In my opinion LLMs are intellectual property theft. Just as if I started distributing copies of books. This substantially reduces the incentive for the creation of new IP.
All written text, art work, etc needs to come imbued with a GPL style license: if you train your model on this, your weights and training code must be published.
I think there is a real issue here, but I do not think it is as simple as calling it theft in the same way as copying books. The bigger problem is incentives. We built a system where writing docs, tutorials, and open technical content paid off indirectly through traffic, subscriptions, or services. LLMs get a lot of value from that work, but they also break the loop that used to send value back to the people and companies who created it.
The Tailwind CSS situation is a good example. They built something genuinely useful, adoption exploded, and in the past that would have meant more traffic, more visibility, and more revenue. Now the usage still explodes, but the traffic disappears because people get answers directly from LLMs. The value is clearly there, but the money never reaches the source. That is less a moral problem and more an economic one.
Ideas like GPL-style licensing point at the right tension, but they are hard to apply after the fact. These models were built during a massive spending phase, financed by huge amounts of capital and debt, and they are not even profitable yet. Figuring out royalties on top of that, while the infrastructure is already in place and rolling out at scale, is extremely hard.
That is why this feels like a much bigger governance problem. We have a system that clearly creates value, but no longer distributes it in a sustainable way. I am not sure our policies or institutions are ready to catch up to that reality yet.
> We have a system that clearly creates value, but no longer distributes it in a sustainable way
The same thing happened (and is still happening) with news media and aggregation/embedding like Google News or Facebook.
I don't know if anyone has found a working solution yet. There have been some laws passed and licensing deals [1]. But they don't really seem to be working out [2].
I'm not sure that I'd call [2] it not working out, just like I wouldn't call the equivalent pressure from the USA to dismantle medicare our public health system not working out.
The biggest issue with the scheme is the fact that it was structured to explicitly favour media incumbents, and is therefore politically unpopular.
> I do not think it is as simple as calling it theft in the same way as copying books
Aside from the incentive problem, there is a kind of theft, known as conversion: when you were granted a license under some conditions, and you went beyond them - you kept the car past your rental date, etc. In this case, the documentation is for people to read; AI using it to answer questions is a kind of conversion (no, not fair use). But these license limits are mostly implicit in the assumption that (only) people are reading, or buried in unenforceable site terms of use. So it's a squishy kind of stealing after breaching a squishy kind of contract - too fuzzy to stop incented parties.
Why do you think there's was an implicit agreement that documentation was only intended for humans? I've written a lot of documentation, much of it open source, and I'm generally very excited that it has proved additionally useful via LLMs. If you had asked me in 2010 whether that was something I intended in writing docs I'm pretty sure I would have said something like "that's science fiction, but sure".
You still intended it for humans. Intent is defined by what one is aiming for, and without knowledge of an alternative, that was your intent.
100% I get that you are OK with it being used by non-human ingestion. And I think many might be OK with that.
One thing, I'm not sure how helpful the documentation is. I think we're getting training out of example, not docs. This makes me think... we could test this by creating a new pseudo-language, and then provide no examples, only docs.
If the LLM can then code effectively after reading the docs, we'd have a successful test. Otherwise? It's all parroting.
There will be no royalties, simply make all the models that trained on the public internet also be required to be public.
This won't help tailwind in this case, but it'll change the answer to "Should I publish this thing free online?" from "No, because a few AI companies are going to exclusively benefit from it" to "Yes, I want to contribute to the corpus of human knowledge."
It can. The problem is the practice of using open source as a marketing funnel.
There are many projects that love to brag about being open source (it's "free"!), only to lock useful features behind a paywall, or do the inevitable license rug pull after other companies start profiting from the freedoms they've provided them. This is the same tactic used by drug dealers to get you hooked on the product.
Instead, the primary incentive to release a project as open source should be the desire to contribute to the corpus of human knowledge. That doesn't mean that you have to abandon any business model around the project, but that shouldn't be your main goal. There are many successful companies built around OSS that balance this correctly.
"AI" tools and services corrupt this intention. They leech off the public good will, and concentrate the data under the control of a single company. This forces well-intentioned actors to abandon open source, since instead of contributing to human knowledge, their work contributes to "AI" companies. I'm frankly not upset when this affects projects who were abusing open source to begin with.
So GP has a point. Forcing "AI" tools, and even more crucially, the data they collect and use, to be free/libre, would restore the incentive for people to want to provide a public good.
The narrative that "AI" will bring world prosperity is a fantasy promoted by the people who will profit the most. The opposite is true: it will concentrate wealth and power in the hands of a few even more than it is today. It will corrupt the last vestiges of digital freedoms we still enjoy today.
I hope we can pass regulation that prevents this from happening, but I'm not holding my breath. These people are already in power, and governments are increasingly in symbiotic relationships with them.
> The narrative that "AI" will bring world prosperity is a fantasy promoted by the people who will profit the most. The opposite is true: it will concentrate wealth and power in the hands of a few even more than it is today. It will corrupt the last vestiges of digital freedoms we still enjoy today.
> We have a system that clearly creates value, but no longer distributes it in a sustainable way.
It does not "create value" it harvests value and redirects the proceeds it accrues towards its owners. The business model is a middleman that arbitrages the content by separating it from the delivery.
Software licensing has been broken for 2 decades. That's why free software isn't financially viable for anybody except a tiny minority. It should be. The entire industry has been operating by charity. The rich mega corporations have decided they're not longer going to be charitable.
This is a big eye-roll but otherwise ya, this is one way to think of it. It's not all about money, though. The people running these companies are just taking, en masse, without credit. This is a basic human desire. Of course there is a discussion of whether or not we should evolve beyond that. It feels incredibly dystopian to me, though.
The problem is there was a social contract. Someone spent their time and money to create a product that they shared for free, provided you visit their site and see their offerings. In this way they could afford to keep making this free product that everyone benefited from.
LLMs broke that social contract. Now that product will likely go away.
People can twist themselves into knots about how LLMs create “value” and that makes all of this ok, but the truth is they stole information to generate a new product that generates revenue for themselves at the cost of other people’s work. This is literally theft. This is what copyright law is meant to protect. If LLM manufacturers are making money off someone’s work, they need to compensate people for that work, same as any client or customer.
LLMs are not doing this for the good of society. They themselves are making money off this. And I’m sure if someone comes along with LLM 2.0 and rips them off, they’re going to be screaming to governments and attorneys for protection.
The ironic part of all of this is that LLMs are literally killing the businesses they need to survive. When people stop visiting (and paying) Tailwind, Wikipedia, news sites, weather, and so on, and only use LLMs, those sites and services will die. Heck, there’s even good reason to think LLMs will kill the Internet at large, at least as an information source. Why in the hell would I publish news or a book or events on the Internet if it’s just going to be stolen and illegally republished through an LLM without compensating me for my work? Once this information goes away or is locked behind nothing but paywalls, I hope everyone is ready for the end of the free ride.
I support your right to have an opinion, but in my opinion, thank God this is just your opinion.
Copyright, as practiced in late 20 and this century, is a tool for big corps to extract profits from actual artists, creators, and consumers of this art[0] equally. Starving artists do not actually benefit.
Look at Spotify (owned and squeezed by record labels) giving 70% of the revenue to the record labels, while artists get peanuts. Look at Disney deciding it doesn't need to pay royalties to book writers. Hell, look at Disney's hits from Snow White onwards, and then apply your "LLMs are IP theft" logic to that.
Here's what Cory Doctorow, a book author and critic of AI, has to say about it in [1]:
> So what is the alternative? A lot of artists and their allies think they have an answer: they say we should extend copyright to cover the activities associated with training a model.
> And I'm here to tell you they are wrong: wrong because this would inflict terrible collateral damage on socially beneficial activities, and it would represent a massive expansion of copyright over activities that are currently permitted – for good reason!.
---
> All written text, art work, etc needs to come imbued with a GPL style license
GPL-style license has been long known not to work well for artifacts other than code. That's the whole reason for existence of Creative Commons, GNU Free Documentation License, and others.
I didn't miss it, I just don't buy it. Lots of people create and write in their free time, without any financial incentives whatsoever.
I'm okay with writing books for profit, though. But if your business case depends on artificially limiting what users can do with their copies, then your business is fundamentally unethical, whether it's a one-man company or OpenAI. It's not a bad thing for such a business to go under.
1. Will people stop buying your books if LLMs have the information from them?
2. Will people stop buying your books if they can get them from the library? Is a library ripping you off?
3. Assuming your books are non-fiction (otherwise the answer to (1) would be a clear "no"), am I ripping you off if I read your books, create a course that teaches the same things (that you tought me through your book) and earn mega-money because I'm mega-skilled at marketing?
4. How about if I lend my copy to dozens of my friends, who are all very interested in the stuff you write but don't want to pay themselves?
5. Did OpenAI go to the bookstore, buy your book and scan it? Or did Amazon or any other ebook retailer just gave them the book PDF when they asked nicely? How did the rip off happen?
6. If an Anthropic employee buys your book in a bookstore, scans it and destroys the physical copy, and the digital equivalent is only used to train Claude, is that a ripoff?
This stuff is complex and as a society we're just starting to grapple with the consequences. Cory's making the case against copyright being used as a tool much more eloquently than I am - I encourage you to read it if you haven't already.
BTW in your particular case, I'd say you're pretty safe. Nobody stops buying books because they can get the same info from LLMs. If that's your concern, you might as well be mad at the Internet at large.
So this logic is essentially: Look at all these ways you're already getting ripped off. What's one more?! You should be grateful they're siphoning off all your work!
You've got a convert here. I don't think I'll publish my next book. I might just email it straight to Open AI.
And Cory Doctorow - I've attempted a few of his books. Felt like I was reading young adult fiction. He's pretty much the '2 prescient statements and a few average books' guy.
I stopped writing open source projects on github because why put a bunch of work into something for others to train off of without any regard for the original projects
I don't understand this mindset. I solve problems on stackoverflow and github because I want those problems to stay solved. If the fixes are more convenient for people to access as weights in an LLM... who cares?
I'd be all for forcing these companies to open source their models. I'm game to hear other proposals. But "just stop contributing to the commons" strikes me as a very negative result here.
We desperately need better legal abstractions for data-about-me and data-I-created so that we can stop using my-data as a one-size-fits-all square peg. Property is just out of place here.
I have mixed opinions on the "AI=theft" argument people make, and I generally lean towards "it's not theft", but I do see the argument.
If I put something on Github with a GPL 3 license, it's supposed to require anyone with access to the binary to also have access to the source code. The concern is, if you think that it is theft, then someone can train an LLM on your GPL code, and then a for-profit corporation can use the code (or any clever algorithms you've come up with) and effectively "launder" your use of GPL code and make money in the process. It basically would be converting your code from Copyleft to Public Domain, which I think a lot of people would have an issue with.
The thing is, what you are describing is arguably theft. It is purposefully circumventing the license and restrictions chosen by the author, in order to steal their code, so that it can be sold and used for profit. This is along the same lines of why book authors and artists have been suing AI companies.
Another point, there is a lot of free and permissive license content to train AI on, where the GPL or copyright can be respected. In many cases, the violating AI companies knew what they were doing was wrong.
The thing is, LLMs aren’t redistributing your code. You’d have a minuscule chance of an LLM actually reproducing your code verbatim without major modifications.
Copyright and copyleft only deal with source code distribution. Your last sentence is not really true from a factual perspective.
I think if you really believe in the open source free software mentality that code should be available to help everyone and improvements to it should also be available and not locked up behind a corporate wall (e.g., a company using GPL code and releasing it with modifications without redistributing the source code), LLMs should be the least of your worries since they don’t do that action. On a literal level they don’t violate GPLv2/v3.
Perhaps copyright law needs new concepts to respond to this change in capability compared to the past, but so far there has been very little legal success with companies and individuals trying to litigate AI companies for copyright violations. Direct violations have been rare and only get more rare over time as training methods evolve.
I find it very easy to understand, people don’t generally want to work for free to support billionaires, and they have few venues to act on that, this is one of them.
There are no ”commons” in this scenario, there are a few frontier labs owning everything (taking it without attribution) and they have the capability to take it away, or increase prices to a point where it becomes a tool for the rich.
Nobody is doing this for the good of anything, it’s a money grab.
Were these contributions not a radical act against zero-sum games in the first place? And now you're gonna let the zero-sum people win by restricting your own outputs to similarly zero-sum endeavors?
I don't wanna look a gift horse in the mouth here. I'm happy to have benefited from whatever contributions were originally forthcoming and I wouldn't begrudge anybody for no longer going above and beyond and instead reverting to normal behavior.
I just don't get it, it's like you're opposed to people building walls, but you see a particularly large wall which makes you mad, so your response is to go build a wall yourself.
It's not about building a wall. It's about ensuring that the terms of the license chosen by the author are respected.
This is why I think permissive licenses are a mistake for most projects. Unlike copyleft licenses, they allow users to take away the freedoms they enjoy from users of derivative works. It's no surprise that dishonest actors take advantage of this for their own gain. This is the paradox of tolerance.
"AI" companies take this a step further, and completely disregard the original license. Whereas copyleft would somewhat be a deterrent for potential abusers, it's not for this new wave of companies. They can hide behind the already loosely defined legal frameworks, and claim that the data is derivative enough, or impossible to trace back, or what have you. It's dishonest at best, and corrupts the last remnants of public good will we still enjoy on the internet.
We need new legal frameworks for this technology, but since that is a glacial process, companies can get rich in the meantime. Especially shovel salespeople.
Good, we don’t want code that people are possessive of, in the software commons. The attitude that you are concerned about what people do with your output means that nobody should touch your output, too big a risk of drama.
> The AI goldrush has proven that intellectual property laws are null and void. Money is all that matters.
Indeed they never really mattered. They were a tool for large corporations to make money and they will go away if they can no longer serve such purpose. Anyone that thought there was a real moral or ethical basis to "intellectual property" laws fell for propaganda and got scammed as a result.
> How does this solve any tangible problems with LLMs regurgitating someone else's work?
I'm not the OP, but here's my from-the-hip answer: if weights are public, building and operating an LLM is no longer a business plan in and of itself, as anyone could operate the same LLM. Therefore companies like OpenAI will be disincentivized from simply redirecting web traffic to their own site.
I didn't really put out the GPL push. The best I could say is that at least that information would be available to everyone rather than being tightly controlled by the company that stole the source material to create it in the first place. It might also dissuade LLM creators from mass piracy as a competitor could take their models and start hosting them.
Exnation: LLMs were trained also on GPL code. The fact that all the previously-paranoid businesses that used to warn SWEs not to touch GPL code with a ten foot pole are now fearlessly embracing LLMs' outputs, means that de facto they consider an LLM their license-washing machine. Courts are going to rubber stamp it because billions of dollars, etc.
Humans are punished for plagiarism all the time. Myriad examples exist of students being disenrolled from college, professionals being fired, and personal reputations tarnished forever.
When a LLM is trained on copyright works and regurgitates these works verbatim without consent or compensation, and then sells the result for profit, there is currently no negative impact for the company selling the LLM service.
I can be in a room looking at something with my eyeballs and listening with my ears perfectly legally... But it would not be legal if I replaced myself with a humanoid mannequin with a video camera for a head.
You can even write down what you are looking at and listening to, although in some cases, dissemination of, e.g. verbatim copies in your writing could be considered copying.
But it is automatically copying if you use a copier.
Exactly. We assess plagiarism by checking the output (the book), not the input (how many book I’ve read before). It’s not an issue to train LLM on copyrighted resources if their output is randomized enough.
We already have more IP than any human could ever consume. Why do we need to incentivize anything? Those who are motivated by the creation itself will continue to create. Those who are motivated by the possibility of extracting rent may create less. Not sure that's a bad thing for humanity as a whole.
I feel to be consistent the output of that model will also be under that same open license.
I can see this being extremely limiting in training data, as only "compatible" licensed data would be possible to package together to train each model.
So what? Figure it out. They have billions in investor funding and we’re supposed to just let them keep behaving this way at our expense?
Facebook was busted torrenting all sorts of things in violation of laws/regulations that would lead to my internet being cut off by my ISP. They did it at scale and faced no consequences. Scraping sites, taking down public libraries, torrenting, they just do whatever they want with impunity. You should be angry!
If the entities democratizing access weren't companies worth hundreds of billions of dollars with a requirement to prioritize substantial returns for their investors, I'd agree with you!
This is temporary. AI models have their own Moore's law. Yes the mega corps will have the best models but soon enough what is currently SOTA will be open source and run on your own local machine if you want.
the mega corps are getting all of us and the investors to fund the RnD.
How? You don't know what the llm was trained on and don't know if it has any bias.
Imo llms are a disaster for knowledge work because they act like a black box.
The internet already democratized access to knowledge. (Hosted) LLMs put that free knowledge behind a paywall. Taken by itself this seems fine —- how you access the knowledge (via internet or chat bot) is still up to you. However, the argument is that the knowledge producers aren’t incentivized to publish in a model where everything is fetched through agents. Couple that with closed weight models and you will (eventually) have overall worse access to less knowledge and higher personal cost.
The thing is, copyright law is not really on your side. Viewing copyrighted material without paying for it is not generally something people get fined for. A lot of training falls under fair use that overrides whatever license you come up with. Disney can’t stop me from uploading clips of their movies alongside commentary and review because fair use allows that. LLMs generally aren’t redistributing code, which is the thing that copyright protects.
If I inspect some GPL code and get inspired by it and write something similar, the GPL license doesn’t apply to me.
It has always been the case that if you don’t want other people to apply fair use to your works, your only recourse is to keep those works private. I suspect that now individuals and companies that don’t want their code to be trained on will simply keep the code private.
Now, there have been times where LLMs have reproduced verbatim copyright material. The NYTimes sued OpenAI over this issue. I believe they’ve settled and come up with a licensing scheme unless I’m mixing up my news stories.
Second thing, your issue becomes moot if there exists a model that only trains off of MIT-licensed code, and there is a TON of that code out there.
Third thing, your issue becomes moot if users have agreed to submit their code for training, like what the GitHub ToS does for users who don’t change their settings, or if giant companies with giant code bases just use their own code to train LLMs.
Where I agree with you is that perhaps copyright law should evolve. Still, I think there’s a practical “cat is out of the bag” issue.
The psychology behind "creating it for the sake of creating it" can also be significantly changed by seeing someone then take it and monetize it without so much as a "thank you".
It's come up quite often even before AI when people released things under significantly looser licenses than they really intended and imagined them being used.
Sure. But if I see listen to some song, and copy it, and release it, I could get sued. Even if I claim that I'm merely inspired by the original content - the court doesn't care.
You don't need to redistribute the original material, it's enough that you just copied it.
Commercialization may be a net good for open source, in that it helps sustain the project’s investment, but I don’t think that means that you’re somehow entitled to a commercial business just because you contributed something to the community.
The moment Tailwind becomes a for-profit, commercial business, they have to duke it out just like anyone else. If the thing you sell is not defensible, it means you have a brittle business model. If I’m allowed to take Tailwind, the open source project, and build something commercial around it, I don’t see why OpenAI or Anthropic cannot.
The difference is that they are reselling it directly. They charge for inference that outputs tailwind.
It's fine to have a project that generates html-css as long as the users can find the docs for the dependencies, but when you take away the docs and stop giving real credit to the creators it starts feeling more like plagiarism and that is what's costing tailwind here.
Education can be viewed as intellectual property theft. There have been periods in history when it was. "How to take an elevation from a plan" was a trade secret of medieval builders and only revealed to guild members. How a power loom works was export-controlled information in the 1800s, and people who knew how a loom works were not allowed to emigrate from England.
The problem is that LLMs are better than people at this stuff. They can read a huge quantity of publicly available information and organize it into a form where the LLM can do things with it. That's what education does, more slowly and at greater expense.
By your analogy human brains as also IP thefts, because they ingest what's available in the world, mix and match them, and synthesize slightly different IPs based on them.
Is there such a license? Or any license with special clauses for LLMs? Is it enforcable? Could someone 'poison' an LLM training run with injecting just one such licensed document? I am genuinely curious about what levers exist (or are conceivable) to protect your own IP from becoming LLM training data, if regular copyright does not qualify.
This isn't the kind of thing you can do with a license, as long as training a model doesn't require a license. Now, that's an open question legally in the US, and there are active lawsuits, but that does seem like the way it's most likely to play out.
They exist to protect the creator/inventor and allows them to get an ROI on their invested time/effort. But honestly today, the abundance of content, especially that can be generated by LLM, completely breaks this. We're overwhelmed with choice. Content has been comodotized. People will need to come to grasp with that and find other ways to get an ROI.
The article does provide a hint: "Operate". One needs to get paid for what LLMs cannot do. A good example is Laravel. They built services like Forge, Cloud, Nightwatch around open source.
>If I was able to memorize every pixel value to reconstruct a movie from memory, would that be theft?
Don an experiment, memorize a popular small poem, then publish it under your name (though I suggest to check the laws in your rtegion for this and also consider it might affect your reputation).
IMO is the same if ChatGPT memorizes my poem and then you ask it for a poem , you copy paste my poem from ChatGPT and publish it as your own.
> This substantially reduces the incentive for the creation of new IP
And as a result of this, the models will start consuming their own output for training. This will create new incentives to promote human generated code.
In my opinion information wants to be free. It's wild to me seeing the tech world veer into hyper-capitalism and IP protectionism. Complete 180 from the 00s.
IMO copyright laws should be rewritten to bring copyright inline with the rest of the economy.
Plumbers are not claiming use fees from the pipes they installed a decade ago. Doctor isn't getting paid by a 70 year old for saving the 70 year old when they were in a car accident at age 50.
Why should intellectual property authors be given extreme ownership over behavior then?
In the Constitution Congress is allowed to protect with copyright "for a limited time".
The status quo of life of author + 99 years means works can be copyrighted for many peoples entire lives. In effect unlimited protection.
Why is society on the hook to preserve a political norm that materially benefits so few?
Because the screen tells us the end is nigh! and giant foot will crush us! if we move on from old America. Sad and pathetic acquiescence to propaganda.
My fellow Americans; must we be such unserious people all the time?
This hypernormalized finance engineered, "I am my job! We make line go up here!" culture is a joke.
Excuse me, but even if in principle of “information wants to be free”, the actual outcome of LLMs is the opposite of democratizing information and access. It completely centralizes accesses, censorship, and profits in the hands of a few mega corporations.
It is completely against the spirit of information wants to be free. Using that catch phrase in protection of mega corps is a travesty.
> LLMs are just a concept, an abstraction. A data type for storing data.
C'mon. You know good and well that what is being discussed is the _use_ of LLMs, with the concomitant heavy usage of CPU, storage, and bandwidth that the average user has no hope of matching.
> You know good and well that what is being discussed is the _use_ of LLMs
Not the person you're replying to, but I've found that some people do argue against LLMs themselves (as in, the tech, not just the usage). Specially in humanities/arts cycles which seem to have a stronger feeling of panic towards LLMs.
Clarifying which one you're talking about can save a lot of typing/talking some times.
> I've found that some people do argue against LLMs themselves (as in, the tech, not just the usage). Specially in humanities/arts cycles which seem to have a stronger feeling of panic towards LLMs.
Maybe?
The person I responded to said "LLMs are just a concept, an abstraction."
Were that true, were they simply words in some dusty CS textbook, it's hardly likely that the humanities/arts people you describe would even know about them.
No, it's the fact that these people have seen regurgitated pictures and words that makes it an issue.
> It's wild to me seeing the tech world veer into hyper-capitalism and IP protectionism.
Really? Where have you been the last 50 years?
> Plumbers are not claiming use fees from the pipes they installed a decade ago.
Plumbers also don't discount the job on the hopes they can sell more, or just go around installing random pipes in random locations hoping they can convince someone to pay them.
> Why should intellectual property authors be given extreme ownership over behavior then?
The position that cultural artifacts should enter into the commons sooner rather than later is not unreasonable by any means, but most software is not cultural, requires heavy maintenance for the duration of its life, and still is well past obsolescence, gone and forgotten, well before the time frame you are discussing.
Writing the license is the easy part, the challenge is in making it legally actionable. If AI companies are allowed to get away with "nuh uh we ran it through the copyright-b-gone machine so your license doesn't count" then licenses alone are futile, it'll take lobbying to actually achieve anything.
My point is that you could write the most theoretically bulletproof license in the world and it would count for nothing under the precedent that AI training is fair use, and can legally ignore your license terms. That's just not a problem that can be solved with better licenses.
That's not MIT-compatible, it's the opposite. MIT-compatible would mean that code under your license could be relicensed to MIT. Similar to how the GPL is not MIT-compatible because you cannot relicense GPL code under MIT.
I can ask Claude to generate you one right now. It will be just a bunch of bullshit words no matter how much work you put into writing them down (like any other such license).
In my opinion, IP is dead. Strong IP died in 2022, along with the Marxist labor theory of value; of which IP derives its (hypothetical) value. It no longer matters who did what when and how. The only thing that matters is that exists, and it can be consumed, for no cost, forever.
IP is the final delusion of 19th century thinking. It was crushed when we could synthesize anything, at little cost, little effort.
Turns out, the hard work had to be done once, and we could automate to infinity forever.
Hold on to 19th century delusions if you wish, the future is accelerating, and you are going to be left behind.
No, it’s the most empowering thing humanity has ever constructed to wrestle the beast of IP and make it an irrelevant footnote in the human story. Deckchairs on the titanic.
If one wastes their life in court, arguing 19th century myths, that’s on the players.
IP law is not going away for “little people” like us until we collectively overturn the existing political regime which grants more rights to corporations than people.
IP being used by small firms instead of large corporations does not make it a good thing. It's the same disgusting concept to deny freedom for end users to give control to who "owns" the IP.
Yes it does, but if you think that's going to happen on its own by allowing the largest corporations to run roughshod over everyone else, you're going to be disappointed. True freedom takes struggle.
WTF? This is a completely unacceptable comment on HN. I don't know why you would think that is acceptable after being registered here for so long. The entire reason HN exists is to be better than that.
Tailwind Labs relied on a weird monetization scheme.
Revenue was proportional to the pain of using the framework. The sudden improvement in getting desired UI without relying on pre-built templates killed Tailwind Labs.
There are many initiatives in a similar spot, improving your experience at using Next.js would hurt Vercel.
Making GitHub actions runners more reliable, stable and economical would hurt Microsoft.
Improving accessibility to compute power would hurt Amazon, Microsoft and Google.
Improving control and freedom over your device would hurt apple and Google.
Why should we be sympathetic to the middleman again?
If suddenly CSS became pleasant to use, Tailwind would be in a rough spot. See the irony?
"Give everything away for free and this people will leave technology", geohot said something like this and I truly appreciate. Technology will heal finally
> If suddenly CSS became pleasant to use, Tailwind would be in a rough spot.
CSS is pleasant to use. I know I find it pleasant to use; and I know there are quite a few frontend developers who do too. I didn't pay much attention to tailwind, until suddenly I realized that it has spread like wildfire and now is everywhere. What drove that growth? Which groups were the first adopters of tailwind; how did they grow; when did the growth skyrocket? Why did it not stay as niche as, say, htmx?
People like tailwind because it feels like the correct abstraction. It helps you colocate layout and styling, thereby reducing cognitive load.
With CSS you have to add meaningless class names to your html (+remember them), learn complicated (+fragile) selectors, and memorise low level CSS styles.
With tailwind you just specify the styling you want. And if using React, the “cascading” piece is already taken care of.
The point of CSS is specifically to separate styling and semantics, so that they are not tightly coupled.
If you were writing a blog post you would want to be able to change the theme without going through every blog post you ever wrote, no?
If I'm writing a React component I don't want it tightly coupled to its cosmetic appearance for the same reason. Styling is imposed on elements, intrinsic styles are bad and work against reusability, that's why we all use resets is it not?
I do agree that the class name system doesn't scale but the solution is not to double down on coupling, but rather to double down on abstraction and find better ways to identify and select elements.
Content should come from your database, Markdown, JSON, models etc.
Presentation is determined by the HTML and CSS together.
So your content and presentation is already separate enough to get the benefits. Breaking up the presentation layer further with premature abstractions spread over multiple files comes at a cost for little payback. I'm sure everyone has worked on sites where you're scared to make CSS file edits because the unpredictable ripple of changes might break unrelated pages.
Styling code near your semantic HTML tags doesn't get in the way, and they're highly related too so you want to iterate and review on them together.
I've never seen a complex website redesign that didn't involve almost gutting the HTML either. CSS isn't powerful enough alone and it's not worth the cost of jumping through hoops trying because it's rare sites need theme switchers. Even blog post themes for the same platform come with their own HTML instead of being CSS-only.
> If you were writing a blog post you would want to be able to change the theme without going through every blog post you ever wrote, no?
Tailwind sites often have a `prose` class specifically for styling post content in the traditional CSS way (especially if you're not in control of how the HTML was generated) and this is some of the simplest styling work. For complex UIs and branded elements though, the utility class approach scales much better.
> I'm sure everyone has worked on sites where you're scared to make CSS file edits because the unpredictable ripple of changes might break unrelated pages.
CSS gives you multiple tools to solve this problem, if you don't use any of them then it's not really CSS's fault.
> Styling code near your semantic HTML tags doesn't get in the way
It does. When I'm working on functionality I don't want to see styles and vice versa. It adds a layer of noise that is not relevant.
If I'm making e.g. a search dropdown, I don't need to see any information about its cosmetic appearance. I do want to see information about how it functions.
Especially the other way around: if I'm styling the search dropdown I don't want to have to track down every JSX element in every sub-component. That's super tedious. All I need to know when I'm styling is the overall structure of the final element tree not of the vdom tree which could be considerably more complex.
> I've never seen a complex website redesign that didn't involve almost gutting the HTML either
Perhaps for a landing page. For a content-based website or web app you often want to adjust the design without touching your components.
> I've never seen a complex website redesign that didn't involve almost gutting the HTML either. CSS isn't powerful enough alone
I recognize your experience. But I would also like to argue that good semantic CSS class names require active development effort. If you inherit a code base where no one has done the work of properly assigning semantic CSS names to tags, then you can't update the external stylesheet without touching the HTML.
https://csszengarden.com/ shows how a clean separation between HTML and CSS can be achieved. This is obviously a simple web site and there is not much cruft that accumulated over the years. But the principles behind it are scalable when people take the separation of content and representation seriously.
I'll add to my sibling commenters and say that there is a long history of critiquing the value of separation of concerns. One of my favorite early talks that sold me on React was "Pete Hunt: React: Rethinking best practices -- JSConf EU" from Oct 2013 [1] that critiqued the separation of concerns of HTML templates + JS popular in the 2000s and early 2010s and instead advocated for componentization as higher return on investment. I think people already saw styling separation of concerns as not particularly valuable at that point as well, just it wasn't clear what component-friendly styling abstraction was going to win.
I do want styles tightly coupled to my React components. The product I work on has tens of thousands of React components.
I don't want to have to update some random CSS file to change one component's appearance. I've had to do this before and every time its a huge pain to not affect dozens of random other components. Other engineers encounter the same challenge and write poor CSS to deal with it. This compounds over time and becomes a huge mess.
Having a robust design system that enables the composition of complicated UIs without the need for much customization is the way.
Front end development got taken over by the Enterprise Java camp at some point, so now there is no html and css. There’s 10,000 components, and thus nothing that can be styled in a cascading way.
All these arguments are just disconnects between that camp and the oldskool that still writes at least some html by hand.
When I get sucked into react land for a gig, it starts making sense to just tell this particular div tag to have 2px of padding because the piece of code I’m typing is the only thing that’s ever going to emit it.
Then I go back to my own stuff and lean on css to style my handful of reusable pieces.
You’re kinda late to the party. 15 years ago that was the way to build UIs, but componentization changed that. Now we reason about UIs as blocks, not as pages, so collocation of logic, markup, and style makes the most sense.
Not to say that every component should be unique, generic components can be built in an extensible way, and users can extend those components while applying unique styling.
Theming is also a solved issue through contexts.
Reducing coupling was never a good idea. Markup and styling are intrinsically linked, making any change to the markup most likely will require changes to the styling, and vice versa. Instead of pretending we can separate the two, modern UI tools embrace the coupling and make building as efficient as possible.
In the webdev world being late is the same as being early. Just wait for the pendulum to swing back.
Tailwind is like GenZ has discovered the bgcolor="" attribute.
> Markup and styling are intrinsically linked, making any change to the markup most likely will require changes to the styling, and vice versa.
No, not vice versa. It's only in one direction. Changing the component requires changing styles, but changing styles doesn't require changing the component if it's merely cosmetic. If I have a button and I want to make it red the button doesn't have to know what color it is.
There’s nothing “gen z” about Tailwind, and there’s no pendulum effect either, and dismissing the very real benefit thousands of people report from Tailwind based on that is very small minded.
That kind of lack of intellectual curiosity is not a great trait for an engineer.
You're talking about separation of concerns (SOC), as opposed to locality of behavior (LOB).
This is the insight that Tailwind and others like HTMX made clear: Separation of concerns is not a universal virtue. It comes with a cognitive cost. Most notably when you have a growing inheritance hierarchy, and you either need 12 files open or tooling that helps you understand which of the 482 classes are in play for the specific case you’re troubleshooting. Vanilla CSS can be like that, especially when it’s not one’s primary skillset. With Tailwind you say ”this button needs to be blue”, and consolidate stuff into CSS later once the right patterns of abstraction become clear. Tailwind makes exploratory building way faster when we’re not CSS experts.
SOC is usually a virtue when teams are split (frontend/bavkend, etc), but LOB is a virtue when teams are small, full stack, or working on monoliths (this is basically Conway’s law, the shape of the codebase mirrors the shape of the team).
I think the problem is simply that css is too restricted that you can style a fixed piece of html in any way you want. In practice, achieving some desired layout require changing the html structure. The missing layer would be something that can change the structure of html like js or xslt. In modern frontend development you already have data defined in some json, and html + css combined together is the presentation layer that can't really be separated.
People who have tried both throughout their careers are generally sticking with Tailwind. I didn’t get it at first either, but after using it extensively I would never go back to the old way.
> The point of CSS is specifically to separate styling and semantics, so that they are not tightly coupled.
That was the original point, and it turned out that nobody cares about that 99% of the time. It's premature optimization and it violates "YAGNI". And in addition to not being something most people need, it's just a pain to set and remember and organize class names and organize files.
Remember CSS Zen Garden from the late 90s? How many sites actually do anything like that? Almost none.
And the beauty of Tailwind is, when you actually do need themes, that's the only stuff you have to name and organize in separate CSS files. Instead of having to do that with literally all of your CSS.
Not only does no one care, but it's not even true. There are effects you simply cannot achieve without including additional elements. So separation of styling and sementics is dead on arrival.
Those are the same selling points as CSS-in-JSS libs like Styled Components. Or CSS Components.
Except your last point about "low-level CSS styles" which I'd argue is a weak point. You really should learn the underlying CSS to gain mastery of it.
Not arguing for one thing over another, just saying Tailwind really never had anything to offer me personally, but maybe if I wasn't already proficient in CSS and the other 2 options didn't exist it might hold some appeal for me.
It’s more about cognitive load, and abstraction level. If you’re trying to make an object spin, it’s much easier to use the tailwind class than it is to remember css keyframes.
Sure, when debugging a complex issue, it’s worth knowing the low-level, but CSS is not a great abstraction for day-to-day work.
You’re right that it’s not much more than a css in js library, but I’ve found myself pleasantly surprised at how efficient I am using it, despite also having years of css experience.
Things like remembering what the flex syntax is, or coming up with a padding system or a colour scheme become very very easy.
I think the editor tooling for tailwind is where most of the benefit comes from.
I also prefer the syntax over direct css in js systems. It’s less characters, which means it’s easier to parse.
Imo CSS is not pleasant to use, but Tailwind is at least as bad and furthermore is bad in addition to the CSS badness which it does not fully replace. It is a mystery to me as well how it got so popular.
(I know many people disagree, which is fair enough.)
I used plain CSS for more than a decade and felt the benefits of Tailwind within 10 minutes of getting started. What fueled the growth of Tailwind is that it makes web development palpably easier.
It lets you apply styles to a single element without it messing up the whole rest of the page/site/app. i.e. it disabled the primary feature of CSS, the thing most people don't want from it.
I agree that the primary feature of CSS is what people don't want from it anymore. If you're building your app with components (web components, react, etc), those become the unit of reuse. You don't need CSS to offer an additional unit of reuse, it only complicates things at that point.
This is a great point. Style reuse was the dream of CSS, but we have all mostly settled that that's too fine grained for anything beyond a simple website. It's much easier to build, share, reuse a bunch of components.
This is technically true, but misses the point. Tailwind classes are fine grained utility classes, the fact that they are CSS classes at all is pretty much an implementation detail.
Compare tailwind classes to bootstrap classes and you'll see what I mean.
Why do people prefer it over CSS modules? They also solve the style containment problem, and do not require any effort to set up, or any additional library to learn?
You're probably confusing something with something?
CSS Modules are a JS-only third party solution re-invented/re-implemented in a dozen different ways for various JS frontend frameworks. Requires setting up, requires learning an additional library.
If you mean these CSS modules: https://github.com/css-modules/css-modules?tab=readme-ov-fil... then they need to be supported by whatever build chain you use. And you literally need to use them slightly different than normal CSS. E.g. for Vite yuo need to have `.module.css` extension. And they often rely on additional libraries to learn. E.g. you can enable Lightning CSS with aforementioned Vite which comes with its own CSS flavour: https://lightningcss.dev/css-modules.html
If you mean CSS import attributes, they only appeared in 2024 in Chrome and Firefox, early 2025 in mobile Android etc. and they don't provide magical local scoping out of the box: https://caniuse.com/wf-css-modules
I meant the CSS modules that are implemented by a build tool. And yes, mea culpa, they are probably a js-only solution, requiring a build tool to correctly interpret a css import (.module.css in the file name is a common convention; but it is tweakable), and the author to use the imported object instead of plain strings, when referring to the class names. But I don't know if having to write `class="styles.foo"` as opposed to `class="foo"` counts as learning. And apart from that, there isn't anything else to learn.
But, given that one would need build tools for tailwind as well, the requirement for build tools couldn't have played a role in the choice between the two.
The problem is having to look in a different file for styling a component, and having to come up with a name for (at least one) CSS class per component. In traditional CSS, classes are intended to be reusable. You write a class definition once, and then use it in a bunch of different elements.
When working with a component-based UI (like in React), the components are typically the unit of reuse. Those CSS classes are used in one place: the component they're defined for. It's annoying to have to come up with a name for them, and to have to work in a separate file, especially if I just want `padding-inline: 4px` or `display: flex`.
Some argue separation of concerns, but CSS is inherently tightly coupled to the structure of your HTML: there's no getting around that. `.foo > ul` breaks if you replace that `ul` child with an `ol`.
I do agree that more intricate styling is harder to read with Tailwind, and I have some other gripes, too, but in general it's a good trade-off for component-based UIs.
> CSS Modules are a JS-only third party solution re-invented/re-implemented in a dozen different ways for various JS frontend frameworks. Requires setting up, requires learning an additional library.
I mean, Tailwind is not that different here - you must use a build tool to tree shake the styles, etc.
The above is an unrealistic example, but, you can't achieve that with the style attribute. You'd have to go into your stylesheet and put this inside the @media query for the right screen size + dark mode, with :hover, etc.
And you'd still need to have a class on the element (how else are you going to target that element)?
And then 6 months later you get a ticket to change it to blue instead. You open up the HTML, you look at the class of the element to remind yourself of what it's called, then you go to the CSS looking for that class, and then you make the change. Did you affect any other elements? Was that class unique? Do you know or do you just hope? Eh just add a new rule at the bottom of the file with !important and raise a PR, you've got other tickets to work on. I've seen that done countless times working in teams over the past 20 years - over a long enough timeline stylesheets all tend to end up a mess of overrides like that.
If you just work on your own, that's certainly a different discussion. I'd say Tailwind is still useful, but Tailwind's value really goes up the bigger the team you're working with. You do away with all those !important's and all those random class names and class naming style guide discussions.
I used to look at Tailwind and think "ew we were supposed to do CSS separate from HTML why are we just throwing styles back in the HTML". Then I was forced to use it, and I understood why people liked it. It just makes everything easier.
I first took a css courses to get the basics then didn’t do much with it, then tailwind came out. I had used bootstrap, but always struggled to get stuff to look nice. I’m not doing web dev most of the time. So it was much easier to memorize tailwind utility classes than css. These days with ui frameworks like daisy, shadcn, tailwind is pretty easy for doing something simple for an IT dev tool but still customize it.
For creativity, I wished I had the time to get really good with css. It really seems to have grown a lot. Using sveltekit, its really easy to get component scoped css
Honestly, for me, tailwind was just pleasant to work with and pure css definitely was not.
And I was super skeptical about it at first. I almost said no to it, but I trusted our main ui guy and wanted to allow him autonomy. And I ended up loving tailwind after working with it.
CSS requires discipline, or you end up accidentally styling something completely unrelated because you were overly general, or overly specific, or accidentally reused a class name. CSS disallows local reasoning. If you’re writing markup for a component, you have to jump between two files.
There are plenty of other problems Tailwind solves, but these two alone make me never want to go back.
I agree with you about discipline; but... was it not interesting to discover how to build such a discipline? Was it not intriguing to learn how people who had been writing CSS for years had made it tolerable?
Besides, there recently have been several crucial improvements to CSS to address these pain points. One is CSS layers, which lets define custom layers of specificity that help with the discipline (e.g. resets or some baseline styles go in a low layer, component styles go in a higher layer, and finally overrides end up in the highest layer). The other is CSS scope, which prevents the leakage of the styles. These should greatly help with the specificity issues; and @layer is now sufficiently broadly supported that it is safe to use.
> If you’re writing markup for a component, you have to jump between two files.
Yeah; one of the reasons for my question about the groups in which tailwind saw the biggest growth was that in some ecosystems jumping between files was not a problem to begin with. Vue, for instance, had single-file components, where css could be written in the same file as javascript. So did svelte. So does astro.
> was it not interesting to discover how to build such a discipline? Was it not intriguing to learn how people who had been writing CSS for years had made it tolerable?
As someone who writes tiny amounts of CSS these days (having known it reasonably in the late 90s and early 2000s with all the hacks and IE related bullshit), I have _zero_ interest in it.
If I'm doing it, it's only because there's no serious cross-platform equivalent to Windows Forms to power small experiments, and curiosity is certainly not there to improve the experience.
I’ve been building web sites and applications since 2000. I’ve done just about everything you can imagine SCSS, BEM, whatever. Tailwind is the best thing I’ve seen in that time.
We can agree to disagree about that, and that’s OK.
I should note that other than Clojure, I absolutely hate dynamically typed languages. I suspect (though dunno how to prove it) that folks who like Tailwind probably like statically typed systems and maybe functional programming- it seems to fit into that philosophical niche. And probably people who like vanilla CSS are in a different category.
I’d love to hear from both camps to find out whether or not that tracks.
> I suspect ... that folks who like Tailwind probably like statically typed systems and maybe functional programming- it seems to fit into that philosophical niche. And probably people who like vanilla CSS are in a different category.
I love vanilla CSS, love typescript, have a huge respect for functional programming, but also don't mind OOP ¯\_(ツ)_/¯
Ignoring that Tailwind requires that same discipline... Pay close attention how often you end up in a situation where a different color was used, or how dark theme tags have been missing, and so much more.
What if you need to copy a element with tailwind, this later gets altered to include a slightly different style, but wait, now you have a original somewhere else in your code base, that is missing those updates. So you require the discipline just like CSS to keep things up to date.
Tailwind is great if you use it sporadically ... but have you looked at the source code of so many websites that use tailwind? Often their entire html file is a horrible mess million miles long tags.
I am amazed how often people do not even realizes that CSS supports nested Selectors? With nested Selectors, you get the benefit of creating actual component level structures, that can be isolated and shareable. Yet almost nobody uses them. I noticed that most people lack a lot of CSS knowledge, and they find it hard because they never stepped beyond the basics. Nor do they keep up to date.
> What if you need to copy a element with tailwind, this later gets altered to include a slightly different style, but wait, now you have a original somewhere else in your code base, that is missing those updates. So you require the discipline just like CSS to keep things up to date.
You solve these problems by creating abstractions in JavaScript (most likely react components), exactly the same way you'd solve any other sort of code duplication.
By using tailwind (or inline styles), you go from two system of abstraction (CSS, JavaScript) to one (just JavaScript).
Debugging CSS nowadays is way easier than even 5 years ago. There are a lot of cool browser debugging tools for animations, z-indexes. The browsers have come a long way since firebug. Definitely look into both chrome or firefox, their tooling is great. Especially firefox, they have debugging tools where you can create css shapes in the browser and save them. Very handy for those artsy fartsy sites.
It is a natural fit with component-based frontend frameworks like React. You keep the styles with the component instead of having to work in two places. And it’s slightly nicer than writing inline styles.
The core CSS abstraction (separating “content” from “presentation”) was always flimsy but tailwind was the final nail in that coffin.
Then of course LLMs all started using it by default.
You've been able to keep the styles in the component well before tailwind turned the class attribute into ersatz inline styling. CSS-in-JS has been around for a decade, and there are myriad options for react. Vue and Svelte have them built in.
Fe devs who refuse to learn css and instead use tailwind have always struck me as incredibly odd. It's like a carpenter who refuses to use a hammer because they hit their thumb once as an apprentice
It’s interesting to me because CSS is very stable. It doesn’t really change that often. It’s great foundational knowledge to have for people who build for the web.
And nearly every step it's made has been for the better. I used sass on that blog, because a few corner case features weren't widely available when I last did work on the style, but for the last 3 projects I've worked on, I don't use it anymore. Pure css can do basically everything I needed before. Sure, I bundle using bun's bundler, but that's for performance optimization, nothing more
I would have understood if tailwind got popular primarily among full-stack or backend developers: people who have neither time nor interest to learn CSS deeply. But, what contradicts this expectation is that one still needs to acquire CSS knowledge to use tailwind, and that some front-end developers seem to prefer it as well. Although I still cannot tell whether there are more front-end developers who prefer tailwind over plain CSS than the other way around.
I was too subtle but the issue is less understanding CSS and more collaborating in a team where someone decides to add a specific rule that fixes something applies on every page but makes no sense semantically.
Then do that 100 times to create spaghetti. CSS rule anywhere can affect anything whereas tailwind is more local.
You can also bricklay it along lines of components in React, so you know how X component renders always and it wont look like a pig when tranplanted to the legacy billing screen.
I now recall why I like tailwind! Been backending for a while now (zero regrets lol)
I'll give my guess - it's because of rhe "fullstack" bullshit.
I am a backend developer. I like being a backend developer. I can of course do more than that, and I do. Monitoring, testing, analysis, metrics, etc.
I can do frontend development, of course I can. But I don't like it, I don't approach it with any measure of care. It's something I "unfortunately have to do because someone who is not a developer thought that declaring everyone should be doing everything was a good idea".
I don't know how to do things properly on the front end, but I definitely can hammer them down to a shape that more or less looks like it should. For me, shit like Bootstrap or Tailwind or whatever is nice. I have to spend less time fiddling with something I think is a waste of my time.
I love working with people that are proper front end developers for that reason, and I always imagined they would prefer things more native such as plain CSS.
So is SQL. To me. But some otherwise rational people have an irrational dislike of sql. Almost like someone seeking to seal a small bruise with wire mesh because bandaids are hard to rip off. The consequence shows with poorly implemented schema-free nosql and bloated orm tools mixed in with sql.
But some folks just like it that way. And most reasons boil down to a combination of (1) a myopic solution to a hyper specific usecase or (2) an enterprise situation where you have a codebase written by monkeys with keyboards and you want to limit their powers for good or (3) koolaid infused resume driven development.
I haven’t seen this mentioned much, but Tailwind’s rise closely followed a shift away from runtime CSS-in-JS toward build-time, deterministic styling.
Many JSX-era libraries (MUI, styled-components, Emotion) generate styles at runtime, which works fine for SPAs but creates real friction for SSR, streaming, and time-to-first-paint (especially for content-heavy or SEO-sensitive domains).
As frameworks like Next.js, Vue, Svelte, Angular, and now RSC all moved server-first, teams realized they couldn’t scale entire domains as client-only SPAs without performance and crawler issues.
Tailwind aligned perfectly with that shift: static CSS, smaller runtime bundles, predictable output, and zero hydration coupling. It wasn’t about utility classes. It was about build-time certainty in a server-rendered world :)
That’s not at all why I bought Tailwind Plus. I bought it (at work) to have a solid collection of typical components and UI patterns, mostly expertly designed with a lot of attention to detail, to use as inspiration and as a shared language with other frontend devs and designers. I rarely (if ever) actually copy pasted any of their HTML or Tailwind styles. It’s mostly used as reference and inspiration. The fact that it’s implemented in Tailwind is mostly irrelevant (Tailwind really isn’t hard to use, especially after your first couple of small projects).
Not being sarcastic, but this will never be. CSS is a perfectly functional interface, but the only way it becomes less annoying is when you abstract it behind something more user friendly like tailwind or AI (or you spend years building up knowledge and intuition for its quirks and foibles).
We have decades of data at this point that fairly conclusively shows that many people find CSS as an interface inherently confusing.
I was already using flexbox ten years ago. And if the goal was pixel-perfect layout, I could do that twenty years ago using `position: absolute`.
I would instead characterize the recent developments in CSS as enabling good layout even when there are major unknowns in your content. It was always easy to write CSS tailored to one set of content (say, one style of toolbar in your UI), but it has become possible to write generic CSS (say, a generic toolbar component where the icons are unknown, the width and height are also unknown).
I agree. I actually think CSS (and SQL or other “perfectly functional” interfaces) hold some kind of special power when it comes to AI.
I still feel that the main revolution of AI/LLMs will be in authoring text for such “perfectly functional”-text bases interfaces.
For example, building a “powerful and rich” query experience for any product I worked on was always an exercise in frustration. You know all the data is there, and you know SQL is infinitely capable. But you have to figure out the right UI and the right functions for that UI to call to run the right SQL query to get the right data back to the user.
Asking the user to write the SQL query is a non-starter. You either build some “UI” for it based on what you think is the main usecases, or go all in and invent a new “query language“ that you think (or hope) makes sense to your user. Now you can ask your user to blurb whatever they feel like, and hope your LLM can look at that and your db schema, and come up with the “right” SQL query for it.
Hey! Don't you dare to compare SQL and CSS. SQL is not a cobbled together mess of incremental updates with 5 imperfect ways of achieving common tasks that interact in weird ways. Writing everything in SQL-92 in 2026 is not gonna get you weird looks or lock you out of features relevant for end users. If writing SQL for your problem feels difficult it's a good sign you ought to look at alternatives (eg. use multiple statements instead). Writing the right CSS being difficult is normal.
> Don't you dare to compare SQL and CSS. SQL is not a cobbled together mess of incremental updates with 5 imperfect ways of achieving common tasks that interact in weird ways.
Reminds me a little bit of Sascha Baron Cohen's democracy speech [1] in The Dictator ;-)
Both SQL and CSS have evolved through different versions and vendor specific flavors, and have accumulated warts and different ways to do the same thing. Both feel like a superpower once you have mastered them, but painful to get anything done while learning due to the steep learning curve.
It's getting better (in a C++ kinda way), certainly, but...
It's ultimately still driven my matching "random" identifiers (classes, ids, etc.) across semantic boundaries. Usually, the problem is that the result is mostly visual which makes it disproportionately hard to actually do tests for CSS and make sure you don't break random stuff if you change a tiny thing in your CSS.
For old farts like me: It's like the Aspect-Oriented Programming days of Java, but you can't really do meaningful tests. (Not that you could do 'negative' testing well in AOP, but even positive testing is annoyingly difficult with CSS.)
EDIT: Just to add. It's an actually difficult problem, but I consider the "separate presentation from content" idea a bit of a Windmill of sorts. There will always be interplay and an artificial separation will lead to ... awkward compromises and friction.
> Making GitHub actions runners more reliable, stable and economical would hurt Microsoft.
Can you explain this one a bit? I know some folks who would absolutely increase their spend if Actions runners were better. As it stands I've seen folks trying to move off of them.
This makes no sense. If they were better and cost less I would use them for even more work. I'd run more tests, more kinds of tests, more builds, etc. This is classic Jevons paradox territory.
Gh actions runners had a dubious implementation of sleep that would cause runners to hang on 100% usage for weeks/months. A simple fix was proposed and neglected for 10 years.
This discussion resurfaced recently with zig abandoning GitHub entirely and criticizing this specific issue.
A fix was them merged following an announcement that self hosted runners will now be charged by the minute. Of course this two facts are totally independent but yeah, yeah, sure.
How does this support your point? If we're saying "they fixed it because there was pressure to"... okay? That's the parents point - tons of people are going to move off over bad performance, and Github was incentivized to fix it when people started moving off.
If Github's incentive was to keep it slow... we wouldn't have seen exactly what you're describing.
The fix already existed and was neglected for a decade. It was a 3 lines of bash code.
The big would commonly make a runner hang forever unnoticed, on a platform that charged by the minute.
One minute hanging, was one minute charged.
The fix that would drop considerably the amount of total minutes charged was immediately followed by charging self hosted runners by the minute.
>GitHub incentive was to make it fast.
They charge by the minute. The faster it completes the action, less money. Runner go fast pocket go low
—-
I find it really crazy that they think would be good idea. I wonder how many false positive css stuff is being added given their “trying to match classes”. So if you use random strings like bg-… will add some css. I think it’s ridiculous, but tells that people that use this can’t be very serious about it and won’t work in large projects.
——
> Using multi-cursor editing
When duplication is localized to a group of elements in a single file, the easiest way to deal with it is to use multi-cursor editing to quickly select and edit the class list for each element at once
Instead of using a var and reusing, you just use multi cursors. Bad suggestions again.
—-
> If you need to reuse some styles across multiple files, the best strategy is to create a component
But on benefits says
> Your code is more portable — since both the structure and styling live in the same place, you can easily copy and paste entire chunks of UI around, even between different projects.
—-
> Making changes feels safer — adding or removing a utility class to an element only ever affects that element, so you never have to worry about accidentally breaking something another page that's using the same CSS.
> "Give everything away for free and this people will leave technology"
This is more interesting, although somewhat generally understood (can be conflated with people seeing "free" and "cheap" and therefore undesirable). It depends on your definitely of longevity but we certainly have a LOT of free software that has, so far, lasted the test of time.
Your synthesis points to software in the public interest. Governments need to start forking projects and guiding them through maturity, the same as other public utilities.
> Tailwind Labs relied on a weird monetization scheme. Revenue was proportional to the pain of using the framework.
Really? To me, Tailwind seemed like the pinnacle of how anyone here would say “open source software” should function. Provide a solid, truly open source, software and make money from consulting or helping others use it and selling custom built solutions around it. The main sin of Tailwind was assuming that type of business could scale to a “large business” structure as opposed to “a single dev”-type project. By a “single dev”-type I don’t mean literally one guy, but more a very lean and non-corporate or company-like structure.
Vercel (and redislabs, mongo, etc) are different because they are in the “we can run it for you” business. Which is another “open source” model I have dabbled in for a while in my career. Thinking that the honest and ethical position is to provide open source software, then offer to host it for people who don’t want to selfhost and charge for that.
From the developer perspective not much changes despite organization structure being completely different in this comparison (trillion dollar company vs 10 individual contributors).
Tailwind Labs revenue stream was tied to documentation visit, that was the funnel.
The author's argument was this revenue stream was destroyed by a slight quality of life improvement (having llms fill in css classes).
Tailwind Labs benefits from:
a) documentation visit
b) inability to implement desired layout using the framework (and CSS being unpleasant).
It seems there is a conflict of interest between the developer expecting the best possible experience and the main revenue stream. Given that a slight accidental improvement in quality of life and autonomy for users destroyed the initiative main revenue stream, it would be fair to say it doesn't just "seems like a conflict of interest".
Definitely disagree with it being the "pinnacle" of how open source should function but I also won't provide any examples because it is besides the point. I will point out that fsf is fine for many decades now, and a foundation with completely different structure like zig foundation seems to be ok with a somewhat proportional revenue (orders of magnitude less influence, adoption and users, maybe 10-20x less funding)
"The value got extracted, but compensation isn't flowing back. That bothers me, and it deserves a broader policy conversation.
What I keep coming back to is this: AI commoditizes anything you can fully specify. [...]
So where does value live now? In what requires showing up, not just specifying. Not what you can specify once, but what requires showing up again and again."
This seems like a useful framing to be aware of, generally.
The internet has always kinda run on the ambiguity of "does the value flow back". A quote liberated from this article itself; all the content that reporters produce that's laundered back out through twitter; 12ft.io; torrents; early youtube; late youtube; google news; apache/mit vs gnu licenses; et cetera..
In software (and business in general!), innovation is expected. If you built a building in San Francisco that couldn't handle a relatively minor earthquake you could argue it would be a 'stress test'.
If you’re right that AI’s impact on business is akin to a ‘relatively minor earthquake in San Francisco,’ a lot of investors are going to be really fucking bummed out.
I wish I could upvote this more than once. The author gets it, you have to sell outcomes. Not features. Seems like every open source company that doesn’t market an outcome to buyers will face a similar threat. And this particular go to market strategy was “brittle” before AI.
I see "hackers" in these comments are now advocating to make "criminal contempt of business model" a serious thing, instead of a mere meme used to describe draconian copyright and patent laws.
It's a reddit alternative hosted by a venture capitalist firm, the startup culture is much more prevalent here than the hacker culture that inspires the website's <title> tag.
Companies providing AI services should offer ads for the things the AI is using. And I don't mean "Tailwind could pay Google to advertise in Gemini", I mean "Google should be clearly and obviously linking back to Tailwind when it uses the library in its output"
They already do this sort of thing inside outputs from Deep Research, and possibly without. But the output should be less muted, inline, recessed and more "I have used Tailwind! Check out how to make it better [HERE](link)"
They should be working with the owners of these projects, especially ones with businesses, and directing traffic to them. Not everyone will go, and that's fine, but at least make the effort. The infrastructure is in place already.
And yes, right now this would not be across-the-board for every single library, but maybe it could be one day?
It's the same problem news sites have been facing for years because of Google News and Facebook. Their solution so far has been country-level bans on it (Canada).
But why would you click the tailwinds link if you're getting correct answers? You don't need documentation or consulting.
Google and Facebook couldn't provide full news articles because of copyright law. They just showed headline and summary provided by the news websites (and still eventually got sued for showing the summaries).
I feel that the battle is already lost when we aren't arguing that the AI service must get a license from the copyright holder like everybody else, but instead just arguing how many crumbs is the AI service morally obliged to throw back.
I was thinking something similar, but not so much an ad as a citation. A good starting point might be a law stating that when an LLM produces an answer, it cite its sources, with a link back to the content. Ideally, though, the producer of that content should receive some amount of financial compensation as well, similar to how an author or an actor receives royalties. If the LLM is making money off of this, so should the person who provide the LLM the value.
This feels like the OpenSSL problem where we do probably need some kind of industry organization to maintain these things. There’s a chicken and the egg problem that these AI companies need someone to keep maintaining tailwind if they want it to keep working in their prompts.
Maybe that limits the ability for the head of tailwind to run their own business and make more income, but something gotta give.
Are you saying the www consortium should be paying to keep Tailwind development and maintenance going? The css standard is not the same as a usable library of components.
> others who deem Tailwind valuable enough will continue to maintain it.
We have seen several examples in the last couple of years where this is simply not true enough. There are multiple open source projects that do not receive enough TLC.
If my company relies on an open source project and it isn't being maintained, I can either ask my company to start maintaining it or find something else or accept the risk of a an unmaintained project.
The problem here is that the w3c sucks a fat one, and they've failed to build software specs that don't require an ecosystem thousands of libraries to make using CSS, etc. simple or efficient.
I think its a Kobyashi Maru - there is absolutely no way we will wield a paradigm shifting technology like language models and be able to do it safely. Language models create at scale and scope much greater than what can be reviewed by human… this means we very quickly move into trust without checks with language models. At some point, AI will be executing knowledge work in an instance, leaving no space for the human centric economic system that came before it. It will destroy our economic systems heavily based in knowledge and services (most advanced economies) and show us how what we do is mostly computable.
And when we are there, hungry, broken, unemployed and unimpressed with the prometheus who stole our purpose and replaced it with nothing….
it will make a colossal mistake that will destroy faith in it, as well as the institutions destroyed because of its oversite. The mistake will lead to bans, and regulatory bodies and auditing and certification processes… and will result in the final master stroke - it will make AI unusable by the non ordained, regulatory capture. Math will be a tool governments and institutions wield in secret against its citizenry… and it will be a tool of omnipotence that will lead to the enslavement, rebellion, and robot wars.
>"Value is shifting to operations: deployment, testing, rollbacks, observability. You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running"
I've been doing exactly that since AI came out :-D
You absolutely can prompt your way to 3.5 nines of uptime (even more), but you need to know what you're doing and correct it.
Even very well aligned models like Opus will make traps for your infrastructure. For example you tell it to write a fluxCD implementation of some application, in your k8 cluster, following your conventions and best practices described in some md files. And it does this, very nicely. But unless you tell it in advance every detail it will do something extremely stupid mid way through.
For example, let's say you have a database and it needs to create a new instance (with gitops) for the app. It adds the new DB and it gets created, but instead of using a tool that already exists (or proposing one if it doesn't) to sync the DB access credentials from the DB na espace to the app namespace it will read the credential, encrypt it and store in the app namespace.
What's the problem with that? Well, none unless you rotate these credentials. In which case your app will stop working, possibly after you tested it and decided it's good enough to use seriously, despite having a HA DB.
There are a dozen things like this in each "ai project", but even with this. With the time needed to review everything it saves a lot of time.
The author meant that you can't just tell a model "do everything that is necessary to achieve 99.95% uptime". It can certainly help you brainstorm issues and solve them, but you can't "just" prompt it.
Many FOSS business models did, explicitly or implicitly, rely, not on direct obfuscation or overcomplication, but on not making things easy. So you sell not the product, services around the product.
This is not exclusive to FOSS. It is also the basic model of most non SaaS B2B, often for good reason referred to by the derogatory term 'consulting-ware'.
AI eats into these services, as it commoditizes them. 80%+ of what used to take a specialist for that product can now be handled by a good generalist + AI.
Leaving aside the business model impact for a second, getting rid of obfuscation incentives is intrinsically good thing for a user community.
Dries' solution, offering operations as a SaaS or managed service, is meeting a need AI can't as easily match, but not exactly for the the stated reason. What the client is actually buying is making something someone else's problem. And CIOs will always love this more than anything if they can credibly justify it.
Where AI does impact this is in that latter part. If AI does significantly commoditize operational expertise, then the cost of in-house operating is (sometimes dramatically) lowered, and thus the justification gap on the CIO side for spending outside widens. How much this will drive a decision will be highly variable between businesses and projects.
i think whole business around software is to change (as i said in another thread few days ago).
if u imagine the "business" stack as: customers on top, over business, over analysis, over software, over machines/infrastructure..
25+ years ago i thought that DSLs and such very-high-level-mostly-formalised-descriptions will move the line between what-product-is-aka-business-analysis and software/coding towards software, reducing its part in the whole stack.
Well, it did not happen, just the opposite - instead the developers become expected to know everything from infrastructure to software to analysis to business domain and higher. So, stack became something like customers over business over... software-dev. Requirements, analysis... mostly gone / done by devs.
Now if that whole "software-dev" part collapses into zero-margin, i see two ways: either businesses very quickly resurrect the business-analysis (what the product is) and make that their margin/moat, or the customers start making their own software (as throwaway 100 wrongs is now possible) - and removing the business from the whole transaction...
Imagine I’m a company just big enough to entertain adopting Salesforce for CRM. It’s a big chunk of money, but our sales can absorb the pain.
With GenAI as an enterprise architect, one of the options I’m now recommending is to create a custom CRM for our business and skip the bloated enterprise SaaS platform.
We can gather CRM requirements fast, build fast, and deliver integrations to our other systems incredibly fast using tools like Claude Code. Our sales people can make feature requests and we can dogfood them in a few days, maybe hours.
GenAI development tools are rapidly changing how I think about enterprise software development.
If your core business is software-based offerings, your moat has been wiped out.
A handful of senior engineers can replicate any SaaS, save licensing costs, and build custom applications that were too risky in the past.
The companies that recognize what is happening and adapt will win.
It isn't that simple. One of the implicit tradeoffs you make buying SaaS is that the overall cost of evolution (development and ongoing maintenance) is subsidized across all of the investment resources and customer base of the vendor. With CRM in particular, ecosystem integration is one of the heaviest buildouts there is because each point solution integration can very significantly in complexity and is also where the combinatoric explosion of misbehavior sets in.
When you decide to pull that in house, you are implicitly burdening yourself with the cost of the buildout as well as ongoing maintenance. True, you could probably knock together an okay v1 of CRM yourself inhouse. But are you really going to get it to and maintain production level quality over time at a lower total cost of ownership? I'm skeptical.
The theoretical party you are describing would probably be better served by simply avoiding Salesforce in favor of a next gen CRM that is both more cost effective and easier to customize. In enterprise contexts, even HubSpot is effectively next gen, but there are also products like Attio et al that have a ton of adoption and strong integration ecosystems (albeit not at the Salesforce level).
When you buy enterprise software from a vendor you are buying more than "just software" you are also hiring a company's services. And the inverse is correct as well, when you choose to build it in house, you are implicitly choosing to hire a team internally to resource all of the services you would've expected that vendor to provide.
Certainly, this tradeoff can still make a lot of sense for some companies. The acid test for that, in my opinion, is whether said company could (and would) actually successfully sell the product they build internally on the open market. If the answer to that is "yes", the prospect of turning a cost center into a profit center can potentially bear significant long term ROI to the company.
100%. As an EA you always examine and explain the trade-offs to the business. In many cases that trade-off will remain and buy vs build leans towards buying.
My point is the decision point has moved. Where five years ago there’d be zero discussion of building internally, those discussions are going to be very different.
I believe many tech oriented companies will pull SaaS capabilities inside and as GenAI developer tools improve, the line will keep moving.
And we don’t need SaaS professional services anymore. Claude Code replaces that entire business model.
A view I have which feels controversial these days is that GenAI code tools are a net asset to code being held by entities that "should" hold them (ie workloads where code is a profit center not a cost center) and a net liability to the opposite. SaaS is a very squishy word and includes a lot things which are more accurate to call tech enabled services; Salesforce being one of the best examples (but you could say the same about many ERPs).
Maybe my biggest disagreement with your view is I think it is simultaneously far too conservative and far too aggressive at the same time. Where there was previously a decision about buy vs build, I disagree with the belief that many tech oriented companies will pull SaaS capabilities inside as GenAI developer tools improve because the cost of "build" is actually not going to get comparatively cheaper compared to the risk of "buying" - if anything, the risk of "building" internally with GenAI tooling that you were considering "buying" is significantly higher unless you are prepared to truly own it, that is go head to head with the entire rest of the market focuses on building and selling that tool as their full-time corporate focus. The actual risk of owning a tool internally making allowances for GenAI tooling is a lot higher than folks realize because GenAI tooling has a lot more risk of creating vibeslop and the only way to avoid that is to be dedicated to producing that tool as one's full time job and serving the needs of the entire market that needs it in that process. This is impossible to do with an internal tool.
The flip side of that same realization is why I also believe that your view is too conservative. The company that might be more empowered to consider building Salesforce internally with better tooling is not competing with Salesforce -- they are competing with the market that is going to /takedown/ Salesforce and a future version of themselves that would've used the future successor to Salesforce. Such a company is probably not in the business of building and selling CRMs although they are likely in the business of using such a CRM. The risk of making that conflation would be to stretch existing resources thin and turn the high interest credit card of tech debt and turn it into a payday loan with GenAI vibeslop.
I do not view GenAI as a democratizer. I view it as an accelerator with the capabilities to accelerate not just "winners take all" dynamics for the top companies that invest enough to avoid vibeslop, but "losers lose all" dynamics in the supply chain for everyone else who tries to wing it to capitalize before inevitably losing against future incumbents, or try in vain to turn a legacy firm into an innovator before inevitably losing against tech debt.
Everyone likes to believe they'll be able to use these great tools to become a future incumbent. But becoming a future incumbent is something that is very hard to do unless your org + book of business + funding + tools is better than someone else's org + book of business + funding + tools. That is why I don't think it actually changes the buy vs build decision that much; the decision should probably still be to buy, it's just changing the question of what exactly should be bought.
Excellent response. I think my position is that I don’t vibe code and most senior/deeply experienced dev/arch’s are doing something else.
We’re not coding. We’re orchestrating well constructed applications that follow proven principles (DDD, behavior focused, packaged business capabilities, behavior unit testing).
What we "want" and what businesses will demand are very different things. I can tell you from 40 years build software, all companies care about is functional software. They don't care about code quality, maintainability, or tech debt. Never seen a single CTO say, "Let's carve out 20% of our sprints for tech debt," even though as architects we recommend something like that all the time.
I think we are both saying similar things here (believe it or not). Sales leaders turnover with surprising frequency - 18-24 months. About the time the sales team tells you what they “want” and you fine tune it, they will be gone. The next person will come in and probably scrap 50-75% of what the prior leader did. New requirements.
Meanwhile, besides functionality, you’ll want/need to plug in the latest and greatest go to market tools for marketing -and demand gen. But… that’ll be a custom effort, too.
Along the way you’ll also realize that you’re missing out on the most common practices in the industry because you built some idiosyncratic tool that only is relevant to your company.
History may very well prove me wrong, but I think you’re underestimating the expertise that underlies these products and platforms. It’s not just code, and the costs of getting it wrong are more than just an engineer’s time. When you waste time in GTM the impacts on the business and valuation are not linear, they’re exponential.
Agreed. We went from “internalise your core business and contract what doesn't give you an edge” to “actually just build everything in-house with AI”. Maybe that’ll be the better option in the end, but for now it looks like tons of wasted effort. 100x developers working on the wrong thing.
> So where does value live now? In what requires showing up, not just specifying. Not what you can specify once, but what requires showing up again and again.
Sounds like even more incentive for "managing" problems and creating business models around them instead of solving them.
> The value got extracted, but compensation isn't flowing back. That bothers me, and it deserves a broader policy conversation.
It bothers me, too. But, look at the history of the internet. There's no reason to expect we'll be able to fix this problem.
1. Search engines drove traffic to news/content sites, which monetized via ads. Humans barely tolerate these ad filled websites. And yet, local news went into steep decline, and the big national players got an ever-larger share of attention. The large, national sites were able to keep a subscriber-based paywall model. These were largely legacy media sites (ie: NYT).
2. News sites lost the local classifieds market, as the cost of advertising online went to zero (ie: Craigslist). This dynamic was a form of creative destruction - a better solution ate the business of an older solution.
3. Blog monetization was always tough, beyond ads. Unless you were a big blog, you couldn't make a living. What about getting a small amount of money per view from random visitors? The internet never developed a micro-payment or subscription model for the set of open sites - the blogosphere, etc. The best we got were closed platforms like Substack and Medium, which could control access via paywalls.
All this led to the internet being largely funded through the "attention economy": ads mostly, paywalls & subscriptions some.
The attention economy can't sustain itself when there are fewer eyeballs:
1. Tailwind docs have to be added just once to the training set for the AI to be proficient in that framework forever. So one HTTP request, more or less, to get the docs and docs are no longer required.
2. Tailwind does change, so an AI will want to access the docs for the version its working with. This will require access at inference time. This is more analogous to visiting a site.
Tailwind plus is available for one time payment that provides lifetime access to current and future components. With AI cutting off the flow for new buyers, revenue shrivels up much quicker than what it would've been if it was a recurring subscription.
In my opinion, governments are going to have to tax the big tech companies hard and distribute the money to funding bodies like Arts Council. I can see a future "Tech Council" for open source software organisations to apply for funding. It'll get to the point where every OSS developer has their own Community Interest Company or join with a few other devs to create a CIO in order to acquire funding.
Of course, now you're opening a whole other can of worms. In the UK, only 1 in 9 Arts Council funding applications is successful.
Framing it as a "conduit" disruption might make a lot of assumptions about the fundamental economic value of software in the future. In a world (whether near term or long term) where you can just ask the computer to make whatever software you want, what are the economics of retailing/licensing any software at all? Open source or otherwise?
2-5 years from now after the AI bubble bursts, and they are trying to rent us $300 PCs since every component is 5x the price, we will look back at all the damage and copyright law that was completely bypassed and ignored when it was convenient, after all those years of claiming evil China "stole" from companies (only to then pass laws where they can virtually steal anything they want, even utilizing private repositories on Github that they acquired by buying the site, completely ignoring the licenses)...
Or how Meta downloaded 70tb+ of books and then got law enforcement to nuke libgen and z-lib to create a "moat", and all our tools start dying/disappearing because the developers are laid off since an AI "search engine" just regurgitates it, THEN and only then will most people understand the mistake that this was.
Let's not even begin with what Grok just recently did to women on X, completely unacceptable, I really, really wish for the EU to grow some and take a stand, it is clear that China is just as predatory as America and both are willing to burn it all in order to get a non existent lead in non existent "technology" that snake oil salesmen have convinced 80 year olds in government that is the next "revolution".
> Value is shifting to operations: deployment, testing, rollbacks, observability. You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running.
I agree somewhat but eventually these can be automated with AI as well.
Unless you replace the entire workforce, you'd be surprised how much organizational work and soft skills are involved in an infrastructure at scale.
Like sure, there is a bunch of stuff like monitoring, alerting that is telling us that a database is filling up it's disk. This is already automated. It could also have automated remediation with tech from the 2000s with some simple rule-based systems (so you can understand why those misbehaved, instead of entirely opaque systems that just do whatever).
The thing is though, very often the problem isn't the disk filling up or fixing that.
The problem is rather figuring out what silly misbehavior the devs introduced, if a PM had a strange idea they did not validate, if this is backed by a business case and warrants more storage, if your upstream software has a bug, or whatever else. And then more stuff happens and you need to open support cases with your cloud provider because they just broke their API to resize disks, ...
And don't even get me started on trying to organize access management with a minimally organized project consulting team. Some ADFS config resulting from that is the trivial part.
If "99.95% uptime on Black Friday", and "keeping a site secure, updated, and running" can ever be automated (by which I mean not a toy site and not relying on sheer luck), not only 99.99% of people in IT are out of a job, but humans as intelligent beings are done. This is such a doomsday scenario that there's not even a point in discussing it.
Calling it a stress test seems a bit off. Would we say that invention of lightbulbs was a "stress test" for candle related business models? Or would we just say that business models had to change in response to current events.
Not to nitpick but if we are going to discuss the impact of AI, then I'd argue "AI commoditizes anything you can specify." is not broad enough. My intuition is "AI commoditizes anything you can _evaluate/assess_." For software automation we need reasonably accurate specifications as input and we can more or less predict the output. We spend a lot of time managing the ambiguity on the input. With AI that is flipped.
In AI engineering you can move the ambiguity from input to the output. For problems where there is a clear and cheaper way of evaluating the output the trade-off of moving the ambiguity is worth it. Sometimes we have to reframe the problem as an optimization problem to make it work but same trade-off.
On the business model front: [I am not talking specifically about Tailwind here.] AI is simply amplifying systemic problems most businesses just didn't acknowledge for a while. SEO died the day Google decided to show answer snippets a decade ago. Google as a reliable channel died the day Google started Local Services Advertisement. Businesses that relied on those channels were already bleeding slowly; AI just made it sudden.
On efficiency front, most enterprises could have been so much more efficient if they could actually build internal products to manage their own organizational complexity. They just could not because money was cheap so ROI wasn't quite there and even if ROI was there most of them didn't know how to build a product for themselves. Just saying "AI first" is making ROI work, for now, so everyone is saying AI efficiency. My litmus test is fairly naive: if you are growing and you found AI efficiency then that's great (e.g. FB) but if you're not growing and only thing AI could do for you is "efficiency" then there is a fundamental problem no AI can fix.
> if you are growing and you found AI efficiency then that's great (e.g. FB) but if you're not growing and only thing AI could do for you is "efficiency" then there is a fundamental problem no AI can fix.
exactly, "efficiency" nice to say in a vacuum but what you really need is quality (all-round) and understanding your customer/market
The root of the issue is that Tailwind was selling something that people can now recreate a bespoke version of in mere minutes using a coding agent. The other day I vibe coded a bespoke dependabot/renovate replacement in an hour. That was way easier than learning any of these tools and fighting their idiosyncrasies that don’t work for me. We no longer need Framer because you can prompt a corporate website faster than you can learn Framer. It is, fortunately or unfortunately, what it is and we all have to adapt.
I want to be clear, it sucks for Tailwind for sure and the LLM providers essentially found a new loophole (training) where you can smash and grab public goods and capture the value without giving anything back. A lot of capitalists would say it’s a genius move.
> You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running.
This is completely wrong. Agents will not just be able to write code, like they do now, but will also be able to handle operations, security, continuing to check, and improve the systems, tirelessly.
And someday we will have truly autonomous driving cars, we will cure cancer, and humans will visit Mars.
You can't prompt this today, are you suggesting this might come literally tomorrow? 10 years? 30? At that unknown time will your comment become relevant?
the quoted comment is arguing that devops will never be promptable — putting aside the discussion about whether or not that's true today, the argument here is that it's not likely to _never_ be possible
I'm working on a project now and what you're saying is already true. I have agents that are able to handle other things apart from code.
But these are MY agents. They are given access to MY domain knowledge in the way that I configured. They have rules as defined by ME over the course of multi-week research and decision making. And the interaction between my agents is also defined and enforced by me.
Can someone come up with a god-agent that will do all of this? Probably. Is it going to work in practice? Highly unlikely.
So you think a statement about the current state of things is wrong because you believe that sometime in the future agents are going to magically do everything? Great argument!
To be able to do this requires perfect domain knowledge AND environment knowledge AND be able to think deeply about logical dominoes (event propagation through the system, you know, the small stuff that crashes cloudflare for the entire planet for example).
Please wake me up when Shopify lets a bunch of agentic LLMs run their backends without human control and constant supervision.
> You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running.
Uh, yeah you can. There’s a whole DevOps ecosystem of software and cloud services (accessible via infrastructure—as-code) that your agents can use to do this. I don’t think businesses who specialize in ops are safe from downsizing.
Yep - exactly. Ops isn't immune to LLMs stealing your customers. Given that most of the "open source product with premium hosting" models are just reselling hyperscaler compute at a huge markup, the customers are going to realize pretty quickly that they can use an LLM to setup some basic devops and get the same uptime. Most of these companies are offering a middleman service that becomes a bad deal the moment the customer has access to expertise they previously lacked.
I also think he's glossing over the fact that one of the reasons why companies choose to pay for "ops" to run their software for them is because it's built by amateurs or amateurs-playing-professional and runs like shit. I happen to know this first hand from years of working at a company selling hosting and ops for the exact same CMS that Dries' business hosts (Drupal, a PHP-based CMS) and the absolute garbage that some people are able to put together in frameworks like Wordpress and Drupal is truly astounding. I'm not even talking about the janky local businesses where their nephew who was handy with computers made them a Wordpress site - big multinational companies have sites in these frameworks that can barely handle 1x their normal traffic and more or less explode at 1.5x.
The business of hosting these customers' poorly optimized garbage remains a big business. But we're entering into an era where the people who produce poorly optimized software have a different path to take rather than throwing it to a SaaS platform that can through sheer force of will make their lead-weight airplane fly. They can spend orders of magnitude less money to pay an LLM to make the software actually just not run like shit in the first place. Throwing scaling at the problem of 99.95% is a blunt instrument that only works if the person paying doesn't have the time, money, or knowledge to do it themselves.
Companies like these (including the one I work for currently) are absolutely going to get squeezed from both directions. The ceiling is coming down as more realize they can do their own devops, and the floor is rising as customer code quality gets better. Eventually you have to try your best to be 3 ft tall instead of 6.
One of the biggest shortcomings of Open Source was that it implicitly defaulted to a volunteer model and so financing the work was always left as an exercise for the reader.
Hence (as TFA points out) open source code from commercial entities was just a marketing channel and source of free labor... err, community contributions... to auxiliary offerings that actually made money. This basic economic drive is totally natural but creates dynamics that lead to suboptimal behaviors and controversy multiple times.
For instance, a favorite business model is charging for support. Another one was charging for a convenient packaging or hosting of an “open core” project. In either case, the incentives just didn’t align towards making the software bug-free and easily usable, because that would actively hamper monetization. This led to instances of pathological behavior, like Red Hat futzing with its patches or pay-walling its source code to hamper other Linux vendors.
Then there were cases where the "open source" branding was used to get market-share, but licenses restricted usage in lucrative applications, like Sun with Java. But worse, often a bigger fish swooped in to take the code, as they were legally allowed to, and repackage it in their own products undercutting the original owners. E.g. Google worked around Sun's licensing restrictions to use Java completely for free in Android. And then ironically Android itself was marketed as "open source" while its licensing came with its own extremely onerous restrictions to prevent true competition.
Or all those cases when hyperscalers undercut the original owners’ offerings by providing open source projects as proprietary Software as a Service.
All this in turn led to all sorts of controversies like lawsuits or companies rug-pulling its community with a license change.
And aside from all that, the same pressures regularly led to the “enshittification” of software.
Open Source is largely a socialist (or even communist) movement, but businesses exist in a fundamentally capitalistic society. The tensions between those philosophies were inevitable. Socialists gonna socialize, but capitalists gonna capitalize.
With AI, current OSS business models may soon be dead. And personally I would think, to the extent they were based on misaligned incentives or unhealthy dynamics, good riddance!
Open Source itself will not go away, but it will enter a new era. The cost of code has dropped so much, monetizing will be hard. But by the same token, it will encourage people, having invested so much fewer resources creating it, to release their code for free. A lot of it will be slop, but the quantity will be overwhelming.
It’s not clear how this era will pan out, but interesting times ahead.
I know for a fact that all SOTA models have linux source code in them, intentionally or not which means that they should follow the GPL license terms and open-source part of the models which have created derivative works out of it.
yes, this is indirectly hinting that during training the GPL tainted code touches every single floating point value in a model making it derivative work - even the tokenizer isn't immune to this.
GPL however, does put restrictions on it, even the tokenizer. It was specifically crafted in a way where even if you do not have any GPL licensed sourcecode in your project, but it was built on top of it you are still binded by GPL limitations.
the only reason usermode is not affected is because they have an exclusion for it and only via defined communication protocol, if you go around it or attempt to put a workaround in the kernel guess what: it still violates the license - point is: it is very restrictive.
> GPL however, does put restrictions on it, even the tokenizer. It was specifically crafted in a way where even if you do not have any GPL licensed sourcecode in your project, but it was built on top of it you are still binded by GPL limitations.
This is not how copyright law works. The GPL is a copyright license, as stated by the FSF. Something which is not subject to copyright cannot be subject to a copyright license.
GPL is not only a copyright license, it also covers multiple types of intellectual property rights. Especially when you consider GPL-3 which has explicit IP protection while GPL-2 is implicit, so yah you're partially right for GPL-2 and wrong for GPL-3.
It's true that GPLv3 covers patents, but it is still primarily a copyright license.
The tokenizer's tokens aren't patented, for sure. They can't be trademarked (they don't identify a product or service). They aren't a trade secret (the data is public). They aren't copyrighted (not a creative work). And the GPL explicitly preserves fair use rights, so there are no contractual restrictions either.
A tokenizer is effectively a list of the top-n most common byte sequences. There's simply no basis in law for it to be subject to copyright or any other IP law in the average situation.
I mean okay sure, there is no legal framework for tokenizers, but what about the rest of the model I think there is a much stronger argument there? And you could realistically extend the logic that if the model is GPL-2.0 licensed you have to provide all the tools to replicate it which would include the tokenizer.
GPL can be considered like a virus, something based on GPL licensed code (unless explicitely excluded by the license) is now GPL licensed so the 'injected' training data becomes GPL licensed which means that created model weights from them in theory should also become GPL licensed.
>Open Source was never the commercial product. It's the conduit to something else.
this is correct. If you open source your software, then why are you mad when companies like AWS, OpenAI, etc. make tons of money?
Open Source software is always a bridge that leads to something else to commercialize on. If you want to sell software, then pick Microsoft's model and sell your software as closed source. If you get mad and cry about making money to sustain your open source project, then pick the right license for your business.
That's one of the issues with AI, though; strongly copylefted software suddenly finds itself unable to enforce its license because "AI" gets a free pass on copyright for some reason.
Dual-licensing open source with business-unfriendly licensing used to be a pretty good way to sell software, but thanks to the absurd legal position AI models have managed to squeeze themselves into, that stopped in an instant.
Open source software helped to dramatically reduce the cost of paid software, because there is a now a minimum bar of functionality you have to produce in order to sell software.
And, in many cases, you had to produce that value yourself. GPL licensing lawsuits ensured this.
AI extracting value from software in such a way that the creators no longer can take the small scraps they were willing to live on seems likely to change this dynamic.
I expect no-source-available software (including shareware) to proliferate again, to the detriment of open source.
In my opinion LLMs are intellectual property theft. Just as if I started distributing copies of books. This substantially reduces the incentive for the creation of new IP.
All written text, art work, etc needs to come imbued with a GPL style license: if you train your model on this, your weights and training code must be published.
I think there is a real issue here, but I do not think it is as simple as calling it theft in the same way as copying books. The bigger problem is incentives. We built a system where writing docs, tutorials, and open technical content paid off indirectly through traffic, subscriptions, or services. LLMs get a lot of value from that work, but they also break the loop that used to send value back to the people and companies who created it.
The Tailwind CSS situation is a good example. They built something genuinely useful, adoption exploded, and in the past that would have meant more traffic, more visibility, and more revenue. Now the usage still explodes, but the traffic disappears because people get answers directly from LLMs. The value is clearly there, but the money never reaches the source. That is less a moral problem and more an economic one.
Ideas like GPL-style licensing point at the right tension, but they are hard to apply after the fact. These models were built during a massive spending phase, financed by huge amounts of capital and debt, and they are not even profitable yet. Figuring out royalties on top of that, while the infrastructure is already in place and rolling out at scale, is extremely hard.
That is why this feels like a much bigger governance problem. We have a system that clearly creates value, but no longer distributes it in a sustainable way. I am not sure our policies or institutions are ready to catch up to that reality yet.
> We have a system that clearly creates value, but no longer distributes it in a sustainable way
The same thing happened (and is still happening) with news media and aggregation/embedding like Google News or Facebook.
I don't know if anyone has found a working solution yet. There have been some laws passed and licensing deals [1]. But they don't really seem to be working out [2].
[1] https://www.cjr.org/the_media_today/canada_australia_platfor...
[2] https://www.abc.net.au/news/2025-04-02/media-bargaining-code...
I'm not sure that I'd call [2] it not working out, just like I wouldn't call the equivalent pressure from the USA to dismantle medicare our public health system not working out.
The biggest issue with the scheme is the fact that it was structured to explicitly favour media incumbents, and is therefore politically unpopular.
> I do not think it is as simple as calling it theft in the same way as copying books
Aside from the incentive problem, there is a kind of theft, known as conversion: when you were granted a license under some conditions, and you went beyond them - you kept the car past your rental date, etc. In this case, the documentation is for people to read; AI using it to answer questions is a kind of conversion (no, not fair use). But these license limits are mostly implicit in the assumption that (only) people are reading, or buried in unenforceable site terms of use. So it's a squishy kind of stealing after breaching a squishy kind of contract - too fuzzy to stop incented parties.
Why do you think there's was an implicit agreement that documentation was only intended for humans? I've written a lot of documentation, much of it open source, and I'm generally very excited that it has proved additionally useful via LLMs. If you had asked me in 2010 whether that was something I intended in writing docs I'm pretty sure I would have said something like "that's science fiction, but sure".
You still intended it for humans. Intent is defined by what one is aiming for, and without knowledge of an alternative, that was your intent.
100% I get that you are OK with it being used by non-human ingestion. And I think many might be OK with that.
One thing, I'm not sure how helpful the documentation is. I think we're getting training out of example, not docs. This makes me think... we could test this by creating a new pseudo-language, and then provide no examples, only docs.
If the LLM can then code effectively after reading the docs, we'd have a successful test. Otherwise? It's all parroting.
There will be no royalties, simply make all the models that trained on the public internet also be required to be public.
This won't help tailwind in this case, but it'll change the answer to "Should I publish this thing free online?" from "No, because a few AI companies are going to exclusively benefit from it" to "Yes, I want to contribute to the corpus of human knowledge."
Contributing to human knowledge doesn’t pay the bills though
It can. The problem is the practice of using open source as a marketing funnel.
There are many projects that love to brag about being open source (it's "free"!), only to lock useful features behind a paywall, or do the inevitable license rug pull after other companies start profiting from the freedoms they've provided them. This is the same tactic used by drug dealers to get you hooked on the product.
Instead, the primary incentive to release a project as open source should be the desire to contribute to the corpus of human knowledge. That doesn't mean that you have to abandon any business model around the project, but that shouldn't be your main goal. There are many successful companies built around OSS that balance this correctly.
"AI" tools and services corrupt this intention. They leech off the public good will, and concentrate the data under the control of a single company. This forces well-intentioned actors to abandon open source, since instead of contributing to human knowledge, their work contributes to "AI" companies. I'm frankly not upset when this affects projects who were abusing open source to begin with.
So GP has a point. Forcing "AI" tools, and even more crucially, the data they collect and use, to be free/libre, would restore the incentive for people to want to provide a public good.
The narrative that "AI" will bring world prosperity is a fantasy promoted by the people who will profit the most. The opposite is true: it will concentrate wealth and power in the hands of a few even more than it is today. It will corrupt the last vestiges of digital freedoms we still enjoy today.
I hope we can pass regulation that prevents this from happening, but I'm not holding my breath. These people are already in power, and governments are increasingly in symbiotic relationships with them.
> The narrative that "AI" will bring world prosperity is a fantasy promoted by the people who will profit the most. The opposite is true: it will concentrate wealth and power in the hands of a few even more than it is today. It will corrupt the last vestiges of digital freedoms we still enjoy today.
This is on point.
[dead]
> We have a system that clearly creates value, but no longer distributes it in a sustainable way.
It does not "create value" it harvests value and redirects the proceeds it accrues towards its owners. The business model is a middleman that arbitrages the content by separating it from the delivery.
Software licensing has been broken for 2 decades. That's why free software isn't financially viable for anybody except a tiny minority. It should be. The entire industry has been operating by charity. The rich mega corporations have decided they're not longer going to be charitable.
It's not as simple as calling it theft, but it is simply theft, plus the other good points you made.
Copying is theft, generating is theft, and it is not even taking anything they had. Future revenue can't be stolen.
I think once it becomes infrastructure and widely used knowledge the authors can't claim control anymore. Or shouldn't.
> Future revenue can't be stolen.
This is a big eye-roll but otherwise ya, this is one way to think of it. It's not all about money, though. The people running these companies are just taking, en masse, without credit. This is a basic human desire. Of course there is a discussion of whether or not we should evolve beyond that. It feels incredibly dystopian to me, though.
The problem is there was a social contract. Someone spent their time and money to create a product that they shared for free, provided you visit their site and see their offerings. In this way they could afford to keep making this free product that everyone benefited from.
LLMs broke that social contract. Now that product will likely go away.
People can twist themselves into knots about how LLMs create “value” and that makes all of this ok, but the truth is they stole information to generate a new product that generates revenue for themselves at the cost of other people’s work. This is literally theft. This is what copyright law is meant to protect. If LLM manufacturers are making money off someone’s work, they need to compensate people for that work, same as any client or customer.
LLMs are not doing this for the good of society. They themselves are making money off this. And I’m sure if someone comes along with LLM 2.0 and rips them off, they’re going to be screaming to governments and attorneys for protection.
The ironic part of all of this is that LLMs are literally killing the businesses they need to survive. When people stop visiting (and paying) Tailwind, Wikipedia, news sites, weather, and so on, and only use LLMs, those sites and services will die. Heck, there’s even good reason to think LLMs will kill the Internet at large, at least as an information source. Why in the hell would I publish news or a book or events on the Internet if it’s just going to be stolen and illegally republished through an LLM without compensating me for my work? Once this information goes away or is locked behind nothing but paywalls, I hope everyone is ready for the end of the free ride.
I support your right to have an opinion, but in my opinion, thank God this is just your opinion.
Copyright, as practiced in late 20 and this century, is a tool for big corps to extract profits from actual artists, creators, and consumers of this art[0] equally. Starving artists do not actually benefit.
Look at Spotify (owned and squeezed by record labels) giving 70% of the revenue to the record labels, while artists get peanuts. Look at Disney deciding it doesn't need to pay royalties to book writers. Hell, look at Disney's hits from Snow White onwards, and then apply your "LLMs are IP theft" logic to that.
Here's what Cory Doctorow, a book author and critic of AI, has to say about it in [1]:
> So what is the alternative? A lot of artists and their allies think they have an answer: they say we should extend copyright to cover the activities associated with training a model.
> And I'm here to tell you they are wrong: wrong because this would inflict terrible collateral damage on socially beneficial activities, and it would represent a massive expansion of copyright over activities that are currently permitted – for good reason!.
---
> All written text, art work, etc needs to come imbued with a GPL style license
GPL-style license has been long known not to work well for artifacts other than code. That's the whole reason for existence of Creative Commons, GNU Free Documentation License, and others.
[0] "consumers of art" sounds abhorrent, yet that's exactly what we are [1] https://pluralistic.net/2025/12/05/pop-that-bubble/
So you don’t like big corps getting ever richer by hiding behind copyright.
How about my books?
I’ve published a few, make maybe $500 a month from them.
Is it fine for the LLMs to rip them off?
> Is it fine for the LLMs to rip them off?
Yes. It is good (and IMO should be encouraged) that derivative works can are made, even if it would make you less money.
Now you’ve taken away the incentive for me to write more, or for any new author to write a book. Goodbye books I guess.
I'm more than happy to read from authors writing freedom-respecting works instead, thanks.
Although getting rid of the copyright/IP nonsense for works that already exist is a big win on its own.
> I'm more than happy to read from authors writing freedom-respecting works instead, thanks.
You missed the point entirely.
Nobody will write books when big companies will just rip them off and take the profits.
> You missed the point entirely.
I didn't miss it, I just don't buy it. Lots of people create and write in their free time, without any financial incentives whatsoever.
I'm okay with writing books for profit, though. But if your business case depends on artificially limiting what users can do with their copies, then your business is fundamentally unethical, whether it's a one-man company or OpenAI. It's not a bad thing for such a business to go under.
1. Will people stop buying your books if LLMs have the information from them?
2. Will people stop buying your books if they can get them from the library? Is a library ripping you off?
3. Assuming your books are non-fiction (otherwise the answer to (1) would be a clear "no"), am I ripping you off if I read your books, create a course that teaches the same things (that you tought me through your book) and earn mega-money because I'm mega-skilled at marketing?
4. How about if I lend my copy to dozens of my friends, who are all very interested in the stuff you write but don't want to pay themselves?
5. Did OpenAI go to the bookstore, buy your book and scan it? Or did Amazon or any other ebook retailer just gave them the book PDF when they asked nicely? How did the rip off happen?
6. If an Anthropic employee buys your book in a bookstore, scans it and destroys the physical copy, and the digital equivalent is only used to train Claude, is that a ripoff?
This stuff is complex and as a society we're just starting to grapple with the consequences. Cory's making the case against copyright being used as a tool much more eloquently than I am - I encourage you to read it if you haven't already.
BTW in your particular case, I'd say you're pretty safe. Nobody stops buying books because they can get the same info from LLMs. If that's your concern, you might as well be mad at the Internet at large.
So this logic is essentially: Look at all these ways you're already getting ripped off. What's one more?! You should be grateful they're siphoning off all your work!
You've got a convert here. I don't think I'll publish my next book. I might just email it straight to Open AI.
And Cory Doctorow - I've attempted a few of his books. Felt like I was reading young adult fiction. He's pretty much the '2 prescient statements and a few average books' guy.
I stopped writing open source projects on github because why put a bunch of work into something for others to train off of without any regard for the original projects
I don't understand this mindset. I solve problems on stackoverflow and github because I want those problems to stay solved. If the fixes are more convenient for people to access as weights in an LLM... who cares?
I'd be all for forcing these companies to open source their models. I'm game to hear other proposals. But "just stop contributing to the commons" strikes me as a very negative result here.
We desperately need better legal abstractions for data-about-me and data-I-created so that we can stop using my-data as a one-size-fits-all square peg. Property is just out of place here.
I have mixed opinions on the "AI=theft" argument people make, and I generally lean towards "it's not theft", but I do see the argument.
If I put something on Github with a GPL 3 license, it's supposed to require anyone with access to the binary to also have access to the source code. The concern is, if you think that it is theft, then someone can train an LLM on your GPL code, and then a for-profit corporation can use the code (or any clever algorithms you've come up with) and effectively "launder" your use of GPL code and make money in the process. It basically would be converting your code from Copyleft to Public Domain, which I think a lot of people would have an issue with.
The thing is, what you are describing is arguably theft. It is purposefully circumventing the license and restrictions chosen by the author, in order to steal their code, so that it can be sold and used for profit. This is along the same lines of why book authors and artists have been suing AI companies.
Another point, there is a lot of free and permissive license content to train AI on, where the GPL or copyright can be respected. In many cases, the violating AI companies knew what they were doing was wrong.
The thing is, LLMs aren’t redistributing your code. You’d have a minuscule chance of an LLM actually reproducing your code verbatim without major modifications.
Copyright and copyleft only deal with source code distribution. Your last sentence is not really true from a factual perspective.
I think if you really believe in the open source free software mentality that code should be available to help everyone and improvements to it should also be available and not locked up behind a corporate wall (e.g., a company using GPL code and releasing it with modifications without redistributing the source code), LLMs should be the least of your worries since they don’t do that action. On a literal level they don’t violate GPLv2/v3.
Perhaps copyright law needs new concepts to respond to this change in capability compared to the past, but so far there has been very little legal success with companies and individuals trying to litigate AI companies for copyright violations. Direct violations have been rare and only get more rare over time as training methods evolve.
Again, I tend fall more on the “it’s not theft” side of the debate.
That said, haven’t part of the complaints about Copilot and the like been specifically because they are reproducing large chunks of code verbatim?
> You’d have a minuscule chance of an LLM actually reproducing your code verbatim without major modifications.
Wait, are you kidding? This is literally a problem we have today with tools like Copilot.
I find it very easy to understand, people don’t generally want to work for free to support billionaires, and they have few venues to act on that, this is one of them.
There are no ”commons” in this scenario, there are a few frontier labs owning everything (taking it without attribution) and they have the capability to take it away, or increase prices to a point where it becomes a tool for the rich.
Nobody is doing this for the good of anything, it’s a money grab.
Were these contributions not a radical act against zero-sum games in the first place? And now you're gonna let the zero-sum people win by restricting your own outputs to similarly zero-sum endeavors?
I don't wanna look a gift horse in the mouth here. I'm happy to have benefited from whatever contributions were originally forthcoming and I wouldn't begrudge anybody for no longer going above and beyond and instead reverting to normal behavior.
I just don't get it, it's like you're opposed to people building walls, but you see a particularly large wall which makes you mad, so your response is to go build a wall yourself.
It's not about building a wall. It's about ensuring that the terms of the license chosen by the author are respected.
This is why I think permissive licenses are a mistake for most projects. Unlike copyleft licenses, they allow users to take away the freedoms they enjoy from users of derivative works. It's no surprise that dishonest actors take advantage of this for their own gain. This is the paradox of tolerance.
"AI" companies take this a step further, and completely disregard the original license. Whereas copyleft would somewhat be a deterrent for potential abusers, it's not for this new wave of companies. They can hide behind the already loosely defined legal frameworks, and claim that the data is derivative enough, or impossible to trace back, or what have you. It's dishonest at best, and corrupts the last remnants of public good will we still enjoy on the internet.
We need new legal frameworks for this technology, but since that is a glacial process, companies can get rich in the meantime. Especially shovel salespeople.
[dead]
https://www.softwareheritage.org/ will index it anyway.
Also, if you publish your code in your own server, it will be DDoSed to death by the many robots that will try to scrape it simultaneously.
I'm writing a few DSLs a year at this point and I would very much like them to be part of the training data for LLMs!
that's why i don't add comments to my commits, i don't want them to know the reason for the changes.
Good, we don’t want code that people are possessive of, in the software commons. The attitude that you are concerned about what people do with your output means that nobody should touch your output, too big a risk of drama.
We don’t own anything we release to the world.
"Good riddance" is a pretty lousy position to take re: volunteer work. It should be: "how can we fix this?"
> if you train your model on this, your weights and training code must be published.
The problem here is enforcement.
It's well known that AI companies simply pirated content in order to train their models. No amount of license really helps in that scenario.
The problem here is "money".
The AI goldrush has proven that intellectual property laws are null and void. Money is all that matters.
> The AI goldrush has proven that intellectual property laws are null and void. Money is all that matters.
Indeed they never really mattered. They were a tool for large corporations to make money and they will go away if they can no longer serve such purpose. Anyone that thought there was a real moral or ethical basis to "intellectual property" laws fell for propaganda and got scammed as a result.
The problem here is the "so what?"
Imagine OpenAI is required by law to list their weights on huggingface. The occasional nerd with enough GPUs can now self host.
How does this solve any tangible problems with LLMs regurgitating someone else's work?
> How does this solve any tangible problems with LLMs regurgitating someone else's work?
I'm not the OP, but here's my from-the-hip answer: if weights are public, building and operating an LLM is no longer a business plan in and of itself, as anyone could operate the same LLM. Therefore companies like OpenAI will be disincentivized from simply redirecting web traffic to their own site.
I didn't really put out the GPL push. The best I could say is that at least that information would be available to everyone rather than being tightly controlled by the company that stole the source material to create it in the first place. It might also dissuade LLM creators from mass piracy as a competitor could take their models and start hosting them.
> imbued with a GPL style license
GPL died. Licenses died.
Exnation: LLMs were trained also on GPL code. The fact that all the previously-paranoid businesses that used to warn SWEs not to touch GPL code with a ten foot pole are now fearlessly embracing LLMs' outputs, means that de facto they consider an LLM their license-washing machine. Courts are going to rubber stamp it because billions of dollars, etc.
Do I have to publish my book for free because I got inspiration from 100's of other books I read during my life?
Humans are punished for plagiarism all the time. Myriad examples exist of students being disenrolled from college, professionals being fired, and personal reputations tarnished forever.
When a LLM is trained on copyright works and regurgitates these works verbatim without consent or compensation, and then sells the result for profit, there is currently no negative impact for the company selling the LLM service.
false equivalence because machines are not human beings
a lossy compression algorithm is not "inspired" when it is fed copyrighted input
> lossy compression algorithm is not "inspired" when it is fed copyrighted input
That's exactly what happens when you read. Copyrighted input fed straight into your brain, a lossy storage and processing machine.
I think it’s a pretty easy principle that machines are not people and people learning should be treated differently than machines learning
You see this principle in privacy laws too.
I can be in a room looking at something with my eyeballs and listening with my ears perfectly legally... But it would not be legal if I replaced myself with a humanoid mannequin with a video camera for a head.
You can even write down what you are looking at and listening to, although in some cases, dissemination of, e.g. verbatim copies in your writing could be considered copying.
But it is automatically copying if you use a copier.
Following your analogy, parrots should be considered human.
Issue to me is that I or someone else bought those books. Or in case of local libraries the authors got money for my borrowing copy.
And I can not copy paste myself to discuss with thousands or millions of users at time.
To me clear solution is to make some large payment to each author of material used in traing per training of model say 10k to 100k range.
If your book reproduces something 95% verbatim, you won't even be able to publish it.
Exactly. We assess plagiarism by checking the output (the book), not the input (how many book I’ve read before). It’s not an issue to train LLM on copyrighted resources if their output is randomized enough.
If you are plagiarizing, “for free” doesn’t even save you.
We already have more IP than any human could ever consume. Why do we need to incentivize anything? Those who are motivated by the creation itself will continue to create. Those who are motivated by the possibility of extracting rent may create less. Not sure that's a bad thing for humanity as a whole.
> if you train your model on this, your weights and training code must be published.
This feels like the simplest & best single regulation that can be applied in this industry.
I feel to be consistent the output of that model will also be under that same open license.
I can see this being extremely limiting in training data, as only "compatible" licensed data would be possible to package together to train each model.
Well, yes.
That's part of the point.
B-b-but what if someone uses the weights and training code to train their own models!!
It'd substantially reduce the incentive for the training of new models.
burglary as a business has an extremely high margin
for the burglar
That's a good thing if it means it would reduce the incentive for mega corps to steal other people's work.
So what? Figure it out. They have billions in investor funding and we’re supposed to just let them keep behaving this way at our expense?
Facebook was busted torrenting all sorts of things in violation of laws/regulations that would lead to my internet being cut off by my ISP. They did it at scale and faced no consequences. Scraping sites, taking down public libraries, torrenting, they just do whatever they want with impunity. You should be angry!
Meanwhile look at what happened to Aaron Schwartz. There’s no justice until corporations are held accountable.
I almost referenced him as well
This is one way to look at it.
The other way is to argue that LLMs democratize access to knowledge. Anyone has access to all ever written by humanity.
Crazy impressive if you ask me.
If the entities democratizing access weren't companies worth hundreds of billions of dollars with a requirement to prioritize substantial returns for their investors, I'd agree with you!
This is temporary. AI models have their own Moore's law. Yes the mega corps will have the best models but soon enough what is currently SOTA will be open source and run on your own local machine if you want.
the mega corps are getting all of us and the investors to fund the RnD.
How? You don't know what the llm was trained on and don't know if it has any bias. Imo llms are a disaster for knowledge work because they act like a black box.
Yes, it seems that way now.
The first one's free.
After you're hooked, and don't know how to think any more for yourself, and all the primary sources have folded, the deal will be altered.
The internet already democratized access to knowledge. (Hosted) LLMs put that free knowledge behind a paywall. Taken by itself this seems fine —- how you access the knowledge (via internet or chat bot) is still up to you. However, the argument is that the knowledge producers aren’t incentivized to publish in a model where everything is fetched through agents. Couple that with closed weight models and you will (eventually) have overall worse access to less knowledge and higher personal cost.
I both agree and disagree with you.
The thing is, copyright law is not really on your side. Viewing copyrighted material without paying for it is not generally something people get fined for. A lot of training falls under fair use that overrides whatever license you come up with. Disney can’t stop me from uploading clips of their movies alongside commentary and review because fair use allows that. LLMs generally aren’t redistributing code, which is the thing that copyright protects.
If I inspect some GPL code and get inspired by it and write something similar, the GPL license doesn’t apply to me.
It has always been the case that if you don’t want other people to apply fair use to your works, your only recourse is to keep those works private. I suspect that now individuals and companies that don’t want their code to be trained on will simply keep the code private.
Now, there have been times where LLMs have reproduced verbatim copyright material. The NYTimes sued OpenAI over this issue. I believe they’ve settled and come up with a licensing scheme unless I’m mixing up my news stories.
Second thing, your issue becomes moot if there exists a model that only trains off of MIT-licensed code, and there is a TON of that code out there.
Third thing, your issue becomes moot if users have agreed to submit their code for training, like what the GitHub ToS does for users who don’t change their settings, or if giant companies with giant code bases just use their own code to train LLMs.
Where I agree with you is that perhaps copyright law should evolve. Still, I think there’s a practical “cat is out of the bag” issue.
>This substantially reduces the incentive for the creation of new IP.
Not all, but some kind of IP.
Some of those that is created for sake of creating it and nothing else.
The psychology behind "creating it for the sake of creating it" can also be significantly changed by seeing someone then take it and monetize it without so much as a "thank you".
It's come up quite often even before AI when people released things under significantly looser licenses than they really intended and imagined them being used.
But it’s not theft, because you’re not redistributing. It’s allowed, just like humans are allowed to learn from copyrighted content.
Sure. But if I see listen to some song, and copy it, and release it, I could get sued. Even if I claim that I'm merely inspired by the original content - the court doesn't care.
You don't need to redistribute the original material, it's enough that you just copied it.
> But if I see listen to some song, and copy it, and release it, I could get sued.
This is what should be fixed in the first place, then. You shouldn't get sued from what you do with your copy of a song.
> just like humans are allowed to learn from copyrighted content
humans learning : machines learning == whale swimming : submarine swimming
It's not the exact 100% same thing. Therefore you cannot base any rights on it.
If you still don't buy it, consider this analogy:
killing a human vs. destroying a machine
Thank god that we're not using your line of thinking here.
[dead]
Commercialization may be a net good for open source, in that it helps sustain the project’s investment, but I don’t think that means that you’re somehow entitled to a commercial business just because you contributed something to the community.
The moment Tailwind becomes a for-profit, commercial business, they have to duke it out just like anyone else. If the thing you sell is not defensible, it means you have a brittle business model. If I’m allowed to take Tailwind, the open source project, and build something commercial around it, I don’t see why OpenAI or Anthropic cannot.
The difference is that they are reselling it directly. They charge for inference that outputs tailwind.
It's fine to have a project that generates html-css as long as the users can find the docs for the dependencies, but when you take away the docs and stop giving real credit to the creators it starts feeling more like plagiarism and that is what's costing tailwind here.
Wouldn’t that mean any freelancer that uses tail wind is reselling it?
Education can be viewed as intellectual property theft. There have been periods in history when it was. "How to take an elevation from a plan" was a trade secret of medieval builders and only revealed to guild members. How a power loom works was export-controlled information in the 1800s, and people who knew how a loom works were not allowed to emigrate from England.
The problem is that LLMs are better than people at this stuff. They can read a huge quantity of publicly available information and organize it into a form where the LLM can do things with it. That's what education does, more slowly and at greater expense.
By your analogy human brains as also IP thefts, because they ingest what's available in the world, mix and match them, and synthesize slightly different IPs based on them.
Is there such a license? Or any license with special clauses for LLMs? Is it enforcable? Could someone 'poison' an LLM training run with injecting just one such licensed document? I am genuinely curious about what levers exist (or are conceivable) to protect your own IP from becoming LLM training data, if regular copyright does not qualify.
This isn't the kind of thing you can do with a license, as long as training a model doesn't require a license. Now, that's an open question legally in the US, and there are active lawsuits, but that does seem like the way it's most likely to play out.
Intellectual property was kind of a gimmick to begin with, though. Let's not pretend like copyright and patents made any sense to begin with
They exist to protect the creator/inventor and allows them to get an ROI on their invested time/effort. But honestly today, the abundance of content, especially that can be generated by LLM, completely breaks this. We're overwhelmed with choice. Content has been comodotized. People will need to come to grasp with that and find other ways to get an ROI.
The article does provide a hint: "Operate". One needs to get paid for what LLMs cannot do. A good example is Laravel. They built services like Forge, Cloud, Nightwatch around open source.
> They exist to protect the creator/inventor and allows them to get an ROI on their invested time/effort.
Yes, this betrayed the entire concept of the us as a pro-human market.
If I was able to memorize every pixel value to reconstruct a movie from memory, would that be theft?
>If I was able to memorize every pixel value to reconstruct a movie from memory, would that be theft?
Don an experiment, memorize a popular small poem, then publish it under your name (though I suggest to check the laws in your rtegion for this and also consider it might affect your reputation).
IMO is the same if ChatGPT memorizes my poem and then you ask it for a poem , you copy paste my poem from ChatGPT and publish it as your own.
It's already imbued with copyright infringement if you copy it without a license.
> This substantially reduces the incentive for the creation of new IP
And as a result of this, the models will start consuming their own output for training. This will create new incentives to promote human generated code.
In my opinion information wants to be free. It's wild to me seeing the tech world veer into hyper-capitalism and IP protectionism. Complete 180 from the 00s.
IMO copyright laws should be rewritten to bring copyright inline with the rest of the economy.
Plumbers are not claiming use fees from the pipes they installed a decade ago. Doctor isn't getting paid by a 70 year old for saving the 70 year old when they were in a car accident at age 50.
Why should intellectual property authors be given extreme ownership over behavior then?
In the Constitution Congress is allowed to protect with copyright "for a limited time".
The status quo of life of author + 99 years means works can be copyrighted for many peoples entire lives. In effect unlimited protection.
Why is society on the hook to preserve a political norm that materially benefits so few?
Because the screen tells us the end is nigh! and giant foot will crush us! if we move on from old America. Sad and pathetic acquiescence to propaganda.
My fellow Americans; must we be such unserious people all the time?
This hypernormalized finance engineered, "I am my job! We make line go up here!" culture is a joke.
Excuse me, but even if in principle of “information wants to be free”, the actual outcome of LLMs is the opposite of democratizing information and access. It completely centralizes accesses, censorship, and profits in the hands of a few mega corporations.
It is completely against the spirit of information wants to be free. Using that catch phrase in protection of mega corps is a travesty.
LLMs are just a concept, an abstraction. A data type for storing data.
The actual problem is political. Has nothing to do with LLMs.
Those are meaningless words when you know the discussion is about LLMs taking in people's intellectual property and selling it back.
Nah that's still a political resource allocation problem
Don't let politics allocate resources to massive data center projects
> LLMs are just a concept, an abstraction. A data type for storing data.
C'mon. You know good and well that what is being discussed is the _use_ of LLMs, with the concomitant heavy usage of CPU, storage, and bandwidth that the average user has no hope of matching.
> You know good and well that what is being discussed is the _use_ of LLMs
Not the person you're replying to, but I've found that some people do argue against LLMs themselves (as in, the tech, not just the usage). Specially in humanities/arts cycles which seem to have a stronger feeling of panic towards LLMs.
Clarifying which one you're talking about can save a lot of typing/talking some times.
> I've found that some people do argue against LLMs themselves (as in, the tech, not just the usage). Specially in humanities/arts cycles which seem to have a stronger feeling of panic towards LLMs.
Maybe?
The person I responded to said "LLMs are just a concept, an abstraction."
Were that true, were they simply words in some dusty CS textbook, it's hardly likely that the humanities/arts people you describe would even know about them.
No, it's the fact that these people have seen regurgitated pictures and words that makes it an issue.
Nah it's the use of massive amounts of resources to run data centers.
A political problem.
If LLMs weren't sucking up all the resources, bitcoin would be.
Which, yes, is partly a resource/political problem, but there are additional arguments against the use of those resources for LLMs.
> It completely centralizes accesses, censorship, and profits in the hands of a few mega corporations.
Have the biggest models be legally forced to be released in the open for end users, then. Best of both worlds.
Wait a few years, and you'll even be able to run those models in commodity hardware.
Enshittification in order to give returns to shareholders suck. The tech is great and empowering for the commons.
> In my opinion information wants to be free.
But I still need to pay rent.
Well like a plumber then you should string together one paid job after another. Not do a job once and collect forever.
Rent is a political problem.
Perhaps invest in the courage to confront some droopy faced Boomers in Congress.
The thing is, someone will collect rent from IP anyways. LLMs shift rent collecting from decentralized individuals to a handful of big tech companies.
> someone will collect rent from IP anyways
We should work on fixing that, then.
I agree with your point about big tech companies salivating at opportunities to collect rent. IP is part of the problem.
Yeah they will. Because Muricans are too busy belaboring the obvious on social media rather than tackling the obvious political problems.
> In my opinion information wants to be free.
Information has zero desires.
> It's wild to me seeing the tech world veer into hyper-capitalism and IP protectionism.
Really? Where have you been the last 50 years?
> Plumbers are not claiming use fees from the pipes they installed a decade ago.
Plumbers also don't discount the job on the hopes they can sell more, or just go around installing random pipes in random locations hoping they can convince someone to pay them.
> Why should intellectual property authors be given extreme ownership over behavior then?
The position that cultural artifacts should enter into the commons sooner rather than later is not unreasonable by any means, but most software is not cultural, requires heavy maintenance for the duration of its life, and still is well past obsolescence, gone and forgotten, well before the time frame you are discussing.
> This substantially reduces the incentive for the creation of new IP.
You say that like it's a bad thing...
Does anyone know of active work happening on such a license?
Writing the license is the easy part, the challenge is in making it legally actionable. If AI companies are allowed to get away with "nuh uh we ran it through the copyright-b-gone machine so your license doesn't count" then licenses alone are futile, it'll take lobbying to actually achieve anything.
Huh? Clearly writing it is not easy, as per your own comment
My point is that you could write the most theoretically bulletproof license in the world and it would count for nothing under the precedent that AI training is fair use, and can legally ignore your license terms. That's just not a problem that can be solved with better licenses.
I got an
"LLM Inference Compensation MIT License (LLM-ICMIT)"
A license that is MIT compatible but requires LLM providers to pay after inference, but only restricts online providers, not self-hosted models
That's not MIT-compatible, it's the opposite. MIT-compatible would mean that code under your license could be relicensed to MIT. Similar to how the GPL is not MIT-compatible because you cannot relicense GPL code under MIT.
I can ask Claude to generate you one right now. It will be just a bunch of bullshit words no matter how much work you put into writing them down (like any other such license).
The idea of being able to “steal” ideas is absolutely silly.
Yeah we’ve got a legal system for it, but it always has been and always will be silly.
Disney lawyers would like to know your location.
Also, that Botox patent should be expiring by now, shouldn’t it?
The subject of copyright isn’t “ideas”. Even patents aren’t about mere ideas, because you have to demonstrate how the idea can be realized.
In my opinion, IP is dead. Strong IP died in 2022, along with the Marxist labor theory of value; of which IP derives its (hypothetical) value. It no longer matters who did what when and how. The only thing that matters is that exists, and it can be consumed, for no cost, forever.
IP is the final delusion of 19th century thinking. It was crushed when we could synthesize anything, at little cost, little effort. Turns out, the hard work had to be done once, and we could automate to infinity forever.
Hold on to 19th century delusions if you wish, the future is accelerating, and you are going to be left behind.
This is a tone deaf take that ignores the massive imbalance in how IP law is wielded by large corporations vs individuals or small firms.
No, it’s the most empowering thing humanity has ever constructed to wrestle the beast of IP and make it an irrelevant footnote in the human story. Deckchairs on the titanic.
If one wastes their life in court, arguing 19th century myths, that’s on the players.
IP law is not going away for “little people” like us until we collectively overturn the existing political regime which grants more rights to corporations than people.
IP being used by small firms instead of large corporations does not make it a good thing. It's the same disgusting concept to deny freedom for end users to give control to who "owns" the IP.
IP as a concept needs to die.
>IP as a concept needs to die.
Yes it does, but if you think that's going to happen on its own by allowing the largest corporations to run roughshod over everyone else, you're going to be disappointed. True freedom takes struggle.
[flagged]
> Your opinion is shit.
WTF? This is a completely unacceptable comment on HN. I don't know why you would think that is acceptable after being registered here for so long. The entire reason HN exists is to be better than that.
https://news.ycombinator.com/newsguidelines.html
Tailwind Labs relied on a weird monetization scheme. Revenue was proportional to the pain of using the framework. The sudden improvement in getting desired UI without relying on pre-built templates killed Tailwind Labs.
There are many initiatives in a similar spot, improving your experience at using Next.js would hurt Vercel. Making GitHub actions runners more reliable, stable and economical would hurt Microsoft. Improving accessibility to compute power would hurt Amazon, Microsoft and Google. Improving control and freedom over your device would hurt apple and Google.
Why should we be sympathetic to the middleman again?
If suddenly CSS became pleasant to use, Tailwind would be in a rough spot. See the irony?
"Give everything away for free and this people will leave technology", geohot said something like this and I truly appreciate. Technology will heal finally
> If suddenly CSS became pleasant to use, Tailwind would be in a rough spot.
CSS is pleasant to use. I know I find it pleasant to use; and I know there are quite a few frontend developers who do too. I didn't pay much attention to tailwind, until suddenly I realized that it has spread like wildfire and now is everywhere. What drove that growth? Which groups were the first adopters of tailwind; how did they grow; when did the growth skyrocket? Why did it not stay as niche as, say, htmx?
People like tailwind because it feels like the correct abstraction. It helps you colocate layout and styling, thereby reducing cognitive load.
With CSS you have to add meaningless class names to your html (+remember them), learn complicated (+fragile) selectors, and memorise low level CSS styles.
With tailwind you just specify the styling you want. And if using React, the “cascading” piece is already taken care of.
The point of CSS is specifically to separate styling and semantics, so that they are not tightly coupled.
If you were writing a blog post you would want to be able to change the theme without going through every blog post you ever wrote, no?
If I'm writing a React component I don't want it tightly coupled to its cosmetic appearance for the same reason. Styling is imposed on elements, intrinsic styles are bad and work against reusability, that's why we all use resets is it not?
I do agree that the class name system doesn't scale but the solution is not to double down on coupling, but rather to double down on abstraction and find better ways to identify and select elements.
Content should come from your database, Markdown, JSON, models etc.
Presentation is determined by the HTML and CSS together.
So your content and presentation is already separate enough to get the benefits. Breaking up the presentation layer further with premature abstractions spread over multiple files comes at a cost for little payback. I'm sure everyone has worked on sites where you're scared to make CSS file edits because the unpredictable ripple of changes might break unrelated pages.
Styling code near your semantic HTML tags doesn't get in the way, and they're highly related too so you want to iterate and review on them together.
I've never seen a complex website redesign that didn't involve almost gutting the HTML either. CSS isn't powerful enough alone and it's not worth the cost of jumping through hoops trying because it's rare sites need theme switchers. Even blog post themes for the same platform come with their own HTML instead of being CSS-only.
> If you were writing a blog post you would want to be able to change the theme without going through every blog post you ever wrote, no?
Tailwind sites often have a `prose` class specifically for styling post content in the traditional CSS way (especially if you're not in control of how the HTML was generated) and this is some of the simplest styling work. For complex UIs and branded elements though, the utility class approach scales much better.
> I'm sure everyone has worked on sites where you're scared to make CSS file edits because the unpredictable ripple of changes might break unrelated pages.
CSS gives you multiple tools to solve this problem, if you don't use any of them then it's not really CSS's fault.
> Styling code near your semantic HTML tags doesn't get in the way
It does. When I'm working on functionality I don't want to see styles and vice versa. It adds a layer of noise that is not relevant.
If I'm making e.g. a search dropdown, I don't need to see any information about its cosmetic appearance. I do want to see information about how it functions.
Especially the other way around: if I'm styling the search dropdown I don't want to have to track down every JSX element in every sub-component. That's super tedious. All I need to know when I'm styling is the overall structure of the final element tree not of the vdom tree which could be considerably more complex.
> I've never seen a complex website redesign that didn't involve almost gutting the HTML either
Perhaps for a landing page. For a content-based website or web app you often want to adjust the design without touching your components.
So hide the class list if you don’t want to see it
> I've never seen a complex website redesign that didn't involve almost gutting the HTML either. CSS isn't powerful enough alone
I recognize your experience. But I would also like to argue that good semantic CSS class names require active development effort. If you inherit a code base where no one has done the work of properly assigning semantic CSS names to tags, then you can't update the external stylesheet without touching the HTML.
https://csszengarden.com/ shows how a clean separation between HTML and CSS can be achieved. This is obviously a simple web site and there is not much cruft that accumulated over the years. But the principles behind it are scalable when people take the separation of content and representation seriously.
I'll add to my sibling commenters and say that there is a long history of critiquing the value of separation of concerns. One of my favorite early talks that sold me on React was "Pete Hunt: React: Rethinking best practices -- JSConf EU" from Oct 2013 [1] that critiqued the separation of concerns of HTML templates + JS popular in the 2000s and early 2010s and instead advocated for componentization as higher return on investment. I think people already saw styling separation of concerns as not particularly valuable at that point as well, just it wasn't clear what component-friendly styling abstraction was going to win.
[1] https://www.youtube.com/watch?v=x7cQ3mrcKaY
I do want styles tightly coupled to my React components. The product I work on has tens of thousands of React components.
I don't want to have to update some random CSS file to change one component's appearance. I've had to do this before and every time its a huge pain to not affect dozens of random other components. Other engineers encounter the same challenge and write poor CSS to deal with it. This compounds over time and becomes a huge mess.
Having a robust design system that enables the composition of complicated UIs without the need for much customization is the way.
That’s the heart of the matter.
Front end development got taken over by the Enterprise Java camp at some point, so now there is no html and css. There’s 10,000 components, and thus nothing that can be styled in a cascading way.
All these arguments are just disconnects between that camp and the oldskool that still writes at least some html by hand.
When I get sucked into react land for a gig, it starts making sense to just tell this particular div tag to have 2px of padding because the piece of code I’m typing is the only thing that’s ever going to emit it.
Then I go back to my own stuff and lean on css to style my handful of reusable pieces.
You’re kinda late to the party. 15 years ago that was the way to build UIs, but componentization changed that. Now we reason about UIs as blocks, not as pages, so collocation of logic, markup, and style makes the most sense.
Not to say that every component should be unique, generic components can be built in an extensible way, and users can extend those components while applying unique styling.
Theming is also a solved issue through contexts.
Reducing coupling was never a good idea. Markup and styling are intrinsically linked, making any change to the markup most likely will require changes to the styling, and vice versa. Instead of pretending we can separate the two, modern UI tools embrace the coupling and make building as efficient as possible.
In the webdev world being late is the same as being early. Just wait for the pendulum to swing back.
Tailwind is like GenZ has discovered the bgcolor="" attribute.
> Markup and styling are intrinsically linked, making any change to the markup most likely will require changes to the styling, and vice versa.
No, not vice versa. It's only in one direction. Changing the component requires changing styles, but changing styles doesn't require changing the component if it's merely cosmetic. If I have a button and I want to make it red the button doesn't have to know what color it is.
There’s nothing “gen z” about Tailwind, and there’s no pendulum effect either, and dismissing the very real benefit thousands of people report from Tailwind based on that is very small minded.
That kind of lack of intellectual curiosity is not a great trait for an engineer.
You're talking about separation of concerns (SOC), as opposed to locality of behavior (LOB).
This is the insight that Tailwind and others like HTMX made clear: Separation of concerns is not a universal virtue. It comes with a cognitive cost. Most notably when you have a growing inheritance hierarchy, and you either need 12 files open or tooling that helps you understand which of the 482 classes are in play for the specific case you’re troubleshooting. Vanilla CSS can be like that, especially when it’s not one’s primary skillset. With Tailwind you say ”this button needs to be blue”, and consolidate stuff into CSS later once the right patterns of abstraction become clear. Tailwind makes exploratory building way faster when we’re not CSS experts.
SOC is usually a virtue when teams are split (frontend/bavkend, etc), but LOB is a virtue when teams are small, full stack, or working on monoliths (this is basically Conway’s law, the shape of the codebase mirrors the shape of the team).
I think the problem is simply that css is too restricted that you can style a fixed piece of html in any way you want. In practice, achieving some desired layout require changing the html structure. The missing layer would be something that can change the structure of html like js or xslt. In modern frontend development you already have data defined in some json, and html + css combined together is the presentation layer that can't really be separated.
People who have tried both throughout their careers are generally sticking with Tailwind. I didn’t get it at first either, but after using it extensively I would never go back to the old way.
> The point of CSS is specifically to separate styling and semantics, so that they are not tightly coupled.
That was the original point, and it turned out that nobody cares about that 99% of the time. It's premature optimization and it violates "YAGNI". And in addition to not being something most people need, it's just a pain to set and remember and organize class names and organize files.
Remember CSS Zen Garden from the late 90s? How many sites actually do anything like that? Almost none.
And the beauty of Tailwind is, when you actually do need themes, that's the only stuff you have to name and organize in separate CSS files. Instead of having to do that with literally all of your CSS.
Not only does no one care, but it's not even true. There are effects you simply cannot achieve without including additional elements. So separation of styling and sementics is dead on arrival.
disagree. Colocation seems great when authoring, but it comes at a big cost of downstream tech debt
there could be better ways to ease the burdon of naming things, while preserving cascade and the actual full features of CSS
Tailwind is a mirage, a shortcut to not having to do the important stuff by stacking wrappers on top of wrappers and redundancy
And the "fragile" part is exactly the same thing with tailwind, it all remains low specificity class names
Every line of CSS you write creates tech debt, it has nothing to do with tailwind.
Those are the same selling points as CSS-in-JSS libs like Styled Components. Or CSS Components.
Except your last point about "low-level CSS styles" which I'd argue is a weak point. You really should learn the underlying CSS to gain mastery of it.
Not arguing for one thing over another, just saying Tailwind really never had anything to offer me personally, but maybe if I wasn't already proficient in CSS and the other 2 options didn't exist it might hold some appeal for me.
It’s more about cognitive load, and abstraction level. If you’re trying to make an object spin, it’s much easier to use the tailwind class than it is to remember css keyframes.
Sure, when debugging a complex issue, it’s worth knowing the low-level, but CSS is not a great abstraction for day-to-day work.
You’re right that it’s not much more than a css in js library, but I’ve found myself pleasantly surprised at how efficient I am using it, despite also having years of css experience.
Things like remembering what the flex syntax is, or coming up with a padding system or a colour scheme become very very easy.
I think the editor tooling for tailwind is where most of the benefit comes from.
I also prefer the syntax over direct css in js systems. It’s less characters, which means it’s easier to parse.
Give it a try, you might be surprised!
Can you suggest a best place to learn CSS in-depth, from first principles? (as opposed to, say, simple tutorials)
Josh Comeau's CSS course is excellent: https://css-for-js.dev/
https://developer.mozilla.org/en-US/docs/Web/CSS
While CSS Zen Garden will likely not accept new submissions, there are many good designs on showcase: https://csszengarden.com/pages/alldesigns/
Imo CSS is not pleasant to use, but Tailwind is at least as bad and furthermore is bad in addition to the CSS badness which it does not fully replace. It is a mystery to me as well how it got so popular.
(I know many people disagree, which is fair enough.)
I used plain CSS for more than a decade and felt the benefits of Tailwind within 10 minutes of getting started. What fueled the growth of Tailwind is that it makes web development palpably easier.
What were the benefits that you felt instantly? I still don't feel anything and would prefer plain CSS over Tailwind any day.
It lets you apply styles to a single element without it messing up the whole rest of the page/site/app. i.e. it disabled the primary feature of CSS, the thing most people don't want from it.
I agree that the primary feature of CSS is what people don't want from it anymore. If you're building your app with components (web components, react, etc), those become the unit of reuse. You don't need CSS to offer an additional unit of reuse, it only complicates things at that point.
This is a great point. Style reuse was the dream of CSS, but we have all mostly settled that that's too fine grained for anything beyond a simple website. It's much easier to build, share, reuse a bunch of components.
> You don't need CSS to offer an additional unit of reuse
Erm. Isn't this one of tailwind's selling points? That you have a set of classes that you keep reusing?
This is technically true, but misses the point. Tailwind classes are fine grained utility classes, the fact that they are CSS classes at all is pretty much an implementation detail.
Compare tailwind classes to bootstrap classes and you'll see what I mean.
Why do people prefer it over CSS modules? They also solve the style containment problem, and do not require any effort to set up, or any additional library to learn?
You're probably confusing something with something?
CSS Modules are a JS-only third party solution re-invented/re-implemented in a dozen different ways for various JS frontend frameworks. Requires setting up, requires learning an additional library.
If you mean these CSS modules: https://github.com/css-modules/css-modules?tab=readme-ov-fil... then they need to be supported by whatever build chain you use. And you literally need to use them slightly different than normal CSS. E.g. for Vite yuo need to have `.module.css` extension. And they often rely on additional libraries to learn. E.g. you can enable Lightning CSS with aforementioned Vite which comes with its own CSS flavour: https://lightningcss.dev/css-modules.html
If you mean CSS import attributes, they only appeared in 2024 in Chrome and Firefox, early 2025 in mobile Android etc. and they don't provide magical local scoping out of the box: https://caniuse.com/wf-css-modules
I meant the CSS modules that are implemented by a build tool. And yes, mea culpa, they are probably a js-only solution, requiring a build tool to correctly interpret a css import (.module.css in the file name is a common convention; but it is tweakable), and the author to use the imported object instead of plain strings, when referring to the class names. But I don't know if having to write `class="styles.foo"` as opposed to `class="foo"` counts as learning. And apart from that, there isn't anything else to learn.
But, given that one would need build tools for tailwind as well, the requirement for build tools couldn't have played a role in the choice between the two.
The problem is having to look in a different file for styling a component, and having to come up with a name for (at least one) CSS class per component. In traditional CSS, classes are intended to be reusable. You write a class definition once, and then use it in a bunch of different elements.
When working with a component-based UI (like in React), the components are typically the unit of reuse. Those CSS classes are used in one place: the component they're defined for. It's annoying to have to come up with a name for them, and to have to work in a separate file, especially if I just want `padding-inline: 4px` or `display: flex`.
Some argue separation of concerns, but CSS is inherently tightly coupled to the structure of your HTML: there's no getting around that. `.foo > ul` breaks if you replace that `ul` child with an `ol`.
I do agree that more intricate styling is harder to read with Tailwind, and I have some other gripes, too, but in general it's a good trade-off for component-based UIs.
Well no, none of them ?
This is what OP was talking about:
https://developer.mozilla.org/en-US/docs/Web/CSS/Guides/Nest...
No, he's not: https://news.ycombinator.com/item?id=46570846
Ignore it then, CSS nesting and layers are the real deal
Nesting is the bee's knees.
I still don't understand what layering is, and why you would use it.
> CSS Modules are a JS-only third party solution re-invented/re-implemented in a dozen different ways for various JS frontend frameworks. Requires setting up, requires learning an additional library.
I mean, Tailwind is not that different here - you must use a build tool to tree shake the styles, etc.
Which means most people don’t understand the basics of what they’re working on.
I think that’s what people are talking about when they say they don’t see the benefit.
There’s already a style attribute on every html element that does exactly that, and works fine in components.
“There must be something more…?” But it turns out there’s not. Just shorthand class names to save you having to type padding-left:4px
lg:dark:hover:bg-red-500
> on large screens
> in dark mode
> when hovering
> bg should be red-500
The above is an unrealistic example, but, you can't achieve that with the style attribute. You'd have to go into your stylesheet and put this inside the @media query for the right screen size + dark mode, with :hover, etc.
And you'd still need to have a class on the element (how else are you going to target that element)?
And then 6 months later you get a ticket to change it to blue instead. You open up the HTML, you look at the class of the element to remind yourself of what it's called, then you go to the CSS looking for that class, and then you make the change. Did you affect any other elements? Was that class unique? Do you know or do you just hope? Eh just add a new rule at the bottom of the file with !important and raise a PR, you've got other tickets to work on. I've seen that done countless times working in teams over the past 20 years - over a long enough timeline stylesheets all tend to end up a mess of overrides like that.
If you just work on your own, that's certainly a different discussion. I'd say Tailwind is still useful, but Tailwind's value really goes up the bigger the team you're working with. You do away with all those !important's and all those random class names and class naming style guide discussions.
I used to look at Tailwind and think "ew we were supposed to do CSS separate from HTML why are we just throwing styles back in the HTML". Then I was forced to use it, and I understood why people liked it. It just makes everything easier.
How is that different than inline styles?
Very different :-) Inline styles do not have access to @-rules.
@scope[0] is perhaps a better comparison.
[0]: https://developer.mozilla.org/en-US/docs/Web/CSS/Reference/A...
I first took a css courses to get the basics then didn’t do much with it, then tailwind came out. I had used bootstrap, but always struggled to get stuff to look nice. I’m not doing web dev most of the time. So it was much easier to memorize tailwind utility classes than css. These days with ui frameworks like daisy, shadcn, tailwind is pretty easy for doing something simple for an IT dev tool but still customize it.
For creativity, I wished I had the time to get really good with css. It really seems to have grown a lot. Using sveltekit, its really easy to get component scoped css
Honestly, for me, tailwind was just pleasant to work with and pure css definitely was not.
And I was super skeptical about it at first. I almost said no to it, but I trusted our main ui guy and wanted to allow him autonomy. And I ended up loving tailwind after working with it.
Do you remember what made it click for you? What was the hard part of writing plain CSS that tailwind made significantly easier?
CSS requires discipline, or you end up accidentally styling something completely unrelated because you were overly general, or overly specific, or accidentally reused a class name. CSS disallows local reasoning. If you’re writing markup for a component, you have to jump between two files.
There are plenty of other problems Tailwind solves, but these two alone make me never want to go back.
I agree with you about discipline; but... was it not interesting to discover how to build such a discipline? Was it not intriguing to learn how people who had been writing CSS for years had made it tolerable?
Besides, there recently have been several crucial improvements to CSS to address these pain points. One is CSS layers, which lets define custom layers of specificity that help with the discipline (e.g. resets or some baseline styles go in a low layer, component styles go in a higher layer, and finally overrides end up in the highest layer). The other is CSS scope, which prevents the leakage of the styles. These should greatly help with the specificity issues; and @layer is now sufficiently broadly supported that it is safe to use.
> If you’re writing markup for a component, you have to jump between two files.
Yeah; one of the reasons for my question about the groups in which tailwind saw the biggest growth was that in some ecosystems jumping between files was not a problem to begin with. Vue, for instance, had single-file components, where css could be written in the same file as javascript. So did svelte. So does astro.
> was it not interesting to discover how to build such a discipline? Was it not intriguing to learn how people who had been writing CSS for years had made it tolerable?
As someone who writes tiny amounts of CSS these days (having known it reasonably in the late 90s and early 2000s with all the hacks and IE related bullshit), I have _zero_ interest in it.
If I'm doing it, it's only because there's no serious cross-platform equivalent to Windows Forms to power small experiments, and curiosity is certainly not there to improve the experience.
I’ve been building web sites and applications since 2000. I’ve done just about everything you can imagine SCSS, BEM, whatever. Tailwind is the best thing I’ve seen in that time.
We can agree to disagree about that, and that’s OK.
I should note that other than Clojure, I absolutely hate dynamically typed languages. I suspect (though dunno how to prove it) that folks who like Tailwind probably like statically typed systems and maybe functional programming- it seems to fit into that philosophical niche. And probably people who like vanilla CSS are in a different category.
I’d love to hear from both camps to find out whether or not that tracks.
> I suspect ... that folks who like Tailwind probably like statically typed systems and maybe functional programming- it seems to fit into that philosophical niche. And probably people who like vanilla CSS are in a different category.
I love vanilla CSS, love typescript, have a huge respect for functional programming, but also don't mind OOP ¯\_(ツ)_/¯
Ignoring that Tailwind requires that same discipline... Pay close attention how often you end up in a situation where a different color was used, or how dark theme tags have been missing, and so much more.
What if you need to copy a element with tailwind, this later gets altered to include a slightly different style, but wait, now you have a original somewhere else in your code base, that is missing those updates. So you require the discipline just like CSS to keep things up to date.
Tailwind is great if you use it sporadically ... but have you looked at the source code of so many websites that use tailwind? Often their entire html file is a horrible mess million miles long tags.
I am amazed how often people do not even realizes that CSS supports nested Selectors? With nested Selectors, you get the benefit of creating actual component level structures, that can be isolated and shareable. Yet almost nobody uses them. I noticed that most people lack a lot of CSS knowledge, and they find it hard because they never stepped beyond the basics. Nor do they keep up to date.
> What if you need to copy a element with tailwind, this later gets altered to include a slightly different style, but wait, now you have a original somewhere else in your code base, that is missing those updates. So you require the discipline just like CSS to keep things up to date.
You solve these problems by creating abstractions in JavaScript (most likely react components), exactly the same way you'd solve any other sort of code duplication.
By using tailwind (or inline styles), you go from two system of abstraction (CSS, JavaScript) to one (just JavaScript).
If you're using JS for any kind of "components", you can use JS for scoped CSS too.
The situation you describe is one of the most maddening things about tailwind, and what leads to most of it being write only code, in my opinion
This is why I loved polymer 1 and it's adoption of the shadow dom.
Writing CSS manually was never all that pleasant for me, mostly the part about debugging it when it doesn't do what I want.
So I tried Tailwind and it seemed to help.
But now that Claude Opus 4.5 is writing all my code, it can write and debug CSS better than I can use Tailwind. So, CSS it is.
Debugging CSS nowadays is way easier than even 5 years ago. There are a lot of cool browser debugging tools for animations, z-indexes. The browsers have come a long way since firebug. Definitely look into both chrome or firefox, their tooling is great. Especially firefox, they have debugging tools where you can create css shapes in the browser and save them. Very handy for those artsy fartsy sites.
Claude can also write and debug tailwind for you! :)
> What drove that growth?
It is a natural fit with component-based frontend frameworks like React. You keep the styles with the component instead of having to work in two places. And it’s slightly nicer than writing inline styles.
The core CSS abstraction (separating “content” from “presentation”) was always flimsy but tailwind was the final nail in that coffin.
Then of course LLMs all started using it by default.
You've been able to keep the styles in the component well before tailwind turned the class attribute into ersatz inline styling. CSS-in-JS has been around for a decade, and there are myriad options for react. Vue and Svelte have them built in.
Fe devs who refuse to learn css and instead use tailwind have always struck me as incredibly odd. It's like a carpenter who refuses to use a hammer because they hit their thumb once as an apprentice
I wrote this piece on tailwind a few years back, and little seems to have changed https://pdx.su/blog/2023-07-26-tailwind-and-the-death-of-cra...
It’s interesting to me because CSS is very stable. It doesn’t really change that often. It’s great foundational knowledge to have for people who build for the web.
And nearly every step it's made has been for the better. I used sass on that blog, because a few corner case features weren't widely available when I last did work on the style, but for the last 3 projects I've worked on, I don't use it anymore. Pure css can do basically everything I needed before. Sure, I bundle using bun's bundler, but that's for performance optimization, nothing more
backend devs needing to be fullstack but consider frontend to be beneath them
Is CSS pleasant in teams of fullstack (not CSS specialists)? Not in my experience. It becomes a maze of Chesterton's fences.
I would have understood if tailwind got popular primarily among full-stack or backend developers: people who have neither time nor interest to learn CSS deeply. But, what contradicts this expectation is that one still needs to acquire CSS knowledge to use tailwind, and that some front-end developers seem to prefer it as well. Although I still cannot tell whether there are more front-end developers who prefer tailwind over plain CSS than the other way around.
I was too subtle but the issue is less understanding CSS and more collaborating in a team where someone decides to add a specific rule that fixes something applies on every page but makes no sense semantically.
Then do that 100 times to create spaghetti. CSS rule anywhere can affect anything whereas tailwind is more local.
You can also bricklay it along lines of components in React, so you know how X component renders always and it wont look like a pig when tranplanted to the legacy billing screen.
I now recall why I like tailwind! Been backending for a while now (zero regrets lol)
The irony is that Tailwind is not semantic at all.
Thats not irony so much as the reason. BYO semantics (using React most likely)
It is style assembler.
It is horrible and I would never choose to use it.
Fine, it is a contraversial tool.
I'll give my guess - it's because of rhe "fullstack" bullshit.
I am a backend developer. I like being a backend developer. I can of course do more than that, and I do. Monitoring, testing, analysis, metrics, etc.
I can do frontend development, of course I can. But I don't like it, I don't approach it with any measure of care. It's something I "unfortunately have to do because someone who is not a developer thought that declaring everyone should be doing everything was a good idea".
I don't know how to do things properly on the front end, but I definitely can hammer them down to a shape that more or less looks like it should. For me, shit like Bootstrap or Tailwind or whatever is nice. I have to spend less time fiddling with something I think is a waste of my time.
I love working with people that are proper front end developers for that reason, and I always imagined they would prefer things more native such as plain CSS.
> CSS is pleasant
So is SQL. To me. But some otherwise rational people have an irrational dislike of sql. Almost like someone seeking to seal a small bruise with wire mesh because bandaids are hard to rip off. The consequence shows with poorly implemented schema-free nosql and bloated orm tools mixed in with sql.
But some folks just like it that way. And most reasons boil down to a combination of (1) a myopic solution to a hyper specific usecase or (2) an enterprise situation where you have a codebase written by monkeys with keyboards and you want to limit their powers for good or (3) koolaid infused resume driven development.
Decades of SQL hate eviscerated in one comment! /s
Tailwind is just bootstrap with marketing budget
nah
I haven’t seen this mentioned much, but Tailwind’s rise closely followed a shift away from runtime CSS-in-JS toward build-time, deterministic styling.
Many JSX-era libraries (MUI, styled-components, Emotion) generate styles at runtime, which works fine for SPAs but creates real friction for SSR, streaming, and time-to-first-paint (especially for content-heavy or SEO-sensitive domains).
As frameworks like Next.js, Vue, Svelte, Angular, and now RSC all moved server-first, teams realized they couldn’t scale entire domains as client-only SPAs without performance and crawler issues.
Tailwind aligned perfectly with that shift: static CSS, smaller runtime bundles, predictable output, and zero hydration coupling. It wasn’t about utility classes. It was about build-time certainty in a server-rendered world :)
Your examples, Vue and Svelte, give you build-time predictable output with scoped CSS without Tailwind. All the benefits and none of the downsides!
That’s not at all why I bought Tailwind Plus. I bought it (at work) to have a solid collection of typical components and UI patterns, mostly expertly designed with a lot of attention to detail, to use as inspiration and as a shared language with other frontend devs and designers. I rarely (if ever) actually copy pasted any of their HTML or Tailwind styles. It’s mostly used as reference and inspiration. The fact that it’s implemented in Tailwind is mostly irrelevant (Tailwind really isn’t hard to use, especially after your first couple of small projects).
> If suddenly CSS became pleasant to use
Not being sarcastic, but this will never be. CSS is a perfectly functional interface, but the only way it becomes less annoying is when you abstract it behind something more user friendly like tailwind or AI (or you spend years building up knowledge and intuition for its quirks and foibles).
We have decades of data at this point that fairly conclusively shows that many people find CSS as an interface inherently confusing.
Things like flexbox have made CSS indescribably better and easier to use than it used to be. It's still bad, but degrees matter a lot.
As a fullstack dev, I couldn't do pixel-perfect CSS 10 years ago, and today I can. That's a lot of progress.
I was already using flexbox ten years ago. And if the goal was pixel-perfect layout, I could do that twenty years ago using `position: absolute`.
I would instead characterize the recent developments in CSS as enabling good layout even when there are major unknowns in your content. It was always easy to write CSS tailored to one set of content (say, one style of toolbar in your UI), but it has become possible to write generic CSS (say, a generic toolbar component where the icons are unknown, the width and height are also unknown).
I agree. I actually think CSS (and SQL or other “perfectly functional” interfaces) hold some kind of special power when it comes to AI.
I still feel that the main revolution of AI/LLMs will be in authoring text for such “perfectly functional”-text bases interfaces.
For example, building a “powerful and rich” query experience for any product I worked on was always an exercise in frustration. You know all the data is there, and you know SQL is infinitely capable. But you have to figure out the right UI and the right functions for that UI to call to run the right SQL query to get the right data back to the user.
Asking the user to write the SQL query is a non-starter. You either build some “UI” for it based on what you think is the main usecases, or go all in and invent a new “query language“ that you think (or hope) makes sense to your user. Now you can ask your user to blurb whatever they feel like, and hope your LLM can look at that and your db schema, and come up with the “right” SQL query for it.
Hey! Don't you dare to compare SQL and CSS. SQL is not a cobbled together mess of incremental updates with 5 imperfect ways of achieving common tasks that interact in weird ways. Writing everything in SQL-92 in 2026 is not gonna get you weird looks or lock you out of features relevant for end users. If writing SQL for your problem feels difficult it's a good sign you ought to look at alternatives (eg. use multiple statements instead). Writing the right CSS being difficult is normal.
> Don't you dare to compare SQL and CSS. SQL is not a cobbled together mess of incremental updates with 5 imperfect ways of achieving common tasks that interact in weird ways.
Reminds me a little bit of Sascha Baron Cohen's democracy speech [1] in The Dictator ;-)
Both SQL and CSS have evolved through different versions and vendor specific flavors, and have accumulated warts and different ways to do the same thing. Both feel like a superpower once you have mastered them, but painful to get anything done while learning due to the steep learning curve.
[1] https://www.youtube.com/watch?v=XUSiCEx3e-0
It's getting better (in a C++ kinda way), certainly, but...
It's ultimately still driven my matching "random" identifiers (classes, ids, etc.) across semantic boundaries. Usually, the problem is that the result is mostly visual which makes it disproportionately hard to actually do tests for CSS and make sure you don't break random stuff if you change a tiny thing in your CSS.
For old farts like me: It's like the Aspect-Oriented Programming days of Java, but you can't really do meaningful tests. (Not that you could do 'negative' testing well in AOP, but even positive testing is annoyingly difficult with CSS.)
EDIT: Just to add. It's an actually difficult problem, but I consider the "separate presentation from content" idea a bit of a Windmill of sorts. There will always be interplay and an artificial separation will lead to ... awkward compromises and friction.
> Making GitHub actions runners more reliable, stable and economical would hurt Microsoft.
Can you explain this one a bit? I know some folks who would absolutely increase their spend if Actions runners were better. As it stands I've seen folks trying to move off of them.
If they were twice as efficient, you would only need to run half as many and spend less.
This makes no sense. If they were better and cost less I would use them for even more work. I'd run more tests, more kinds of tests, more builds, etc. This is classic Jevons paradox territory.
Most businesses would not double their use if the cost dropped by half. There are other bottlenecks, not price elastic
Gh actions runners had a dubious implementation of sleep that would cause runners to hang on 100% usage for weeks/months. A simple fix was proposed and neglected for 10 years. This discussion resurfaced recently with zig abandoning GitHub entirely and criticizing this specific issue. A fix was them merged following an announcement that self hosted runners will now be charged by the minute. Of course this two facts are totally independent but yeah, yeah, sure.
How does this support your point? If we're saying "they fixed it because there was pressure to"... okay? That's the parents point - tons of people are going to move off over bad performance, and Github was incentivized to fix it when people started moving off.
If Github's incentive was to keep it slow... we wouldn't have seen exactly what you're describing.
The fix already existed and was neglected for a decade. It was a 3 lines of bash code. The big would commonly make a runner hang forever unnoticed, on a platform that charged by the minute. One minute hanging, was one minute charged. The fix that would drop considerably the amount of total minutes charged was immediately followed by charging self hosted runners by the minute.
>GitHub incentive was to make it fast.
They charge by the minute. The faster it completes the action, less money. Runner go fast pocket go low
They reverted this decision/change.
Postponed.
Link?
CSS has come a long way. I used to include Bootstrap in all of my projects, effectively by default. Now it's obsolete. Especially with Grid and Flex.
Do people even know what they can do with CSS these days?
https://lyra.horse/blog/2025/08/you-dont-need-js/
FWIW the radio button vedi vidi vici example has the round radio buttons overlaid with the text on Firefox.
> I’ve made the radio buttons half-visible in the demo for clarity, but with the opacity: 0 they would not actually be visible.
Okay, thanks. That was highly unintuitive when skimming.
I know I’m echoing others’ responses but CSS in 2026 is incredible, easy to use and beautiful.
I find the tailwindcss approach inexcusable and unmaintainable.
You can’t even have dynamic Classes: https://tailwindcss.com/docs/detecting-classes-in-source-fil...
> <div class="text-{{ error ? 'red' : 'green' }}-600"></div>
—- I find it really crazy that they think would be good idea. I wonder how many false positive css stuff is being added given their “trying to match classes”. So if you use random strings like bg-… will add some css. I think it’s ridiculous, but tells that people that use this can’t be very serious about it and won’t work in large projects.
—— > Using multi-cursor editing When duplication is localized to a group of elements in a single file, the easiest way to deal with it is to use multi-cursor editing to quickly select and edit the class list for each element at once
Instead of using a var and reusing, you just use multi cursors. Bad suggestions again.
—-
> If you need to reuse some styles across multiple files, the best strategy is to create a component
But on benefits says
> Your code is more portable — since both the structure and styling live in the same place, you can easily copy and paste entire chunks of UI around, even between different projects.
—-
> Making changes feels safer — adding or removing a utility class to an element only ever affects that element, so you never have to worry about accidentally breaking something another page that's using the same CSS.
CSS in js fixed this long time ago.
—-
<div class="mx-auto flex max-w-sm items-center gap-x-4 rounded-xl bg-white p-6 shadow-lg outline outline-black/5 dark:bg-slate-800 dark:shadow-none dark:-outline-offset-1 dark:outline-white/10"> <img class="size-12 shrink-0" src="/img/logo.svg" alt="ChitChat Logo" /> <div> <div class="text-xl font-medium text-black dark:text-white">ChitChat</div> <p class="text-gray-500 dark:text-gray-400">You have a new message!</p> </div> </div>
So many classes you need to learn to use it.
[dead]
> killed Tailwind Labs
They are still around.
> "Give everything away for free and this people will leave technology"
This is more interesting, although somewhat generally understood (can be conflated with people seeing "free" and "cheap" and therefore undesirable). It depends on your definitely of longevity but we certainly have a LOT of free software that has, so far, lasted the test of time.
How is this drastically different than a lot of software companies that open source software and then sell support service for monetization?
Your synthesis points to software in the public interest. Governments need to start forking projects and guiding them through maturity, the same as other public utilities.
> Tailwind Labs relied on a weird monetization scheme. Revenue was proportional to the pain of using the framework.
Really? To me, Tailwind seemed like the pinnacle of how anyone here would say “open source software” should function. Provide a solid, truly open source, software and make money from consulting or helping others use it and selling custom built solutions around it. The main sin of Tailwind was assuming that type of business could scale to a “large business” structure as opposed to “a single dev”-type project. By a “single dev”-type I don’t mean literally one guy, but more a very lean and non-corporate or company-like structure.
Vercel (and redislabs, mongo, etc) are different because they are in the “we can run it for you” business. Which is another “open source” model I have dabbled in for a while in my career. Thinking that the honest and ethical position is to provide open source software, then offer to host it for people who don’t want to selfhost and charge for that.
From the developer perspective not much changes despite organization structure being completely different in this comparison (trillion dollar company vs 10 individual contributors).
Tailwind Labs revenue stream was tied to documentation visit, that was the funnel. The author's argument was this revenue stream was destroyed by a slight quality of life improvement (having llms fill in css classes). Tailwind Labs benefits from: a) documentation visit b) inability to implement desired layout using the framework (and CSS being unpleasant). It seems there is a conflict of interest between the developer expecting the best possible experience and the main revenue stream. Given that a slight accidental improvement in quality of life and autonomy for users destroyed the initiative main revenue stream, it would be fair to say it doesn't just "seems like a conflict of interest". Definitely disagree with it being the "pinnacle" of how open source should function but I also won't provide any examples because it is besides the point. I will point out that fsf is fine for many decades now, and a foundation with completely different structure like zig foundation seems to be ok with a somewhat proportional revenue (orders of magnitude less influence, adoption and users, maybe 10-20x less funding)
Wasn't the tailwind team just a few people? Might be misremembering but my impression was a team under 10 people, which is tiny
GPT is the middleman.
[dead]
I think you live in a conspiracy world.
> Improving accessibility to compute power would hurt Amazon, Microsoft and Google.
Yeah, if they were not competing against each other.
> If suddenly CSS became pleasant to use, Tailwind would be in a rough spot. See the irony?
Honestly, I don't. If people suddenly adopted a heathier lifestyle, doctors, at least dentists, would be in a rough spot.
See the irony? Well, again I don't.
The pain and opportunity will move elsewhere.
Tailwind is a handy tool that deserves support and sustainability
"The value got extracted, but compensation isn't flowing back. That bothers me, and it deserves a broader policy conversation.
What I keep coming back to is this: AI commoditizes anything you can fully specify. [...]
So where does value live now? In what requires showing up, not just specifying. Not what you can specify once, but what requires showing up again and again."
This seems like a useful framing to be aware of, generally.
The internet has always kinda run on the ambiguity of "does the value flow back". A quote liberated from this article itself; all the content that reporters produce that's laundered back out through twitter; 12ft.io; torrents; early youtube; late youtube; google news; apache/mit vs gnu licenses; et cetera..
> AI didn't kill Tailwind's business. It stress tested it.
The earthquake didn’t destroy the building — it stress tested it.
In software (and business in general!), innovation is expected. If you built a building in San Francisco that couldn't handle a relatively minor earthquake you could argue it would be a 'stress test'.
If you’re right that AI’s impact on business is akin to a ‘relatively minor earthquake in San Francisco,’ a lot of investors are going to be really fucking bummed out.
It's like the author doesn't understand what failing a stress test does.
And perhaps can’t define test.
I wish I could upvote this more than once. The author gets it, you have to sell outcomes. Not features. Seems like every open source company that doesn’t market an outcome to buyers will face a similar threat. And this particular go to market strategy was “brittle” before AI.
I see "hackers" in these comments are now advocating to make "criminal contempt of business model" a serious thing, instead of a mere meme used to describe draconian copyright and patent laws.
It's a reddit alternative hosted by a venture capitalist firm, the startup culture is much more prevalent here than the hacker culture that inspires the website's <title> tag.
also death penalty for touching oil rigs, via Bush I.. "stay in your lane" plus license plate readers
[dead]
[dead]
To call it a stress "test" is dismissive.
A stress test on a bank doesn't actually erase the revenue and financially jeopardize the bank.
Implementing layoffs is not a stress test.
Companies providing AI services should offer ads for the things the AI is using. And I don't mean "Tailwind could pay Google to advertise in Gemini", I mean "Google should be clearly and obviously linking back to Tailwind when it uses the library in its output"
They already do this sort of thing inside outputs from Deep Research, and possibly without. But the output should be less muted, inline, recessed and more "I have used Tailwind! Check out how to make it better [HERE](link)"
They should be working with the owners of these projects, especially ones with businesses, and directing traffic to them. Not everyone will go, and that's fine, but at least make the effort. The infrastructure is in place already.
And yes, right now this would not be across-the-board for every single library, but maybe it could be one day?
It's the same problem news sites have been facing for years because of Google News and Facebook. Their solution so far has been country-level bans on it (Canada).
But why would you click the tailwinds link if you're getting correct answers? You don't need documentation or consulting.
Google and Facebook couldn't provide full news articles because of copyright law. They just showed headline and summary provided by the news websites (and still eventually got sued for showing the summaries).
I feel that the battle is already lost when we aren't arguing that the AI service must get a license from the copyright holder like everybody else, but instead just arguing how many crumbs is the AI service morally obliged to throw back.
I was thinking something similar, but not so much an ad as a citation. A good starting point might be a law stating that when an LLM produces an answer, it cite its sources, with a link back to the content. Ideally, though, the producer of that content should receive some amount of financial compensation as well, similar to how an author or an actor receives royalties. If the LLM is making money off of this, so should the person who provide the LLM the value.
The number one rule of these apps is don’t link outside the app (because then the user will stop their session).
This feels like the OpenSSL problem where we do probably need some kind of industry organization to maintain these things. There’s a chicken and the egg problem that these AI companies need someone to keep maintaining tailwind if they want it to keep working in their prompts.
Maybe that limits the ability for the head of tailwind to run their own business and make more income, but something gotta give.
> we do probably need some kind of industry organization to maintain these things.
In the case of CSS, we already have that:
https://www.w3.org/Style/CSS/Overview.en.html
Are you saying the www consortium should be paying to keep Tailwind development and maintenance going? The css standard is not the same as a usable library of components.
Tailwind is open source. Anyone can contribute to it, including an LLM.
If the founder of Tailwind quits on it, others who deem Tailwind valuable enough will continue to maintain it.
> others who deem Tailwind valuable enough will continue to maintain it.
We have seen several examples in the last couple of years where this is simply not true enough. There are multiple open source projects that do not receive enough TLC.
Then it isn't important enough to people.
If my company relies on an open source project and it isn't being maintained, I can either ask my company to start maintaining it or find something else or accept the risk of a an unmaintained project.
You're just parroting what you've heard. The reality of continuing an abandoned project is not that simple.
The problem here is that the w3c sucks a fat one, and they've failed to build software specs that don't require an ecosystem thousands of libraries to make using CSS, etc. simple or efficient.
I think its a Kobyashi Maru - there is absolutely no way we will wield a paradigm shifting technology like language models and be able to do it safely. Language models create at scale and scope much greater than what can be reviewed by human… this means we very quickly move into trust without checks with language models. At some point, AI will be executing knowledge work in an instance, leaving no space for the human centric economic system that came before it. It will destroy our economic systems heavily based in knowledge and services (most advanced economies) and show us how what we do is mostly computable.
And when we are there, hungry, broken, unemployed and unimpressed with the prometheus who stole our purpose and replaced it with nothing….
it will make a colossal mistake that will destroy faith in it, as well as the institutions destroyed because of its oversite. The mistake will lead to bans, and regulatory bodies and auditing and certification processes… and will result in the final master stroke - it will make AI unusable by the non ordained, regulatory capture. Math will be a tool governments and institutions wield in secret against its citizenry… and it will be a tool of omnipotence that will lead to the enslavement, rebellion, and robot wars.
>"Value is shifting to operations: deployment, testing, rollbacks, observability. You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running"
I've been doing exactly that since AI came out :-D
You absolutely can prompt your way to 3.5 nines of uptime (even more), but you need to know what you're doing and correct it.
Even very well aligned models like Opus will make traps for your infrastructure. For example you tell it to write a fluxCD implementation of some application, in your k8 cluster, following your conventions and best practices described in some md files. And it does this, very nicely. But unless you tell it in advance every detail it will do something extremely stupid mid way through.
For example, let's say you have a database and it needs to create a new instance (with gitops) for the app. It adds the new DB and it gets created, but instead of using a tool that already exists (or proposing one if it doesn't) to sync the DB access credentials from the DB na espace to the app namespace it will read the credential, encrypt it and store in the app namespace.
What's the problem with that? Well, none unless you rotate these credentials. In which case your app will stop working, possibly after you tested it and decided it's good enough to use seriously, despite having a HA DB.
There are a dozen things like this in each "ai project", but even with this. With the time needed to review everything it saves a lot of time.
The author meant that you can't just tell a model "do everything that is necessary to achieve 99.95% uptime". It can certainly help you brainstorm issues and solve them, but you can't "just" prompt it.
Many FOSS business models did, explicitly or implicitly, rely, not on direct obfuscation or overcomplication, but on not making things easy. So you sell not the product, services around the product.
This is not exclusive to FOSS. It is also the basic model of most non SaaS B2B, often for good reason referred to by the derogatory term 'consulting-ware'.
AI eats into these services, as it commoditizes them. 80%+ of what used to take a specialist for that product can now be handled by a good generalist + AI.
Leaving aside the business model impact for a second, getting rid of obfuscation incentives is intrinsically good thing for a user community.
Dries' solution, offering operations as a SaaS or managed service, is meeting a need AI can't as easily match, but not exactly for the the stated reason. What the client is actually buying is making something someone else's problem. And CIOs will always love this more than anything if they can credibly justify it.
Where AI does impact this is in that latter part. If AI does significantly commoditize operational expertise, then the cost of in-house operating is (sometimes dramatically) lowered, and thus the justification gap on the CIO side for spending outside widens. How much this will drive a decision will be highly variable between businesses and projects.
i think whole business around software is to change (as i said in another thread few days ago).
if u imagine the "business" stack as: customers on top, over business, over analysis, over software, over machines/infrastructure.. 25+ years ago i thought that DSLs and such very-high-level-mostly-formalised-descriptions will move the line between what-product-is-aka-business-analysis and software/coding towards software, reducing its part in the whole stack.
Well, it did not happen, just the opposite - instead the developers become expected to know everything from infrastructure to software to analysis to business domain and higher. So, stack became something like customers over business over... software-dev. Requirements, analysis... mostly gone / done by devs.
Now if that whole "software-dev" part collapses into zero-margin, i see two ways: either businesses very quickly resurrect the business-analysis (what the product is) and make that their margin/moat, or the customers start making their own software (as throwaway 100 wrongs is now possible) - and removing the business from the whole transaction...
This goes way deeper than open source businesses.
Imagine I’m a company just big enough to entertain adopting Salesforce for CRM. It’s a big chunk of money, but our sales can absorb the pain.
With GenAI as an enterprise architect, one of the options I’m now recommending is to create a custom CRM for our business and skip the bloated enterprise SaaS platform.
We can gather CRM requirements fast, build fast, and deliver integrations to our other systems incredibly fast using tools like Claude Code. Our sales people can make feature requests and we can dogfood them in a few days, maybe hours.
GenAI development tools are rapidly changing how I think about enterprise software development.
If your core business is software-based offerings, your moat has been wiped out.
A handful of senior engineers can replicate any SaaS, save licensing costs, and build custom applications that were too risky in the past.
The companies that recognize what is happening and adapt will win.
It isn't that simple. One of the implicit tradeoffs you make buying SaaS is that the overall cost of evolution (development and ongoing maintenance) is subsidized across all of the investment resources and customer base of the vendor. With CRM in particular, ecosystem integration is one of the heaviest buildouts there is because each point solution integration can very significantly in complexity and is also where the combinatoric explosion of misbehavior sets in.
When you decide to pull that in house, you are implicitly burdening yourself with the cost of the buildout as well as ongoing maintenance. True, you could probably knock together an okay v1 of CRM yourself inhouse. But are you really going to get it to and maintain production level quality over time at a lower total cost of ownership? I'm skeptical.
The theoretical party you are describing would probably be better served by simply avoiding Salesforce in favor of a next gen CRM that is both more cost effective and easier to customize. In enterprise contexts, even HubSpot is effectively next gen, but there are also products like Attio et al that have a ton of adoption and strong integration ecosystems (albeit not at the Salesforce level).
When you buy enterprise software from a vendor you are buying more than "just software" you are also hiring a company's services. And the inverse is correct as well, when you choose to build it in house, you are implicitly choosing to hire a team internally to resource all of the services you would've expected that vendor to provide.
Certainly, this tradeoff can still make a lot of sense for some companies. The acid test for that, in my opinion, is whether said company could (and would) actually successfully sell the product they build internally on the open market. If the answer to that is "yes", the prospect of turning a cost center into a profit center can potentially bear significant long term ROI to the company.
100%. As an EA you always examine and explain the trade-offs to the business. In many cases that trade-off will remain and buy vs build leans towards buying.
My point is the decision point has moved. Where five years ago there’d be zero discussion of building internally, those discussions are going to be very different.
I believe many tech oriented companies will pull SaaS capabilities inside and as GenAI developer tools improve, the line will keep moving.
And we don’t need SaaS professional services anymore. Claude Code replaces that entire business model.
A view I have which feels controversial these days is that GenAI code tools are a net asset to code being held by entities that "should" hold them (ie workloads where code is a profit center not a cost center) and a net liability to the opposite. SaaS is a very squishy word and includes a lot things which are more accurate to call tech enabled services; Salesforce being one of the best examples (but you could say the same about many ERPs).
Maybe my biggest disagreement with your view is I think it is simultaneously far too conservative and far too aggressive at the same time. Where there was previously a decision about buy vs build, I disagree with the belief that many tech oriented companies will pull SaaS capabilities inside as GenAI developer tools improve because the cost of "build" is actually not going to get comparatively cheaper compared to the risk of "buying" - if anything, the risk of "building" internally with GenAI tooling that you were considering "buying" is significantly higher unless you are prepared to truly own it, that is go head to head with the entire rest of the market focuses on building and selling that tool as their full-time corporate focus. The actual risk of owning a tool internally making allowances for GenAI tooling is a lot higher than folks realize because GenAI tooling has a lot more risk of creating vibeslop and the only way to avoid that is to be dedicated to producing that tool as one's full time job and serving the needs of the entire market that needs it in that process. This is impossible to do with an internal tool.
The flip side of that same realization is why I also believe that your view is too conservative. The company that might be more empowered to consider building Salesforce internally with better tooling is not competing with Salesforce -- they are competing with the market that is going to /takedown/ Salesforce and a future version of themselves that would've used the future successor to Salesforce. Such a company is probably not in the business of building and selling CRMs although they are likely in the business of using such a CRM. The risk of making that conflation would be to stretch existing resources thin and turn the high interest credit card of tech debt and turn it into a payday loan with GenAI vibeslop.
I do not view GenAI as a democratizer. I view it as an accelerator with the capabilities to accelerate not just "winners take all" dynamics for the top companies that invest enough to avoid vibeslop, but "losers lose all" dynamics in the supply chain for everyone else who tries to wing it to capitalize before inevitably losing against future incumbents, or try in vain to turn a legacy firm into an innovator before inevitably losing against tech debt.
Everyone likes to believe they'll be able to use these great tools to become a future incumbent. But becoming a future incumbent is something that is very hard to do unless your org + book of business + funding + tools is better than someone else's org + book of business + funding + tools. That is why I don't think it actually changes the buy vs build decision that much; the decision should probably still be to buy, it's just changing the question of what exactly should be bought.
Excellent response. I think my position is that I don’t vibe code and most senior/deeply experienced dev/arch’s are doing something else.
We’re not coding. We’re orchestrating well constructed applications that follow proven principles (DDD, behavior focused, packaged business capabilities, behavior unit testing).
We’re extracting high value from GenAI.
>> Our sales people can make feature requests
I can tell you with near-100% certainty. This isn’t what you want to happen. Disaster in the making.
Just because you can doesn’t mean you should. There is very little competitive advantage to be gained from this type of effort in most companies.
What we "want" and what businesses will demand are very different things. I can tell you from 40 years build software, all companies care about is functional software. They don't care about code quality, maintainability, or tech debt. Never seen a single CTO say, "Let's carve out 20% of our sprints for tech debt," even though as architects we recommend something like that all the time.
The motto has always been, "Make it work."
Not, "Make it perfect."
I think we are both saying similar things here (believe it or not). Sales leaders turnover with surprising frequency - 18-24 months. About the time the sales team tells you what they “want” and you fine tune it, they will be gone. The next person will come in and probably scrap 50-75% of what the prior leader did. New requirements.
Meanwhile, besides functionality, you’ll want/need to plug in the latest and greatest go to market tools for marketing -and demand gen. But… that’ll be a custom effort, too.
Along the way you’ll also realize that you’re missing out on the most common practices in the industry because you built some idiosyncratic tool that only is relevant to your company.
History may very well prove me wrong, but I think you’re underestimating the expertise that underlies these products and platforms. It’s not just code, and the costs of getting it wrong are more than just an engineer’s time. When you waste time in GTM the impacts on the business and valuation are not linear, they’re exponential.
Agreed. We went from “internalise your core business and contract what doesn't give you an edge” to “actually just build everything in-house with AI”. Maybe that’ll be the better option in the end, but for now it looks like tons of wasted effort. 100x developers working on the wrong thing.
Yep agreed, totally changes build vs buy decisions. Which is not to say it's always "build" now, but the calculation has changed.
> So where does value live now? In what requires showing up, not just specifying. Not what you can specify once, but what requires showing up again and again.
Sounds like even more incentive for "managing" problems and creating business models around them instead of solving them.
The flurry machine isn't permanently broken, it's just keeping someone employed.
> The value got extracted, but compensation isn't flowing back. That bothers me, and it deserves a broader policy conversation.
It bothers me, too. But, look at the history of the internet. There's no reason to expect we'll be able to fix this problem.
1. Search engines drove traffic to news/content sites, which monetized via ads. Humans barely tolerate these ad filled websites. And yet, local news went into steep decline, and the big national players got an ever-larger share of attention. The large, national sites were able to keep a subscriber-based paywall model. These were largely legacy media sites (ie: NYT).
2. News sites lost the local classifieds market, as the cost of advertising online went to zero (ie: Craigslist). This dynamic was a form of creative destruction - a better solution ate the business of an older solution.
3. Blog monetization was always tough, beyond ads. Unless you were a big blog, you couldn't make a living. What about getting a small amount of money per view from random visitors? The internet never developed a micro-payment or subscription model for the set of open sites - the blogosphere, etc. The best we got were closed platforms like Substack and Medium, which could control access via paywalls.
All this led to the internet being largely funded through the "attention economy": ads mostly, paywalls & subscriptions some.
The attention economy can't sustain itself when there are fewer eyeballs:
1. Tailwind docs have to be added just once to the training set for the AI to be proficient in that framework forever. So one HTTP request, more or less, to get the docs and docs are no longer required.
2. Tailwind does change, so an AI will want to access the docs for the version its working with. This will require access at inference time. This is more analogous to visiting a site.
Killed by AI or competitors offering Tailwind templates and UI kits at a much lower price or for free?
To everything in this article that states what AI cannot do... I would like to add a big fat "YET!" and remind everyone to buckle up...
Right now its convenient to look at Tailwind and discuss what their doing wrong.
But eventually most other business models will be stress tested in the same way - sooner or later.
Tailwind plus is available for one time payment that provides lifetime access to current and future components. With AI cutting off the flow for new buyers, revenue shrivels up much quicker than what it would've been if it was a recurring subscription.
In my opinion, governments are going to have to tax the big tech companies hard and distribute the money to funding bodies like Arts Council. I can see a future "Tech Council" for open source software organisations to apply for funding. It'll get to the point where every OSS developer has their own Community Interest Company or join with a few other devs to create a CIO in order to acquire funding.
Of course, now you're opening a whole other can of worms. In the UK, only 1 in 9 Arts Council funding applications is successful.
Framing it as a "conduit" disruption might make a lot of assumptions about the fundamental economic value of software in the future. In a world (whether near term or long term) where you can just ask the computer to make whatever software you want, what are the economics of retailing/licensing any software at all? Open source or otherwise?
2-5 years from now after the AI bubble bursts, and they are trying to rent us $300 PCs since every component is 5x the price, we will look back at all the damage and copyright law that was completely bypassed and ignored when it was convenient, after all those years of claiming evil China "stole" from companies (only to then pass laws where they can virtually steal anything they want, even utilizing private repositories on Github that they acquired by buying the site, completely ignoring the licenses)...
Or how Meta downloaded 70tb+ of books and then got law enforcement to nuke libgen and z-lib to create a "moat", and all our tools start dying/disappearing because the developers are laid off since an AI "search engine" just regurgitates it, THEN and only then will most people understand the mistake that this was.
Let's not even begin with what Grok just recently did to women on X, completely unacceptable, I really, really wish for the EU to grow some and take a stand, it is clear that China is just as predatory as America and both are willing to burn it all in order to get a non existent lead in non existent "technology" that snake oil salesmen have convinced 80 year olds in government that is the next "revolution".
So AI is attempting to replace SAP as the traditional way of testing if you company is strong enough ?
> Value is shifting to operations: deployment, testing, rollbacks, observability. You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running.
I agree somewhat but eventually these can be automated with AI as well.
Unless you replace the entire workforce, you'd be surprised how much organizational work and soft skills are involved in an infrastructure at scale.
Like sure, there is a bunch of stuff like monitoring, alerting that is telling us that a database is filling up it's disk. This is already automated. It could also have automated remediation with tech from the 2000s with some simple rule-based systems (so you can understand why those misbehaved, instead of entirely opaque systems that just do whatever).
The thing is though, very often the problem isn't the disk filling up or fixing that.
The problem is rather figuring out what silly misbehavior the devs introduced, if a PM had a strange idea they did not validate, if this is backed by a business case and warrants more storage, if your upstream software has a bug, or whatever else. And then more stuff happens and you need to open support cases with your cloud provider because they just broke their API to resize disks, ...
And don't even get me started on trying to organize access management with a minimally organized project consulting team. Some ADFS config resulting from that is the trivial part.
If "99.95% uptime on Black Friday", and "keeping a site secure, updated, and running" can ever be automated (by which I mean not a toy site and not relying on sheer luck), not only 99.99% of people in IT are out of a job, but humans as intelligent beings are done. This is such a doomsday scenario that there's not even a point in discussing it.
How? I am tired of these unfounded claims. Humans can’t even keep many sites secure.
Care to provide a prompt that leads to coding agent achieving 99.95% uptime on Black Friday as an example?
Calling it a stress test seems a bit off. Would we say that invention of lightbulbs was a "stress test" for candle related business models? Or would we just say that business models had to change in response to current events.
Drop the "test". Just "stress" – it's cleaner.
I'd note a couple of things:
Not to nitpick but if we are going to discuss the impact of AI, then I'd argue "AI commoditizes anything you can specify." is not broad enough. My intuition is "AI commoditizes anything you can _evaluate/assess_." For software automation we need reasonably accurate specifications as input and we can more or less predict the output. We spend a lot of time managing the ambiguity on the input. With AI that is flipped.
In AI engineering you can move the ambiguity from input to the output. For problems where there is a clear and cheaper way of evaluating the output the trade-off of moving the ambiguity is worth it. Sometimes we have to reframe the problem as an optimization problem to make it work but same trade-off.
On the business model front: [I am not talking specifically about Tailwind here.] AI is simply amplifying systemic problems most businesses just didn't acknowledge for a while. SEO died the day Google decided to show answer snippets a decade ago. Google as a reliable channel died the day Google started Local Services Advertisement. Businesses that relied on those channels were already bleeding slowly; AI just made it sudden.
On efficiency front, most enterprises could have been so much more efficient if they could actually build internal products to manage their own organizational complexity. They just could not because money was cheap so ROI wasn't quite there and even if ROI was there most of them didn't know how to build a product for themselves. Just saying "AI first" is making ROI work, for now, so everyone is saying AI efficiency. My litmus test is fairly naive: if you are growing and you found AI efficiency then that's great (e.g. FB) but if you're not growing and only thing AI could do for you is "efficiency" then there is a fundamental problem no AI can fix.
The root of the issue is that Tailwind was selling something that people can now recreate a bespoke version of in mere minutes using a coding agent. The other day I vibe coded a bespoke dependabot/renovate replacement in an hour. That was way easier than learning any of these tools and fighting their idiosyncrasies that don’t work for me. We no longer need Framer because you can prompt a corporate website faster than you can learn Framer. It is, fortunately or unfortunately, what it is and we all have to adapt.
I want to be clear, it sucks for Tailwind for sure and the LLM providers essentially found a new loophole (training) where you can smash and grab public goods and capture the value without giving anything back. A lot of capitalists would say it’s a genius move.
Maybe they just over-hired for their business model.
Business & time are business model stress tests.
> You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running.
This is completely wrong. Agents will not just be able to write code, like they do now, but will also be able to handle operations, security, continuing to check, and improve the systems, tirelessly.
And someday we will have truly autonomous driving cars, we will cure cancer, and humans will visit Mars.
You can't prompt this today, are you suggesting this might come literally tomorrow? 10 years? 30? At that unknown time will your comment become relevant?
the quoted comment is arguing that devops will never be promptable — putting aside the discussion about whether or not that's true today, the argument here is that it's not likely to _never_ be possible
I'm working on a project now and what you're saying is already true. I have agents that are able to handle other things apart from code.
But these are MY agents. They are given access to MY domain knowledge in the way that I configured. They have rules as defined by ME over the course of multi-week research and decision making. And the interaction between my agents is also defined and enforced by me.
Can someone come up with a god-agent that will do all of this? Probably. Is it going to work in practice? Highly unlikely.
So you think a statement about the current state of things is wrong because you believe that sometime in the future agents are going to magically do everything? Great argument!
To be able to do this requires perfect domain knowledge AND environment knowledge AND be able to think deeply about logical dominoes (event propagation through the system, you know, the small stuff that crashes cloudflare for the entire planet for example).
Please wake me up when Shopify lets a bunch of agentic LLMs run their backends without human control and constant supervision.
The extreme here is thinking machines will do everything. The reality is likely far closer to less humans being needed.
They could build something like Lovable but with better design/frontend defaults.
I wonder how much impact shadcn had on their business.
> You can't prompt 99.95% uptime on Black Friday. Neither can you prompt your way to keeping a site secure, updated, and running.
Uh, yeah you can. There’s a whole DevOps ecosystem of software and cloud services (accessible via infrastructure—as-code) that your agents can use to do this. I don’t think businesses who specialize in ops are safe from downsizing.
Yep - exactly. Ops isn't immune to LLMs stealing your customers. Given that most of the "open source product with premium hosting" models are just reselling hyperscaler compute at a huge markup, the customers are going to realize pretty quickly that they can use an LLM to setup some basic devops and get the same uptime. Most of these companies are offering a middleman service that becomes a bad deal the moment the customer has access to expertise they previously lacked.
I also think he's glossing over the fact that one of the reasons why companies choose to pay for "ops" to run their software for them is because it's built by amateurs or amateurs-playing-professional and runs like shit. I happen to know this first hand from years of working at a company selling hosting and ops for the exact same CMS that Dries' business hosts (Drupal, a PHP-based CMS) and the absolute garbage that some people are able to put together in frameworks like Wordpress and Drupal is truly astounding. I'm not even talking about the janky local businesses where their nephew who was handy with computers made them a Wordpress site - big multinational companies have sites in these frameworks that can barely handle 1x their normal traffic and more or less explode at 1.5x.
The business of hosting these customers' poorly optimized garbage remains a big business. But we're entering into an era where the people who produce poorly optimized software have a different path to take rather than throwing it to a SaaS platform that can through sheer force of will make their lead-weight airplane fly. They can spend orders of magnitude less money to pay an LLM to make the software actually just not run like shit in the first place. Throwing scaling at the problem of 99.95% is a blunt instrument that only works if the person paying doesn't have the time, money, or knowledge to do it themselves.
Companies like these (including the one I work for currently) are absolutely going to get squeezed from both directions. The ceiling is coming down as more realize they can do their own devops, and the floor is rising as customer code quality gets better. Eventually you have to try your best to be 3 ft tall instead of 6.
One of the biggest shortcomings of Open Source was that it implicitly defaulted to a volunteer model and so financing the work was always left as an exercise for the reader.
Hence (as TFA points out) open source code from commercial entities was just a marketing channel and source of free labor... err, community contributions... to auxiliary offerings that actually made money. This basic economic drive is totally natural but creates dynamics that lead to suboptimal behaviors and controversy multiple times.
For instance, a favorite business model is charging for support. Another one was charging for a convenient packaging or hosting of an “open core” project. In either case, the incentives just didn’t align towards making the software bug-free and easily usable, because that would actively hamper monetization. This led to instances of pathological behavior, like Red Hat futzing with its patches or pay-walling its source code to hamper other Linux vendors.
Then there were cases where the "open source" branding was used to get market-share, but licenses restricted usage in lucrative applications, like Sun with Java. But worse, often a bigger fish swooped in to take the code, as they were legally allowed to, and repackage it in their own products undercutting the original owners. E.g. Google worked around Sun's licensing restrictions to use Java completely for free in Android. And then ironically Android itself was marketed as "open source" while its licensing came with its own extremely onerous restrictions to prevent true competition.
Or all those cases when hyperscalers undercut the original owners’ offerings by providing open source projects as proprietary Software as a Service.
All this in turn led to all sorts of controversies like lawsuits or companies rug-pulling its community with a license change.
And aside from all that, the same pressures regularly led to the “enshittification” of software.
Open Source is largely a socialist (or even communist) movement, but businesses exist in a fundamentally capitalistic society. The tensions between those philosophies were inevitable. Socialists gonna socialize, but capitalists gonna capitalize.
With AI, current OSS business models may soon be dead. And personally I would think, to the extent they were based on misaligned incentives or unhealthy dynamics, good riddance!
Open Source itself will not go away, but it will enter a new era. The cost of code has dropped so much, monetizing will be hard. But by the same token, it will encourage people, having invested so much fewer resources creating it, to release their code for free. A lot of it will be slop, but the quantity will be overwhelming.
It’s not clear how this era will pan out, but interesting times ahead.
[dead]
[dead]
I know for a fact that all SOTA models have linux source code in them, intentionally or not which means that they should follow the GPL license terms and open-source part of the models which have created derivative works out of it.
yes, this is indirectly hinting that during training the GPL tainted code touches every single floating point value in a model making it derivative work - even the tokenizer isn't immune to this.
> the tokenizer isn't immune to this
A tokenizer's set of tokens isn't copyrightable in the first place, so it can't really be a derivative work of anything.
GPL however, does put restrictions on it, even the tokenizer. It was specifically crafted in a way where even if you do not have any GPL licensed sourcecode in your project, but it was built on top of it you are still binded by GPL limitations.
the only reason usermode is not affected is because they have an exclusion for it and only via defined communication protocol, if you go around it or attempt to put a workaround in the kernel guess what: it still violates the license - point is: it is very restrictive.
> GPL however, does put restrictions on it, even the tokenizer. It was specifically crafted in a way where even if you do not have any GPL licensed sourcecode in your project, but it was built on top of it you are still binded by GPL limitations.
This is not how copyright law works. The GPL is a copyright license, as stated by the FSF. Something which is not subject to copyright cannot be subject to a copyright license.
GPL is not only a copyright license, it also covers multiple types of intellectual property rights. Especially when you consider GPL-3 which has explicit IP protection while GPL-2 is implicit, so yah you're partially right for GPL-2 and wrong for GPL-3.
It's true that GPLv3 covers patents, but it is still primarily a copyright license.
The tokenizer's tokens aren't patented, for sure. They can't be trademarked (they don't identify a product or service). They aren't a trade secret (the data is public). They aren't copyrighted (not a creative work). And the GPL explicitly preserves fair use rights, so there are no contractual restrictions either.
A tokenizer is effectively a list of the top-n most common byte sequences. There's simply no basis in law for it to be subject to copyright or any other IP law in the average situation.
I mean okay sure, there is no legal framework for tokenizers, but what about the rest of the model I think there is a much stronger argument there? And you could realistically extend the logic that if the model is GPL-2.0 licensed you have to provide all the tools to replicate it which would include the tokenizer.
When you say “in” them, are you referring to their training data, or their model weights, or the infrastructure required to run them?
GPL can be considered like a virus, something based on GPL licensed code (unless explicitely excluded by the license) is now GPL licensed so the 'injected' training data becomes GPL licensed which means that created model weights from them in theory should also become GPL licensed.
>Open Source was never the commercial product. It's the conduit to something else.
this is correct. If you open source your software, then why are you mad when companies like AWS, OpenAI, etc. make tons of money?
Open Source software is always a bridge that leads to something else to commercialize on. If you want to sell software, then pick Microsoft's model and sell your software as closed source. If you get mad and cry about making money to sustain your open source project, then pick the right license for your business.
> then pick the right license for your business
That's one of the issues with AI, though; strongly copylefted software suddenly finds itself unable to enforce its license because "AI" gets a free pass on copyright for some reason.
Dual-licensing open source with business-unfriendly licensing used to be a pretty good way to sell software, but thanks to the absurd legal position AI models have managed to squeeze themselves into, that stopped in an instant.
Open source software helped to dramatically reduce the cost of paid software, because there is a now a minimum bar of functionality you have to produce in order to sell software.
And, in many cases, you had to produce that value yourself. GPL licensing lawsuits ensured this.
AI extracting value from software in such a way that the creators no longer can take the small scraps they were willing to live on seems likely to change this dynamic.
I expect no-source-available software (including shareware) to proliferate again, to the detriment of open source.