2/3 down the article I gave up. What are you trying to tell me? What is this revolution about? How are agent things fundamentally different and how are they being solved? What is this "second coming"?
Also, why "second"? Was git the first? But then what about all the other things before it? CVS was huge before, for better or worse.
The author has a severe case of being oblivious to the fact that not everyone knows what they know, or views the world the way they do. As shown for example by their assertion that "we all stopped coding manually around December 2025", as if AI critics don't exist.
As a result you have to look at what they share about themselves to infer where they're going with things, because they never explicitly state it. All traits amplified for the worse when people use chat bots too much, in my experience.
Scrolling to the end, they write:
> In 2005 a dozen version control systems fought to replace Bitkeeper, and the winner ended up ruling software development for two decades. Nobody in that race predicted that the decisive factor would be a hosting site with a social network on top.
They had a VCM start-up around their own technology, then Git displaced all other technologies, and then Github ate everyone's lunch. So the first revolution not-so-coincidentally overlaps with the time they ran a SCM startup, and they ignore anything that happened before they entered the space. Funny that.
So the implicit context is the business case and programmer culture around VCM systems.
Meaning the second revolution is anything that challenges Git and Github.
They also keep bringing up gigantic mono-repos and how Git can't handle the scale at which AI increases the amount of code bloat or commits.
So what I read between the lines is that they believe that there are enough people who don't want to actually address the automated hyperscaled Wirth's Law in the room, and are willing to pay services and technology to adapt around it instead.
That's their new business case that will power the second "revolution".
Which sounds horrifying to me, and the worst part is that I can actually see it happening if this AI bubble doesn't pop fast enough.
> They also keep bringing up gigantic mono-repos and how Git can't handle the scale at which AI increases the amount of code bloat or commits.
I don't understand why everyone keeps saying "AI" can write code and somehow the rest of the world can't keep up when we haven't even solved a more basic problem that "AI" itself can't keep up with the code it writes. Claude on the web has to search and scan to find the relevant parts and for whatever reason there is a size limit to how much context I can give to claude and it is laughably low. Shouldn't we be trying to see if it can keep up with itself first? Also not all project files fit in the "context" so once you are in search mode anyway, why is the limit so arbitrarily small? Why not allow at least about 1GiB of plain text? Claude tells me 1 GiB is 250 million tokens and it is too much for current capacity but I thought the whole point of project files was not all of it is in the context window or whatever. It says it is in search mode anyway... So only search for the relevant bit and put only that in the context. You can limit your search to a particular submodule or something but still have broader context where required
>> > They also keep bringing up gigantic mono-repos and how Git can't handle the scale at which AI increases the amount of code bloat or commits.
I also doubt very much that the average monorepo is bigger than the linux kernel. While the latter is a single software as things go, it contains quasi independent subsystems. And those repos don't usually have a merge flow that is as smooth as the Linux kernel. If you have chaotic process, then the result won't be very good and it's not git's fault.
I read it yesterday and had a discussion on Git Discord. The author believes people will slop Google-scale repos and Google-type SCM is needed now.
Some say a USB-C cable has more computing power than the Apollo program, so maybe.
Banks, airlines, power infra, and other high-stakes customers got computerized in the 60s and 70s. IBM still holds a chunk of that market. Microsoft and Google are not going anywhere. So I guess all that slop will serve something less critical than all of the above.
My Claude says that a decent cable has an ARM Cortex-M0 running at tens of MHz, while the Apollo Guidance Computer was 1MHz. I am afraid to think what a charger packs.
Well, I like a good story, there is big potential in this long form stuff that's more than a tweet or two. There can be a great reward when the crux is revealed after building a foundation and then looking at it from different angles. But if that crux never comes then it's really just leaving a feeling of disappointment and waste of time that I invested into reading all this. And that does a disservice to everybody else writing blog posts.
I think that's part of why tiktok and yt shorts really took off. It's so short that if you realize it's bad and move on then your wasted investment in terms of time and energy was negligible. (Times 200 that's a different story but signal-to-noise is still high enough that people accept this.)
FWIW, I flagged and downvoted both of your comments. The article is a very interesting overview of the space and recent developments, and having to scroll past multiple paragraphs about your low attention span and reading difficulties to get to some actual technical discussion was quite frustrating!
Is flagging and downvoting really the right tool to express disagreement of opinion here? I disagree very much with what you just wrote, but I would not flag it.
For the next time, instead of scrolling, you can just press the little "-" next to a post which collapses it and its whole subtree. Out of sight.
Besides, I explicitly stated that I read 2/3 of the article. How is that "low attention span"? And I didn't just bash it, I formulated what I find missing in it. Do you have answers to those questions? And how does asking such questions indicate "reading difficulties"?
It's quite fascinating that if my comment is so appaling to you, why you'd take the time to respond. I thank you for that though.
Plan 9 had such a powerful model for networked systems using these virtual file systems, it sounds like a fairytale!
Oh, want to use that other machine as a gateway? Just mount its /net.
Oh, want to route audio through another machine? Just mount their soundcard into your /dev.
Oh, your machine is too puny to do the task at hand? Just run “cpu thebigmachine” which transplanted your entire environment over there (all the virtual file systems) so that you can continue doing what you were doing, but using that machine’s CPU and memory.
This solved the problem of having to transplant your setup to the remote machine, which you have with modern SSH. If you wanted a different environment you instead created it locally. Each process har its own virtual file tree with mounts.
There were cool things at the local level too: All the programs would expose virtual file systems to interact with. Text editor? Each window had a directory with files containing window content, current selection, even the UI “tagline” with commands. This meant you could write scripts for your programs in any language, because you just had to interact with files.
A modern take on plan 9 is definitely on my Christmas wishlist!
Have a look at Objective-Smalltalk[1][2], with polymorphic identifiers[3] (all identifiers are URIs), storage combinators[4] (virtual filesystems on steroids), and polymorphic write streams[5] (streams everywhere).
Objective-Smalltalk basically provides the sorts of things you write about locally at the language level. You could also move specific instances behind an operating system boundary.
Could you explain the connection you see a bit more?
Part of the appeal of the Plan 9 approach was that you could use any program in your distributed environment, written in any language, because the abstraction layer was the file system – the lingua Franca of IO.
I wonder if we're doing virtual filesystems wrong.
There is a good reason why traditionally filesystem access was mediated by the OS layer, but there are many use cases where you just want to give processes a different view of what they already can access and it could be done as a library in the same userspace process.
However, for that to work across all the processes in a session we'd need a standard way to install such a hook in all peocesses and that's achievable to some extent using LD preload but falls apart quite rapidly with statically built binaries or different libcs
Sending -asScheme to a reference is just a shorthand that actually constructs a composition[2] of a "path relative" store with the underlying store of the original reference. So the following two are identical:
This composition mechanism can be carried further with post-processing, so for example an img-scheme can be constructed by composing an image-decoder store with the previous store
Git is incredibly mediocre. But it's all most people know. It's a version control tool that can't handle binary files; and no GitLFS does not count. The end result is a version control tool that is unable to actually version control all the things you need for a project.
This results in a Meta VCS layer where a ton of critical assets are stored in Docker files and other misery. If you want to re-compile a project for 2015 then good luck and god speed.
Personally I think full toolchains belong in source control. And that you should be able to clone / materialize a repro, yank your network cable, and build. This is how big tech monorepos work. It is TheWay imho.
IMO screw that. It's maybe a good way to build software in exactly one environment for exactly one environment, deployment to a corporate server fleet.
Consider a Linux desktop distro: if every little binary (out of order of magnitude 1000) acted like the center of the universe with gigabytes of build environment and "opinions" galore instead of portability, builds would take much more resources than they already do and parts wouldn't necessarily work together.
Visual Studio and Xcode take up tens of gigabytes, are updated often, and include system components. Storing them in VCS is impossible, and would be a waste of disk space.
Disagree. They’re stored _somewhwre_ anyway, and they may as well be versioned.
Putting toolchains in perforce is how it works for lots of C++ shops, the setup instructions are “sync and hit build”, whether there’s a toolchain upgrade required or not
It sounds like the author went through a lot of pain to avoid using git+lfs or perforce.
I use git+lfs for unity projects and it works out great. If I had a real studio I'd buy some perforce seats.
Reinventing the wheel like this is quite exhausting. There are options that are proven to work. AI authorship does not fundamentally violate the idea of some thing owning a specific commit. We don't need new schemas in our source control system. "Provenance" is a bullshit word used to make the AI sound like it's some kind of oracular source with superhuman capabilities.
From my experience Git LFS is extremely brittle and will break your local check out if you so much as breath on it wrong. Perforce is an expensive solution to the problems of LFS, and I've yet to find a workflow as powerful as my git workflows for managing code.
I don't really see what the article is talking about as the future, but if P4 or Git LFS is the best we can do as a species then we're doomed. All VCS options suck for one reason or another, I hope we don't stop trying to make something better. If only to save me from perforce.
Love the enthusiasm but I don't buy the two macro tailwinds he's counting on: ever larger monorepos & more centralisation.
What you version together you build and release together, and there are architectural tensions pushing that size down. E.g. dependency indirection and change frequency. Mileage will surely vary by domain, but the idea that the "future is monorepo because agents" doesn't track with me.
The centralisation aspect has less to do with connectedness and more with topology, I'd say. Here, the agentic ways might actually push towards more hierarchical and distributed topologies than the centralised hub-and-spoke.
I helped swap a mess of ClearCase & Subversion over to git (early 2010s). What a fucking cursed piece of software!
Will admit it got exactly one thing right, which could absolutely justify everything else (including the price) for a long time: Software Bill of Materials, which is non negotiable in a lot of more serious/heavily regulated development contexts
That still didn't excuse the hacked together pile of ruby scripts our QA lead maintained, each one designed to unfuck a specific weird thing clearcase did. Once everything was on git, they just wrote another script that grabbed a bunch of tags from git and shoved them into clearcase to spit out the bom.
The author praises the enterprise version vs. the more simpler one. Having had the pleasure coordinating one clearcase site in a very large distributed setup with significant replication traffic and windows and unix clients: It was an unstable mess.
> If only 2 years ago somebody said GitHub would be no longer relevant soon, nobody would believe them. GitHub was the undisputed leader in repository hosting [..]
> GitHub has done too many great things over the years, so I hope it remains, but there is obviously an earthquake going on.
Is the 'earthquake' referring to the frequent outages over the last couple of years, and if so are those outages because Microsoft can't keep up with the demand?
It's closer to 'too relevant' than it is to 'no longer relevant'.
The author mentions it briefly, but I'd like to see more exploration of what we can do differently now that we're all online all the time. I think that could potentially enable a Git-level revolution in the way we do things, similar to the way unlimited disk space did for Git.
I half agree here. Now that the primary usage of Git is increasingly centered on providing auditable and revertible history for organizations that use a mix of AI (mostly) and people to do stuff with it, the requirements are going to shift away from being user friendly towards just being fit for purpose and efficient.
Git is so far good enough for this. It's not particularly user friendly. But that's not a problem for AI agents. What is a problem is that GitHub is a shared resource that is bottle necked on massively increased usage. That's nice if you are sharing code with other people but it becomes a bottleneck otherwise with a clear solution in the form of maybe using faster and compatible (or completely different) alternatives that do things faster/better.
If you sit back and watch what agents do with Git, it involves a lot of agents going through the moves of creating lots of pull requests, waiting for whatever CI systems to kick in, dealing with failures, etc. All that takes a lot of time and tokens and it's designed to compensate for human failures to properly follow processes. So, at least some of that is kind of becoming redundant. With AI we can compensate with more complicated processes instead.
There's definitely some optimization potential lurking there. If you have tens of thousands of agents working on a thing, it might be more efficient to share the burden of integration testing instead of each agent trying to do this independently and testing each micro change in isolation. Also you could question the logic of needing some centralized hub to dump and integrate code. Git is decentralized by design. GitHub is nice as a backup strategy but there are probably cheaper or different ways to do QA and integration with agents. As the development process changes and adapts to all this, the role of Git and Github also needs to be rethought.
As for the rest of the article, it seems a bit too people centric. Virtual file systems are cool. But do AI agents really need them?
The distributed repos and "commit-then-push" metaphor might be due for a replacement since we're always online anyways and repos grow larger and larger. Perhaps using VFS where all files are always instantly available with copy-on-write semantics.
There are several players working on systems that work on thousands of commits per second scale, some based on Git and others not.
Now that the revolution has apparently started, does anyone know of a version control system which uses encrypted storage? (I'm not concerned about performance, for I will use it for small projects only.)
I can accept the authors bias, but all credibility was gone when I read “AccuRev was a fantastic system”. AccuRevs UX was akin to putting your face next to a farting anus and breathing deeply.
Honestly felt like I was missing some sort of satirical masterpiece as the author gleefully declared all the up and coming projects that are going to scatter open source projects to the four corners of the Earth.
Maybe they’ve been so locked into version control tooling as a career they’ve missed the bigger picture, but version control before git ubiquity sucked. Not saying Git is perfect but come on, are our memories so short?
> GitHub is still gigantic in both repos and minds, and its impact in the world of software development and collaboration in general has been enormous, and most likely it will continue to be.
Well - Microsoft worsened GitHub in the last few months with regards to reliability. The next corporate slayer move is to worsen it feature wise. GitHub will indeed most likely remain strong, but the outer shell has some cracks and that means people will be more eager to look for alternatives. It is also a problem that Microsoft has a say in open source projects via GitHub - I never liked that, and many others did not like that either, even more so with Trump acting in a political and ideological manner with his TechBros (who all have nothing to do with Epstein ... right? because what if some of them do ...).
TL;DR better scale and concurrency. But really, just for for human-assisted AI swarms? I guess this is a cool concept if you think everyone will be handling 100+ agent swarms regularly?
It's great to see some moves to break up this great calcification around git. Gut greatest contribution to version control was stagnation. The space was evolving with great fresh ideas before git became a quasi-religion among the early adopters because Linus made it in a day and therefore it must be great or something. I'm exaggerating somewhat, but the zeal of some people back then was next level annoying.
Beyond that, git has a great deal of shortcomings, some obvious, some subtle. It was a regression against SVN in some ways and inferior to Mercurial in others. But the strengths of SCN and Mercurial are again vastly different. There is a reason SVN isn't dead.
My biggest gripe with the open source VCSes is that in the last 20 years, no meaningful evolution happened in the established tools, especially around any weaknesses. Commercial systems like Plastic and Perforce as well as proprietary solutions like piper/jujutsu and sapling show that the tools can still improve drastically. I'm excited for a future where we get better open tools for the masses.
I don't think jujutsu is proprietary (unless you want to redefine that term). Source is freely available, and it uses Apache License 2.0.
While you're right about the disadvantages of git, pretending that it became ubiquitous because it "became a quasi-religion because Linus made it in a day" is selling short its advantages. If you think about it for even just a little bit, it should be obvious what a simplistic statement that it. Also, git was not the stagnation you make it out to be. Even with its warts, it was a breath of fresh air, not unlike jujutsu is now a breath of fresh air vs git.
I remember working with SVN, and all things considered, git was a vast net improvement. Git took a lot of pain away. It made working with a versioned code base faster and simpler, to the point of enabling much better collaborative software development. There's a reason we got GitHub and not SVNhub. And GitHub was what helped git become so dominant.
Would it have been better if Mercurial had beat out git in the propularity contest? Possibly? There's trade-offs between the two, but Mercurial's easier interface counts for a lot. But if it had won, I'm sure we'd be griping about its shortcomings by now.
So yeah, I'm also happy to see some movement around the ergonomics of version control, but I don't understand the need to disparage the tools that got us where we are. It just seems that you're more bitter than happy, and like you're letting that bitterness cloud your judgment.
The combination jujutsu/piper is proprietary. As I understand it, sapling as released is also not the entire internal system used at Meta.
Git was an improvement over SVN in some aspects, but a monumental regression in others. There are no two ways around that.
Mercurial is better than git in a bunch of ways. It is much more usable because it has way fewer footguns. But it has some of the same shortcomings as git when compared to SVN. I won't call it perfect.
The reason we didn't get svnhub is that some git fanboy nabbed the domain and essentially said "you shall not have it". Joking aside, there was Sourceforge with working SVN support. But I would say that Sourceforge lost market share for a lot of reasons, not all of them technical. Wrapping Windows downloads in adware installers was only one of those many crazy unforced errors.
Am I bitter? Maybe, because I have to use tools that could be so much better. I've experienced better and every time I am forced back to git, it feels a bit painful because I know what we could have instead. If am bitter, it is because of my experience with a wide range of tools.
2/3 down the article I gave up. What are you trying to tell me? What is this revolution about? How are agent things fundamentally different and how are they being solved? What is this "second coming"?
Also, why "second"? Was git the first? But then what about all the other things before it? CVS was huge before, for better or worse.
The author has a severe case of being oblivious to the fact that not everyone knows what they know, or views the world the way they do. As shown for example by their assertion that "we all stopped coding manually around December 2025", as if AI critics don't exist.
As a result you have to look at what they share about themselves to infer where they're going with things, because they never explicitly state it. All traits amplified for the worse when people use chat bots too much, in my experience.
Scrolling to the end, they write:
> In 2005 a dozen version control systems fought to replace Bitkeeper, and the winner ended up ruling software development for two decades. Nobody in that race predicted that the decisive factor would be a hosting site with a social network on top.
They had a VCM start-up around their own technology, then Git displaced all other technologies, and then Github ate everyone's lunch. So the first revolution not-so-coincidentally overlaps with the time they ran a SCM startup, and they ignore anything that happened before they entered the space. Funny that.
So the implicit context is the business case and programmer culture around VCM systems.
Meaning the second revolution is anything that challenges Git and Github.
They also keep bringing up gigantic mono-repos and how Git can't handle the scale at which AI increases the amount of code bloat or commits.
So what I read between the lines is that they believe that there are enough people who don't want to actually address the automated hyperscaled Wirth's Law in the room, and are willing to pay services and technology to adapt around it instead.
That's their new business case that will power the second "revolution".
Which sounds horrifying to me, and the worst part is that I can actually see it happening if this AI bubble doesn't pop fast enough.
> They also keep bringing up gigantic mono-repos and how Git can't handle the scale at which AI increases the amount of code bloat or commits.
I don't understand why everyone keeps saying "AI" can write code and somehow the rest of the world can't keep up when we haven't even solved a more basic problem that "AI" itself can't keep up with the code it writes. Claude on the web has to search and scan to find the relevant parts and for whatever reason there is a size limit to how much context I can give to claude and it is laughably low. Shouldn't we be trying to see if it can keep up with itself first? Also not all project files fit in the "context" so once you are in search mode anyway, why is the limit so arbitrarily small? Why not allow at least about 1GiB of plain text? Claude tells me 1 GiB is 250 million tokens and it is too much for current capacity but I thought the whole point of project files was not all of it is in the context window or whatever. It says it is in search mode anyway... So only search for the relevant bit and put only that in the context. You can limit your search to a particular submodule or something but still have broader context where required
>> > They also keep bringing up gigantic mono-repos and how Git can't handle the scale at which AI increases the amount of code bloat or commits.
I also doubt very much that the average monorepo is bigger than the linux kernel. While the latter is a single software as things go, it contains quasi independent subsystems. And those repos don't usually have a merge flow that is as smooth as the Linux kernel. If you have chaotic process, then the result won't be very good and it's not git's fault.
> Wirth's law is an adage on computer performance which states that software is getting slower more rapidly than hardware is becoming faster.
So, datacenters will fix that.
Is this an ironic "tell me you don't learn from history without telling me you don't learn from history" joke?
I read it yesterday and had a discussion on Git Discord. The author believes people will slop Google-scale repos and Google-type SCM is needed now. Some say a USB-C cable has more computing power than the Apollo program, so maybe.
Banks, airlines, power infra, and other high-stakes customers got computerized in the 60s and 70s. IBM still holds a chunk of that market. Microsoft and Google are not going anywhere. So I guess all that slop will serve something less critical than all of the above.
USB-C charger, not cable.
My Claude says that a decent cable has an ARM Cortex-M0 running at tens of MHz, while the Apollo Guidance Computer was 1MHz. I am afraid to think what a charger packs.
Same. I had no idea if the article was for me and what I was going to get out of it, so I just gave up.
I gave up after 5 seconds.
Blog authors need a straight up TL;DR because I ain't reading all of that.
Well, I like a good story, there is big potential in this long form stuff that's more than a tweet or two. There can be a great reward when the crux is revealed after building a foundation and then looking at it from different angles. But if that crux never comes then it's really just leaving a feeling of disappointment and waste of time that I invested into reading all this. And that does a disservice to everybody else writing blog posts.
I think that's part of why tiktok and yt shorts really took off. It's so short that if you realize it's bad and move on then your wasted investment in terms of time and energy was negligible. (Times 200 that's a different story but signal-to-noise is still high enough that people accept this.)
FWIW, I flagged and downvoted both of your comments. The article is a very interesting overview of the space and recent developments, and having to scroll past multiple paragraphs about your low attention span and reading difficulties to get to some actual technical discussion was quite frustrating!
Is flagging and downvoting really the right tool to express disagreement of opinion here? I disagree very much with what you just wrote, but I would not flag it.
For the next time, instead of scrolling, you can just press the little "-" next to a post which collapses it and its whole subtree. Out of sight.
Besides, I explicitly stated that I read 2/3 of the article. How is that "low attention span"? And I didn't just bash it, I formulated what I find missing in it. Do you have answers to those questions? And how does asking such questions indicate "reading difficulties"?
It's quite fascinating that if my comment is so appaling to you, why you'd take the time to respond. I thank you for that though.
There is no reason to flag this.
This is my dilemma. Do I feel like giving up because it isn’t written well or because my attention span is shot? Hard to tell.
It isn't written well.
fully agree. cant even started to read this. just skipped directly.
You should try TikTok, I bet you’ll like it.
Dreaming of "virtual filesystems everywhere". Hmm, sorta sounds like Plan 9.
Great to have an inside view of wrangling technologies for these behemoth data sets.
Plan 9 had such a powerful model for networked systems using these virtual file systems, it sounds like a fairytale!
Oh, want to use that other machine as a gateway? Just mount its /net.
Oh, want to route audio through another machine? Just mount their soundcard into your /dev.
Oh, your machine is too puny to do the task at hand? Just run “cpu thebigmachine” which transplanted your entire environment over there (all the virtual file systems) so that you can continue doing what you were doing, but using that machine’s CPU and memory.
This solved the problem of having to transplant your setup to the remote machine, which you have with modern SSH. If you wanted a different environment you instead created it locally. Each process har its own virtual file tree with mounts.
There were cool things at the local level too: All the programs would expose virtual file systems to interact with. Text editor? Each window had a directory with files containing window content, current selection, even the UI “tagline” with commands. This meant you could write scripts for your programs in any language, because you just had to interact with files.
A modern take on plan 9 is definitely on my Christmas wishlist!
Have a look at Objective-Smalltalk[1][2], with polymorphic identifiers[3] (all identifiers are URIs), storage combinators[4] (virtual filesystems on steroids), and polymorphic write streams[5] (streams everywhere).
Objective-Smalltalk basically provides the sorts of things you write about locally at the language level. You could also move specific instances behind an operating system boundary.
[1] https://objective.st
[2] https://dl.acm.org/doi/10.1145/3689492.3690052
[3] https://dl.acm.org/doi/10.1145/2508168.2508169
[4] https://2019.splashcon.org/details/splash-2019-Onward-papers...
[5] https://dl.acm.org/doi/10.1145/3359619.3359748
Could you explain the connection you see a bit more?
Part of the appeal of the Plan 9 approach was that you could use any program in your distributed environment, written in any language, because the abstraction layer was the file system – the lingua Franca of IO.
Oh wow, I didn't know it had things like that.
Especially the CPU functionality is really interesting. I'm pondering similar problems and the current solutions just aren't good enough.
Redox seems to take inspiration from it and has all things exposed as schemes so /home/user is actually /scheme/file/home/user, etc.
I wonder if we're doing virtual filesystems wrong.
There is a good reason why traditionally filesystem access was mediated by the OS layer, but there are many use cases where you just want to give processes a different view of what they already can access and it could be done as a library in the same userspace process.
However, for that to work across all the processes in a session we'd need a standard way to install such a hook in all peocesses and that's achievable to some extent using LD preload but falls apart quite rapidly with statically built binaries or different libcs
I think there is a good case to be made for these things not to be mediated by the operating system by default.
In Objective-Smalltalk[1], I can access a file as follows:
This is structurally the same way I would access a local variable, environment variable, database, remote http server etc. etc.And you can also introduce shortcuts
Or Sending -asScheme to a reference is just a shorthand that actually constructs a composition[2] of a "path relative" store with the underlying store of the original reference. So the following two are identical: This composition mechanism can be carried further with post-processing, so for example an img-scheme can be constructed by composing an image-decoder store with the previous store And so on and so forth, caching is also a nice example.[1] https://objective.st
[2] https://dl.acm.org/doi/10.1145/3359591.3359729
> virtual filesystems everywhere
Please universe I beg you.
Git is incredibly mediocre. But it's all most people know. It's a version control tool that can't handle binary files; and no GitLFS does not count. The end result is a version control tool that is unable to actually version control all the things you need for a project.
This results in a Meta VCS layer where a ton of critical assets are stored in Docker files and other misery. If you want to re-compile a project for 2015 then good luck and god speed.
Personally I think full toolchains belong in source control. And that you should be able to clone / materialize a repro, yank your network cable, and build. This is how big tech monorepos work. It is TheWay imho.
IMO screw that. It's maybe a good way to build software in exactly one environment for exactly one environment, deployment to a corporate server fleet.
Consider a Linux desktop distro: if every little binary (out of order of magnitude 1000) acted like the center of the universe with gigabytes of build environment and "opinions" galore instead of portability, builds would take much more resources than they already do and parts wouldn't necessarily work together.
Visual Studio and Xcode take up tens of gigabytes, are updated often, and include system components. Storing them in VCS is impossible, and would be a waste of disk space.
Disagree. They’re stored _somewhwre_ anyway, and they may as well be versioned.
Putting toolchains in perforce is how it works for lots of C++ shops, the setup instructions are “sync and hit build”, whether there’s a toolchain upgrade required or not
You could consider ZFS a VCS, and it can easily store multiple versions (snapshots) of Visual Studio.
It's not impossible, there just isn't that much demand for it.
What second coming? If I had the option I would still be using either Mercurial or SVN.
In fact, the way I use Git is hardly any different, I have no interest in getting a black belt in git magic.
It sounds like the author went through a lot of pain to avoid using git+lfs or perforce.
I use git+lfs for unity projects and it works out great. If I had a real studio I'd buy some perforce seats.
Reinventing the wheel like this is quite exhausting. There are options that are proven to work. AI authorship does not fundamentally violate the idea of some thing owning a specific commit. We don't need new schemas in our source control system. "Provenance" is a bullshit word used to make the AI sound like it's some kind of oracular source with superhuman capabilities.
From my experience Git LFS is extremely brittle and will break your local check out if you so much as breath on it wrong. Perforce is an expensive solution to the problems of LFS, and I've yet to find a workflow as powerful as my git workflows for managing code.
I don't really see what the article is talking about as the future, but if P4 or Git LFS is the best we can do as a species then we're doomed. All VCS options suck for one reason or another, I hope we don't stop trying to make something better. If only to save me from perforce.
Love the enthusiasm but I don't buy the two macro tailwinds he's counting on: ever larger monorepos & more centralisation.
What you version together you build and release together, and there are architectural tensions pushing that size down. E.g. dependency indirection and change frequency. Mileage will surely vary by domain, but the idea that the "future is monorepo because agents" doesn't track with me.
The centralisation aspect has less to do with connectedness and more with topology, I'd say. Here, the agentic ways might actually push towards more hierarchical and distributed topologies than the centralised hub-and-spoke.
People complained about having to use ClearCase back then, it wasn't just the price that was wrong with it.
I helped swap a mess of ClearCase & Subversion over to git (early 2010s). What a fucking cursed piece of software!
Will admit it got exactly one thing right, which could absolutely justify everything else (including the price) for a long time: Software Bill of Materials, which is non negotiable in a lot of more serious/heavily regulated development contexts
That still didn't excuse the hacked together pile of ruby scripts our QA lead maintained, each one designed to unfuck a specific weird thing clearcase did. Once everything was on git, they just wrote another script that grabbed a bunch of tags from git and shoved them into clearcase to spit out the bom.
> Software Bill of Materials
Yet another thing that came from PRIDE, despite that PRIDE itself is little known in today's software industry.
The author praises the enterprise version vs. the more simpler one. Having had the pleasure coordinating one clearcase site in a very large distributed setup with significant replication traffic and windows and unix clients: It was an unstable mess.
> If only 2 years ago somebody said GitHub would be no longer relevant soon, nobody would believe them. GitHub was the undisputed leader in repository hosting [..]
> GitHub has done too many great things over the years, so I hope it remains, but there is obviously an earthquake going on.
Is the 'earthquake' referring to the frequent outages over the last couple of years, and if so are those outages because Microsoft can't keep up with the demand?
It's closer to 'too relevant' than it is to 'no longer relevant'.
The author mentions it briefly, but I'd like to see more exploration of what we can do differently now that we're all online all the time. I think that could potentially enable a Git-level revolution in the way we do things, similar to the way unlimited disk space did for Git.
“and a sudden urge to replace GitHub.” Says someone making a blog post on github…
what everyone wants is the code in those repos to train their AI, or train someone else's AI
I half agree here. Now that the primary usage of Git is increasingly centered on providing auditable and revertible history for organizations that use a mix of AI (mostly) and people to do stuff with it, the requirements are going to shift away from being user friendly towards just being fit for purpose and efficient.
Git is so far good enough for this. It's not particularly user friendly. But that's not a problem for AI agents. What is a problem is that GitHub is a shared resource that is bottle necked on massively increased usage. That's nice if you are sharing code with other people but it becomes a bottleneck otherwise with a clear solution in the form of maybe using faster and compatible (or completely different) alternatives that do things faster/better.
If you sit back and watch what agents do with Git, it involves a lot of agents going through the moves of creating lots of pull requests, waiting for whatever CI systems to kick in, dealing with failures, etc. All that takes a lot of time and tokens and it's designed to compensate for human failures to properly follow processes. So, at least some of that is kind of becoming redundant. With AI we can compensate with more complicated processes instead.
There's definitely some optimization potential lurking there. If you have tens of thousands of agents working on a thing, it might be more efficient to share the burden of integration testing instead of each agent trying to do this independently and testing each micro change in isolation. Also you could question the logic of needing some centralized hub to dump and integrate code. Git is decentralized by design. GitHub is nice as a backup strategy but there are probably cheaper or different ways to do QA and integration with agents. As the development process changes and adapts to all this, the role of Git and Github also needs to be rethought.
As for the rest of the article, it seems a bit too people centric. Virtual file systems are cool. But do AI agents really need them?
AI makes running a deterministic workflow after a code change useless, because unlike humans, agents are not fallible, got it.
TLDR for those who gave up reading:
The distributed repos and "commit-then-push" metaphor might be due for a replacement since we're always online anyways and repos grow larger and larger. Perhaps using VFS where all files are always instantly available with copy-on-write semantics.
There are several players working on systems that work on thousands of commits per second scale, some based on Git and others not.
Not worth reading.
Funny this is hosted on GitHub pages
> While we all stopped coding manually around December 2025
Come on at least say "many people". There are plenty of people still coding by hand.
Now that the revolution has apparently started, does anyone know of a version control system which uses encrypted storage? (I'm not concerned about performance, for I will use it for small projects only.)
Git on an encrypted volume?
I can accept the authors bias, but all credibility was gone when I read “AccuRev was a fantastic system”. AccuRevs UX was akin to putting your face next to a farting anus and breathing deeply.
Honestly felt like I was missing some sort of satirical masterpiece as the author gleefully declared all the up and coming projects that are going to scatter open source projects to the four corners of the Earth.
Maybe they’ve been so locked into version control tooling as a career they’ve missed the bigger picture, but version control before git ubiquity sucked. Not saying Git is perfect but come on, are our memories so short?
> GitHub is still gigantic in both repos and minds, and its impact in the world of software development and collaboration in general has been enormous, and most likely it will continue to be.
Well - Microsoft worsened GitHub in the last few months with regards to reliability. The next corporate slayer move is to worsen it feature wise. GitHub will indeed most likely remain strong, but the outer shell has some cracks and that means people will be more eager to look for alternatives. It is also a problem that Microsoft has a say in open source projects via GitHub - I never liked that, and many others did not like that either, even more so with Trump acting in a political and ideological manner with his TechBros (who all have nothing to do with Epstein ... right? because what if some of them do ...).
TL;DR better scale and concurrency. But really, just for for human-assisted AI swarms? I guess this is a cool concept if you think everyone will be handling 100+ agent swarms regularly?
It's great to see some moves to break up this great calcification around git. Gut greatest contribution to version control was stagnation. The space was evolving with great fresh ideas before git became a quasi-religion among the early adopters because Linus made it in a day and therefore it must be great or something. I'm exaggerating somewhat, but the zeal of some people back then was next level annoying.
Beyond that, git has a great deal of shortcomings, some obvious, some subtle. It was a regression against SVN in some ways and inferior to Mercurial in others. But the strengths of SCN and Mercurial are again vastly different. There is a reason SVN isn't dead.
My biggest gripe with the open source VCSes is that in the last 20 years, no meaningful evolution happened in the established tools, especially around any weaknesses. Commercial systems like Plastic and Perforce as well as proprietary solutions like piper/jujutsu and sapling show that the tools can still improve drastically. I'm excited for a future where we get better open tools for the masses.
I don't think jujutsu is proprietary (unless you want to redefine that term). Source is freely available, and it uses Apache License 2.0.
While you're right about the disadvantages of git, pretending that it became ubiquitous because it "became a quasi-religion because Linus made it in a day" is selling short its advantages. If you think about it for even just a little bit, it should be obvious what a simplistic statement that it. Also, git was not the stagnation you make it out to be. Even with its warts, it was a breath of fresh air, not unlike jujutsu is now a breath of fresh air vs git.
I remember working with SVN, and all things considered, git was a vast net improvement. Git took a lot of pain away. It made working with a versioned code base faster and simpler, to the point of enabling much better collaborative software development. There's a reason we got GitHub and not SVNhub. And GitHub was what helped git become so dominant.
Would it have been better if Mercurial had beat out git in the propularity contest? Possibly? There's trade-offs between the two, but Mercurial's easier interface counts for a lot. But if it had won, I'm sure we'd be griping about its shortcomings by now.
So yeah, I'm also happy to see some movement around the ergonomics of version control, but I don't understand the need to disparage the tools that got us where we are. It just seems that you're more bitter than happy, and like you're letting that bitterness cloud your judgment.
The combination jujutsu/piper is proprietary. As I understand it, sapling as released is also not the entire internal system used at Meta.
Git was an improvement over SVN in some aspects, but a monumental regression in others. There are no two ways around that.
Mercurial is better than git in a bunch of ways. It is much more usable because it has way fewer footguns. But it has some of the same shortcomings as git when compared to SVN. I won't call it perfect.
The reason we didn't get svnhub is that some git fanboy nabbed the domain and essentially said "you shall not have it". Joking aside, there was Sourceforge with working SVN support. But I would say that Sourceforge lost market share for a lot of reasons, not all of them technical. Wrapping Windows downloads in adware installers was only one of those many crazy unforced errors.
Am I bitter? Maybe, because I have to use tools that could be so much better. I've experienced better and every time I am forced back to git, it feels a bit painful because I know what we could have instead. If am bitter, it is because of my experience with a wide range of tools.
... next, GNU please. /s