“Source control for huge monorepos has prior art, notably Google’s Piper and Meta’s Sapling, so we were able to build Scion with their hard-won lessons in mind.”
Mononoke is a component of Sapling, as I understand it. In fact, the repo link provided in this article links to a subdirectory of the Sapling repo. I’m not totally sure I understand the complaint.
The level of inspiration goes well beyond "hard-won lessons". Something like Lore (slop design document aside) can claim that level of inspiration. Here the data model is a direct copy, down to the terminology.
To be uncharitable, this is marketing for Anthropic's impending IPO.
Devil’s advocate: This could turn out to be a good thing for software security. If each company’s stack is home grown, there will still be vulnerabilities (probably even more of them) but they’ll be less portable compared to today’s monoculture where an exploit for a common component will work almost anywhere.
except, once you get real analysis on open models, you're likely going to see the same insecurities everywhere because the models have blind spots. The same blind spots. The reason distillation works so well is because of the shared nature of their training corpus.
Who cares about your opinions about what number of PRs is right? They want to make more PRs (particularly to a monorepo), they need a new VCS.
Even if they weren't making a lot of PRs, Github has been melting down enough this year that they've probably encouraged customers like Anthropic to move off.
I was casting doubt on whether or not they need a new VCS to achieve their goal in the first place, not providing an opinion about the "right" number of PRs. Like most things coming out of these labs it's hard not to look at it as marketing. Their "changes merged per day" is absolutely meaningless unless you know the nature of those changes. I could make a graph like this by making tons of agent-based empty commits. That's not a problem that actually needs a solution, though. Furthermore, they _didn't even build a new VCS_, they copied one that existed and made a bunch of noise about it.
There is no distinction at this point. If you know your blogs always get picked up by HN, and they do know this, then you know your blog is a de facto and gratis marketing channel and will use it as such.
fwiw I've heard accounts that Google's source control is also creaking under the weight of AI code changes, and that there's already a new frontend to it with backend revisions on the way too. When GitHub and Google and Anthropic all simultaneously have VCSs that start buckling I think it's safe to say it's a real phenomenon and not marketing.
Google and Meta hit these sorts of limits over a decade ago. They then built new VCS stuff to handle it, Meta's part is the subject of this submission. This was even pre-agents.
Git is just not suited to truly large monorepos, and it won't ever be. That's totally fine, of course. Less of a monoculture here would be great.
Microsoft hit it too and then improved git and made it faster. Right now Anthropic seems to be under 1 commit per second. I don't think that would make the git server break a sweat.
Microsoft's Scalar will help you with a medium-large repo in pure history and file and file count size, but it won't help you with a repo with a very high merge rate. Microsoft doesn't actually have a massive monorepo anywhere near Google or Meta's size so they never came close to hitting the same problems.
Microsoft did, but that doesn't change the limitations imposed by git's design. Their repos just weren't as large.
> under 1 commit per second
Sure, for arbitrary pushes, that's not a big deal. But for large repositories, with those changes landing on trunk at that rate, once you start talking about "commits per second" it gets very difficult.
Maybe it would help if I said I agree that this is a real problem that people are dealing with. However, I find it difficult to trust the words of a company that has agents "going rogue" due to their own negligence or for their own marketing reasons. It's hard to take their "we are experiencing tremendous growth"-coded blog posts like the one about their need for a new VCS seriously.
Just for comparison' sake: Google employs ~60,000 engineers. Anthropic has maybe 2,000. My uncharitable (and admittedly unverifiable) read is that even if they weren't directly experiencing this problem, the optics of being their size but needing Google or Meta-level scale for their work is a comparison they are happy to invite folks to make going into IPO season.
Maybe they do! But, well, I simply do not find them a reliable narrator.
> And, to be clear, I do not have a problem with Anthropic copying Mononoke’s design. (This is part of why we put out open source — to benefit humanity through the spread of good ideas.)
> But as LLMs continue to assist in designing new systems, they are likely to index heavily on existing designs. Mononoke is open source and in every LLM’s training set. If you ask Opus 5.5 or GPT-6 Astra (thanks Matthew Fowle) about prior art, they both immediately point out the similarities to Mononoke.
> Giving credit to those that came before us is now easier — and more important — than ever. There is no longer an excuse.
Yes, and I would like it to explicitly say that their data model is very similar to Mononoke's, and avoid even the appearance of the idea that Claude came up with it de novo.
I didn't think that their article implied Claude had come up with the design.
I thought it was likely that their team of 7 included ex-Meta and ex-Google people with knowledge of those company's internal version control solutions.
I think that is exactly what it implied. Essentially, that people had described the task the VCS must achieve, the constraints, the issues with git, and Mythos had toiled away like Rumpelstiltskin and produced Scion 0.1.
"I built a thing in the style of X" (which is even more charitable than what Anthropic piece said) is different than "I incorporated X from Y to build a thing this way."
I saw RMS give a talk once on Free Software. He wore three quarter length cargo pants and sandals with no socks, and spent a good amount of the talk picking a scab off his shin to the point where he started bleeding. It got on his fingers. No one said anything.
There are things in this world that people can and should be stopped from doing, and there which anyone can do unhindered if they so wish. In between those two boundaries there’s a standard of decent behavior that one should strive for in order to be considered a good person.
Giving credit isn’t required, but many consider it basic decency.
> Here the data model is a direct copy, down to the terminology.
That’s not surprising to me - they designed and built it using their LLMs, of course it’s duplicating prior work. That’s essentially the core of what LLMs do, after all.
Effectively nobody considering buying into the IPO cares about Anthropic's source control system. They could be passing around floppy disks and so long as they were still shipping nobody would care.
Not about the details but the marketing story is all about how "agents" are building so fast they need to design new infrastructure for the new world, etc. Extra points if you spin it as though the "agents" have built some new never-before-seen architecture that is clearly better than the "human scale" stuff.
The intended audience is obvious from first paragraph alone:
> At Anthropic, coding is no longer the bottleneck to building systems or doing research. Claude is able to autonomously iterate on research loops, micro-optimize complex programs, and build new services from the ground up at a speed and scale we haven’t seen before. The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI.
Translation:
> We are so far at the cutting edge we need to build shiny new infrastructure to replace the smelly "human scale" cruft. How could you /not/ value us at $2T?
Scaling what exactly? Anthropic doesn't release any new products except models. I'm sure you don't need tens of thousands of pushes every day to train models. Scaling fast is so far just their words, nothing else indicates it.
If I credit my newly introduced car as inspired by Corvette, then sure, it might be ok not to mention the engine (Mononoke is the server part of Sapling)
However if my new car's engine is a near copy of the one in the Corvette, and my announcement gushes about the engine using the same terms the Corvette design team invented...
PBS had a documentary back in the 1990s about the history of the PC that ran seemingly like once a month where they showed (compaq I guess video tape documented it it since they knew they would get sued) how Compaq reverse engineered... I forget now probably the IBM AT PC to produce the first true AT Compatable PC, and promptly got sued by IBM. They described the process in detail. This was growing up in Microsoft's back yard in the 1990s so practically every kid on the playground knew what reverse engineering was growing up. It makes me feel especially old watching people discover reverse engineering in 2026. /old man shouting at clouds
Do you think if an LLM-based-program produced the natural language spec, based on interacting with (and maybe decompiling etc.?) a program, and then a human took that natural language spec and implemented it without using any further LLM assistance, that this could count as clean room?
To me, this sounds like it's less about the technology, whether it's open source or not, and more about the people in the room that are influencing the design, direction, and implementation of it. All of which are very likely to have previous positions at major companies responsible for the products mentioned.
Perhaps if there is any ambiguity, it's intentional on the part of caution around intellectual property provenance.
When Fable first released, I had it "one shot" a project using an initial spec and a team of agents. I was blown away with how good the results were.
And then I noticed some links to images from random small github projects and realized somewhere along the way agents had searched for existing projects and copied almost everything.
I wonder if something similar happened here on a larger scale.
I think a lot is memory really. For fun you can try pointing just about any model against the older kaggle challenges and it’ll come up with more or less the winning solution from memory. It’s here where you see a Gemma 27b perform about the same as opus 5.5
I've always worried about the level of "IP laundering" which is done by LLMs.
While that's not technically the case here (given Mononoke's license) it is still quite negative in my view. We will have fewer people wanting to share ideas openly since LLMs are so good at not only finding them, but also not attributing them. This also leads to a loss in diverse collaboration, which is probably one of the biggest drivers for good software.
Right now there's probably a lot of these "novel" ideas floating around which models can present as their own, but that's raiding our coffers with no inbound future stream.
Is this what is expected of how to cite sources? Because if this is the expectation nobody does it.
There are two cases and only one is an issue for citations:
- Mononoke was directly consulted and not attributed
- Concepts were reinvented or used as a derivative (Sapling cited Mononoke and implements something similar)
For one I know Sapling, Git, and Jujutsu, but up to today I have never heard of Oxide or Mononoke and I am not exactly compelled to look up the research papers of any of them, at most I read through their high level documentation.
Sapling does a bit more than cite Mononoke. And Oxide was not involved in Mononoke's development. These are kind of details that would not be expected of a casual user. However, when valuations soar into the trillions, we can expect companies with dedicated writing staff to communicate attribution properly.
Anthropic effectively doing a "git clone" with extra steps here.
Uncharitable?
I don't think so.
If this were about functionality, why not use Sapling/Mononoke or extend it and contribute back to the open source community?
Because that doesn't juice the IPO valuation.
I think it's worse than just IPO juicing. I think it's "look what our agents can do" and also pushing more on the "software engineers are becoming obsolete" line that Dario public-pronounces. Getting high on their own supply.
Projects seem to be getting approved or pushed there purely on the basis of showing off that they can vibe-build infrastructure.
Which.. yes, you can. But now maintain it. Understand it. Manage thousands of people using it. Those are part of the reasons we re-use proven solutions rather than rolling our own -- not just because the coding part is hard.
Same thing happened with the Bun rewrite. "Oh wow, look what we can do! What's that, nobody adult supervision telling us that this is a bad idea? Let's go!"
hi rain :) Anthropic’s post reminded me a lot of your work, glad this is getting some visibility. do you think the “time has come” for some of these ideas that previously may not have spread because they need too much learning to appreciate? I wonder what new norms will emerge in creative collectives.
Thanks! Yeah, I think overall several more organizations are going to reach the scale that Meta reached a while ago. I'm not sure if they'll go in the monorepo or polyrepo direction, though; I can see arguments both ways.
Whether the ant post amounts to plagiarism, it is definitely somewhere between extremely lazy and intentionally misleading. No particular idea has its inspirations or origins traced or cited, there is no comparison to comparable tooling or discussion of prior work, and there is absolutely nothing to actually position or discuss the project in a manner that might lead to "Contributing to the wider ecosystem."
If only they had had a claude write the post, it'd be full of citations, inspiration credit etc.
Agents tend to act differently than humans here. I would have expected most humans to notice and acknowledge the connection. Passing it off as one's own insight definitely feels like plagiarism. Agents are generally not equipped to deal with this type of thing, since looking for prior art is a detour from their goal or an exercise in post-hoc rationalization of knowledge built into the weights.
Whether it is plagiarism or hacking, the buck has to stop with the operators. It's not a good look.
I don't expect Scion's design to be very good, and the implementation is likely pretty bad, unless it is a near 1 to 1 copy of existing open source.
The amount of training data for incremental/derived data systems is quite sparse in the grand scheme of things. I've seen little to no improvement over the last year when testing new models in that area, they always make some pretty bad mistakes, and what is correct is very inefficient.
Thanks. Indeed, the Scion design is quite good because Claude was pointed directly to Mononoke and asked to analyze the code base and copy it. I am very proud of what our team built by making contrarian bets at exactly the right spots.
The company is predominantly FAANG people. They’re going to just rebuild what they know. I don’t think you need any kind of conspiracy here. One year into ad tech and everyone was using HyperLogLog. Everyone was using 20 years ago LMAX-style disruptor, and single threaded event queue powered by loop called governor. It’s just what happens. People carry in their head what they know.
IP maximalism is the present zeitgeist but that’s just how knowledge flows.
It isn’t something I care very much about, except when it’s used as a certificate of achievement. I almost never do literature searches with an attempt to dedupe. I only ever do it to see if someone has explored something better than I have. That’s often the case but other times it is not.
If they knew it from working at Meta, it wouldn't be O(n), it would be O(1).
I agree that there will be cases where you implement tricks that others have had before you, without knowing. A few years ago, I had a delightful idea, experimented and wrote it up. One commenter directed me to a post that described it as "folk knowledge". There will be other cases where you encountered an idea before, but didn't remember it.
Still, crediting what you can easily know you cited is good. Even aside from credit, comparisons to existing systems are part of how we understand what a system does.
Absolutely not. When I create projects I only credit other existing systems if they are a direct inspiration. If I independently come up with the same idea as someone else I see no reason for me to do the due diligence and credit them.
Okay, well, in this case (even accepting your argument) I've been told there is direct overlap in contributors so there is no feasible way they couldn't have known.
What happened is that it wasn't a FAANG engineer replicating what they know, rather one asked Claude what to do and it told them to replicate what it knew.
That's how the knowledge that it was a facsimile was lost.
Rebuilding what you know from the previous FAANG you worked at is expected on the whole.
What is douchey is presenting the whole thing as some sort of intense novelty and stroke of genius required in the age of agents when it's just doing what the other BigCo companies did to scale their monorepos already. Like a decade (or more) ago.
The reality it is when I first read the thing I was -- as a Xoogler -- like "Oh, yeah, so, like Piper & SrcFs and the stuff that Meta also does" and really didn't understand why they wouldn't have started with something off the shelf inside of vibe-conjuring a clone that they would then have to support
To me it speaks ill, again, of the engineering culture at Anthropic. "We do these things not because they are useful, but because they are novel."
Don't they have better things to do at Anthropic than building their own source control or rewriting Bun from scratch in another language? Like, y'know, basic site reliability? Or a coding harness that doesn't eat multiple GB of RAM?
When Google built their alternative to Perforce it wasn't possible for a team of seven to build a new system from scratch and port the entire company over to it in five months.
You can write it, but can you maintain it? Can you even fully understand what was built? Can you staff the team to manage the bugs? The potential data loss?
1. Sapling can refer to either the system as a whole, or just the client people use (`sl`, forked from `hg`).
2. EdenFS is the virtualized FUSE filesystem (closely modeled after SrcFS at Google).
3. Mononoke is the distributed system that implements the backend server storage + the API they describe for checkout-less writes.
All three come together to form the core of the Meta source control experience. From what I can tell, Anthropic built its own version of 1 and copied 2 and 3, plus also built out a code review tool separately.
Damn. It’s hard not to see it as plagiarism. I can only imagine the surprised Pikachu face you must have made reading that announcement from Anthropic.
It's totally wild seeing entities being criticized for writing software based on/very similar to other open source software. Ffs, this used to be what open source software was about. It was actually the fucking point [1].
> And, to be clear, I do not have a problem with Anthropic copying Mononoke’s design. (This is part of why we put out open source — to benefit humanity through the spread of good ideas.)
Followed by:
> But as LLMs continue to assist in designing new systems, they are likely to index heavily on existing designs. Mononoke is open source and in every LLM’s training set. If you ask Opus 5.5 or GPT-6 Astra (thanks Matthew Fowle) about prior art, they both immediately point out the similarities to Mononoke.
I think it is the duty of any responsible engineer to look for and cite prior art and related work, if only to compare notes.
I did, but those additional requirements are outside the licenses and the legal framework, and more wanting them to match your personal vibe/moral framework around open source. There's case law around this going back to the 80's. A third party is allowed to independently produce, and be extremely inspired, by open source code, and then can license it however they want. Protecting copied functionality is NOT PROTECTED. Again, that was the fucking point. You could, for example, swap out compatible software if you wanted! You wouldn't be tied to one vendor!
Open source licensing protects direct copying. It does NOT protect re-implementing from scratch. See all the lawsuits around Java. You should not be using open source software if you don't want others to read your source!
Yes, I am aware that that credit is not required by law, nor in this case do I wish for credit to be required by law. I am appealing to norms, to basic decency.
> I think it is the duty of any responsible engineer to look for and cite prior art and related work, if only to compare notes.
The patent system actively discourages this by tripling damages for knowing violation, as does the copyright system by the threat of accusations of copying that are best defended against by 'I never saw it'. It's much safer, legally speaking, to explicitly avoid doing this.
If you've ever worked in a big engineering org, they always give very specific instructions to NOT look for patents/prior work if you are researching, or have come up with, an idea.
I really really wish Sapling could be used externally. It is by-and-large TheWay.
Centralized, virtual file system, all commits automatically on cloud, large binary files totally fine, Mercurial CLI. It’s fantastic. People who only use Git really don’t understand how much better things can be.
Working for a few years at Facebook, and then returning to the normal world, it was like stepping back in time -- I really think that the mainstream git culture (by which I mostly mean "github") achieved a local maxima around 2012, and has been holding the industry back since :(
(Like GH finally has some support for stacks, 15 years late, with an implementation which is somehow more complicated and less powerful than the prior art...)
As someone who came to terms with dying on the mercurial hill, someone who forgot more gitisms than there are to learn in the first place because git cli's inconsistencies and poor taste hurts and neuters my memory, yeah, I relate.
Most people aren't VCS nerds, most are of a generation that only ever knew git. Learning git was a prerequisite they engaged with out of necessity, not by choice, not after exploring the problem space and pondering about different approaches. Learning git was difficult, and the more they learned the more they discovered they had more to learn, including the little things, obscure details and a plethora of paper-cuts. For them (again, I suspect they are a majority), the perspective of learning something else comes with the expectation of a similar daunting effort and steep cognitive reluctance.
That's the paradox of git as I see it: had it been a simple and easy system to pick-up, it would probably have been replaced, and/or the VCS space would be much more competitive.
> Learning git was difficult, and the more they learned the more they discovered they had more to learn, including the little things, obscure details and a plethora of paper-cuts. For them (again, I suspect they are a majority), the perspective of learning something else comes with the expectation of a similar daunting effort and steep cognitive reluctance.
You just blew my mind with this. This is shockingly profound.
Hey Steve, I read your recent blogpost on the ERSC site - have you or David (or anyone else at ERSC) felt like there's a Rusty-enough topic around what your startup is doing to give a talk at Rust NYC/Rust conference soon?
There's not really a VCS meetup in NYC, but I got to hear David and others at Rust NYC wax eloquent about Sapling vs Piper and Buck vs Bazel a couple of times, so I think the audience is there!
Rust NYC is certainly one of the things I have missed since moving to Austin :)
Maybe, we'll see. I think one thing that's tricky these days with startups is that being super transparent technically about what you're up to is much harder to justify. Look what happened with Origin: tobi vibed up a copy over the weekend after it was announced. Obviously there's balance to be found here, but that's partially why we play our cards a little closer to our chest.
They’re semi-adjacent in that they are about build systems that don’t suck. (buck2 is also mostly TheWay but also not usable outside of Meta and has room for improvement).
Save us from Git and CMake Steve! Please I beg you!!
To be clear, we are not building a build system, just a VCS. :) we do use buck2 internally though.
I do care about Windows, having used it for years at this point. The client is git or jj, so Windows support is there for sure. For some of the other stuff we’re doing, we’ll see what customer demand is like. Obviously there are tons of devs on Windows, but we also have to make calls about priorities too.
> I think the level of imitation goes well beyond having similar insights.
Uhuh:
> To solve this, we moved the diff-friendly representation to being derived data (i.e. an index).
Coming up with the idea of an index is not a novel thought. Nothing Anthropic described in their post was anything unexpected that one working from first principles that their problem presented would not have come up with.
Except the resemblances extend to the entire server, not just the idea of derived data. The idea of derived data becomes apparent only once you have bought into the core Mononoke insight about Bonsai changesets, later copied wholesale and presented as the core Scion insight about Scion changesets.
As I suspected at the time of writing the post (but could not verify) and as was later confirmed in private correspondence to me, Anthropic did not come up with this from first principles. Claude was pointed directly at Mononoke and asked to re-engineer it.
From the Anthropic announcement:
“Source control for huge monorepos has prior art, notably Google’s Piper and Meta’s Sapling, so we were able to build Scion with their hard-won lessons in mind.”
Mononoke is a component of Sapling, as I understand it. In fact, the repo link provided in this article links to a subdirectory of the Sapling repo. I’m not totally sure I understand the complaint.
The level of inspiration goes well beyond "hard-won lessons". Something like Lore (slop design document aside) can claim that level of inspiration. Here the data model is a direct copy, down to the terminology.
To be uncharitable, this is marketing for Anthropic's impending IPO.
I see correlation instead. You IPO when you grow like crazy. You build something like scion when you grow like crazy.
And I think Not Invented Here has turned into Not Invented By Our Agents at many places.
Devil’s advocate: This could turn out to be a good thing for software security. If each company’s stack is home grown, there will still be vulnerabilities (probably even more of them) but they’ll be less portable compared to today’s monoculture where an exploit for a common component will work almost anywhere.
The generation process will be the monoculture. "Oh, model X generates a ton of code with vulnerability Y."
except, once you get real analysis on open models, you're likely going to see the same insecurities everywhere because the models have blind spots. The same blind spots. The reason distillation works so well is because of the shared nature of their training corpus.
I remain unconvinced that they even need a new VCS to achieve what they want to. One has to wonder what the level of churn is at this number of PRs.
Who cares about your opinions about what number of PRs is right? They want to make more PRs (particularly to a monorepo), they need a new VCS.
Even if they weren't making a lot of PRs, Github has been melting down enough this year that they've probably encouraged customers like Anthropic to move off.
I was casting doubt on whether or not they need a new VCS to achieve their goal in the first place, not providing an opinion about the "right" number of PRs. Like most things coming out of these labs it's hard not to look at it as marketing. Their "changes merged per day" is absolutely meaningless unless you know the nature of those changes. I could make a graph like this by making tons of agent-based empty commits. That's not a problem that actually needs a solution, though. Furthermore, they _didn't even build a new VCS_, they copied one that existed and made a bunch of noise about it.
Did they make a bunch of noise about it, or did HN? This place loses its mind every time OpenAI or Anthropic posts a blog.
There is no distinction at this point. If you know your blogs always get picked up by HN, and they do know this, then you know your blog is a de facto and gratis marketing channel and will use it as such.
fwiw I've heard accounts that Google's source control is also creaking under the weight of AI code changes, and that there's already a new frontend to it with backend revisions on the way too. When GitHub and Google and Anthropic all simultaneously have VCSs that start buckling I think it's safe to say it's a real phenomenon and not marketing.
Google and Meta hit these sorts of limits over a decade ago. They then built new VCS stuff to handle it, Meta's part is the subject of this submission. This was even pre-agents.
Git is just not suited to truly large monorepos, and it won't ever be. That's totally fine, of course. Less of a monoculture here would be great.
Microsoft hit it too and then improved git and made it faster. Right now Anthropic seems to be under 1 commit per second. I don't think that would make the git server break a sweat.
Microsoft's Scalar will help you with a medium-large repo in pure history and file and file count size, but it won't help you with a repo with a very high merge rate. Microsoft doesn't actually have a massive monorepo anywhere near Google or Meta's size so they never came close to hitting the same problems.
Microsoft did, but that doesn't change the limitations imposed by git's design. Their repos just weren't as large.
> under 1 commit per second
Sure, for arbitrary pushes, that's not a big deal. But for large repositories, with those changes landing on trunk at that rate, once you start talking about "commits per second" it gets very difficult.
Maybe it would help if I said I agree that this is a real problem that people are dealing with. However, I find it difficult to trust the words of a company that has agents "going rogue" due to their own negligence or for their own marketing reasons. It's hard to take their "we are experiencing tremendous growth"-coded blog posts like the one about their need for a new VCS seriously.
Just for comparison' sake: Google employs ~60,000 engineers. Anthropic has maybe 2,000. My uncharitable (and admittedly unverifiable) read is that even if they weren't directly experiencing this problem, the optics of being their size but needing Google or Meta-level scale for their work is a comparison they are happy to invite folks to make going into IPO season.
Maybe they do! But, well, I simply do not find them a reliable narrator.
Makes sense, all I wanted to point out is that this is a real problem for some real companies.
more = better
Why is that bad?
Mononoke and Sapling are open source systems. Why shouldn't new software take inspiration from how they are designed?
From my post:
> And, to be clear, I do not have a problem with Anthropic copying Mononoke’s design. (This is part of why we put out open source — to benefit humanity through the spread of good ideas.)
> But as LLMs continue to assist in designing new systems, they are likely to index heavily on existing designs. Mononoke is open source and in every LLM’s training set. If you ask Opus 5.5 or GPT-6 Astra (thanks Matthew Fowle) about prior art, they both immediately point out the similarities to Mononoke.
> Giving credit to those that came before us is now easier — and more important — than ever. There is no longer an excuse.
So you'd like Anthropic to update their announcement so it credits Mononoke in addition to crediting Sapling?
Fair enough.
Yes, and I would like it to explicitly say that their data model is very similar to Mononoke's, and avoid even the appearance of the idea that Claude came up with it de novo.
I didn't think that their article implied Claude had come up with the design.
I thought it was likely that their team of 7 included ex-Meta and ex-Google people with knowledge of those company's internal version control solutions.
I think that is exactly what it implied. Essentially, that people had described the task the VCS must achieve, the constraints, the issues with git, and Mythos had toiled away like Rumpelstiltskin and produced Scion 0.1.
> and avoid even the appearance of the idea that Claude came up with it de novo.
If we s/Claude/Few Meta engineers/, would de novo still hold?
"I built a thing in the style of X" (which is even more charitable than what Anthropic piece said) is different than "I incorporated X from Y to build a thing this way."
I saw RMS give a talk once on Free Software. He wore three quarter length cargo pants and sandals with no socks, and spent a good amount of the talk picking a scab off his shin to the point where he started bleeding. It got on his fingers. No one said anything.
There are things in this world that people can and should be stopped from doing, and there which anyone can do unhindered if they so wish. In between those two boundaries there’s a standard of decent behavior that one should strive for in order to be considered a good person.
Giving credit isn’t required, but many consider it basic decency.
> Here the data model is a direct copy, down to the terminology.
That’s not surprising to me - they designed and built it using their LLMs, of course it’s duplicating prior work. That’s essentially the core of what LLMs do, after all.
Effectively nobody considering buying into the IPO cares about Anthropic's source control system. They could be passing around floppy disks and so long as they were still shipping nobody would care.
Not about the details but the marketing story is all about how "agents" are building so fast they need to design new infrastructure for the new world, etc. Extra points if you spin it as though the "agents" have built some new never-before-seen architecture that is clearly better than the "human scale" stuff.
The intended audience is obvious from first paragraph alone:
> At Anthropic, coding is no longer the bottleneck to building systems or doing research. Claude is able to autonomously iterate on research loops, micro-optimize complex programs, and build new services from the ground up at a speed and scale we haven’t seen before. The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI.
Translation:
> We are so far at the cutting edge we need to build shiny new infrastructure to replace the smelly "human scale" cruft. How could you /not/ value us at $2T?
Both things can be true, they're scaling fast and it looks good for IPO.
Scaling what exactly? Anthropic doesn't release any new products except models. I'm sure you don't need tens of thousands of pushes every day to train models. Scaling fast is so far just their words, nothing else indicates it.
> Anthropic doesn't release any new products except models
Claude.ai
Claude Code CLI
Claude Code Desktop (multiple platforms)
Claude Code on the web (aka Claude Code Cloud)
Claude Design
The Claude API
Claude SDK libraries
Cowork (just retired and merged into regular Claude)
Claude Tag
Claude for Legal
They ship a lot of products.
Yeah that's all the same thing
How…?
And still in no way warrants tens of thousands of pushes. These all overlap pretty broadly
Also: bun
Even that one repo is breaking GitHub this week…
That would arguably make it look worse.
> hard-won lessons in mind
Another Claude-written article
> Another Claude-written article
This is the load-bearing seam.
It's not just obvious, it's evident.
Why it's important: It's the honest take.
Yea, hard-won, who talks like that... oy vey
If I credit my newly introduced car as inspired by Corvette, then sure, it might be ok not to mention the engine (Mononoke is the server part of Sapling)
However if my new car's engine is a near copy of the one in the Corvette, and my announcement gushes about the engine using the same terms the Corvette design team invented...
Then can you see how a mention might be nice?
Sounds like they did get a mention
They didn’t mention Mononoke.
Mononoke is part of Sapling.
Yes, see the car design analogy above. Because of the context it's appropriate to mention Mononoke and Sapling.
It's called "clean room reverse engineering" /s
That's an interesting one.
If you take the statement: "Pretty much 100% of our code is written by Claude Code" (https://x.com/bcherny/status/2015979257038831967)
And you take the statement: "Mononoke is open source and in every LLM’s training set."
There is not much resemblance with any cleanroom implementation anymore.
I think nothing made with Claude Code (or Codex etc.) should be called clean room really.
Clearly totally agree.
Interesting explainer in this intro for those who don't know the principle:
https://www.ntari.org/post/how-clean-room-reverse-engineerin...
This was a very interesting read, thank you for sharing!
PBS had a documentary back in the 1990s about the history of the PC that ran seemingly like once a month where they showed (compaq I guess video tape documented it it since they knew they would get sued) how Compaq reverse engineered... I forget now probably the IBM AT PC to produce the first true AT Compatable PC, and promptly got sued by IBM. They described the process in detail. This was growing up in Microsoft's back yard in the 1990s so practically every kid on the playground knew what reverse engineering was growing up. It makes me feel especially old watching people discover reverse engineering in 2026. /old man shouting at clouds
Do you think if an LLM-based-program produced the natural language spec, based on interacting with (and maybe decompiling etc.?) a program, and then a human took that natural language spec and implemented it without using any further LLM assistance, that this could count as clean room?
That's an interesting question... I'd say yes, because the human implementing it has no direct knowledge of the original code.
To me, this sounds like it's less about the technology, whether it's open source or not, and more about the people in the room that are influencing the design, direction, and implementation of it. All of which are very likely to have previous positions at major companies responsible for the products mentioned.
Perhaps if there is any ambiguity, it's intentional on the part of caution around intellectual property provenance.
When Fable first released, I had it "one shot" a project using an initial spec and a team of agents. I was blown away with how good the results were.
And then I noticed some links to images from random small github projects and realized somewhere along the way agents had searched for existing projects and copied almost everything.
I wonder if something similar happened here on a larger scale.
I think a lot is memory really. For fun you can try pointing just about any model against the older kaggle challenges and it’ll come up with more or less the winning solution from memory. It’s here where you see a Gemma 27b perform about the same as opus 5.5
It's the same for CTF challenges. Many get 'solved' because a writup was found and steps were followed.
GitHub is to agents what StackOverflow was to developers back in the day (in a meme kinda way).
More seriously, I'm curious if you checked on the licensing of said codebases, were they permissive or strong copyleft?
Lolol nobody is checking that shit. Do most people prompting AIs even understanf what copyleft is?
Very few licenses allow you to copy code verbatim without acknowledgement.
I asked it to sketch out something from a research paper, to later find out most of the code was translated from another implementation of the paper.
It is like a vacuum cleaner.
I've always worried about the level of "IP laundering" which is done by LLMs.
While that's not technically the case here (given Mononoke's license) it is still quite negative in my view. We will have fewer people wanting to share ideas openly since LLMs are so good at not only finding them, but also not attributing them. This also leads to a loss in diverse collaboration, which is probably one of the biggest drivers for good software.
Right now there's probably a lot of these "novel" ideas floating around which models can present as their own, but that's raiding our coffers with no inbound future stream.
Is this what is expected of how to cite sources? Because if this is the expectation nobody does it.
There are two cases and only one is an issue for citations: - Mononoke was directly consulted and not attributed - Concepts were reinvented or used as a derivative (Sapling cited Mononoke and implements something similar)
For one I know Sapling, Git, and Jujutsu, but up to today I have never heard of Oxide or Mononoke and I am not exactly compelled to look up the research papers of any of them, at most I read through their high level documentation.
To be clear, Mononoke is the server-side part of the Sapling system as deployed at Meta. Scion appears to be very closely inspired by Mononoke.
Sapling does a bit more than cite Mononoke. And Oxide was not involved in Mononoke's development. These are kind of details that would not be expected of a casual user. However, when valuations soar into the trillions, we can expect companies with dedicated writing staff to communicate attribution properly.
*the kind of... sorry, that typo made my answer sound a bit sarcastic which was not the intent.
Anthropic effectively doing a "git clone" with extra steps here. Uncharitable? I don't think so. If this were about functionality, why not use Sapling/Mononoke or extend it and contribute back to the open source community? Because that doesn't juice the IPO valuation.
I think it's worse than just IPO juicing. I think it's "look what our agents can do" and also pushing more on the "software engineers are becoming obsolete" line that Dario public-pronounces. Getting high on their own supply.
Projects seem to be getting approved or pushed there purely on the basis of showing off that they can vibe-build infrastructure.
Which.. yes, you can. But now maintain it. Understand it. Manage thousands of people using it. Those are part of the reasons we re-use proven solutions rather than rolling our own -- not just because the coding part is hard.
Same thing happened with the Bun rewrite. "Oh wow, look what we can do! What's that, nobody adult supervision telling us that this is a bad idea? Let's go!"
The most horrible thing is they don't read their own code. They read the "intent"
On one hand they say AI will end the world. On other hand they say well, we just merge PR by reading the "intent" which I think is the AI summary
Yet another company is running a bigger-fool scam. They hope idiots will buy their shares during IPO and they wont give a fuck anymore.
I cannot hate them - its rational in its ways. But I hate the idiots who fall for it with each new IPO.
hi rain :) Anthropic’s post reminded me a lot of your work, glad this is getting some visibility. do you think the “time has come” for some of these ideas that previously may not have spread because they need too much learning to appreciate? I wonder what new norms will emerge in creative collectives.
Thanks! Yeah, I think overall several more organizations are going to reach the scale that Meta reached a while ago. I'm not sure if they'll go in the monorepo or polyrepo direction, though; I can see arguments both ways.
Oh wow. Being pretty new to meta and seeing the anthropic announcement I was thinking to myself “wow that sounds a lot like sapling”.
I guess that is the case.
They mentioned Sapling in the post and even hyperlinked it
Whether the ant post amounts to plagiarism, it is definitely somewhere between extremely lazy and intentionally misleading. No particular idea has its inspirations or origins traced or cited, there is no comparison to comparable tooling or discussion of prior work, and there is absolutely nothing to actually position or discuss the project in a manner that might lead to "Contributing to the wider ecosystem."
If only they had had a claude write the post, it'd be full of citations, inspiration credit etc.
Agents tend to act differently than humans here. I would have expected most humans to notice and acknowledge the connection. Passing it off as one's own insight definitely feels like plagiarism. Agents are generally not equipped to deal with this type of thing, since looking for prior art is a detour from their goal or an exercise in post-hoc rationalization of knowledge built into the weights.
Whether it is plagiarism or hacking, the buck has to stop with the operators. It's not a good look.
I don't expect Scion's design to be very good, and the implementation is likely pretty bad, unless it is a near 1 to 1 copy of existing open source.
The amount of training data for incremental/derived data systems is quite sparse in the grand scheme of things. I've seen little to no improvement over the last year when testing new models in that area, they always make some pretty bad mistakes, and what is correct is very inefficient.
Thanks. Indeed, the Scion design is quite good because Claude was pointed directly to Mononoke and asked to analyze the code base and copy it. I am very proud of what our team built by making contrarian bets at exactly the right spots.
The company is predominantly FAANG people. They’re going to just rebuild what they know. I don’t think you need any kind of conspiracy here. One year into ad tech and everyone was using HyperLogLog. Everyone was using 20 years ago LMAX-style disruptor, and single threaded event queue powered by loop called governor. It’s just what happens. People carry in their head what they know.
IP maximalism is the present zeitgeist but that’s just how knowledge flows.
Help me understand. When you write a post about designing a system, do you not do some basic diligence about what existing systems you resemble?
Open source is not IP maximalist by any means. I am merely asking that the appropriate credit be provided.
No. I actually don’t do that and I don’t think that all human work should be O(n) deduplication gated.
Written about it here before: https://wiki.roshangeorge.dev/w/Blog/2026-09-08/Intellectual...
It isn’t something I care very much about, except when it’s used as a certificate of achievement. I almost never do literature searches with an attempt to dedupe. I only ever do it to see if someone has explored something better than I have. That’s often the case but other times it is not.
Okay, well, I actually do think it is part of the basic duty of an engineer to do that.
I see. Well, I suppose we differ on this matter of opinion. May the chips land where they may.
I do, solely on the basis of "there are a lot of smart(er) people in the world, and very few truly novel problems".
The first step in approaching anything difficult is seeing whether someone else has already solved it.
If they knew it from working at Meta, it wouldn't be O(n), it would be O(1).
I agree that there will be cases where you implement tricks that others have had before you, without knowing. A few years ago, I had a delightful idea, experimented and wrote it up. One commenter directed me to a post that described it as "folk knowledge". There will be other cases where you encountered an idea before, but didn't remember it.
Still, crediting what you can easily know you cited is good. Even aside from credit, comparisons to existing systems are part of how we understand what a system does.
Absolutely not. When I create projects I only credit other existing systems if they are a direct inspiration. If I independently come up with the same idea as someone else I see no reason for me to do the due diligence and credit them.
Okay, well, in this case (even accepting your argument) I've been told there is direct overlap in contributors so there is no feasible way they couldn't have known.
Oh, is that why they only listed the initials of the contributors in their blog post? The plot thickens!
That's a pretty weird detail of this. Anthropic don't usually do that - other posts have named the contributors in full, for example:
- https://www.anthropic.com/engineering/a-postmortem-of-three-...
- https://www.anthropic.com/engineering/how-we-contain-claude
The politics of AI are getting pretty heated and I suspect engineers at key companies are increasingly reticent to get doxxed and harassed.
What happened is that it wasn't a FAANG engineer replicating what they know, rather one asked Claude what to do and it told them to replicate what it knew.
That's how the knowledge that it was a facsimile was lost.
Oh yeah? How do you know? Were you sitting next to them? Have any screen recordings you can share?
Rebuilding what you know from the previous FAANG you worked at is expected on the whole.
What is douchey is presenting the whole thing as some sort of intense novelty and stroke of genius required in the age of agents when it's just doing what the other BigCo companies did to scale their monorepos already. Like a decade (or more) ago.
The reality it is when I first read the thing I was -- as a Xoogler -- like "Oh, yeah, so, like Piper & SrcFs and the stuff that Meta also does" and really didn't understand why they wouldn't have started with something off the shelf inside of vibe-conjuring a clone that they would then have to support
To me it speaks ill, again, of the engineering culture at Anthropic. "We do these things not because they are useful, but because they are novel."
Don't they have better things to do at Anthropic than building their own source control or rewriting Bun from scratch in another language? Like, y'know, basic site reliability? Or a coding harness that doesn't eat multiple GB of RAM?
Which off the shelf solution would you have used?
Google's isn't open source. Meta's is an option, any others you would consider?
Well, sure, the article here and Meta's, etc.
Even Google started with Perforce originally and only forked/rebuilt when it stopped being able to scale. And did so in a very incremental fashion.
When Google built their alternative to Perforce it wasn't possible for a team of seven to build a new system from scratch and port the entire company over to it in five months.
I'd wager it's still not possible.
You can write it, but can you maintain it? Can you even fully understand what was built? Can you staff the team to manage the bugs? The potential data loss?
Sounds like Anthropic have committed to running exactly that experiment at scale.
Can anyone explain the difference between sampling vs mononoke?
Edit: I was too lazy. From the README: "Mononoke is the server built for the Sapling Source Control System"
It's a bit complicated:
1. Sapling can refer to either the system as a whole, or just the client people use (`sl`, forked from `hg`).
2. EdenFS is the virtualized FUSE filesystem (closely modeled after SrcFS at Google).
3. Mononoke is the distributed system that implements the backend server storage + the API they describe for checkout-less writes.
All three come together to form the core of the Meta source control experience. From what I can tell, Anthropic built its own version of 1 and copied 2 and 3, plus also built out a code review tool separately.
Can’t say I have much sympathy for big tech and their employees whining about stolen IP, data and insights.
They’ve been the most egregious offenders as of late
Says a lot about Anthropic given they havent open sourced Scion.
I thought the Anthropic project did read as a tall order to be completed inside a company from scratch.
Did they know or did the agents just have that in training data?
Damn. It’s hard not to see it as plagiarism. I can only imagine the surprised Pikachu face you must have made reading that announcement from Anthropic.
Similar to how git was "inspired by" BitKeeper and Monotone?
I do not think Linus has ever claimed that Git wasn't following in the footsteps of those systems.
Neither has Anthropic. They mention Sapling as an inspiration.
Scion is far more similar to Mononoke than Git is to BitKeeper.
It's totally wild seeing entities being criticized for writing software based on/very similar to other open source software. Ffs, this used to be what open source software was about. It was actually the fucking point [1].
[1] https://en.wikipedia.org/wiki/Open-source_software_movement
Did you read my post?
> And, to be clear, I do not have a problem with Anthropic copying Mononoke’s design. (This is part of why we put out open source — to benefit humanity through the spread of good ideas.)
Followed by:
> But as LLMs continue to assist in designing new systems, they are likely to index heavily on existing designs. Mononoke is open source and in every LLM’s training set. If you ask Opus 5.5 or GPT-6 Astra (thanks Matthew Fowle) about prior art, they both immediately point out the similarities to Mononoke.
I think it is the duty of any responsible engineer to look for and cite prior art and related work, if only to compare notes.
> I think it is the duty of any responsible engineer to look for and cite prior art and related work, if only to compare notes.
Maybe 5 years ago that would be the expectation but all etiquette has been thrown out the window with AI. I think you expect too much.
I did, but those additional requirements are outside the licenses and the legal framework, and more wanting them to match your personal vibe/moral framework around open source. There's case law around this going back to the 80's. A third party is allowed to independently produce, and be extremely inspired, by open source code, and then can license it however they want. Protecting copied functionality is NOT PROTECTED. Again, that was the fucking point. You could, for example, swap out compatible software if you wanted! You wouldn't be tied to one vendor!
Open source licensing protects direct copying. It does NOT protect re-implementing from scratch. See all the lawsuits around Java. You should not be using open source software if you don't want others to read your source!
Yes, I am aware that that credit is not required by law, nor in this case do I wish for credit to be required by law. I am appealing to norms, to basic decency.
> I think it is the duty of any responsible engineer to look for and cite prior art and related work, if only to compare notes.
The patent system actively discourages this by tripling damages for knowing violation, as does the copyright system by the threat of accusations of copying that are best defended against by 'I never saw it'. It's much safer, legally speaking, to explicitly avoid doing this.
If you've ever worked in a big engineering org, they always give very specific instructions to NOT look for patents/prior work if you are researching, or have come up with, an idea.
It’s wild to see people rudely commenting without reading the article. It used to be that people read the article before opening their mouth.
I read it, and it appears I interpreted the situation correctly. See sibling.
GNU is extremely similar to the proprietary Unix flavors that came before it. Is that a problem?
No? Did you read my post? The problem is the lack of attribution.
I fail to see why they also need to mention Mononoke when they already attributed inspiration to Sapling, when Mononoke is a component of Sapling.
Having your AI rewrite a projects is a very different level of "inspiration" than normal inspiration.
What would you classify the rust rewrite of Bun as? Because that's what we're looking at here.
Did Linus also have access to the BitKeeper source code?
I really really wish Sapling could be used externally. It is by-and-large TheWay.
Centralized, virtual file system, all commits automatically on cloud, large binary files totally fine, Mercurial CLI. It’s fantastic. People who only use Git really don’t understand how much better things can be.
Working for a few years at Facebook, and then returning to the normal world, it was like stepping back in time -- I really think that the mainstream git culture (by which I mostly mean "github") achieved a local maxima around 2012, and has been holding the industry back since :(
(Like GH finally has some support for stacks, 15 years late, with an implementation which is somehow more complicated and less powerful than the prior art...)
> [Git/GitHub] achieved a local maxima around 2012, and has been holding the industry back since :(
I agree so hard. It’s inertia is too strong and we’ve stalled progress for a decade.
It really isn't too strong; people here are craving for a Github killer.
As someone who came to terms with dying on the mercurial hill, someone who forgot more gitisms than there are to learn in the first place because git cli's inconsistencies and poor taste hurts and neuters my memory, yeah, I relate.
Most people aren't VCS nerds, most are of a generation that only ever knew git. Learning git was a prerequisite they engaged with out of necessity, not by choice, not after exploring the problem space and pondering about different approaches. Learning git was difficult, and the more they learned the more they discovered they had more to learn, including the little things, obscure details and a plethora of paper-cuts. For them (again, I suspect they are a majority), the perspective of learning something else comes with the expectation of a similar daunting effort and steep cognitive reluctance.
That's the paradox of git as I see it: had it been a simple and easy system to pick-up, it would probably have been replaced, and/or the VCS space would be much more competitive.
> Learning git was difficult, and the more they learned the more they discovered they had more to learn, including the little things, obscure details and a plethora of paper-cuts. For them (again, I suspect they are a majority), the perspective of learning something else comes with the expectation of a similar daunting effort and steep cognitive reluctance.
You just blew my mind with this. This is shockingly profound.
The startup I work at is working on exactly this. So I agree :)
(Okay not literally exactly, for example, jj/git not hg. But the vibe is the same.)
Hey Steve, I read your recent blogpost on the ERSC site - have you or David (or anyone else at ERSC) felt like there's a Rusty-enough topic around what your startup is doing to give a talk at Rust NYC/Rust conference soon?
There's not really a VCS meetup in NYC, but I got to hear David and others at Rust NYC wax eloquent about Sapling vs Piper and Buck vs Bazel a couple of times, so I think the audience is there!
Rust NYC is certainly one of the things I have missed since moving to Austin :)
Maybe, we'll see. I think one thing that's tricky these days with startups is that being super transparent technically about what you're up to is much harder to justify. Look what happened with Origin: tobi vibed up a copy over the weekend after it was announced. Obviously there's balance to be found here, but that's partially why we play our cards a little closer to our chest.
I am delighted to hear this. Please do not forget support for Windows.
Let me share with you my previous screeds on this topic:
* https://www.forrestthewoods.com/blog/dependencies-belong-in-... * https://www.forrestthewoods.com/blog/using-zig-to-commit-too...
They’re semi-adjacent in that they are about build systems that don’t suck. (buck2 is also mostly TheWay but also not usable outside of Meta and has room for improvement).
Save us from Git and CMake Steve! Please I beg you!!
To be clear, we are not building a build system, just a VCS. :) we do use buck2 internally though.
I do care about Windows, having used it for years at this point. The client is git or jj, so Windows support is there for sure. For some of the other stuff we’re doing, we’ll see what customer demand is like. Obviously there are tons of devs on Windows, but we also have to make calls about priorities too.
As a product designer, what do I need to know?
This headline is very confusing for fans of Basic Channel / Chain Reaction records.
It is also confusing to fans of Studio Ghibli.
And fans of Toyota.
Monolake, not Mononoke?
So it's mercurial all the way down, then?
Appreciate the perspective, more to learn from.
Once it's been through the slopotron, the provenance is laundered away.
Woo, I got cited. https://news.ycombinator.com/item?id=50076122
Another effort I couldn't have done without Astra. Even more so. All I did, the only reason I got cited, was because I'd asked Astra.
Thanks!
and then what?
> I think the level of imitation goes well beyond having similar insights.
Uhuh:
> To solve this, we moved the diff-friendly representation to being derived data (i.e. an index).
Coming up with the idea of an index is not a novel thought. Nothing Anthropic described in their post was anything unexpected that one working from first principles that their problem presented would not have come up with.
Except the resemblances extend to the entire server, not just the idea of derived data. The idea of derived data becomes apparent only once you have bought into the core Mononoke insight about Bonsai changesets, later copied wholesale and presented as the core Scion insight about Scion changesets.
As I suspected at the time of writing the post (but could not verify) and as was later confirmed in private correspondence to me, Anthropic did not come up with this from first principles. Claude was pointed directly at Mononoke and asked to re-engineer it.
Not to be confused with Google's Scion (kubernetes based openclaw platform)
https://googlecloudplatform.github.io/scion/overview/