This is what happens when you give people tools that let them achieve an outcome, without necessarily giving them the judgement or expertise to know whether the outcome is any good.
If you asked me to build a house, I could probably assemble something that would stand for a few months. Hopefully. It might even keep the rain out. But it might also fall on my head, because I do not know enough about building houses to be confident that it won’t.
And even if it didn’t fall on my head under normal conditions, I also would not know when I needed to design for earthquakes. Or floods. Or fire. Or wind. Or grandmother-cosplaying wolves with very strong lungs.
But if all I need is shelter for a day, would I necessarily care whether it lasts more than a week?
That is effectively what a website like this is. It is not really a product. People don’t depend on it. Tan’s visitors are probably using MacBooks and iPhones on fast networks, and most of them will never notice how bad it is under the surface.
That does not mean it is good. It means it is good enough for the context.
AI feels to me like having access to someone who got a D in literally every single course offered at a university. If you don't know anything about the subject they are smarter than you. If you do know the subject its unsettling how bad they are. Basically the Gell-Mann effect:
> The phenomenon of a person trusting newspapers for topics which that person is not knowledgeable about, despite recognizing the newspaper as being extremely inaccurate on certain topics which that person is knowledgeable about.
They've improved from someone who failed every single university course a couple years ago. Maybe they'll get to a C or even a B in the future; maybe not.
We're trying to up the GPA on the code quality exam by giving them access to structural metrics calculated on graph representations of the programs they write. Hoping this is the study guide template they need to start getting an A in minimizing slop, even if they're actually a D student. They're good at optimizing against scores...
>AI feels to me like having access to someone who got a D in literally every single course offered at a university. If you don't know anything about the subject they are smarter than you. If you do know the subject, its unsettling how bad they are. Basically the Gell-Mann effect:
I don't know how you can say that with a straight face when AI's capable of matching the best human students on the hardest exams we have, like the IMO, the Putnam and the bar exam. I can only assume you've only ever used the free tier of any AI service.
Exams are usually discussed and part of the corpus though right? Performance on novel problems are really the only good metric and those decay really quickly across models once they’re known, like the pelican on a bicycle test
I like that you start by saying that you don’t know how someone could draw that conclusion about AI and end by saying they probably got there by using AI
There are no studies to suggest that "Gell-Mann" is a real phenomenon. It was invented by an author, Michael Crichton, on a whim.
Any rational reader will adjust their faith in a publication based on the identified errors it makes. Too many and they'll reject it as a source for anything beyond "a thing may have happened".
I think this is where aggregate effects have to be considered. One person building an idiosyncratic house out of found materials: neat little project. One million people doing so: shanty town that can be seen from orbit and is a disaster waiting to happen.
The Web already had a problem with externalizing costs onto users. Both the simple cost of poorly executing websites (power, mobile data, time), and more subtle ones (social media). AI is a huge accelerant for that.
Correct. It's "good enough" for now and if it needs to be industrialized you can do that later. If it is brand new and has no traffic and is not exposing user data or opening up your keys to someone.... who cares how perfect the code is if it mostly works? Users don't have to use it...
That is, until you have silent corruption of data because the admin tool is not aware of the new field you added to the website, and it's not written with forward compatibility in mind. (Have seen this lots of times)
Or a downstream service is overwhelmed and suddenly all the retired you added to different places DDOS your own service. (Also have seen this lots of times)
Some quality problems you can fix later. With others, once they happen, there's no "later"
In the thread the developer is only able to look at the front-end code. The real problem is more likely in the back-end code where there'll be hilariously bad security issues. Hopefully Tan isn't running this on the same server or network as anything important.
... and engineers don't have to consider this a good look coming from the head of YC. These are rookie mistakes - generated by both the human and AI model in a collaborative process. It's also the kind of mistake most of us aren't willing to make b/c we don't have the capital to sweep it under the rug if it goes pear-shaped.
It's not really a question about if GStack is 'good enough' - it's the claims about 37K LOC being legitimate.
Because it's producing 'functional slop' that has little bearing on how anything else would be made the 37K LOC is totally out of context and therefore a misrepresentation.
He has bragging rights over GStack - not 37K LOC, that's bad.
That's a great way to put it: functional slop. That's what I mean tho - as long as it is minimally functional it does its job.
I totally get the shade thrown at ycombinator. But I guess my perception is that it has never been about the tech for them and more about getting quick hacks to get to market quickly. So it feels on brand?
>>if all I need is shelter for a day, would I necessarily care whether it lasts more than a week
An interesting choice of a metaphor here. If I needed a shelter for a day, I would have bought (or rented) a nice tent. Or booked a room in a proper hotel
I recently did a feasibility test for porting complex c++ project to a stripped down DSP processor with no c or c++ stdlib without making changes to the actual project - sonnet was able to do it in half a day and stumbled for another half a day on a minor error until I pointed out its mistake, but the code was absolutely bad. Just the same, I would’ve taken two weeks to just know it was feasible and I have a reference to know how it avoided certain pitfalls. It’s taken about a month for two of us to do it right, but honestly I don’t think we would’ve attempted it without proof it was possible. I still wouldn’t put any AI code in production for actual end users (beyond the “shed built for a day” use you mention), but that doesn’t mean as a tool it’s not magical. It’s a false binary between the “it’s always amazing” viewpoint of Garry Tan and many others and the “it’s all slop” of the original author, it’s a tool that has specific uses and high costs right now. Its irresponsible to blindly assume it’ll continue to get better, but that doesn’t mean there aren’t valid use cases today
> give people tools that let them achieve an outcome, without necessarily giving them the judgement or expertise to know whether the outcome is any good
This could be fine if all you're looking is to get the quick and dirty result at any cost, or private use, etc. When anything is better than nothing.
The problem starts when this extremely low bar becomes the baseline for anything. When you're willing to attach your name to a stream of absolute slop and you're even proud of it.
Measuring success in LoC is so wrong… It's like bragging that you took 10,000 steps to reach the store by walking in circles when you could have taken 500 steps if you just walked straight. The end result is the same.
AI has a tendency to generate more code than necessary. It keeps re-inventing things, and every time you ask it to add a new feature or fix something, it just keeps on piling the code. I now periodically ask AI to refactor the code by simplifying, removing unused things, factoring out, and reusing.
> It's like bragging that you took 10,000 steps to reach the store by walking in circles when you could have taken 500 steps if you just walked straight.
My wife does exactly that. For the exercise. She makes large detours to go anywhere. The end result is a healthier body.
How this translates to software, I don't know. I don't think AI benefits from this exercise.
"found numerous examples of bloat and inefficiencies in Tan’s site code, and used a single (Anthropic) Claude session to review the files he downloaded from the website to confirm his observations"
1. I hope they never get hold of the code of MS Office or almost any other piece of real-world business software.
2. So anyone with claude access could arrive at the same conclusions ... and ask claude to fix it?
Personally, I've used my newfound productivity to build higher quality software at the same rate as I produced mediocre quality software before. Some people, like Gary Tan for example, seem to instead produce software at the same quality or worse than before, but just produce a ton more software regardless of quality.
I guess the article in reality is just these two perspectives pitted against each other for some cheap views.
That was the strangest line for me as well. The article tries to play up this "Polish engineer with an MSc in computer science who has 13 years of industry experience" who ... used claude to review the code. My mom could have done this without an MSc or years of industry experience.
Note that Rails was built as a framework for making blogs, I'm having trouble understanding what 78,000 lines of ruby in the context of a Rails blog could ... do.
I'm sure there's some powerful ugly stuff in Office but in a good code that's calcified kind of way. It got that way over like 30 years of releasing to the public across platforms, not over a weekend.
I'd be surprised if microsoft.com is shipping their entire test suite unminified and their back-end posting rich text editor with index.html (with two title tags in the head) and rendering the entire DOM for desktop and mobile regardless of your platform.
I'm not critiquing Garry or the site. I think it's great people are using AI to build things that bring them joy, or that they find useful. I certainly do.
I am opposed to the idea that we've decided to go back to measuring work in terms of lines of code. It has always been the worst metric on earth as a proxy for productivity. Every line is a liability, and it always was. AI has not changed that, if anything it's amplifying it.
The best PRs remove code, not add, and the only companies that seem to have exponentially grown their revenues in line with AI-generated LOC are OpenAI and Anthropic. Everyone else seems to be rummaging around for an ROI.
> Note that Rails was built as a framework for making blogs
This doesn't seem right. Yes, the "build a blog in 15 minutes" demo was pretty memorable, but I don't think it's ever really been pitched as a sweet spot for Rails. IIRC Rails was essentially extracted out of the 37Signals codebases for Basecamp et al.
> I hope they never get hold of the code of MS Office or almost any other piece of real-world business software.
Except he's not building Office, a software with decades of legacy, used by hundreds of millions of users. He's coding a website, effectively writing bloat with a silver lining of useful features. AI automated and inflated the worst of practices too. Anyone outputting 37K LoC daily is creating bloat and inefficiencies at unprecedented rate.
And enterprise software was already the butt of all jokes in this regard. We found ways to make that worse at scale even for the basic things. It's not a good look when you need to use this "whatabout".
> So anyone with claude access could arrive at the same conclusions ... and ask claude to fix it?
Begs the question of why the original "author" of the code hasn't just asked Claude to fix it? Or asked Claude to generate good code from the beginning. I suppose the answer is that nobody cares about good or efficient code anymore. But that's been the case since long before LLM coding though (as stated in your point 1).
Why write clever code when we can just write JS slop and ask customers "upgrade" their hardware every year...
Hmm the list is a bit underwhelming. Basically, it's unnecessary requests, bloated JS, unoptimized images and generally poorly structured code. I would hate if that was where the average website is headed, but realistically, we were already headed there before LLMs. From the headline I was expecting CVEs, broken UX flows / business logic, leaked secrets.
It's good that no leaked keys were found in the front end code but the developer wasn't able to look at the backend where, if the quality of the front end is any indication, there are likely to be many security issues. Hopefully it's not all running on the same servers/network as anything important.
I think we're on the same team, I find the general attitude towards security in the current AI scene scary. I was just hoping for a bit more ammunition than what that article gave us.
Whilst I don't _really_ consider 37k LoC a day to be a particularly extreme number, my issue with all the high profile high output usage of AI, is that it doesn't really prove much of value.
I use AI in a pretty single-threaded way and my primary challenge is figuring out how to keep the LoC as minimal as possible, as well as minimise my spend. I am, unintentionally, one of the highest spenders at work; this is possible because I do more fiddly UI enhancements where the desired behaviour is often subjective and invariably hard to articulate. When I work on big backend features, my spend tends to be much lower.
What I really want from the people who have effectively unlimited tokens to spend, is to use that finding ways for the rest of us to produce higher quality output at lower costs, rather than focusing on output alone.
Garry Tan's point still stands: he never pretended to be building nice software. But his point was that he can now build AT ALL! Shipping a webpage at all is the firs step ; making it load under 7 Mo is just a refinement, an important one of course (who tf wants bloated webpages) but still only a refinement. Tan is right to be amazed and to be shipping.
Bloat is not "since generative AI coding" new. We always had it. As the author says:
> “It does sound like Facebook’s ‘move fast and break things,’ which didn’t age well either.”
Which indeed did not age well, but it did help the company to grow to a certain point at which it is now a staple in our lives (and can do super expensive BS experiments like "Metaverse" and still show profits).
This may be what Tan does as well: first profitable, then correct.
This is an approach, which may work for some, it may also allow some companies to become irrelevant (like: once the bloat-app is profitable, a clone-app emerges that is bloat-free and overtakes the bloat-app in every dimension, while the bloat-app is figuring out how to scale up with a shitty db schema).
Seems to be a frequently occurring issue, most (good) software engineers know that LOC basically means nothing, if anything, less LOC is a better goal over more LOC, if you absolutely have to have a goal. As soon as people/companies starts bragging about how many LOC they can ship, you need to start being very suspicious, mostly because they just admitted to not actually understand software engineering at all.
What I don't understand, isn't there a single engineer working with these people who ask them what the fuck they're doing, before they hit that publish button? Or is there just such a constant pressure to publish anything that quality just simply doesn't matter at all to these people?
> 169 server requests for various assets totaling 6.42 megabytes in size.
Ahh … remember the good ol’ days, when we were forced to cap our Web pages at 32K?
I have found similar issues with LLM-generated code. Lots of code … lots of issues.
Just yesterday, I went through a whole bunch of LLM-generated code, and switched the threading model. It sped up my code significantly, and the jetsam crashes stopped immediately. Until I used LLM code, my apps had never jetsam-crashed before.
And with that 37k a day, or 185k per 5-day work-week, what exactly has he achieved?
I don't see any actual output coming from these AI tools, despite how many are saying it's greatly increased their productivity. Where are all the new businesses and tools? I only see more shovels being sold.
> The larger point is that while AI coding tools make it easy to pump out lots of code, it’s really (still) the quality of the code that matters. Quantity, in other words, doesn’t necessarily equal quality.
They didn't provide any evidence for this point. They showed that the code is bad quality, but not that it matters that it's bad.
My hot take is that the entire history of the personal internet was inefficient, badly built and chaotic and it was much more fun to exist in.
I really don't see a problem if it breathes life into the internet for non-technical people to be creative and interesting again. Wordpress etc gave people a cheap voice, but at the cost of immovable guard-rails.
And even at a small business level, most stuff people build get's torn down by the next person or next agency 6, 12, 18 months later anyway so why care about engineering rigour (this point does not extend to security concerns) in the front end when you can spin up and tear down something in a day or two for incredibly cheap.
I'm honestly tired at the pearl-clutching when we ourselves have helped shape an industry that constantly raises the bar to entry with unnecessary techincal complexity and bloat. Ever built a brochureware site in react or similar? you've done exactly what the this guy did, you just did it with pure intent rather than an AI doing it
> The larger point is that while AI coding tools make it easy to pump out lots of code, it’s really (still) the quality of the code that matters
This conclusion is unsupported from the observations. The code makes lots of requests, has too much CSS, and 6 different logo formats. So what? Lots of real, production codebases have just as many warts.
Folks need to stop dealing in absolutes with AI coding. Code quality always mattered, and still does, in certain circumstances. In others, it's less important and speed of iteration has more value. That's still true even with AI code.
> In others, it's less important and speed of iteration has more value.
Maybe in some sort of topsy-turvy techbro world.
I've never, in three decades of developing, had a user say "what I want are new features updated frequently and I'm happy to accept sluggish, buggy code".
What they do say is "I want more features, sooner, without bugs".
I'd guess you're maybe working on more established codebases then? I see this regularly with my clients. I work in sprints and they pay for my time, so they will often deprioritise engineering/robustness work if it's taking up time that could be used some shiny new feature they want.
Also it's rarely "I want new features and I'm happy for them to be buggy or sluggish", it's more "We want to solve this problem and we're not exactly sure what the solution will look like, so lets try stuff and accept it's not going to be perfect"
Ironically the fastcompany page hosting the article that should only be read only paragraphs of text and a simple navigation seems to do 300+ requests, with 29.4mb of resources.
I mean honestly the numbers and issues talked about there seem to be on most modern web apps. It just doesn't seem to affect business results enough for people to care about fixing them.
It's 100% fine for Gary to write software in a way that works for him - and for us to push boundaries, and not 'gatekeep'.
BUT - it's still a pile of steaming garbage.
It's not just a 'bit odd' - it's just massive slop.
The total lack of self awareness is comical and disturbing.
GStack + Gary's Tweets about 30K LOC a day is 'Trumpian' in self delusion - it looks like basically he and YC are frauds that don't know what technology is.
This is like Elon and his 'salute' or his 'I'm a top 10 Diablo player - hey watch me play and expose myself unwittingly' type situations.
A bit of AI hype is fine.
Someone needs to take Gary and have a side-discussion.
He can do GStack - he just needs to characterize it properl for what it is.
Talking about that amount of code with the assumption 'it's sensible' code, is basically a lie - it's fraudulent' levels of hype.
Just describe it as an unwieldy but productive experiment, not for mainline consumption etc. then it's fine.
"Tan/AI built the website so that when a user visits, their browser makes 169 server requests for various assets totaling 6.42 megabytes in size. For comparison, the minimalist Hacker News homepage (also run by Y Combinator) makes seven requests for data totaling just 12 kilobytes.
The website ships 28 actual test files (code developers use to reality-check their work) straight to every visitor’s browser. That’s 300 kilobytes of pure developer scaffolding that users never asked for.
It loads 78 different JavaScript controllers for features like AI image generation, voice extraction, video tools, etc., none of which appear on the homepage. The browser still has to download all of them “just in case.”
The site’s logo is an illustration of a bear. The site downloads the logo in eight different formats, including a completely empty 0-byte file that somehow made it to production, Gregorein found.
The website uses huge, uncompressed old-school PNGs (some nearly 2 megabytes each), even though the browser literally asks for modern tiny formats. Two images alone waste about 4 MB; with newer formats they could have been just 300KB.
Gregorein also found duplicate page content, an empty CSS (Cascading Style Sheets) file, a huge rich-text editor loaded on a read-only page, missing image descriptions, and analytics code that deliberately routes through a proxy to dodge people’s ad blockers (with a comment in the code admitting it), Gregorein reports,"
Brave new world indeed... It might be elliptically relevant that in the original novel by Aldous Huxley, the protagonist John the Savage hangs himself at the end when his search for truth fails.
Impressive. Now, let's see the link in this post to fastcompany.com; It opens a website that also pulls in 6MB using more "server requests" than 169, in fact the number keeps going up, and all it does is display a paywall. So, let's not pretend this is something unique.
Do you remember when you had to write essays in school and the word count mattered?
That’s when you’d write "it is" instead of "it's" and "will not" instead of "won't."
This is worse.
Not the first time that I read quantity over quality related to YC
This is what happens when you give people tools that let them achieve an outcome, without necessarily giving them the judgement or expertise to know whether the outcome is any good.
If you asked me to build a house, I could probably assemble something that would stand for a few months. Hopefully. It might even keep the rain out. But it might also fall on my head, because I do not know enough about building houses to be confident that it won’t.
And even if it didn’t fall on my head under normal conditions, I also would not know when I needed to design for earthquakes. Or floods. Or fire. Or wind. Or grandmother-cosplaying wolves with very strong lungs.
But if all I need is shelter for a day, would I necessarily care whether it lasts more than a week?
That is effectively what a website like this is. It is not really a product. People don’t depend on it. Tan’s visitors are probably using MacBooks and iPhones on fast networks, and most of them will never notice how bad it is under the surface.
That does not mean it is good. It means it is good enough for the context.
Most people also tolerated the hilarious gigabyte JSON parsing bug in Grand Theft Auto for years, until a hacker patched it and cut GTA Online loading times by around 70%: https://nee.lv/2021/02/28/How-I-cut-GTA-Online-loading-times...
It was good enough, even if people noticed how bad it was.
Business applications, and typical software really doesn’t have to be super tuned or perform fast. It just needs to work.
At least until your product category has been commoditized, and _then_ you’re competing on experience.
AI feels to me like having access to someone who got a D in literally every single course offered at a university. If you don't know anything about the subject they are smarter than you. If you do know the subject its unsettling how bad they are. Basically the Gell-Mann effect:
> The phenomenon of a person trusting newspapers for topics which that person is not knowledgeable about, despite recognizing the newspaper as being extremely inaccurate on certain topics which that person is knowledgeable about.
They've improved from someone who failed every single university course a couple years ago. Maybe they'll get to a C or even a B in the future; maybe not.
We're trying to up the GPA on the code quality exam by giving them access to structural metrics calculated on graph representations of the programs they write. Hoping this is the study guide template they need to start getting an A in minimizing slop, even if they're actually a D student. They're good at optimizing against scores...
https://github.com/Krv-Labs/topos
>AI feels to me like having access to someone who got a D in literally every single course offered at a university. If you don't know anything about the subject they are smarter than you. If you do know the subject, its unsettling how bad they are. Basically the Gell-Mann effect:
I don't know how you can say that with a straight face when AI's capable of matching the best human students on the hardest exams we have, like the IMO, the Putnam and the bar exam. I can only assume you've only ever used the free tier of any AI service.
Exams are usually discussed and part of the corpus though right? Performance on novel problems are really the only good metric and those decay really quickly across models once they’re known, like the pelican on a bicycle test
I like that you start by saying that you don’t know how someone could draw that conclusion about AI and end by saying they probably got there by using AI
There are no studies to suggest that "Gell-Mann" is a real phenomenon. It was invented by an author, Michael Crichton, on a whim.
Any rational reader will adjust their faith in a publication based on the identified errors it makes. Too many and they'll reject it as a source for anything beyond "a thing may have happened".
I thought you generally knew what you were talking about till i read this. Now I am wondering if you know anything at all.
I think this is where aggregate effects have to be considered. One person building an idiosyncratic house out of found materials: neat little project. One million people doing so: shanty town that can be seen from orbit and is a disaster waiting to happen.
The Web already had a problem with externalizing costs onto users. Both the simple cost of poorly executing websites (power, mobile data, time), and more subtle ones (social media). AI is a huge accelerant for that.
Correct. It's "good enough" for now and if it needs to be industrialized you can do that later. If it is brand new and has no traffic and is not exposing user data or opening up your keys to someone.... who cares how perfect the code is if it mostly works? Users don't have to use it...
That is, until you have silent corruption of data because the admin tool is not aware of the new field you added to the website, and it's not written with forward compatibility in mind. (Have seen this lots of times)
Or a downstream service is overwhelmed and suddenly all the retired you added to different places DDOS your own service. (Also have seen this lots of times)
Some quality problems you can fix later. With others, once they happen, there's no "later"
In the thread the developer is only able to look at the front-end code. The real problem is more likely in the back-end code where there'll be hilariously bad security issues. Hopefully Tan isn't running this on the same server or network as anything important.
... and engineers don't have to consider this a good look coming from the head of YC. These are rookie mistakes - generated by both the human and AI model in a collaborative process. It's also the kind of mistake most of us aren't willing to make b/c we don't have the capital to sweep it under the rug if it goes pear-shaped.
It's not really a question about if GStack is 'good enough' - it's the claims about 37K LOC being legitimate.
Because it's producing 'functional slop' that has little bearing on how anything else would be made the 37K LOC is totally out of context and therefore a misrepresentation.
He has bragging rights over GStack - not 37K LOC, that's bad.
That's a great way to put it: functional slop. That's what I mean tho - as long as it is minimally functional it does its job.
I totally get the shade thrown at ycombinator. But I guess my perception is that it has never been about the tech for them and more about getting quick hacks to get to market quickly. So it feels on brand?
>>if all I need is shelter for a day, would I necessarily care whether it lasts more than a week
An interesting choice of a metaphor here. If I needed a shelter for a day, I would have bought (or rented) a nice tent. Or booked a room in a proper hotel
I recently did a feasibility test for porting complex c++ project to a stripped down DSP processor with no c or c++ stdlib without making changes to the actual project - sonnet was able to do it in half a day and stumbled for another half a day on a minor error until I pointed out its mistake, but the code was absolutely bad. Just the same, I would’ve taken two weeks to just know it was feasible and I have a reference to know how it avoided certain pitfalls. It’s taken about a month for two of us to do it right, but honestly I don’t think we would’ve attempted it without proof it was possible. I still wouldn’t put any AI code in production for actual end users (beyond the “shed built for a day” use you mention), but that doesn’t mean as a tool it’s not magical. It’s a false binary between the “it’s always amazing” viewpoint of Garry Tan and many others and the “it’s all slop” of the original author, it’s a tool that has specific uses and high costs right now. Its irresponsible to blindly assume it’ll continue to get better, but that doesn’t mean there aren’t valid use cases today
> give people tools that let them achieve an outcome, without necessarily giving them the judgement or expertise to know whether the outcome is any good
This could be fine if all you're looking is to get the quick and dirty result at any cost, or private use, etc. When anything is better than nothing.
The problem starts when this extremely low bar becomes the baseline for anything. When you're willing to attach your name to a stream of absolute slop and you're even proud of it.
“…without necessarily giving them the judgement or expertise to know whether the outcome is any good.”
Ahem…this man is A CHIEF EXECUTIVE OFFICER how DARE you cast aspersions on his judgment abilities! /s
The biggest takeaway for me is “failing upward” is much easier in the 21st Century.
Measuring success in LoC is so wrong… It's like bragging that you took 10,000 steps to reach the store by walking in circles when you could have taken 500 steps if you just walked straight. The end result is the same.
AI has a tendency to generate more code than necessary. It keeps re-inventing things, and every time you ask it to add a new feature or fix something, it just keeps on piling the code. I now periodically ask AI to refactor the code by simplifying, removing unused things, factoring out, and reusing.
> It's like bragging that you took 10,000 steps to reach the store by walking in circles when you could have taken 500 steps if you just walked straight.
My wife does exactly that. For the exercise. She makes large detours to go anywhere. The end result is a healthier body.
How this translates to software, I don't know. I don't think AI benefits from this exercise.
People like using the "AI is like bringing a forklift to the gym" metaphor. I suppose for the shopping one it's using a SUV to make the 500m trip.
Using an SUV to make a 500m trip and then bragging about how you now drive 37 kilometers per day.
"found numerous examples of bloat and inefficiencies in Tan’s site code, and used a single (Anthropic) Claude session to review the files he downloaded from the website to confirm his observations"
1. I hope they never get hold of the code of MS Office or almost any other piece of real-world business software.
2. So anyone with claude access could arrive at the same conclusions ... and ask claude to fix it?
Personally, I've used my newfound productivity to build higher quality software at the same rate as I produced mediocre quality software before. Some people, like Gary Tan for example, seem to instead produce software at the same quality or worse than before, but just produce a ton more software regardless of quality.
I guess the article in reality is just these two perspectives pitted against each other for some cheap views.
Yes but for that you need to focus on the quality. There are probably horrible issues on backend as well.
It's worse than anything in MS really.
That was the strangest line for me as well. The article tries to play up this "Polish engineer with an MSc in computer science who has 13 years of industry experience" who ... used claude to review the code. My mom could have done this without an MSc or years of industry experience.
Your mom couldn’t distinguish Claude‘s BS analysis and hallucinations from real bloat and errors.
You need experience to check Claude‘s findings
I'd suggest looking at the review itself, there's an X-the-everything-app thread on it.
https://x.com/Gregorein/status/2038953944475472316
Note that Rails was built as a framework for making blogs, I'm having trouble understanding what 78,000 lines of ruby in the context of a Rails blog could ... do.
I'm sure there's some powerful ugly stuff in Office but in a good code that's calcified kind of way. It got that way over like 30 years of releasing to the public across platforms, not over a weekend.
I'd be surprised if microsoft.com is shipping their entire test suite unminified and their back-end posting rich text editor with index.html (with two title tags in the head) and rendering the entire DOM for desktop and mobile regardless of your platform.
I'm not critiquing Garry or the site. I think it's great people are using AI to build things that bring them joy, or that they find useful. I certainly do.
I am opposed to the idea that we've decided to go back to measuring work in terms of lines of code. It has always been the worst metric on earth as a proxy for productivity. Every line is a liability, and it always was. AI has not changed that, if anything it's amplifying it.
The best PRs remove code, not add, and the only companies that seem to have exponentially grown their revenues in line with AI-generated LOC are OpenAI and Anthropic. Everyone else seems to be rummaging around for an ROI.
> Note that Rails was built as a framework for making blogs
This doesn't seem right. Yes, the "build a blog in 15 minutes" demo was pretty memorable, but I don't think it's ever really been pitched as a sweet spot for Rails. IIRC Rails was essentially extracted out of the 37Signals codebases for Basecamp et al.
Ah yeah that is indeed what I was thinking of. [1]
[1] https://www.youtube.com/watch?v=Gzj723LkRJY
>2. So anyone with a claude subscription could arrive at the same conclusions?
It's symmetrical: just like anyone with a claude subscription could arrive at the same vibe slop!
> I hope they never get hold of the code of MS Office or almost any other piece of real-world business software.
Except he's not building Office, a software with decades of legacy, used by hundreds of millions of users. He's coding a website, effectively writing bloat with a silver lining of useful features. AI automated and inflated the worst of practices too. Anyone outputting 37K LoC daily is creating bloat and inefficiencies at unprecedented rate.
And enterprise software was already the butt of all jokes in this regard. We found ways to make that worse at scale even for the basic things. It's not a good look when you need to use this "whatabout".
> So anyone with claude access could arrive at the same conclusions ... and ask claude to fix it?
Begs the question of why the original "author" of the code hasn't just asked Claude to fix it? Or asked Claude to generate good code from the beginning. I suppose the answer is that nobody cares about good or efficient code anymore. But that's been the case since long before LLM coding though (as stated in your point 1).
Why write clever code when we can just write JS slop and ask customers "upgrade" their hardware every year...
Absolutely. For him the bloat is not a problem; it's a negative externality.
Hmm the list is a bit underwhelming. Basically, it's unnecessary requests, bloated JS, unoptimized images and generally poorly structured code. I would hate if that was where the average website is headed, but realistically, we were already headed there before LLMs. From the headline I was expecting CVEs, broken UX flows / business logic, leaked secrets.
It's good that no leaked keys were found in the front end code but the developer wasn't able to look at the backend where, if the quality of the front end is any indication, there are likely to be many security issues. Hopefully it's not all running on the same servers/network as anything important.
I think we're on the same team, I find the general attitude towards security in the current AI scene scary. I was just hoping for a bit more ammunition than what that article gave us.
That's because only frontend was reviewed, who knows what's on backend.
I love that they made the comparison to Hacker News in terms of request sizes lmao. Yea definitely a representative sample
Whilst I don't _really_ consider 37k LoC a day to be a particularly extreme number, my issue with all the high profile high output usage of AI, is that it doesn't really prove much of value.
I use AI in a pretty single-threaded way and my primary challenge is figuring out how to keep the LoC as minimal as possible, as well as minimise my spend. I am, unintentionally, one of the highest spenders at work; this is possible because I do more fiddly UI enhancements where the desired behaviour is often subjective and invariably hard to articulate. When I work on big backend features, my spend tends to be much lower.
What I really want from the people who have effectively unlimited tokens to spend, is to use that finding ways for the rest of us to produce higher quality output at lower costs, rather than focusing on output alone.
A lot of this stuff is simply Veblen: https://en.wikipedia.org/wiki/Conspicuous_consumption
They're not trying to consume less, they're trying to consume more (and I would count lines of code as "spend", a liability rather than an asset).
Tokenmaxxing is not dissimilar to NFTs. A means of laundering electricity into stock prices and social status, which works for a while.
Garry Tan's point still stands: he never pretended to be building nice software. But his point was that he can now build AT ALL! Shipping a webpage at all is the firs step ; making it load under 7 Mo is just a refinement, an important one of course (who tf wants bloated webpages) but still only a refinement. Tan is right to be amazed and to be shipping.
Has Garry produced any software that is actually useful to anybody? Cause that kinda needs to be be lowest we can set the bar, right?
His point is about the volume, and that to us (software engineers) it sounds like he actively just adding bloat, not putting out useful software.
Nice article...
Bloat is not "since generative AI coding" new. We always had it. As the author says:
> “It does sound like Facebook’s ‘move fast and break things,’ which didn’t age well either.”
Which indeed did not age well, but it did help the company to grow to a certain point at which it is now a staple in our lives (and can do super expensive BS experiments like "Metaverse" and still show profits).
This may be what Tan does as well: first profitable, then correct.
This is an approach, which may work for some, it may also allow some companies to become irrelevant (like: once the bloat-app is profitable, a clone-app emerges that is bloat-free and overtakes the bloat-app in every dimension, while the bloat-app is figuring out how to scale up with a shitty db schema).
Seems to be a frequently occurring issue, most (good) software engineers know that LOC basically means nothing, if anything, less LOC is a better goal over more LOC, if you absolutely have to have a goal. As soon as people/companies starts bragging about how many LOC they can ship, you need to start being very suspicious, mostly because they just admitted to not actually understand software engineering at all.
Cursor did something similar months ago, bragging about producing millions of LOC while what they actually made barely worked and could have been built with an order of magnitude less LOC: https://emsh.cat/en/one-human-one-agent-one-browser/ (https://news.ycombinator.com/item?id=46779522 | 324 points | 5 months ago | 156 comments)
What I don't understand, isn't there a single engineer working with these people who ask them what the fuck they're doing, before they hit that publish button? Or is there just such a constant pressure to publish anything that quality just simply doesn't matter at all to these people?
> 169 server requests for various assets totaling 6.42 megabytes in size.
Ahh … remember the good ol’ days, when we were forced to cap our Web pages at 32K?
I have found similar issues with LLM-generated code. Lots of code … lots of issues.
Just yesterday, I went through a whole bunch of LLM-generated code, and switched the threading model. It sped up my code significantly, and the jetsam crashes stopped immediately. Until I used LLM code, my apps had never jetsam-crashed before.
First time for everything, I guess.
[dead]
But web dev is extremely bloat and inefficient already.
left-pad comes to mind
And with that 37k a day, or 185k per 5-day work-week, what exactly has he achieved?
I don't see any actual output coming from these AI tools, despite how many are saying it's greatly increased their productivity. Where are all the new businesses and tools? I only see more shovels being sold.
> The larger point is that while AI coding tools make it easy to pump out lots of code, it’s really (still) the quality of the code that matters. Quantity, in other words, doesn’t necessarily equal quality.
They didn't provide any evidence for this point. They showed that the code is bad quality, but not that it matters that it's bad.
My hot take is that the entire history of the personal internet was inefficient, badly built and chaotic and it was much more fun to exist in.
I really don't see a problem if it breathes life into the internet for non-technical people to be creative and interesting again. Wordpress etc gave people a cheap voice, but at the cost of immovable guard-rails.
And even at a small business level, most stuff people build get's torn down by the next person or next agency 6, 12, 18 months later anyway so why care about engineering rigour (this point does not extend to security concerns) in the front end when you can spin up and tear down something in a day or two for incredibly cheap.
I'm honestly tired at the pearl-clutching when we ourselves have helped shape an industry that constantly raises the bar to entry with unnecessary techincal complexity and bloat. Ever built a brochureware site in react or similar? you've done exactly what the this guy did, you just did it with pure intent rather than an AI doing it
Does anyone know which CDN this site is using? I got an unusual type of CAPTCHA asking me to slide right.
https://www.fastly.com/
You can make decent code with AI assistance. You can churn out 37kLOC per day.
However, I’ve seen no evidence to suggest that you can do both.
An opinionated agent harness is all it takes to correct the mentioned issues.
https://archive.ph/KHM9E
my cpu fan spinning a lot when visiting that website
Focus on a single metric instead of outcome and you win on that metric instead of the outcome.
I remember that for, uh, Key Quarterly Objectives, was that the name? Aeons ago.
Same shit new decade.
(I love working with AI though. It has many of the benefits of good pair programming.)
> The larger point is that while AI coding tools make it easy to pump out lots of code, it’s really (still) the quality of the code that matters
This conclusion is unsupported from the observations. The code makes lots of requests, has too much CSS, and 6 different logo formats. So what? Lots of real, production codebases have just as many warts.
Folks need to stop dealing in absolutes with AI coding. Code quality always mattered, and still does, in certain circumstances. In others, it's less important and speed of iteration has more value. That's still true even with AI code.
> In others, it's less important and speed of iteration has more value.
Maybe in some sort of topsy-turvy techbro world.
I've never, in three decades of developing, had a user say "what I want are new features updated frequently and I'm happy to accept sluggish, buggy code".
What they do say is "I want more features, sooner, without bugs".
I'd guess you're maybe working on more established codebases then? I see this regularly with my clients. I work in sprints and they pay for my time, so they will often deprioritise engineering/robustness work if it's taking up time that could be used some shiny new feature they want.
Also it's rarely "I want new features and I'm happy for them to be buggy or sluggish", it's more "We want to solve this problem and we're not exactly sure what the solution will look like, so lets try stuff and accept it's not going to be perfect"
Ironically the fastcompany page hosting the article that should only be read only paragraphs of text and a simple navigation seems to do 300+ requests, with 29.4mb of resources.
I mean honestly the numbers and issues talked about there seem to be on most modern web apps. It just doesn't seem to affect business results enough for people to care about fixing them.
The cult of 10x got their magical tool that lets them feel 100x. Don't take their toy away from an excited kid.
How can anyone in their right mind think that growing a codebase by 200K LOC every week could possibly be a good thing?
Garry is buying horses for all the defenders in this thread.
No he wont fund your startup.
Still better than when he was making death threats against public officials.
Letting this guy run the show is a public declaration of the moral bankruptcy of the tech sphere.
It's 100% fine for Gary to write software in a way that works for him - and for us to push boundaries, and not 'gatekeep'.
BUT - it's still a pile of steaming garbage.
It's not just a 'bit odd' - it's just massive slop.
The total lack of self awareness is comical and disturbing.
GStack + Gary's Tweets about 30K LOC a day is 'Trumpian' in self delusion - it looks like basically he and YC are frauds that don't know what technology is.
This is like Elon and his 'salute' or his 'I'm a top 10 Diablo player - hey watch me play and expose myself unwittingly' type situations.
A bit of AI hype is fine.
Someone needs to take Gary and have a side-discussion.
He can do GStack - he just needs to characterize it properl for what it is.
Talking about that amount of code with the assumption 'it's sensible' code, is basically a lie - it's fraudulent' levels of hype.
Just describe it as an unwieldy but productive experiment, not for mainline consumption etc. then it's fine.
"Tan/AI built the website so that when a user visits, their browser makes 169 server requests for various assets totaling 6.42 megabytes in size. For comparison, the minimalist Hacker News homepage (also run by Y Combinator) makes seven requests for data totaling just 12 kilobytes.
The website ships 28 actual test files (code developers use to reality-check their work) straight to every visitor’s browser. That’s 300 kilobytes of pure developer scaffolding that users never asked for.
It loads 78 different JavaScript controllers for features like AI image generation, voice extraction, video tools, etc., none of which appear on the homepage. The browser still has to download all of them “just in case.”
The site’s logo is an illustration of a bear. The site downloads the logo in eight different formats, including a completely empty 0-byte file that somehow made it to production, Gregorein found.
The website uses huge, uncompressed old-school PNGs (some nearly 2 megabytes each), even though the browser literally asks for modern tiny formats. Two images alone waste about 4 MB; with newer formats they could have been just 300KB.
Gregorein also found duplicate page content, an empty CSS (Cascading Style Sheets) file, a huge rich-text editor loaded on a read-only page, missing image descriptions, and analytics code that deliberately routes through a proxy to dodge people’s ad blockers (with a comment in the code admitting it), Gregorein reports,"
Brave new world indeed... It might be elliptically relevant that in the original novel by Aldous Huxley, the protagonist John the Savage hangs himself at the end when his search for truth fails.
Impressive. Now, let's see the link in this post to fastcompany.com; It opens a website that also pulls in 6MB using more "server requests" than 169, in fact the number keeps going up, and all it does is display a paywall. So, let's not pretend this is something unique.
The difference is that nobody brags about churning out pages like the fastcompany.com link landing page.
If anything, they're ashamed of it - and they only do it because managers and shady SEO types push for that shit.
Nor are such bloated pages regarded (or were regarded pre-AI) as anything else than as a symptom of the sickness of modern web practices.
I’m not a huge AI fan, and was ready for some Schadenfreude. But with every paragraph I thought "so, just like modern human-created websites?"
Do you remember when you had to write essays in school and the word count mattered? That’s when you’d write "it is" instead of "it's" and "will not" instead of "won't."
This is worse.
Not the first time that I read quantity over quality related to YC
[flagged]
[flagged]
[dead]