystems there should be low or zero overhead to that.
> Not including the unused features both makes the code you are adding easier to read and understand, but it also may be more efficient for your specific use case, since you don't have to take into account all the other possible use cases you don't care about.
Maybe. I find generic code is often easier to read than specialised custom implementations, because there is necessarily a proper separation of concerns in the generic version.
Right, but you do avoid worries like "will I have to update this dependency every week and deal with breaking changes?" or "will the author be compromised in a supply-chain attack, or do a deliberate protestware attack?" etc. As for performance, a lot of npm packages don't have proper tree-shaking, so you might be taking on extra bloat (or installation cost). Your point is well-taken, though.
> you do avoid worries like "will I have to update this dependency every week and deal with breaking changes?
This is not a worry with NPM. You can just specify a specific version of a dependency in your package.json, and it'll never be updated ever.
I have noticed for years that the JS community is obsessed with updating every package to the latest version no matter what. It's maddening. If it's not broke, don't fix it!
Wouldn't call it a risk in itself, but part of the benefit of using a library, a good and tailored one at least, is that it'll get modernised without my intervention. Even if the code produced for you was state-of-the-art at the moment of inclusion, will it remain that way 5 years from now?
> and you just dump it in your project with limited vetting
Well yes there’s your problem. But people have been doing this with random snippets found on the internet for a while now. The truth is that irresponsibles developers will produce irresponsible code, with or without LLMs
> The truth is that irresponsibles developers will produce irresponsible code, with or without LLMs
True. But the difference is the scale and ease of doing with code generators. With a few clicks you can add hundreds of lines of code which supposedly does the right thing. While in the past, you would get code snippets for a particular aspect of the problem that you are trying to solve. You still had to figure out how to add it to your code base and somehow make it “work”
Node.js's core library is purposefully minimal. Purposefully or not, it's still an issue. Wether you add many small things to your project using an LLM or NPM, it still requires work and the annoying thing is that you'll have to do it for every new Node.js project.
Node.js is very good for IO and it has decent performance even for CPU-intensive work considering it's a dynamic language, but it would sure be nice to have a rich core library like Ruby or Clojure has.
The fact that ClojureScript can do it proves that it's doable even for front-end javascript (using advanced optimisations).
This seems to be specific to the JavaScript ecosystem, and to some extent orthogonal to AI coding. Micro-libraries with trivial-ish functionality are generally more difficult to justify to add as a dependency. I’m sure AI coding agents will eventually learn as well that there’s a balance to strike between implementing stuff yourself and the use of libraries. I’m also sure that people who are interested in actual coding will continue to open-source what they find useful, and other people like them will read their code.
More likely, what we will see is the decline of low effort projects. The JavaScript/ Typescript ecosystem has been plagued with such packages. But that’s more anomalous to the JS community than it is a systemic problem with open source in general.
So if fewer people are including silly dependencies like isEven or leftPad, then I see that as a positive outcome.
> ….but I do think it’s a future where we prize instant answers over teaching and understanding
It depends. For stuff I don’t care about I’m happy to treat it as a black box. Conversely, AI now allows me to do deep dive on essentially anything I’m interested in, which has been a massive boon to my learning ability.
Right now I tend to not use an external library unless the code would be large e.g. http server and/or the library is extremely popular.
Otherwise, writing it myself is much better. It's more customizable and often much smaller in size. This is because the library has to generalize, and it comes with bloats.
Using AI to help write is great because I should understand that anyway whether AI writes it or not or whether it's in an external library.
One example recently is that I built a virtual list myself. The code is much smaller and simpler compared to other popular libraries. But of course it's not as generalizable.
If you're not learning to code, then you want efficient code, so the comments are wasted bytes (ok, not a huge expense, but still).
If you are learning to code, or just want to understand how this code works, then asking an LLM is going to get a lot better result. LLMs are fantastic tutors. Endlessly patient, go down any rabbit hole with you, will continue explaining a concept until you get it, etc. I think they're going to revolutionise education, especially for practical subjects like coding.
Respect to the author for trying to educate, though.
Small open source is still valuable, but the bar is higher. If your project is something that's trivial and nobody just thought to do it before you and bothered to do it after, that's probably not going to survive, but if your project is a small focused tool that handles something difficult really well, it's 100% got a future.
Earlier this year I wrote an interpreter for a niche, proprietary binary format. Someone asked if I could open source it so that they could (more easily) run it on NixOS. I declined as I'm just that strongly opposed to my work being used to train AI models and further entrench the enshittification of the internet.
Curious: So you only release your interpreter as a binary? Is there some license that explicitly forbids use for AI training? (Although it would probably be unclear if that is really enforceable until someone challenges it.)
TLDR: [AI promises] a future where we prize instant answers over teaching and understanding
But what this article and the comments don't say: open-source is mainly a quality metric. I re-use code from popular open-source repo's in part because others have used it without complaints (or document the bugs), in part because people are embarrassed to write poor-quality open-source so it's above-par code, and in part because if there are issues in this corner of the world, this dependency will solve them over time (and I can watch and wait when I don't have time to fix and contribute).
The quality aspect drives me to prefer dependencies over AI when I don't want full ownership, so I'll often ask AI to show open-source projects that do something well.
(As an aside, this article is about AI, but AI is so pervasive now as an issue that it doesn't even need saying in the title.)
I totally agree that the goal should be teaching. But I do wonder which code someone is more likely to actually read and understand: the LLM-generated one in their own codebase, or the one hidden in some npm package they installed to just ship things. I honestly don’t know. Careless devs probably won’t read either as long as it seems to work
I’ve been trying to encourage forking my libraries, or have people just copy them into their codebase and adapt them, e.g. https://github.com/mastrojs/mastro/ (especially the “extension” libs.) But it’s an uphill battle against the culture of convenience over understanding.
>Many software developers will argue that asking a candidate to reverse a binary tree is pointless
Is "reversing a binary tree" actually a thing, or is this a cute kind of "rocket surgery" phrase intentionally mixing reversing a linked list and searching a binary tree?
I think it's a reference to the Google interview problem that the author of Homebrew (IIRC) failed. They were quite upset about it since they have proved their worth through their famous open-source contributions, but got rejected in a LeetCode-like interview.
AI offering the solution for a small problem that probably doesn’t deserve yet another dependency suggests to me that there’s a middle ground that we’ve failed to sufficiently cover: how to socialize code snippets that you’re meant to just inline into your project. Stack Overflow is probably the closest we’ve gotten to a generalized solution and it doesn’t exactly feel like a good one.
I came across this once before in the form of a react hooks library that had no dependency to install. It was just a website and when you found the hook you wanted you were meant to paste it into your project.
I suppose some people would see this as progress: fewer dependencies, more robust code (even if it’s a bit more verbose), quicker turnaround time than the old “search npm, find a package, read the docs, install it” approach.
Why would randomized code be more robust?
Also, how is searching/reading the docs slower than checking the correctness of a randomized function?
Open source exists because coding was a significant effort and code was a thing of high value. Unsurprisingly companies hesitated to make the code public and free. All of this is changing now as coding has suddenly become trivial. So, yes, the mission of open source, in general, will be challenged.
Why wouldn't you? Your codebase (if you're a business) exists to make you money, people being able to copy some unknown portions of it without further license if they somehow legally get their hands on a copy of it seems entirely irrelevant.
PS. I think this is much less clear and much less settled law than you are suggesting.
It's more nuanced. If I even have a few lines I can prove are mine, those parts are copywritable in the same way Pride and Prejudice is public domain but pride and prejudice and zombies is copyrighted.
> Even now there’s a movement toward putting documentation in an llms.txt file, so you can just point an agent at it and save your brain cells the effort of deciphering English prose. (Is this even documentation anymore? What is documentation?)
I look at it as why not have the best of both worlds? The docs for my JS framework all have the option of being returned as LLM-friendly text [1].
When I utilize this myself, it's to get help fleshing out skeleton code inside of an app built w/ my framework (e.g, Claude Sonnet w/ these docs in context build a nearly ~90-100% accurate implementation for most stuff I throw at it—anything from little lib stuff up to full-blown API endpoint design and even framework-level implementations of small stuff like helping to refactor the built-in middleware loading). It's not for a lack of desire to read, but rather a lack of desire to build every little thing from scratch when I know an LLM is perfectly capable (and much faster than me).
Will we ever even get a widespread new programming language? AIs work great for coding because they have massive codebases in C/C++/Python/JS/etc. to train on. If I create a language called "foo" tomorrow, no such codebase will exist for it and the lack of AI support for the language will be a huge barrier for adoption so it's going to be much less likely to catch on like Python did back in the day.
I'm working on my own lisp. I've used ChatGPT to generate some cases for my test suite. Somehow it can generate decent code if I instruct it thoroughly enough and provide samples. Maybe the fact all lisps have similar syntax helps?
1. Reducing dependencies is a wrong success metric. You just end up doing more work yourself, except you can't be an expert in everything, so your code is often strictly worse.
2. Regenerating the same solutions with a probabilistic machine will produce bugs a certain percentage of the time. Dependencies are always the same code (when versioned).
3. Cognitive overhead for human review is higher with LLM-generated libs, for no additional benefit.
> Reducing dependencies is a wrong success metric. You just end up doing more work yourself
Except it's just not true in many cases because of social systems we've built. If I want to ship software to Debian I have to make sure that every single of my 3rdparty dependencies is registered and packaged as a proper debian package - a lot of time it will take much less work to rewrite some code than to get 25 100-lines-of-code micro-libraries accepted into debian.
Chances are, even if you deliberately and strategically pick to work on an OSS project that you are positively sure an LLM can’t just spit out on command, it will be capable of doing so by the time you are close to completion. In that sense, one has to either be not inclined to question “what’s the point” or have a bit of gambling mentality in order to work on anything substantial.
That’s not automatically a problem, however. The problem is that even if you do come up with a really cool idea that LLM is not capable of autocompleting, and you release it under a copyleft license (to ensure the project survives and volunteer contributor’s work is not adopted and extinguished by some commercial interest), it will get incorporated into its dataset regardless of the licensing, and thereafter the LLM will be capable of spitting it out and its large corporate operator will be able to monetise your code (allowing anyone with money wishing to build a commercial product based on it).
Now way more than before, i value my feed collection because i know all these blogs are human written, some of them got a post every year, but it is always good quality and human written.
LLMs give us the opportunity to work on more complex projects and gain fuller understanding of the problem space and concepts. Or create tons of slop. Take your pick.
> Claude’s version is pretty close to the blob-util version (unsurprising, since it was probably trained on it!).
AI are thieves!
> I don’t know which direction we’re going in with AI (well, ~80% of us; to the remaining holdouts, I salute you and wish you godspeed!), but I do think it’s a future where we prize instant answers over teaching and understanding.
Google ruined its search engine years ago before AI already.
The big problem I see is that we have become WAY too dependent on these mega-corporations. Which browser are people using? Typically chrome. An evil company writes the code. And soon it will fire the remaining devs and replace them with AI. Which is kind of fitting.
> Even now there’s a movement toward putting documentation in an llms.txt file, so you can just point an agent at it and save your brain cells the effort of deciphering English prose. (Is this even documentation anymore? What is documentation?)
Documentation in general sucks. But documentation is also a hard problem.
I love examples. Small snippets. FAQs. Well, many projects barely have these.
Look at ruby webassembly/wasm or ruby opal. Their documentation is about 99% useless. Or, even worse - rack in ruby. And I did not notice this in the past, in part because e. g. StackOverflow still worked, there were many blogs which helped fill up missing information too. Well all of that is largely gone now or has been slurped up by AI spam.
> the era of small, low-value libraries like blob-util is over. They were already on their way out thanks to Node.js and the browser taking on more and more of their functionality (see node:glob, structuredClone, etc.), but LLMs are the final nail in the coffin.
I still think they have value, but looking at organisations such as rubygems.org disrupt the ecosystem and bleeding it dry by kicking out small hobbyists, I think there is indeed a trend towards eliminating the silly solo devs who think their unpaid spare time is not worthy of anything at all, yet the big organisations eagerly throw down more and more restrictions onto them (my favourite example is the arbitrary 100k download limit for gems hosted at rubygems.org, but look at the new shiny corporate rules on rubygems.org - this is when corporations take over the infrastructure and control it. Ironically this also happened to pypi and they admit this indirectly: https://blog.pypi.org/posts/2023-05-25-securing-pypi-with-2f... - of course they deny that corporations control pypi now, but by claiming otherwise they admit it, because this is how hobbyists get eliminated. Just throw more and more restrictions at them without paying them. Sooner or later they decide to do something better with their time.)
He almost got it right. It's not just the fate of small open source. It's the fate of all programmers now. Why hire a programmer when an LLM costs less, works faster and makes less mistakes (OP compliments better error handling, read the article).
Unless you are a product owner, you have paying clients that love you and your product and won't simply ditch it in favour of a new clone, you are really screwed.
"when an LLM costs less, works faster and makes less mistakes"... indeed, but it doesn't follow at all that it's the fate of all programmers _now_... at least in my experience none of these things are true ATM.
Well, at the very least it costs less than asking intern to look for a lib doing something particular and give some examples... still about as accurate as the intern tho.
So far I've never seen yet a non-programmer release production-grade code using only LLMs. There's just so much to care about (from security, deployments, secret management, event-driven architectures, and a large etc.) that "just" providing a prompt to create an "app" doesn't cut it. You need infra and non-engineers just don't know shit about infra (even if it's 99% managed), you need to deploy your llm-generated code in that infra; that should happen in a ci/cd probably. And what about migrations? Git? Who's setting up the api gateway? I don't mean to say that LLMs don't know how to do that, but you need to instruct them to do so, and even there, they will make silly mistakes and you need to re-instruct them or fix it.
Prompting is just 50% of the work (and the easy part actually). Ask the Head of Product or whoever is there to deploy something valuable to production and maintain it for 6 months while not losing money. It's just not going to happen, not even with truly AGI.
An LLM might be able to replace the majority of the code Sindre Sorhus has put out there, but it's probably a stretch to think that it could replace someone like John Carmack.
Trivial NPM libraries were never needed, but LLMs really are the nail in the coffin for them even when it comes to the most incompetent programmers because now they can literally just ask an LLM to spit out the exact same thing.
That type check is honestly not pointless at all. You can never be certain of your inputs in a web app. The likelihood of that parameter being something other than an arraybuffer is non-zero, and you generally want to have code coverage for that kind of undefined behavior. TypeScript doesn't complain without a reason.
Letting LLM write utility code is a sword that cuts both ways. You often create a throw-away code that is unproven and requires maintenance. It's not a guarantee that the blobutil or toString or whatever created by AI won't fail at some edge cases. That's why e.g. in Java there is Apache commons which is perceived as an industry standard nowadays.
Exactly. When you assume blob-util to be a utility library that has been in use for quite a while by many people in many different contexts, hasn't seen much changes and just "works", IMHO the risk of weird bugs is a lot larger with LLM-generated code. Code generated by LLM's often have the problem that the code seems logical, but then contain weird bugs that aren't immediately obvious.
Agreed. The code available in e.g. Apache commons is not perfect, but it's seasoned and has the edge cases documented
This mostly sounds like a good thing to me from a utilitarian standpoint. Getting all your utility classes from somewhere like npm and creating dependencies on 20 different people and organizations who may or may not maintain their software has been a security nightmare with many highly public examples. If a LLM writes a utility class for me then my supply chain is smaller, meaning less surface area to attack plus I probably benefit from some form of security through obscurity for whatever non-trivial amount that's worth. "Downside" is I don't have some rando, probably unpaid labor out there updating a piece of my app for me...
Your supply chain is superficially fewer, but not smaller. The way you're counting the number of suppliers is heterogeneous: ChatGPT has a bigger surface area than 20 individuals.
2 replies →
The proper way to do it, would be to have an industry standard on the default things people blindly but massivly pull through dependencies.
I also don't get how code can be so massivly inefficient. left-pad needs 9kb to download and the code is a handful of lines: https://www.npmjs.com/package/left-pad?activeTab=code
If my unit tests run through, i don't have 'unproven' code. I have well working code which doesn't need to go through a dependency hell upgrade cycle just because one function in that lib, i don't use, has some CVE too high to be ignored.
It's not a new thing either, many years ago there was already the debate whether you should trust utility code copied from SO or use an NPM library. In fact, I'm 99% confident that the slew of single function NPM libraries became a thing because of that mindset.
Actually sounds like something someone that wrote a really small utility that is surprisingly (to him) used by a lot of people would say.
But the benefit he provided is significantly more than he realizes/acknowledges.
Less incentive to write small libraries. Less incentive to write small tutorials on your own website. Unless you are a hacker or a spammer where your incentives have probably increased. We are entering the era of cheap spam of everything with little incentive for quality. All this for the best case outcome of most people being made unemployed and rolling the dice on society reorganising to that reality.
> or a spammer where your incentives have probably increased.
Slight pushback on this. The web has been spammed with subpar tutorials for ages now. The kind of medium "articles" that are nothing more than "getting started" steps + slop that got popular circa 2017-2019 is imo worse than the listy-boldy-emojy-filled articles that the LLMs come up with. So nothing gained, nothing lost imo. You still have to learn how to skim and get signals quickly.
I'd actually argue that now it's easier to winnow the slop. I can point my cc running in a devcontainer to a "tutorial" or lib / git repo and say something like "implement this as an example covering x and y, success condition is this and that, I want it to work like this, etc.", and come back and see if it works. It's like a litmus test of a tutorial/approach/repo. Can my cc understand it? Then it'll be worth my time looking into it. If it can't, well, find a different one.
I think we're seeing the "low hanging fruit" of slop right now, and there's an overcorrection of attitude against "AI". But I also see that I get more and more workflows working for me, more or less tailored, more or less adapted for me and my uses. That's cool. And it's powered by the same underlying tech.
The thing is, what is the actual point of this approach? Is it for leaning? I strongly believe there’s no learning without inmersion and practice. Is it for automation? The whole idea of automation is to not think about the thing again unless there’s a catastrophic error, it’s not about babysitting a machine. Is it about judgment? Judgment is something you hone by experiencing stuff then deciding whether it’s bad or not. It’s not something you delegate lightly.
The problem isn't that AI slop is doing something new. Phishing, blogspam, time wasting PRs, website scraping, etc have all existed before.
The problem is that AI makes all of that far, far easier.
Even using tooling to filter articles doesn't scale as slop grows to be a larger and larger percentage of content, and it means I'm going to have to consider prompt injections and running arbitrary code. All of this is a race to the bottom of suck.
2 replies →
The difference is that the cost of slop has decreased by orders of magnitude. What happens when only 1 in 10,000 of those tutorials you can find is any good, from someone actually qualified to write it?
23 replies →
> We are entering the era of cheap spam of everything with little incentive for quality
Correction -- sadly, we're already well within this era
I was searching a specific niche on Youtube today, and scrolled endlessly trying to find something that wasn't AI generated. Youtube is being completely spammed.
That was true even before llms
On that - naive proposition; shouldn't we establish say "humanTube" - service which would strictly prohibit AI content? With all this AI slop engulfing our web 2.0 - maybe this is the time and place to establish the new web for "nerds" i.e. people who care for the real thing? Just as our current web once was a place for scientists and engineers mostly, maybe we now need something as this? I feel the flaws in my own point, but maybe it's not all hopeless?
2 replies →
But some webdev said they are 10x faster now so it cant be bad for humanity /s
They must be slackers, I heard from multiple sources that AI makes you a 100x developer.
2 replies →
[flagged]
Upvoting because it's a salient point and downvoters are mad. HN has a complex lately.
4 replies →
I see this as an absolute win. The state of micro dependencies of js was a nightmare that only happened because a lot of undereducated developers flooded the market to get that sweet faang money.
Now that both have dried up I hope we can close the vault door on js and have people learn how to code again.
Your first sentence cheers that we're moving from NPM micropackages to LLM-generated code, and then you say this will result in people having to learn to code again.
I don't see how the conclusion follows from this.
There will be many LLM-generated functions purporting to do the same thing, and a bug in one of them that gets fixed means only one project gets fixed instead of every project using an NPM package as a dependency.
There will be many LLM-generated functions doing the same thing, in the same project, unless the human pays attention.
I've been playing a lot recently with various models, lately with the expensive claude models (API, for the large context windows), and in every attempt things are really impressive at the beginning, and start going south once the codebase reaches about 10k to 15k lines of code. Even with tools split out into a separate library and separate documentation at that point it has a tendency to generate tool functions again in the module it's currently working on over taking the already defined one in the helper library.
You mention that you don’t see how this could result in developers that write code. Then you immediately follow that with fixing bugs.
Yes, use of LLMs is still developers not writing original code, but it’s still an improvement (a minor one) of the copy/paste of micro dependencies.
Ultimately developers aren’t going to figure out writing original code until they are forced to do so from changed conditions within their employer.
The best outcome was things like jquery and then lodash where a whole collection of small util functions get rolled in to one package.
Oh god, without tree shaking, lodash is such a blight.
I've seen so many tiny packages pull in lodash for some little utility method so many times. 400 bytes of source code becomes 70kb in an instant, all because someone doesn't know how to filter items in an array. And I've also seen plenty of projects which somehow include multiple copies of lodash in their dependency tree.
Its such a common junior move. Ugh.
Experienced engineers know how to pull in just what they need from lodash. But ... most experienced engineers I know & work with don't bother with it. Javascript includes almost everything you need these days anyway. And when it doesn't, the kind of helper functions lodash provides are usually about 4 lines of code to write yourself. Much better to do that manually rather than pull in some 70kb dependency.
18 replies →
> Now that both have dried up
They have??
And is it not going to be orders of magnitude worse with the vibe-coded crud hitting the internet?
I don't quite understand your argument. Wasn't the post about how users might replace transparent dependencies with transparent LLM drop-ins? I don't see how having an LLM to do the same job would enable someone to learn more. They're probably the kind of person who will ask the LLM to perform a refactor when problems arise, so they won't learn that much through osmosis.
Now you got vibe coder
Yeah, you're sort of swapping in one issue for another. I agree that micro dependencies needs to be tackled, but perhaps not using LLMs.
I could fear that rather than keep pushing for a Javascript standard library, which would encompass all these smaller function, we now just get more or less the same defective implementations generated by LLMs, but hidden in thousands of repos, where tools won't find security issues. At least with NPM we can pull in updated versions with NPM tells us that we're running an outdated version. Who is going to traverse your proprietary code base and let you know that the vibe coded left-pad Claude put in three years ago is buggy?
> the era of small, low-value libraries like blob-util is over.
Thankfully (not against blob-util specifically because I've never intentionally used it), I wouldn't completely blame llms either since languages like Go never had this dependency hell.
npm is a security nightmare not just because of npm the package manager, because the culture of the language rewards behavior such as "left-pad".
Instead of writing endless utilities for other project to re-use, write actual working things instead - that's where the value/fun is.
But as Go puts it:
“A little copying is better than a little dependency.”
https://go-proverbs.github.io/
Copying is just as much dependency, you just have to do maintenance through manual find-and-replace now.
20 replies →
Yep something like blob util could be a blog post or a gist (or several stack overflow answers). And a lot of NPM library falls under that. They always have the anemic standard library of JavaScript forgetting that the C library is even smaller.
> since languages like Go never had this dependency hell
What is the feature of Go that this is referring to?
It's a cultural thing.
And libraries try harder not to have absurd dependencies, than finished products (correctly, IMO).
Another thing about Go, and more generally, the statically-typed languages, is that these utility functions are generally safer to get a once-over and a commit into the code base than in dynamically-typed languages. Something that joins an array of strings with a joiner in some non-trivial way is a lot easier to review when you don't have to worry about it getting an array of DOM nodes and a NaN for the join character, and all the other things that could come in. Well-tested little utilities for Javascript are valuable, and I don't 100% trust LLMs to not give me code based on common misconceptions rather than the actually correct code in that case. Or at least, correct for some definition of correct, which in the dynamic languages can itself be some work to figure out.
Bad take. In the first place, the appropriate time order of magnitude for how long static type checkers have existed for JS is "decades". In the second place, the function that "joins an array of strings with a joiner" is the wrong place to worry about whether that array is actually an array of DOM nodes or the joiner is NaN.
I don’t think open source is going anywhere. It’s posed to get significantly stronger — as the devs which care about it learn how to leverage AI tools to make things that corporate greasemonkeys never had the inspiration to. Low quality code spammers are just marketing themselves for jobs where they can be themselves: soulless and devoid of creative impulse.
That’s the thing: open source is the only place where the true value (or lack of value) of these tools can be established — the only place where one can test mettle against metal in a completely unconstrained way.
Did you ever want to build a compiler (or an equally complex artifact) but got stuck on various details? Try now. It’s going to stand up something half-baked, and as you refine it, you will learn those details — but you’ll also learn that you can productively use AI to reach past the limits of your knowledge, to make what’s beyond a little more palatable.
All the things people say about AI is true to some degree: my take is that some people are rolling the slots to win a CRUD app, and others are trying to use it to do things that they could only imagine before —- and open source tends to be the home of the latter group.
True innovation will come from open source for sure. As the developers don't have the same economic incentives to be "safe", "ethical" "profitable" or whatever. large corporations know this and fear this development. That's why i expect a significant lobbying to take hold in USA that will try and make local AI systems illegal. And I think they will be very convincing to the government. Because the government also fears the "peasants" and giving them any true semblance of real AGI like systems. I bet very soon we will start seeing various classifications that will define what is legal and what is not for a citizen to possess or use.
> That's why i expect a significant lobbying to take hold in USA that will try and make local AI systems illegal.
I think they're going to be using porn and terrorism (as usual) to do that, but also child suicide. I also think they're going to leverage this rhetoric to lock down OSes in general, by making them uninstallable on legally-available hardware unless approved, because approved OSes will only be able to run approved LLMs.
Meaning that I think LLMs/generative AI will be the lever to eliminate general-purpose computing. As mobile went, so will desktop.
I think this is inevitable. The real question for me is whether China will partner with the west on this, or whether we will be trading Chinese CPUs with each other like contraband in order to run what we want.
> any true semblance of real AGI like systems.
This is the only part I don't agree with. This isn't going to happen, but I'm not even sure it would be more useful than what we have. We have billions of full AGI machines walking around, and most of them aren't great. I'm talking about restrictions on something technically barely better than what we have now; maybe only a significant bit more compute-efficient. Training techniques will probably be where we get the most improvements.
> It’s posed to get significantly stronger
It's really not. Every project of any significance is now fending off AI submissions from people who have not the slightest fucking clue about what is involved in working on long-running, difficult projects or how offensive it is to just slather some slop on a bug report and demand it is given scrutiny.
Even at the 10,000 feet view it has wasted people's time because they have to sit down and have a policy discussion about whether to accept AI submissions, which involves people reheating a lot of anecdotal claims about productivity.
Having learned a bit about how to write compilers I know enough to know that I can guarantee you that an AI cannot help you solve the difficult problems that compiler-building tools and existing libraries cannot solve.
It's the same as it is with any topic: the tools exist and they could be improved, but instead we have people shoehorning AI bollocks into everything.
This isn't an AI issue. It is a care issue. People shouldn't submit PRs to project where they don't care enough to understand the project they are submitting to or the code they are submitting. This has always been a problem, there is nothing new. The thing that is new is more people can get to a point where they can submit regardless of their care or understanding. A lot of people are trying to gild their resume by saying they contributed to a project. Blaming AI is blaming the wrong problem. AI is a a tool like a spreadsheet. Project owners should instead be working ways to filter out careless code more efficiently.
7 replies →
yeah we are getting lots of "I don't know how to do this and AI gave me this code that doesn't work, can you fix it" or "AI said it can do this" and the feature doesn't exist... some people will even argue and say "but AI said it doesn't take long, why won't you add it"
3 replies →
>Every project of any significance is now fending off AI submissions
Not anything with a cathedral model.
'open source' is too ambiguous to be useful.
> Every project of any significance is now fending off AI submissions from people who have not the slightest fucking clue
I'm kinda hoping that Github will provide an Anubis-equivalent for issue submissions by default.
Sounds like a lot of FUD to me — if major projects balk at the emergence of new classes of tools, perhaps the management strategy wasn’t resilient in the first place?
Further: sitting down to discuss how your project will adapt to change is never a waste of time, I’m surprised you stated it like that.
In such a setting, you’re working within a trusted party — and for a major project, that likely means extremely competent maintainers and contributors.
I don’t think these people will have any difficulty adapting to the usage of these tools …
4 replies →
I am sure I am not the only one who thinks these micro-dependencies are worthless anyway. You'd be better off just listing the functions in a markdown file for people to copy over than ship an entire package for it.
This isn't "small" open source, "small" would be something you put together in a week or weekend. These are like "micro" projects, where more work goes into actually publishing and maintaining the repository than actually writing the library.
I like the approach C sometimes takes, with the "tiny header file" type of libraries. Though I guess that also stems from the lack of a central build system.
Vendoring dependencies is how I remember doing it for web projects pre-NPM. Find an open source, well tested library and copy the source into your project.
You have to manually update for any releases you care about, but that is also an incentive to keep dependency count low.
Why aren't those tiny header file libraries just part of the standard C library?
Wait sorry, I don't mean that. I read too many bog-standard HN comments about NPM above.
This. The code is like 10 lines long. I might as well copy that from Stackoverflow and adapt that for my own use. Why bother packaging that?
What's your copy& paste solution to security updates?
The security maintenance of the ten lines of code I have read and copied into my code is the same as the ten lines of code next to it, that I have written myself.
1 reply →
Does left-pad have security updates? You may as well ask what's the security update solution for Stack Overflow answers.
1 reply →
> I’m still trying to figure out what kinds of open source are worth writing in this new era
Is there any upside to opensourcing anything anymore? Anything published today becomes training data for the next model, with no attribution to the original work.
If the goal is to experiment, share ideas, or let others learn from the work, maybe the better default now is "source available", instead of FOSS in the classic sense. It gives people visibility while setting clearer boundaries on how the work can be used.
I learned most of what I know thanks to FOSS projects so I'm still on the fence on this.
I keep seeing this attitude and I don't really understand it at all; there's no upside to publishing open source work because it might be utilized by more people, is that correct?
Or is it the attribution? There are many many libraries I have used and continue to use and I don't know the author's internet handle or Christian name. Does that matter? Why?
I have written a lot of code that my name is no longer attached to. I don't care and I don't know why anyone does. If it were valuable I would have made more money off of it in the first place, and I don't have the ego to just care that people know it's my code either.
I want the things I do today to have an upside for people in the future. If that means I write code that gets incorporated into a model that people use to build things N number of years from now, that's great. That's awesome. Why the hell is that apparently so demotivating to some people?
> there's no upside to publishing open source work because it might be utilized by more people, is that correct?
I believe the perspective here is "I make code for fellow hackers to look into, critique, be educated on, or simply play with". If you see the hacker scene as a social one, LLM's are an awful black hole that sucks up everything around it and ruins this collaboration.
Not to mention that the hacker scene was traditionally thought to be a rejection of what we now call "Big Tech". Corporate was free to grab the code, but it didn't matter much as long as the scene was kept. Now even that invisible social contract is broken.
But I suppose if you're of a diehard FOSS mentality, "Free" means "Free". Free to be used to build, or destroy society at its whim. a hivemind to meld into and progress the overall understanding of science, for science's sake.
I'll admit the last few years have had me questioning what I truly want to do within the on these two mentalities.
> I want the things I do today to have an upside for people in the future.
I think most would agree with this, but the way things work today don't support it. As of now, AI gains are privatized while the losses are socialized. Until that one-sided imbalance is addressed, LLM's "use" of open source is unbounded and nonreciprocal.
Attribution is a big part of the human experience. Your response frames it as ego driven, but it's also what motivates people to maintain code that is not usually compensated, it's also what builds reputation, trust, communities, and even careers.
Until that’s figured out, we can still share, but maybe in ways that are closer to one another, or under distribution models that reflect the reality we’re in rather than the one we used to have.
People were publishing "open source" not because they gave a shit, but to pad their resumes to get hired at FAANG - like they're paid by the number of projects published on Github It was always human slop. I'm glad this practice is dying. Soon we might be able to reinstate the assumption that the remaining open source projects contain interesting ideas that are worth your eyeball time to look at.
Sorry but source-available is probably going to get slurped up for training data as well
Microsoft already did this for all code in every public repo.
When are they going to start doing it for private repos too...
1 reply →
This is kinda how I've felt for months. I don't have any interest in continuing existing open source projects and don't want to create any new ones.
What's the point?
All of my personal projects for the past few months have been entirely private, I don't even host them on Github anymore, I have a private Forgejo instance I use instead.
I also don't trust any new open source project I stumble upon anymore unless I know it was started at least a year ago.
You're not obligated to give away your mind for free. You're free to share, of course. But sharing implies reciprocity, a back and forth. The internet used to be like that, but if the environment changes, you adapt your behavior accordingly.
In the long run I think it's time to starve the system from input until it's attitude reverts to reciprocal. It's not what I'd want, but it seems necessary. People learn from consequences, not from words alone
Or use AGPL licenses. As a bonus, nobody in JavaScript checks licenses so you might be able to sue Amazon for money.
Staying true to free software principles. It's unethical to publish nonfree code or binaries.
Code is only useful if it's used. I could write a ton of code and be buried with it, or publish it for people (or AI software, or dolphins or aliens) to use. Who has the energy to have Anubis measure whether my code, or yours, is ethical enough? I'm going to die someday!
> It gives people visibility while setting clearer boundaries on how the work can be used.
... Because those would be respected?
This author assumes that open sourcing a package only delivers value if is added as a dependency. Publicly sharing code with a permissive license is still useful and a radical idea.
Yep. Even just sharing code with any license is valuable. Much I have learned from reading an implementation of code I have never run even once. Solutions to tough problems are an under-recognized form of generosity.
This is a point where the lack of alignment between the free beer crowd and those they depend on is all too clear. The free beer enthusiast cannot imagine benefiting from anything other than a finished work. They are concerned about the efficient use of scarce development bandwidth without consciousness of why it is scarce or that it is not theirs to direct. They view solutions without a hot package cache as a form of waste, oblivious to how such solutions expedite the development of all other tools they depend on, commercial or free.
Yeah if I find some (small) unmaintained code I need, I will just copy it (then add in my metrics and logging standards :)
It shouldn't be a radical idea, it is how science overall works.
Also, as per the educational side, I find in modern software ecosystem, I don't want to learn everything. Excellent new things or dominantly popular new things, sure, but there are a lot of branching paths of what to learn next, and having Claude code whip up a good enough solution is fine and lets me focus on less, more deeply.
(Note: I tried leaving this comment on the blog but my phone keyboard never opened despite a lot of clicking, and on mastodon but hit the length limit).
A copyleft license is much better because it ensures that it will remain open and in most cases large companies won't use it, making sure they will have to shell out the money to hire someone to do it instead.
I do agree with this, but there are some caveats. At the end of the day it is time people invest into a project. And that is often unpaid time.
Now, that does not mean it has no value, but it is a trade-off. After about 14 years, for instance, I retired permanently from rubygems.org in 2024 due to the 100k download limit (and now I wouldn't use it after the shameful moves RubyCentral did, as well as the new shiny corporate rules with which I couldn't operate within anyway; it is now a private structure owned by Shopify. Good luck finding people who want to invest their own unpaid spare time into anything tainted by corporations here).
> Sure, you could use blob-util, but then you’d be taking on an extra dependency, with unknown performance, maintenance, and supply-chain risks.
Use of an AI to write your code is also a form of dependency. When the LLM spits out code and you just dump it in your project with limited vetting, that's not really that different from vendoring a dependency. It has a different set of risks, but it still has risks.
Part of the benefit over a dependency is that the code added will (hopefully) be narrowly tailored to your specific need, rather than the generic implementation from a library that likely has support for unused features.
Not including the unused features both makes the code you are adding easier to read and understand, but it also may be more efficient for your specific use case, since you don't have to take into account all the other possible use cases you don't care about.
But in a lot of cases you can't know all the dependencies, so you lean on the community trusting that a package solves the problem well enough that you can abstract it.
You can pin the dependency and review the changes for security reasons, but fully grasping the logic is non-trivial.
Smaller dependencies are fine to copy at first, but at some point the codebase becomes too big, so you abstract it and at that point it becomes a self-maintained dependency. Which is a fair decision, but it is all about tradeoffs and sometimes too costly.
2 replies →
> Part of the benefit over a dependency is that the code added will (hopefully) be narrowly tailored to your specific need, rather than the generic implementation from a library that likely has support for unused features.
In decent ecosystems there should be low or zero overhead to that.
> Not including the unused features both makes the code you are adding easier to read and understand, but it also may be more efficient for your specific use case, since you don't have to take into account all the other possible use cases you don't care about.
Maybe. I find generic code is often easier to read than specialised custom implementations, because there is necessarily a proper separation of concerns in the generic version.
Right, but you do avoid worries like "will I have to update this dependency every week and deal with breaking changes?" or "will the author be compromised in a supply-chain attack, or do a deliberate protestware attack?" etc. As for performance, a lot of npm packages don't have proper tree-shaking, so you might be taking on extra bloat (or installation cost). Your point is well-taken, though.
You can avoid all those worries by vendoring the code anyway. you only 'need' to update it if you are pulling it in as a separate dependency.
> you do avoid worries like "will I have to update this dependency every week and deal with breaking changes?
This is not a worry with NPM. You can just specify a specific version of a dependency in your package.json, and it'll never be updated ever.
I have noticed for years that the JS community is obsessed with updating every package to the latest version no matter what. It's maddening. If it's not broke, don't fix it!
Wouldn't call it a risk in itself, but part of the benefit of using a library, a good and tailored one at least, is that it'll get modernised without my intervention. Even if the code produced for you was state-of-the-art at the moment of inclusion, will it remain that way 5 years from now?
> and you just dump it in your project with limited vetting
Well yes there’s your problem. But people have been doing this with random snippets found on the internet for a while now. The truth is that irresponsibles developers will produce irresponsible code, with or without LLMs
> The truth is that irresponsibles developers will produce irresponsible code, with or without LLMs True. But the difference is the scale and ease of doing with code generators. With a few clicks you can add hundreds of lines of code which supposedly does the right thing. While in the past, you would get code snippets for a particular aspect of the problem that you are trying to solve. You still had to figure out how to add it to your code base and somehow make it “work”
2 replies →
Node.js's core library is purposefully minimal. Purposefully or not, it's still an issue. Wether you add many small things to your project using an LLM or NPM, it still requires work and the annoying thing is that you'll have to do it for every new Node.js project.
Node.js is very good for IO and it has decent performance even for CPU-intensive work considering it's a dynamic language, but it would sure be nice to have a rich core library like Ruby or Clojure has.
The fact that ClojureScript can do it proves that it's doable even for front-end javascript (using advanced optimisations).
This seems to be specific to the JavaScript ecosystem, and to some extent orthogonal to AI coding. Micro-libraries with trivial-ish functionality are generally more difficult to justify to add as a dependency. I’m sure AI coding agents will eventually learn as well that there’s a balance to strike between implementing stuff yourself and the use of libraries. I’m also sure that people who are interested in actual coding will continue to open-source what they find useful, and other people like them will read their code.
More likely, what we will see is the decline of low effort projects. The JavaScript/ Typescript ecosystem has been plagued with such packages. But that’s more anomalous to the JS community than it is a systemic problem with open source in general.
So if fewer people are including silly dependencies like isEven or leftPad, then I see that as a positive outcome.
> ….but I do think it’s a future where we prize instant answers over teaching and understanding
It depends. For stuff I don’t care about I’m happy to treat it as a black box. Conversely, AI now allows me to do deep dive on essentially anything I’m interested in, which has been a massive boon to my learning ability.
Right now I tend to not use an external library unless the code would be large e.g. http server and/or the library is extremely popular.
Otherwise, writing it myself is much better. It's more customizable and often much smaller in size. This is because the library has to generalize, and it comes with bloats.
Using AI to help write is great because I should understand that anyway whether AI writes it or not or whether it's in an external library.
One example recently is that I built a virtual list myself. The code is much smaller and simpler compared to other popular libraries. But of course it's not as generalizable.
I don't buy the education angle.
If you're not learning to code, then you want efficient code, so the comments are wasted bytes (ok, not a huge expense, but still).
If you are learning to code, or just want to understand how this code works, then asking an LLM is going to get a lot better result. LLMs are fantastic tutors. Endlessly patient, go down any rabbit hole with you, will continue explaining a concept until you get it, etc. I think they're going to revolutionise education, especially for practical subjects like coding.
Respect to the author for trying to educate, though.
> so the comments are wasted bytes
Is there any modern compiler where the output code has anything to do with the comments in the source?
Small open source is still valuable, but the bar is higher. If your project is something that's trivial and nobody just thought to do it before you and bothered to do it after, that's probably not going to survive, but if your project is a small focused tool that handles something difficult really well, it's 100% got a future.
Earlier this year I wrote an interpreter for a niche, proprietary binary format. Someone asked if I could open source it so that they could (more easily) run it on NixOS. I declined as I'm just that strongly opposed to my work being used to train AI models and further entrench the enshittification of the internet.
Curious: So you only release your interpreter as a binary? Is there some license that explicitly forbids use for AI training? (Although it would probably be unclear if that is really enforceable until someone challenges it.)
1 reply →
TLDR: [AI promises] a future where we prize instant answers over teaching and understanding
But what this article and the comments don't say: open-source is mainly a quality metric. I re-use code from popular open-source repo's in part because others have used it without complaints (or document the bugs), in part because people are embarrassed to write poor-quality open-source so it's above-par code, and in part because if there are issues in this corner of the world, this dependency will solve them over time (and I can watch and wait when I don't have time to fix and contribute).
The quality aspect drives me to prefer dependencies over AI when I don't want full ownership, so I'll often ask AI to show open-source projects that do something well.
(As an aside, this article is about AI, but AI is so pervasive now as an issue that it doesn't even need saying in the title.)
I totally agree that the goal should be teaching. But I do wonder which code someone is more likely to actually read and understand: the LLM-generated one in their own codebase, or the one hidden in some npm package they installed to just ship things. I honestly don’t know. Careless devs probably won’t read either as long as it seems to work
I’ve been trying to encourage forking my libraries, or have people just copy them into their codebase and adapt them, e.g. https://github.com/mastrojs/mastro/ (especially the “extension” libs.) But it’s an uphill battle against the culture of convenience over understanding.
>Many software developers will argue that asking a candidate to reverse a binary tree is pointless
Is "reversing a binary tree" actually a thing, or is this a cute kind of "rocket surgery" phrase intentionally mixing reversing a linked list and searching a binary tree?
I think it's a reference to the Google interview problem that the author of Homebrew (IIRC) failed. They were quite upset about it since they have proved their worth through their famous open-source contributions, but got rejected in a LeetCode-like interview.
It’s probably a mistake on the author’s end, but the problem comes across anyway.
I can only imagine reversing a binary tree would imply changing the “<“ comparison in nodes to “>” which would be a useless exercise
AI offering the solution for a small problem that probably doesn’t deserve yet another dependency suggests to me that there’s a middle ground that we’ve failed to sufficiently cover: how to socialize code snippets that you’re meant to just inline into your project. Stack Overflow is probably the closest we’ve gotten to a generalized solution and it doesn’t exactly feel like a good one.
I came across this once before in the form of a react hooks library that had no dependency to install. It was just a website and when you found the hook you wanted you were meant to paste it into your project.
GitHub gists are probably the best technical option. They come with versioning (I think), and have comment threads.
And a decent API as well.
I think what’s missing is some amount of organization to make them more discoverable.
I suppose some people would see this as progress: fewer dependencies, more robust code (even if it’s a bit more verbose), quicker turnaround time than the old “search npm, find a package, read the docs, install it” approach.
Why would randomized code be more robust? Also, how is searching/reading the docs slower than checking the correctness of a randomized function?
> it’s a future where we prize instant answers over teaching and understanding
It doesn't have to be this way- coupled with verification (to mitigate hallucination), llms can help so much with personal education.
It's a purely judgement free environment and for a lot of people that's incredibly meaningful.
Open source exists because coding was a significant effort and code was a thing of high value. Unsurprisingly companies hesitated to make the code public and free. All of this is changing now as coding has suddenly become trivial. So, yes, the mission of open source, in general, will be challenged.
In the U.S., anything machine generated is uncopyrightable.
Why would you put uncopyrightable code into your codebase?
Why wouldn't you? Your codebase (if you're a business) exists to make you money, people being able to copy some unknown portions of it without further license if they somehow legally get their hands on a copy of it seems entirely irrelevant.
PS. I think this is much less clear and much less settled law than you are suggesting.
It's more nuanced. If I even have a few lines I can prove are mine, those parts are copywritable in the same way Pride and Prejudice is public domain but pride and prejudice and zombies is copyrighted.
Even worse...unmaintained code. Only the human-written one has a maintainer. The other one plagiariased by AI is instant legacy code
> The other one plagiariased by AI is instant legacy code
I have used this "instant legacy code" concept before. It's absolutely true, IMO. But people really, really, really hate hearing it.
Autocomplete has been around for decades
I think if you dig a little deeper you will find that the answer is not so black and white.
> Even now there’s a movement toward putting documentation in an llms.txt file, so you can just point an agent at it and save your brain cells the effort of deciphering English prose. (Is this even documentation anymore? What is documentation?)
I look at it as why not have the best of both worlds? The docs for my JS framework all have the option of being returned as LLM-friendly text [1].
When I utilize this myself, it's to get help fleshing out skeleton code inside of an app built w/ my framework (e.g, Claude Sonnet w/ these docs in context build a nearly ~90-100% accurate implementation for most stuff I throw at it—anything from little lib stuff up to full-blown API endpoint design and even framework-level implementations of small stuff like helping to refactor the built-in middleware loading). It's not for a lack of desire to read, but rather a lack of desire to build every little thing from scratch when I know an LLM is perfectly capable (and much faster than me).
[1] https://docs.cheatcode.co/joystick/ui/component/dynamic-page...
> I don’t know which direction we’re going in with AI
Maybe programming languages will be designed for AIs in the future. Maybe they'll have features that make grafting unknown generated code easier.
Will we ever even get a widespread new programming language? AIs work great for coding because they have massive codebases in C/C++/Python/JS/etc. to train on. If I create a language called "foo" tomorrow, no such codebase will exist for it and the lack of AI support for the language will be a huge barrier for adoption so it's going to be much less likely to catch on like Python did back in the day.
I'm working on my own lisp. I've used ChatGPT to generate some cases for my test suite. Somehow it can generate decent code if I instruct it thoroughly enough and provide samples. Maybe the fact all lisps have similar syntax helps?
I would have and did write this instead of including it anyway. These small npms you would spend more time to look up than write are a pest.
Several issues:
1. Reducing dependencies is a wrong success metric. You just end up doing more work yourself, except you can't be an expert in everything, so your code is often strictly worse.
2. Regenerating the same solutions with a probabilistic machine will produce bugs a certain percentage of the time. Dependencies are always the same code (when versioned).
3. Cognitive overhead for human review is higher with LLM-generated libs, for no additional benefit.
> Reducing dependencies is a wrong success metric. You just end up doing more work yourself
Except it's just not true in many cases because of social systems we've built. If I want to ship software to Debian I have to make sure that every single of my 3rdparty dependencies is registered and packaged as a proper debian package - a lot of time it will take much less work to rewrite some code than to get 25 100-lines-of-code micro-libraries accepted into debian.
The main point isn't about dependencies but loosing the mindset to learn from small domain problem
It would be a cool feature of AI to include only the subset of the library you use and nothing else.
The level of apologetism with regard to ai is depressing in this article.
I think this applies to more than just small open source projects.
Chances are, even if you deliberately and strategically pick to work on an OSS project that you are positively sure an LLM can’t just spit out on command, it will be capable of doing so by the time you are close to completion. In that sense, one has to either be not inclined to question “what’s the point” or have a bit of gambling mentality in order to work on anything substantial.
That’s not automatically a problem, however. The problem is that even if you do come up with a really cool idea that LLM is not capable of autocompleting, and you release it under a copyleft license (to ensure the project survives and volunteer contributor’s work is not adopted and extinguished by some commercial interest), it will get incorporated into its dataset regardless of the licensing, and thereafter the LLM will be capable of spitting it out and its large corporate operator will be able to monetise your code (allowing anyone with money wishing to build a commercial product based on it).
Now way more than before, i value my feed collection because i know all these blogs are human written, some of them got a post every year, but it is always good quality and human written.
Bleak
LLMs give us the opportunity to work on more complex projects and gain fuller understanding of the problem space and concepts. Or create tons of slop. Take your pick.
> Claude’s version is pretty close to the blob-util version (unsurprising, since it was probably trained on it!).
AI are thieves!
> I don’t know which direction we’re going in with AI (well, ~80% of us; to the remaining holdouts, I salute you and wish you godspeed!), but I do think it’s a future where we prize instant answers over teaching and understanding.
Google ruined its search engine years ago before AI already.
The big problem I see is that we have become WAY too dependent on these mega-corporations. Which browser are people using? Typically chrome. An evil company writes the code. And soon it will fire the remaining devs and replace them with AI. Which is kind of fitting.
> Even now there’s a movement toward putting documentation in an llms.txt file, so you can just point an agent at it and save your brain cells the effort of deciphering English prose. (Is this even documentation anymore? What is documentation?)
Documentation in general sucks. But documentation is also a hard problem.
I love examples. Small snippets. FAQs. Well, many projects barely have these.
Look at ruby webassembly/wasm or ruby opal. Their documentation is about 99% useless. Or, even worse - rack in ruby. And I did not notice this in the past, in part because e. g. StackOverflow still worked, there were many blogs which helped fill up missing information too. Well all of that is largely gone now or has been slurped up by AI spam.
> the era of small, low-value libraries like blob-util is over. They were already on their way out thanks to Node.js and the browser taking on more and more of their functionality (see node:glob, structuredClone, etc.), but LLMs are the final nail in the coffin.
I still think they have value, but looking at organisations such as rubygems.org disrupt the ecosystem and bleeding it dry by kicking out small hobbyists, I think there is indeed a trend towards eliminating the silly solo devs who think their unpaid spare time is not worthy of anything at all, yet the big organisations eagerly throw down more and more restrictions onto them (my favourite example is the arbitrary 100k download limit for gems hosted at rubygems.org, but look at the new shiny corporate rules on rubygems.org - this is when corporations take over the infrastructure and control it. Ironically this also happened to pypi and they admit this indirectly: https://blog.pypi.org/posts/2023-05-25-securing-pypi-with-2f... - of course they deny that corporations control pypi now, but by claiming otherwise they admit it, because this is how hobbyists get eliminated. Just throw more and more restrictions at them without paying them. Sooner or later they decide to do something better with their time.)
[dead]
He almost got it right. It's not just the fate of small open source. It's the fate of all programmers now. Why hire a programmer when an LLM costs less, works faster and makes less mistakes (OP compliments better error handling, read the article).
Unless you are a product owner, you have paying clients that love you and your product and won't simply ditch it in favour of a new clone, you are really screwed.
"when an LLM costs less, works faster and makes less mistakes"... indeed, but it doesn't follow at all that it's the fate of all programmers _now_... at least in my experience none of these things are true ATM.
Well, at the very least it costs less than asking intern to look for a lib doing something particular and give some examples... still about as accurate as the intern tho.
2 replies →
So far I've never seen yet a non-programmer release production-grade code using only LLMs. There's just so much to care about (from security, deployments, secret management, event-driven architectures, and a large etc.) that "just" providing a prompt to create an "app" doesn't cut it. You need infra and non-engineers just don't know shit about infra (even if it's 99% managed), you need to deploy your llm-generated code in that infra; that should happen in a ci/cd probably. And what about migrations? Git? Who's setting up the api gateway? I don't mean to say that LLMs don't know how to do that, but you need to instruct them to do so, and even there, they will make silly mistakes and you need to re-instruct them or fix it.
Prompting is just 50% of the work (and the easy part actually). Ask the Head of Product or whoever is there to deploy something valuable to production and maintain it for 6 months while not losing money. It's just not going to happen, not even with truly AGI.
An LLM might be able to replace the majority of the code Sindre Sorhus has put out there, but it's probably a stretch to think that it could replace someone like John Carmack.
Trivial NPM libraries were never needed, but LLMs really are the nail in the coffin for them even when it comes to the most incompetent programmers because now they can literally just ask an LLM to spit out the exact same thing.
He also points out a pointless type check in a type checked language...
Your name is very accurate I must say.
That type check is honestly not pointless at all. You can never be certain of your inputs in a web app. The likelihood of that parameter being something other than an arraybuffer is non-zero, and you generally want to have code coverage for that kind of undefined behavior. TypeScript doesn't complain without a reason.