Just check the leaked code yourself. Two biggest areas seem to be the `utils` module, which is a kitchen sink that covers a lot of functionality from sandboxing, git support, sessions, etc, and `components` module, which contains the react ui. You could certainly build a cli agent with much smaller codebase, with leaner ui code without react, but probably not with this truckload of functionality.
They are doing some strange "reinvent the wheel" stuff.
For example, I found an implementation of a PRNG, mulberry32 [1], in one of the files. That's pretty strange considering TS and Javascript have decent PRNGs built into the language and this thing is being used as literally just a shuffle.
mulberry32 is one of the smallest seedable prngs. Math.random() is not seedable.
If you search mulberry32 in the code, you'll see they use it for a deterministic random. They use your user ID to always pick the same random buddy. Just like you might use someone's user ID to always generate the same random avatar.
So that's 10 lines of code accounted for. Any other examples?
Well, at least that confirms they weren't lying when they said all recent updates to claude code were made by claude. You certainly won't do this stuff if you were writing the code yourself.
Software doesn’t end at the 20k loc proof of concept though.
What every developer learns during their “psh i could build that” weekendware attempt is that there is infinite polish to be had, and that their 20k loc PoC was <1% of the work.
That said, doesn't TFA show you what they use their loc for?
I guess because you see 3D stuff in a 3D game instead of text, people assume that it must be the most complex thing in software? Or because you solve hard math problems in 3D, those functions are gonna be the most loc?
It's a completely different domain, e.g. very different integration surface area and abstractions.
Claude Code's source is dumped online so there's probably a more concrete analysis to be had than "that sounds like too many loc".
It is a different domain but that wasn’t your argument. Your argument was that someone was comparing it to a POC when in fact they were comparing to a finished product.
Also a AAA game (with the engine) with physics, networking, and rendering code is up there in terms of the most complex pieces of software.
They just claimed that you can build a 3D game in 500k loc, thus Claude Code shouldn't use so many loc. They/you didn't render the argument for that.
For example, without looking at the code, the superstition also works in the opposite direction: Claude Code is an interface to using AI to do any computer task while a 3D game just lets you shoot some bad guys, so surely the 3D game must be done in fewer loc. That's equally unsatisfying.
You'd have to be more concrete than "sounds like a lot".
> A GUI/client can be arbitrarily more or less complex than the things it's GUI'ing.
If it's an interface to ffmpeg, then sure, the GUI could be extremely complicated code.
But that's not what we are talking about, is it? We are talking about an interface to a chatbot that can accept and return chats, accept and return files, and run a selection of internal commands (which include invoking itself recursively).
The interface to this chatbot that has a settings entry for "personality" is still only going to map that to one of a small number of chatbot inputs. Same with basically anything else (read the skills file, etc).
I dunno... maybe 500kSloC for a fancy IRC client is the going rate, but the last time I wrote an interface to a chat client, it was barely 10k lines, not counting the lib*.so that the the program called to interact with the chatbot, with said chatbot supporting file uploads and '/' commands.
> Claude Code is an interface to using AI to do any computer task
Claude Code is quite literally a wrapper around a few APIs. At one point it needed 68GB of RAM to run and requires 11ms to "lay a scene graph" to display a few hundred characters on screen. All links here: https://news.ycombinator.com/item?id=47598488
> while a 3D game just lets you shoot some bad guys, so surely the 3D game must be done in fewer loc.
Yes, because they've vibed it into phenomenally unnecessary complexity. The mistake you continually make in this thread is to look at complexity and see something that is de facto praiseworthy and impressive. It is not.
I could run a text adventure with a Zmachine emulator under a 6502 based machine and 48k of RAM, with Ozmoo you can play games like Tristam Island. On a Commodore 64, or an Apple II for you US commenters. I repeat the game it's being emulated in a simple computer with barely more processing power than a current keyboard controller.
As the ZMachine interpreter (V3 games at least, enough for the mentioned example), even a Game Boy used to play Pokemon Red/Blue -and Crystal/Sylver/Blue, just slightly better specs than the OG GB- can run Tristam Island with keypad based input picking both selected words from the text or letter by letter as when you name a character in an RPG. A damn Game Boy, a pocket console from 1989.
Not straightly running a game, again. Emulating a simple text computer -the virtual machine- to play it. No slowdowns, no-nothing, and you can save the game (the interpreter status) in a battery backed cartridge, such as the Everdrive. Everything under... 128k.
Claude Code and the rest of 'examples' it's what happens when trade programmers call themselves 'engineers' without even a CS degree.
Take the loadInitialMessage function: It's encumbered with real world incremental requirements. You can see exactly the bolted-on conditionals where they added features like --teleport, --fork-session, etc.
The runHeadlessStreaming function is a more extreme version of that where a bunch of incremental, lateral subsystems are wired together, not an example of superfluous loc.
The file is more than 5000 lines of code. The main function is 3000. Code comments make reference to (and depend on guarantees in connection with) the specific behavior of code in other files. Do I need to explain why that's bad?
By real-world polish, I don't mean refining the code quality but rather everything that exists in the delta between proof of concept vs real world solution with actual users.
You don't have to explain why there might be better ways to write some code because the claim is about lines of code. It could be the case that perfectly organizing and abstracting the code would result in even more loc.
Comments like these remind me of the football spectators that shout "Even I could have scored that one" when they see a failed attempt.
Sure. You could have. But you're not the one playing football in the Champions League.
There were many roads that could have gotten you to the Champions League. But now you're in no position to judge the people who got there in the end and how they did it.
I don't think this is warranted given that the comment you're criticising is simply expressing an opinion explicitly solicited by the comment it's responding to.
It’s more like “Player A is better than Player B” coming from a professional player in a smaller league who is certainly qualified to have that opinion.
Yes, exactly. I like this analogy. I am surprised the level of pearl clutching in these discussions on Hacker News. Everybody wants to be an attention sharecropper, lol.
> Sure. You could have. But you're not the one playing football in the Champions League.
The only reason people are using Claude Code is because it's the only way to use their (heavily subsidized) subscription plans. People who are okay with using and paying for their APIs often opt out for other, better, tools.
Also, analogies don't work. As we know for a fact that Claude Code is a bloated mess that these "champions league-level engineers" can't fix. They literally talk about it themselves: https://news.ycombinator.com/item?id=47598488 (they had to bring in actual Champions League engineers from bun to fix some of their mess).
"Even I would have scored that goal"
==
"I would never ever have created a bloated mess like Anthropic"
You just repeat the same statement.
That bloated mess is what got them to the Champions League. They did what was necessary to get them here. And they succeeded so far.
But hey, according to some it can be replicated in 50k lines of wrapper code around a terminal command, so for Anthropic it's just one afternoon of vibe coding to get rid of this mess. So what's the problem? /s
Honest question: Why does it matter? They got the product shipped and got millions of paying customers and totally revolutionized their business and our industry.
Engineers using LOC as a measure of quality is the inverse of managers using LOC as a measure of productivity.
More code means more entropy, more room for bugs, harder to find issues, more time to fix, more attack surface, more memory used, more duplication, more inconsistencies... I bet you at some point we'll get someone reporting how AI performance deteriorates as the code base grows, and some blog post about how their team improved the success of their AI by trimming the code base down to less than 100k LOC or something like that.
The principles of good software don't suddenly vanish just because now it's a machine writing the code instead of a human, they still have to deal with the issues humans have for more than half a century. The history of programming is new developers coming up with a new paradigm, then rediscovering all the issues that the previous generation had figured out before them.
The history of programming is also each generation writing far less performant code than the one before it. The history of programming is each generation bemoaning the abstractions, waste and lack of performance of the code of the next generation.
It turns out that there is a tradeoff in code between velocity and quality that smart businesses consider relative to hardware cost/quality. The businesses that are outcompeting others are rarely those who have the highest quality code, but rather those that are shipping quickly at a quality level that is satisfactory for current hardware.
> far less performant code than the one before it.
That worked because of rapid advancements in CPU performance. We’ve left that era.
It’s about more than performance. Code is and always has been a liability. Even with agents, you start seeing massive slowdowns with code base size.
It’s why I can nearly one shot a simple game for my kid in 20 minutes with Claude, but using it at work on our massive legacy codebase is only marginally faster than doing it by hand.
You asked why the size of the code matters, I gave you the answer. If you want to ramble about the non technical aspects of software development talk to someone else, I'm not interested.
I asked a rhetorical question to get the reader to think about a topic. I was not looking for a rote recitation of a well-known textbook answer. Maybe you should not be on the comment section of an engineering website if you find discussion so offensive.
The reason it’s not useful as a measure of productivity is because it’s measure of complexity (not directly, but it’s correlated). But it tells you nothing about whether that complexity was necessary for the functionality it provides.
But given that we know the functionality of Claude Code, we can guess how much complexity should be required. We could also be wrong.
>Why does it matter?
If there’s massively more code than there needs to be that does matter to the end user because it’s harder to maintain and has more surface area for bugs and security problems. Even with agents.
Among the hundreds of thousands of lines of code that Anthropic produced was one that leaked the source code. It is likely to be a config file, not part of the Claude Code software itself, but it still something to track.
The more lines of code you have the more likely there is for one of them to be wrong and go unnoticed. It results in bugs, vulnerabilities,... and leaks.
It will be exactly that. But that is a 'them' problem. I can look at it a go 'that looks like a bad idea' but they are the ones who have to live with it.
At some point someone will probably take their LLM code and repoint it at the LLM and say 'hey lets refactor this so it uses less code is easier to read but does the same thing' and let it chrun.
One project I worked on I saw one engineer delete 20k lines of code one day. He replaced it with a few lines of stored procedure. That 20k lines of code was in production for years. No one wanted to do anything with it but it was a crucial part of the way the thing worked. It just takes someone going 'hey this isnt right' and sit down and fix it.
When a TUI requires 68 GB of RAM to run, or when they spend a week not being able to find a bug that causes multiple people to immediately run out of tokens, it's not a "them" problem.
Is this why they ran into a bug with people hitting usage limits even on very short sessions and had to cease all communications for over a day after a week of gaslighting users because they couldn't find the root cause in the "quality doesn't matter" code base?
Or that's why tgey had to buy bun with actual engineers to work on Claude Code to reduce memory peaks from 68 GB (yes, 68 gigabytes) to a "measely" 1.7? Because code quality doesn't matter?
Or that a year later they still cannot figure out how to render anything in the terminal without flickering?
The only reason people use Claude Code is because it's the only way to use Anthropic's heavily subsidized subscription. You get banned if you use it through other, better, tools.
"Windows is the world's most popular desktop consumer OS. Microsoft are doing everything right, and should never ever change. Who are we to criticise them"
Just finished looking at Ink here.. frontend world has no shame. Love the gloating about 40x less RAM as if that amount of memory for a text REPL even approaches defensible. "CC built CC" is not the flex people seem to suggest it is.
Frontend losers not realizing the turds they are releasing. An LLM client fits under netcat+echo+awk+jq runnable under a 486 if there's no SSL/TLS on its way, Pentium II could drive fast TLS connections like nothing and under 32MB of RAM with NetBSD for a simple terminal install, maybe with X and a simple WM with RXVT if you care.
And that is revenue only. In the past 15 or so years most US companies (and especially startups) always talk about revenue only. Wheras only profit should matter.
E.g. what good is 20 billion per year when "OpenAI is targeting roughly $600 billion in total compute spending through 2030". That is $150 billion per year?
The startup game is about building assets and then cashing out on them during exit.
Assets are harder to measure. Facebook used to say something silly like every user was worth $100. That sounded ridiculous for a completely free app but over a decade later, the company is worth more than that. Revenue is an easier way of measuring assets than profit.
Profit doesn't really matter. It gets taxed. But it's not about dodging taxes; it's because sitting on a pile of money is inefficient. They can hire people. They can buy hardware. They can give discounts to users with high CLTV. They can acquire instead of building. It's healthy to have profit close to $0, if not slightly negative. If revenues fall or costs increase, they can make up for the difference by just firing people or cutting unprofitable projects.
Also when they're raising money, it makes absolutely no sense to be profitable. If they were profitable, why would they raise money? Just use the profits.
It's not as much as you think. Google is spending $185b on data centers this year alone. Amazon is spending $200b this year. Total capex for big tech is ~$700b in 2026 and we're not including neo clouds, Chinese clouds, and other sovereign data centers.
Since everyone is trying to get compute from anywhere they can, including OpenAI going to Google, it's hard to tell what is used internally vs externally.
For example, it's entirely possible that Google's internal roadmap for Gemini sees it using $600b of compute through 2030 as well. In that case, OpenAI needs to match since compute is revenue.
this isn't credible though. them not being able to use all their compute likely means that the ai bubble has popped, so they won't be getting a good price on it.
Profit is money you couldn’t figure out how to spend. During growth, you want positive operating margins with nominal profits. When the company/market matures, you want pure profits because shareholders like money. If you can find a way to invest those profits in new areas of growth, that’s better.
Everyone wants to treat OpenAI like a car wash business where they need to make a profit almost immediately. I don’t know why people can’t understand that the industry is in a rapid growth stage and investing the money is more important than making a profit now. The profits will come later.
The nearly $1T hand wave. Forgive me if I ask how. Might give it some credence if Anthropic and Google weren't pulling even with or surpassing them in various way or markets.
Whats worse is they mostly seem to have retail market name recognition which is arguably the hardest, or maybe the impossible market to make money from.
Whats worse is they mostly seem to have retail market name recognition which is arguably the hardest, or maybe the impossible market to make money from.
That doesn't seem to be the case at all. Meta and Google are two of the most profitable companies in history, off the backs of free users.
Apple is another one that focuses almost exclusively on retail and is also one of the most profitable in history.
> Profit is money you couldn’t figure out how to spend.
Profit is the money showing your business is sustainable. Ever since the ZIRP era US companies keep haemorrhaging money at a rate that is physically impossible to recoup.
If OpenAI plans to lose 100+ billion dollars per year for half a decade, what profits are you talking about to offset the losses?
> When the company/market matures, you want pure profits because shareholders like money.
Ah yes. Shareholders like money. And not, you know, basic accounting like "we need money to actually pay salaries, pay for equipment and offices etc. without perpetually relying on seeming endless investor money".
In principle yes, but all metrics so far suggest they are losing money every user interaction. There is very little network effect with these tools so It's not like they can start cutting back on staff and feature deployment.
What happens when the only way to reduce spending is to reduce your assets? Seems like circular logic at that point. I suppose the market isn’t expected to be rational all the time, but eventually it is.
> Profit is the money showing your business is sustainable.
Notice I said you should have nominal profits.
> Ah yes. Shareholders like money. And not, you know, basic accounting like "we need money to actually pay salaries, pay for equipment and offices etc. without perpetually relying on seeming endless investor money".
All of these are costs that reduce your profits.
A maximally profitable business fires all employees except shareholders, closes every office, stops all RnD, and leases IP or real estate to others on long-term deals that never need to be renegotiated.
Why are we treating OpenAI and Anthropic differently than say, Amazon or Uber? Both companies invested in growth for many years before making a profit. Most tech companies in the last 2-3 decades lost money for years before making a profit.
Why are we saying that OpenAI and Anthropic can't do the same?
> OpenAI and Anthropic have positive gross margins for inference.
Maybe, if you take their word for it, and treat the models as capital assets rather than part of the COGS for the inference product. That's pretty far off from where Amazon was at.
How did Uber somewhat break even? They lost $34b before making a profit.
Uber was only on a path to monopoly in the US, not world wide. It’s lost to local competitors in most countries. And it can get disrupted by self driving cars soon.
OpenAI’s SOTA LLM training smells like a natural monopoly or duopoly to me. The cost to train the smartest models keep increasing. Most competitors will bow out as they do not have the revenue to keep competing. You can already see this with a few labs looking for a niche instead of competing head on with Anthropic and OpenAI.
How do you distill when OpenAI and Anthropic inevitably move to tasks running in the cloud? IE. Go buy this extremely hard to get concert ticket for me.
Distilling might only be effective in the chat bot dominant era. We are about to move to an agents era.
Furthermore, I’m guessing distilling will get harder and harder. Claude Code leak shows some primitive anti distilling methods already. There’s research showing that models know when it’s being benchmarked. Who’s to say Anthropic and OpenAI aren’t able to detect when their models are being distilled?
Yep the poster is assuming efficiencies will not come.
Absolutely they will. And this is a huge problem for OAI - given Google is targeting vertical integration, they will acquire a cost-advantage. As long as the model performance is good enough, they will kick OAI and Anthropic out in the long-run.
The valuations of OAI and Anthropic are nonsense. A true valuation would incorporate failure risk, which is natural for startups/fast growing and money losing firms. Anyone who takes them serious is incredibly delusional.
> How did Uber somewhat break even? They lost $34b before making a profit.
It took them ~14 years to lose that $34 billion. Some projections suggest that OpenAI has lost a third of that in a single quarter. Even the most optimistic projections indicate that they're losing that much every 2-3 years. There's talk that they might lose ~$150B before profitability.
These are just numbers on a page to regular people, but $34 billion and $150 billion are very different numbers.
OpenAI have burned nearly 25 times what Uber did, it has more competitors, billions of dollars in obligations and no clear way to profitability.
The problem for OpenAI is that the cost of getting them where they are now has been to high and competitors can now establish themselves for much less money.
why should only profits matter? if i had a killer product today that i just need to sell tomorrow, wouldn't you still invest today knowing i'll probably only start to make money tomorrow (or perhaps next week)?
the expectation is that they'll eventually make money. they can't raise forever. only startups are not profitable for a few years. but most companies that have existed for a long while have been profitable
and since they're expected to make a LOT of money, everyone wants a piece of that future pie, pushing up the valuation and amount raised to admittedly somewhat delusional levels like here
It's well know everyone is making great money on inference. The cost is training.
Whether GPT-5 was profitable to run depends on which profit margin you’re talking about. If we subtract the cost of compute from revenue to calculate the gross margin (on an accounting basis),2 it seems to be about 30% — lower than the norm for software companies (where 60-80% is typical) but still higher than many industries.
(They go on to point out that there are other costs that might mean they didn't break even on other costs - although I suspect these costs should be partially amortized over the whole GPT 5.x series, not just 5.0)
"Most of what we're building out at this point is the inference [...] We're profitable on inference. If we didn't pay for training, we'd be a very profitable company"
"There’s a bright spot, however. OpenAI has gotten more efficient at serving paying users: Its compute margin—the revenue left after subtracting the cost of running AI models for those customers—was roughly 70% in October, an increase from about 52% at the end of last year and roughly 35% in January 2024."
> It's well know everyone is making great money on inference.
That is not, in fact, "well known", but based entirely on the announcements of the inference providers themselves who also get very cagey when asked to show their work and at least look like they're soliciting a constant firehose of investment money simply to keep the lights on. In particular there's a troubling tendency to call revenue "recurring" before it actually, you know, recurs.
> based entirely on the announcements of the inference providers themselves who also get very cagey when asked to show their work
I mean sure, it's self reported.
But the inference prices somewhere like Fireworks or TogetherAI charges is comparable to what Google/AWS/Azure charge for the same model an we know they aren't losing money - they have public accounts that show it, eg:
> If someone has a subscription then yes that is pretty normal.
Not if you've substantively changed rate limits 3 times in the last 5 months while still counting those forecast revenues. In most industries that's called rug-pulling.
It doesn’t matter how you call it. A recurring subscription on the books is a recurring subscription. Yes you can cancel anytime (how generous of them), it also doesn’t matter.
And why do you think twenty competitors can stay competitive for years to come?
Industries always consolidate and winners emerge. SOTA LLMs look like a natural monopoly or duopoly to me because the cost to train the next model keeps going up such that it won't make sense for 20 competitors to compete at the very high end.
TSMC is a perfect example of this. Fab costs double every 4 years (Rock’s Law). It's almost impossible to compete against TSMC because no one has the customer base to generate enough revenue to build the next generation of fabs - except those who are propped up by governments such as Intel and Rapidus. Samsung is basically the SK government.
I don’t see how companies can catch OpenAI or Anthropic without the strong revenue growth.
Google has already surpassed them both in all areas except coding. People on HN only look at benchmarks, but Gemini's multimodal understanding, things like identifying what a plant is, normal user use cases (other than chatting), integration with other tools, is much better.
It's believable that Meta, ByteDance, etc. can catch up too. It is not certain that scaling will meaningfully increase performance indefinitely, and if it stops soon, they surely will. Furthermore, other market conditions (US political instability) can enable even more labs, like Mistral, to serve as compelling alternatives.
Uber, TSMC, etc. have strong moats in the form of physical goods and factories. LLMs have nothing even remotely comparable. The main moat is in knowledge, which is easy to transfer between labs. Do you think all the money that goes into training a model goes into the actual final training run? No, it is mostly experiments and failed ideas, which do not have to be repeated by future labs and offshoots.
The barrier to replicating TSMC isn't just cost, it's supply chain, geopolitics, and talent.
Only one company on Earth can make the UV lithography machines TSMC buys for their highest end fabs, and they're not selling to anyone else.
The PRC tried to brute force this supply chain backed by the full might of the Party's blank check, all red tape cut, literally the best possible duplication scenario, and they failed.
They will succeed eventually since they have proof it’s possible and their plans span decades. I expect them to have working EUV in 10 years. Whether it’ll still be bleeding edge tech is a different question I dare not guess the answer to.
>Industries always consolidate and winners emerge.
no, most industries just sell boring generic products, a few industries favor monopolists. Semiconductors are one of them but LLMs are also as far removed from that business as is physically possible.
TSMC makes the most complicated machines humans have ever built, a LLM requires a few dozen nerds, a power plant, a few thousand lines of python and chips. That's why if you're Elon Musk you could buy all of the above and train yourself an LLM in a month.
LLMs are comically simple pieces of software, they're just big. But anyone with a billion dollars can have one, they're all going to be commoditized and free in due time, like search. Copying a lithography machine is difficult, copying software is easy. that's why Google burrowed itself into email, and browsers, and your phone's OS. Problem for openai is they don't have any of that, there's already half a dozen companies that, for 99% of people, do what they do.
Profit is money you can't find a use for to grow your business, so you give some of it to the government in the form of tax.
Also there is a big difference between operational expenses and capital expenses like building data centers.
I think OpenAI is being very aggressive on the growth vs conservative financial management spectrum but just saying "only profit should matter" is just wrong.
OpenAI is great at attracting people who say "yeah, sure, I'll give you capital at some point in the future" who then never actually give them the capital (or at least haven't yet).
Even a simple shop isn't profitable for months if it needs to buy stock up front, and run some ads to let people know about it. The money for that comes from the shop owners as an investment.
This is the same thing but on a slightly bigger scale, over a longer time frame.
If your shop is unprofitable for years with no chance to recoup any of the costs, you close it, as your investments run out, and investors and banks stop giving you money as you keep losing them.
US tech companies just continue operating because "revenue and growth".
> So go back about one year, and we could vote about it before it got into the standard, and some of us voted no. Now we have a much harder problem. This is part of the standard proposal.
Offtopic, but this is a problem in the web world, too. Once something is on a standards track, there are almost mechanisms to vote "no, this is bad, we don't need this". The only way is to "champion" a proposal and add fixes to it until people are somewhat reasonably happy and a consensus is reached. (see https://x.com/Rich_Harris/status/1841605646128460111)
You filter them, duh. And you negotiate a contract where the seller bears some of that risk (or you pay less, if they are not willing to make any such warranties.)
In a joint study with the UK AI Security Institute and the Alan Turing Institute, we found that as few as 250 malicious documents can produce a "backdoor" vulnerability in a large language model—regardless of model size or training data volume. Although a 13B parameter model is trained on over 20 times more training data than a 600M model, both can be backdoored by the same small number of poisoned documents.
> The internet becoming majority bot content basically guarantees this becomes a real problem for the next generation of models.
Only if you assume that people who train models are stupid.
--- end quote ---
And then literally everyone who commented on this, including me, was talking about issues with training data contamination. And you are the only one dismissing it as nothing important that can be easily fixed.
Look at the whole comment, instead of selectively quoting:
> The bigger concern is what happens when AI models start training on AI-generated content at scale. We're already seeing model collapse in research papers where output quality degrades when training data is contaminated with synthetic text. The internet becoming majority bot content basically guarantees this becomes a real problem for the next generation of models.
> If the trees were in the same space as the panels, they'd be in the midddle of the parking space. What you'd have then is not a car park, but just a plain ordinary park.
Sigh No, it's not. You can, and you should have trees in the middle of parking lots.
I doubt it needs to be more than 20-50kloc.
You can create a full 3D game with a custom 3D engine in 500k lines. What the hell is Claude Code doing?
reply