There’s nothing open model specific about this paper. They do leave it pretty open to people who are concerned about the Chinese government, but then again you don’t see anyone saying that people who advise people visiting China to use burner phones are Anthropic shills.
Or, for that matter, accusing people who recommend burner phones for visiting the US of shilling for China. Sometimes geopolitical caution is warranted.
FWIW, my biggest use case for agents like this is ongoing research. I.e.: check Humble Bundle weekly and tell me if there’s a book bundle I might be interested in. Has the Music Box announced a film series I might want to visit Chicago for? That sort of thing; I want filters for the firehose.
I like this and makes me think of making your own content algorithm profiles for the firehose. Let your agents scour the net for the things you already like, across platforms.
I use Claude for this stuff all the time. The difference is I have Claude create repeatable scripts which do the actual things on a schedule. Claude is never invoked in on a schedule, only by me. One example is job board search filtered for roles I'd be interested in. The script itself includes all the things I care about like being remote. Clear salary ranges. Type of work I want to be doing. Claude built a good script to scrape multiple sources and aggregate them into a sqlite database. It filters these out until a number of relevant roles (or amount of time has passed) that the script sends me a message and adds a reminder to my Apple account to check in on new job listings. At that point I'll invoke Claude myself to summarize and prioritize the listings. If there are gaps in what Claude finds vs the script, that's a point where I iterate again on the script. My goal is for the job alerts and filtering to be fully useful even if Claude and all LLM tools disappear tomorrow. And also not give such powerful tools the ability to run on a schedule on vague instructions.
Not really. I’ve got an RSS feed generator that I wrote to give Humble Bundle an RSS feed and it’s useful, but it doesn’t do any legwork for me. It doesn’t go out and search the Web for relevant reviews and send me a summary. It can’t serve as a preliminary attention filter and detect D&D bundles and screen those out.
The Internet firehose is vast. Tools that can help me filter it are valuable. Tools that filter poorly, of course, are bad, so I’m wary… but that’s still my potential use case.
Please re-read the policy. You will notice that it refers to “bodies of water” and “international geopolitical disputes.” Denali is neither a body of water or the subject of any international disagreement, given that it’s contained within the borders of the US. I believe Google just changed it, but you can see why the question is not in fact relevant to the current discussion.
The name of the mountain was changed. Note that mountains aren’t bodies of water, so the link is not relevant to this question, and Denali isn’t bordered by multiple countries, so Google presumably feels that the US has the right to name it.
And you’re getting downvoted because, as per the above, it’s an irrelevant question.
But I’ll indulge you just a bit more. The principle here is not “renaming things is bad!” The principle is that indulging an authoritarian who’s trying to unilaterally rename territory his country doesn’t own is bad, particularly when he’s doing so in a fit of pique (Lake Ontario).
As you may or may not be aware (I wasn’t), the Alaskan state legislature passed a bill in 1975 requesting that the mountain be renamed. Trump didn’t even try to get Congress to pass any bills before renaming things.
Hope this all helps clarify and gives you some reason to rethink the MAGA propaganda that tries to equate Denali and the Gulf of Mexico.
Not GP, but thank you those are interesting points.
However I think you're assuming/reading into the question a whole lot more than is there. Unfortunately your otherwise good answer comes across very hostile because you are assuming (rightly or wrongly) that the asker is parroting MAGA propaganda. I can assure you I am very much not MAGA, yet I thought it was a good question because it's interesting to see if/when Google stopped their 2008 policy and why. Your "why" makes sense to me, but you almost lost me with stuff like, "Hope this all helps clarify and gives you some reason to rethink the MAGA propaganda that tries to equate Denali and the Gulf of Mexico" which feels completely unwarranted.
I don’t think there’s a right or wrong about whether or not the asker is parroting MAGA propaganda; the Denali false equivalence is absolutely something that comes up all the time in right-wing circles. Possibly the asker didn’t intend to parrot MAGA propaganda. If so, well, perhaps there’s a lesson here for them about reading carefully before posting. I don’t think it’s unresonable to ask HN posters to read the article with some care and again, Denali is in no way an equivalent situation.
Well technically it also says that in Japan they only display "Sea of Japan" in the local language (Japanese), so displaying just "Gulf of America" in the local language follows. This blog never expected that the USA would be involved in this kind of dispute.
No because americans actually commonly call it the Gulf of Mexico (despite US government fiat dictating it should be called the Super Intelligent Gulf and anyone against that being double plus ungood), Japanese people don't commonly call it the East Sea. That's the difference.
Google Maps shows Gulf of Mexico (Gulf of America) to Canadians - it does properly label Lake Ontario, though. The idiocy isn't limited just to the US.
The Eugene CAHOOTS team has existed since 1989 until it was shut down recently. Dispatchers make the call as to whether or not to send police officers along with CAHOOTS responders on a per call basis.
Quoting vera.org: “ Of the estimated 24,000 calls CAHOOTS responded to in 2019, only 311 required police backup, and in Eugene, CAHOOTS teams resolved almost 20 percent of all calls coming through the city’s public safety communications center.”
To the best of my knowledge, no CAHOOTS staff have died or been seriously injured on the job. I’m not sure the evidence shows that finding the right balance is dangerous. Hard for sure, CAHOOTS put a lot of effort into building the right relationship with the police, but the risk to responders is overstated.
I lived in Eugene previously and CAHOOTS is (seems they are still operating in Springfield) quite effective in the limited response work they do, particularly de-escalation and triage of mental health crises. These are classic situations that (American) police are not needed nor are most properly trained for, and several crisis intervention teams modelled after CAHOOTS have been stood up elsewhere in the US in the past several years (e.g. BCR in Minneapolis) with seemingly good results.
However, CAHOOTS is selectively dispatched by 911 and is only sent to a small percentage of situations which would otherwise overlap with police. (Most of their calls are more general assistance to the houseless folks/public where no police response would occur.) This is largely due to very strict “low-risk” criteria to not send police. This may be appropriate to protect all parties since CAHOOTS is not as equipped or qualified as other first responders — CAHOOTS only receives partial city funding, notably through the police department, and they are not equipped as EMTs/paramedics.
But this also means if you witness someone in crisis and call 911, you cannot be certain that CAHOOTS will show up and not the police, because as the article points out, 911 is set up to send police as the default and will likely do so whenever in doubt. This can clearly have life-or-death consequences for the person in crisis, and we won’t understand the full potential of these intervention teams until they are decoupled from the police and treated as peers with a discrete scope of responsibility.
“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.
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.
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.
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.
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.
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.
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.
> 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.
> 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 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.
"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."
This article asks a bunch of questions without ever bothering to try and answer them. That makes it less than useless; it’s mostly insinuations.
There’s also a lack of historical knowledge. Sure, mass market paperbacks were a new format — but not completely new in 1939. Cheap paperbacks date back to the previous century. The author of this piece would have done well to spend some time with Wikipedia.
For that matter, showing some curiosity about reading habits in general would have helped. Reading wasn’t a new habit: dime novels and pulp magazines were deeply popular in the early 20th century. The jump to mass market paperbacks was not particularly long.
A funny thing to vote down as it’s so well known. Dime novels, TV, video games, social media…anything new is a candidate to be attacked as corrupting “kids these days”. Even Socrates (recounted in Phaedrus) said that writing would be the downfall of students.
Yeah, this was my thought as well. 1939 would be a very late date for books to become affordable for the first time.
Virtually all the 19th century novels on the Polish school curriculum first appeared in chapters published in newspapers, appealing to mass reader and yet still managing to touch important social and political issues, often under disguise.
Here in Germany where I live now, the situation is different because of protectionist policies towards publishing (it's illegal to sell a book below the asking price of the publisher). There are publishers specializing in affordable editions of classics (Reclam, Nikol) and some publishers run their titles in multiple paperback editions with different prices, but obtaining any recent work is a costly endeavor.
You’re assuming the “listening in” aspect, which as far as I know is not proven (even in the less loaded “the model was trained on sessions including the ones in question” form of the claim.
Plausible? Absolutely. Did OpenAI behave badly in other ways regarding this issue? Yes. Does it help to assume unproven facts and then accuse people of reaching emotional decisions? Nope.
Maybe the conditions of the singularity are not going to be recognized, at first, for having produced new original decisions, and thus thought.
Perhaps it is going to be more like, we will see the singularity predict the future.
While “the model was trained on sessions including the ones in question” is one aspect; ‘the model produced an accurate prediction of future human thought’, I think, is another very interesting facet.
Models predicting future things might be how we see, actually, how we ourselves formulate thought - by saying, in big and small words, ‘something is about to happen’.
>unproven facts
This isn’t a court room. We’re discussing ways by which we humans both succeed and fail at reigning in our creations. The OpenAI kerfuffle is pretty much irrelevant already. Of course AI will fill in the gaps of human thought - it is literally constructed from the stuff, in every squeeze of the curd and whey.
Or, for that matter, accusing people who recommend burner phones for visiting the US of shilling for China. Sometimes geopolitical caution is warranted.
reply