Anthropic's Scion is heavily inspired by Mononoke
72 points by mitsuhiko
72 points by mitsuhiko
It's hard for me not to see this as Anthropic effectively doing a "git clone" with extra steps.
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? I guess that doesn't juice the IPO valuation.
There's a reason no one uses Sapling at scale besides Meta. Just because it's open source doesn't mean it's practical to deploy on your own infrastructure. Using it and contributing back to the open source community (does it have a community?) would be much more work than forking. This is not because Sapling is bad, it's very impressive tech, but it's designed to meet Meta's needs specifically. The README says directly, "Not yet supported publicly, OSS is buildable for unsupported experimentation" for two of its three components (Mononoke and EdenFS).
If Anthropic vibe-coded a million lines of improvements to Sapling and dropped them in a PR (or many) I expect the Sapling project would not be overjoyed to accept those contributions.
I don't find this defense of these lab's continued vibe stunts compelling. They certainly could behave in a way that benefits rather than strip mines. That is a choice.
I think the idea that Anthropic had seven engineers spend five months building a custom VCS as a "vibe stunt" when they could just have easily used and contributed to an existing open source solution is unnecessarily cynical, personally.
It suggests that those engineers are doing bad work. I don't think that's a well supported conclusion.
I think that last statement is a poorly supported non sequitur. Whether they are doing good or bad engineering is entirely beside the point and (again) unverifiable.
Why would you imply that the only other way they could have handled this is be a bad open source citizen with a 1M line PR? And you are the one who first used "vibe-coded" as the hypothetical alternative.
We are talking about a company who's entire business model is intellectual labor arbitrage, hoovering up the entirety of human knowledge, distilling it, and renting it back to us.
You don't have to carry their water in every post.
For me, Occam's razor says that Anthropic looked at existing VCS options and decided none of them were the right fit for what they needed.
Two years ago, adopting an open source system and helping improve it may have been the best option.
This year, with unlimited Claude credits, I would likely have made the same decision they did. Contributing to an existing project is a lot more time consuming than building from scratch (thanks to the human coordination overhead), and time is the only resource they didn't have in abundance.
You don't have to carry their water in every post.
Contributing to an existing project is a lot more time consuming than building from scratch
This is yet another false dichotomy.
And time consuming for what?
What about bug fixes, training, growing the system, understanding, long term maintenance, shared value, openness, building something that lasts? These all have virtue and value. Subverting those concerns with yet another pull of the magic slot machine lever has a cost too.
LISP programmers know the value of everything and the cost of nothing. - Alan Perlis
It's this pervasive, frenetic, short term thinking that is hollowing out the heart and soul of our industry.
But, so what if we (literally) destroy a few thousand libraries worth of books, strip mine every open source repo, and extract every ounce of unmonetized value from knowledge work of the last 4 thousand years. Saying we are obfuscating another copy operation all while claiming novelty is just "unnecessarily cynical."
How does that hypothetical justify what they did? And how do we know the alternative was a 1M line slop PR? No way to know that because Scion isnt open source, is it?
I read this post and was angry on the author's behalf. Then I learned - not from the post - that in fact one of the things Anthropic did credit is the larger system of which Mononoke is a component. This has made me much less upset with Anthropic and moderately upset with this post's author for not stating this up front. I'm willing to believe that this was omitted only because it's hard to remember that these things aren't background knowledge to everyone, rather than e.g. to make this seem like a bigger deal to readers than it would otherwise (which is the effect this omission had on me), but I think this post ought to be clearer about that.
With this context, and having re-read Anthropic's post, I don't think Anthropic behaved badly here, unless there's even more context I'm still missing. Can someone who is upset even with this context spell out in a little more detail what you think the problem is?
From my post:
The fact that Sapling was only credited for “hard-won lessons”
The resemblance goes well beyond "hard-won lessons". It is a direct copy.
And yes, there is further information that isn't public that I was later made aware of. Sometimes you write to the world with the expectation that the world writes back to you, and sometimes that expectation is fulfilled.
They said "we were able to build Scion with [Piper's and Sapling's] hard-won lessons in mind". You would have liked them to say "we based Scion's design on Sapling's design"? Because to me, the former covers the latter, but if you don't read it that way I guess I can see why you'd like them to be more explicit. Still, this seems a minor faux pas at best (in the absence of the nonpublic information you allude to, which obviously I can't really update on unless you want to gesture at its shape somehow).
Since you're here, would you consider editing your post to make the connection between Sapling and Mononoke explicit so that future readers are not misled the way I was?
You would have liked them to say "we based Scion's design on Sapling's design"?
I would have liked them to say "we based the server design closely on (specifically) Mononoke's design", especially given the overlap in personnel. The reality is closer to "Claude was pointed directly to Mononoke to take its design from", though I didn't have confirmation of this before I wrote the post — reading the Anthropic post, nothing like that even remotely comes through. (It's a lot less impressive when the hard design work is already done and you copied it wholesale in five months!)
I was charitable when I assumed Claude came up with it from its training data.
Given that Anthropic credited Sapling, and Mononoke is a component of Sapling (the server part of it), I don't find this complaint particularly convincing.
If Anthropic hadn't credited Sapling I'd see their behavior as extremely bad form.
(I'd I was Anthropic I'd update the original announcement to credit Mononoke, too.)
Enough, Simon. You have carried enough water for these companies. This is nonsense. This is well beyond "hard-won lessons" — this is a direct copy made by (what I've been told) is an overlapping group of people. Given how contrarian the design was back then, and still is in many ways, it is basic decency to say that the data model is a direct copy (the word "ripoff" has been used).
I agree with you that they should credit the data model to Sapling and Mononoke.
I do not think that their failure to do so in the original post (that credited Sapling) was an outrageous lapse of judgement.
I have since privately received information that this was done in bad faith, and that I was far too generous in my post. Does that change your beliefs?
It would certainly change my beliefs if I had any evidence beyond your bare statement.
We do not have a trust relationship, so obviously I'm not going to share private correspondence with you. But the public evidence is what it is, and beyond that you'll need to take my word for it.
Yes. That's a scandal.
Thanks.
FYI in your post you say:
it looks like the core concept of "derived data" was heavily inspired by Mononoke, the server component of Sapling
I want to be clear that it isn't just the concept of derived data that is inspired by Mononoke. A Scion changeset is almost exactly the same as what we in Mononoke called a Bonsai changeset. I led the design and initial implementation work for Bonsai changesets back in 2018, including writing the four canonical design documents that are still the reference. The advantages of Bonsai changesets are very non-obvious until you spend a great deal of time thinking about the problem, which I had given my years of experience on version control.
This is not surprising given the additional information I received later, which includes the fact that Claude was pointed directly at Mononoke and told to copy it. The fact that none of this is even remotely hinted at is shameful; taking an existing battle-tested (over a decade!) design and copying it in 5 months is much less impressive than coming up with a brand new design and productionizing it in 5 months.
edit to add that there are also other parts such as an API to do source control queries and mutations with that are copied from Mononoke. Those have been done elsewhere before Mononoke, too. In general it is always good to situate your work in the larger context of what came before you.
First they brought noise and air pollution, and increased power bills to backwater towns, but I did not complain, because I didn't live in one of those towns, and my personal convinience outweighs the externalities.
The they came for the fruits of my intellectual labor, but there was no one left to speak out about it.
Those damn leopards, who would have thonk'd...
The fact that Sapling was only credited for “hard-won lessons”, and that Mononoke was not mentioned at all
These days Mononoke lives in sapling/eden/mononoke. Is it really a scandal to credit the parent project and not name the server component directly?
How different is Mononoke from Google's Piper (and the layers on top of it, of which there are several)? Is there anyone in the world familiar with both? I feel like it should be easy for some amount of conference evolution to happen here.