Anthropic Ditched Git. Here’s Why Your Agents Will Too

0:09Chris Romeo Good morning. Good evening. Good afternoon. Welcome to the Starship Enterprise, where Captain Izar Terran-Denosh is on a mission to— what was their mission that they were on? I can never remember. Why were they flying?

0:25Izar Tarandach Oh, that was the continuing Voyager mission. To boldly go where no man has ever gone before.

0:32Chris Romeo That's right. That's our tagline for AI. Go where, where nobody's cared to go. Mm-hmm. Where nobody's cared to go. That's what it is. Isar, that's, this is a record for Isar to start typing in an episode.

0:51Izar Tarandach I need to sign in again.

0:53Chris Romeo He gets offended and then he starts typing to his AI for support.

0:58Izar Tarandach I wasn't offended. I like Star Trek.

1:01Chris Romeo It's much more than I like the other

1:03Izar Tarandach star thing.

1:03Chris Romeo Yeah. The other star thing. It's ridiculous. Doesn't make any sense. No logic about it.

1:08Matt Coles Who would leave that giant hole in the middle of that giant thing? Right. Come on.

1:12Chris Romeo Listen, to, to, to sum it up, they made how many Death Stars? 3. And they all had the same problem. Come on. Threat model that thing.

1:22Izar Tarandach I'm sorry, but the 3—

1:23Chris Romeo or even R2-D2.

1:25Izar Tarandach I, the, the, the whole, The whole universe is held by a magic force that lives in your mitochondria, and you are worried about a hole in a Death Star.

1:34Chris Romeo I just look for the plot problems in something, and it's like,

1:41Izar Tarandach how come that? I'm just gonna say books have been written on the subject.

1:45Chris Romeo True.

1:45Matt Coles How come, how come the, how come the, all the guys with, how come all the guys in the white suits with laser blasters can't hit anything?

1:53Chris Romeo That's also a good point. Why don't they improve their training regimen? Do they ever actually shoot anybody? I don't think they do.

2:00Matt Coles Very rarely. They, they shoot each other, I think more so. Yeah.

2:04Izar Tarandach Let me tell you this. I have seen things.

2:06Matt Coles Okay. Did anybody watch the Mythbusters episode where they actually tested that out? What?

2:13Izar Tarandach I missed the Mythbusters.

2:14Matt Coles Well, the laser blasters, first off, lasers don't shoot in like little things that travel, right? They, they would shoot a laser, which is

2:22Chris Romeo like, it would be you wouldn't be able to see it.

2:24Matt Coles You wouldn't be able to see it. No, but the, the, and it would be continuous.

2:26Izar Tarandach The camera was faster than light.

2:29Matt Coles Right. Okay. So it, it acts like a wave more like a, more than a particle.

2:35Izar Tarandach Um, but depends on how you're looking at it.

2:37Matt Coles You should, you should watch the, you should watch the episode if you haven't. It, it's excellent.

2:41Chris Romeo There's a lot of Mythbusters I should watch, but I never do and I'm probably never going to.

2:46Matt Coles Okay. So how,

2:47Izar Tarandach that was a

2:47Matt Coles good, have your, have your favorite, favorite AI watch it for you and then Summarize it.

2:52Chris Romeo That's even better idea. And then I'm gonna tell it just not to tell me about it, cuz then it's been watched. I can say, yeah, part of my team watched it. What'd you get out of it? I don't know. Let me ask.

3:03Matt Coles Busted. That's, that's all you need is busted.

3:07Chris Romeo My favorite MythBusters episode is gonna be the most bizarre thing that anybody could list as their favorite episode. It's the one where they investigate whether running in the rain Is actually fast, is actually, do you get as wet, more wet or less wet if you run across a rainstorm?

3:27Matt Coles Uh-huh.

3:28Chris Romeo And they found out it doesn't matter. It doesn't do any good to run. You still get the same amount of, of rain on you.

3:35Izar Tarandach Okay.

3:35Matt Coles So yeah.

3:36Izar Tarandach Let me ask you this. Have you ever heard of the Ig Nobel Awards?

3:41Chris Romeo No,

3:41Matt Coles I, I have actually, but I don't see them there.

3:44Izar Tarandach No, the Ig Nobel Awards are something that happens once a year.

3:48Chris Romeo Okay.

3:49Izar Tarandach And it takes on serious academic research and gives premiums for the most outrageous. So one of the ones that I saw was research on, is it true that buttered bread always falls with the buttered side down?

4:05Matt Coles That was also a MythBusters episode.

4:07Chris Romeo That's important.

4:09Izar Tarandach Wait, let me—

4:10Matt Coles I mean, that's one of, uh, that's one of Murphy's laws, right? So.

4:14Izar Tarandach Uh, give me—

4:16Chris Romeo I think it's one of the laws of thermodynamics actually, isn't it?

4:19Izar Tarandach So then,

4:20Chris Romeo fourth

4:20Izar Tarandach law, best winners of the law.

4:25Matt Coles Technically the bread is, is flat on one side and, and, and curved on the other side with a little bit of weight. So it has a changed aerodynamics as it falls. But

4:34Chris Romeo true. Where, how did we get here?

4:36Matt Coles I have no idea. Isar did it.

4:39Izar Tarandach It's always me.

4:41Matt Coles And actually, it may be Isaur, or it may have been Claude.

4:45Izar Tarandach Okay. So really, Isaur? What? In the year 2000. Okay.

4:50Matt Coles He's still going. There must be something else because—

4:52Chris Romeo an LLM.

4:53Matt Coles You can't turn him off.

4:55Izar Tarandach Wait, there was an Ig Nobel Award for levitating a live frog with magnets.

5:00Chris Romeo That's made it swallow magnets.

5:02Izar Tarandach Now, the guy that won that went 10 years later to win the real Nobel Prize. For graphene. So he's the only person, he's the only person with both prizes, the Ig Nobel and the Nobel.

5:15Chris Romeo Oh wow. Now can we submit our podcast for one of these awards or?

5:21Izar Tarandach How about this one from 2009? A brother turns into a pair of gas masks.

5:25Chris Romeo He's literally like an LLM. He

5:26Matt Coles didn't even hear me. He is. Yep. Whatever question you ask, it's going to answer

5:31Chris Romeo you. We have a new feature on the podcast here, folks. It's Izar LLM.

5:36Izar Tarandach Not yet.

5:37Chris Romeo Command line, Izar. You just type in Izar on your local computer.

5:41Izar Tarandach So what is it that we were talking about today?

5:45Chris Romeo See, he's just always trying to summarize the conversation.

5:48Matt Coles We're trying to do the Turing test on the Izar before us.

5:52Chris Romeo Let's see if we can get his context window to fill. How do we do it?

5:56Matt Coles We're the only ones who can do this. Can you do this, Izar?

5:59Chris Romeo Yeah, that's right.

6:01Izar Tarandach I don't have a beard.

6:03Chris Romeo See, the LLM's becoming alive. It's learning.

6:07Matt Coles Oh, speaking of which, actually, I know it's not, I know it's not on our topic list for today, but I was reading recently that somebody, somebody sent their AI to an interview with an AI HR person.

6:19Chris Romeo Yeah, I did see that. I saw that.

6:21Izar Tarandach That's so funny.

6:22Matt Coles And it negotiated a lower salary for him.

6:24Chris Romeo A lower salary.

6:26Izar Tarandach That actually reminds me of a funny, forget about it. Okay.

6:30Chris Romeo Wait, did that one, did that go through the whole interview process?

6:33Matt Coles I think so. I don't know if it went through the whole interview process, but it went to the point of negotiations. So

6:38Chris Romeo I, I'll tell you what, I was waiting for you.

6:41Matt Coles I tell you,

6:42Izar Tarandach I'm waiting for you.

6:43Chris Romeo Okay. I was, no, a friend of mine's been going through the job interview process right now, and I was, I was asking him questions, just trying to understand how the world has changed. And he described having to go through an interview with a model. An AI performed the interview in text chat and he had to like, and it kept like asking more questions. So he would give an answer and then it would ask a refining question about something that he, that he, he had answered. But I did, I guess I just didn't realize that the world had gone to this.

7:21Izar Tarandach That's not an interview, that's an interrogation.

7:24Chris Romeo True.

7:25Izar Tarandach and it's load-bearing.

7:26Chris Romeo He needs a job. And so, but I was like, is that, it's not even the, the, the filter, the HR using AI to filter resumes. Now, if you get through the first filter, then you have to get through the model itself, which also made me think about how I would prepare to be able to defeat the model. In that case.

7:52Matt Coles So unfortunately, I know somebody else who's going through the interview process and, or the filtering process. Apparently it looks for keywords.

8:00Chris Romeo Yep.

8:01Matt Coles But if it finds all those keywords, it rejects the, the resume because

8:05Chris Romeo it thinks it's

8:05Matt Coles LLM-generated. But you need all the keywords to pass the filter.

8:10Chris Romeo Well, you also need 15 years experience in something that's only been out for 2 years.

8:16Matt Coles Yeah.

8:16Chris Romeo That was the funny one. I saw some guy that was an open source library.

8:20Izar Tarandach That means that you're experiencing very fast.

8:22Chris Romeo But somebody who created an open source library was, that had been out for like 4 years, was interviewing for a job that required 6 years experience in the library that he created. I don't remember what his name was.

8:38Izar Tarandach You just said, you just said that you've been thinking about it for 2 years before you actually talked to somebody.

8:41Chris Romeo Well, no, he's the one who invented the library. He was the person who authored and created it, and they wanted 2 years more experience than when he created it in it.

8:49Izar Tarandach So you can say that he was thinking about it for 2 years. Yeah.

8:52Chris Romeo He's probably the, be the only one, the only one really that was, uh, you know, potential for the job that had the experience, but,

8:59Matt Coles and he still failed.

9:00Chris Romeo Yeah. They probably still filtered him out with the model. But people do an interview with a human being, please, please don't let AI make your hiring decisions. We, we're really going off. We literally do have a plan for this episode, but I feel a soapbox moment coming on here. Have a human being interview a human being with another human being, please. Why do we have to be— why do you have to be interviewed by a model?

9:29Matt Coles Unless it's a, unless it's a model hand model or, or—

9:35Chris Romeo careful, Matt. Careful, Matt.

9:37Matt Coles I was gonna say something, but I won't.

9:41Izar Tarandach Careful. Is that a train coming down the—

9:44Chris Romeo Yes, it was a train.

9:45Izar Tarandach It's gonna crash.

9:47Chris Romeo Pushed Matt outta the way and pushed Isak in front of the train.

9:50Matt Coles Is that, is that one of those, is that one of those tests? Is it, there's like 6 people on one side and a, and a dog on the other side and that's the trolley.

9:57Chris Romeo That's a trolley problem that you're describing there. Not the—

10:00Izar Tarandach wait, you just used 6 people and a dog?

10:03Matt Coles Yeah.

10:04Izar Tarandach For 6 people.

10:05Chris Romeo I was gonna say, I think that's where Matt was going. As we know, he's, he's a, a dog guy and yeah.

10:13Matt Coles Yeah.

10:13Chris Romeo I think he was alluding to the fact that he was gonna redirect the trolley towards the 6 people.

10:19Izar Tarandach That's not even a thought.

10:21Chris Romeo Will the, will the L, we think we tried this before. We did this already. Will the model do the, it'll, it won't do, it won't answer.

10:27Izar Tarandach It won't answer.

10:28Matt Coles Either it won't answer or it will kill the dog.

10:30Chris Romeo Yeah. Interesting. All right, let's talk about something having to do with AI.

10:34Matt Coles It follows the three laws of robotics and the three laws of robotics don't care about animals.

10:38Chris Romeo That's right. We did figure that out. That's good. Cause I follow the three laws of robotics as well. So in general, all right, big news. So kind of big news. Anthropic doesn't use Git. Kind of an interesting idea. I don't know. For some reason I read this and I was like, Oh, that's cool. Wow. Didn't, I didn't think about this. I guess I just don't sit around and think about Git repos at mass scale.

11:05Matt Coles Remember the day when Git was new and nobody used Git? Because once this thing— when was that?

11:11Chris Romeo When did it come out? I don't even remember.

11:13Izar Tarandach CVS was the thing.

11:16Chris Romeo Did you say CVS?

11:17Matt Coles CVS. Yeah. Or S, or SSVC.

11:20Chris Romeo Oh, you mean CVS, the source code?

11:22Matt Coles SSVC. No, not SSVCS.

11:24Chris Romeo What does CVS stand for?

11:26Izar Tarandach Concurrent Version System.

11:28Chris Romeo Yeah. Oh man. I hadn't thought of that in a long time.

11:33Matt Coles And then you had a visual, you had something by Microsoft that was like, oh my God, I can't remember. I used to work on it. Uh, it's been such a long time. Uh, the Microsoft Source Control System.

11:42Chris Romeo Yeah. Yep.

11:44Matt Coles Which was, which was, had its own challenges.

11:47Chris Romeo So. It makes sense why Anthropic has created this new thing. They call it Scion, which is a server-centric replacement for Git.

12:00Matt Coles Visual Source Safe. That's what I was looking for.

12:02Chris Romeo Visual Source Safe. That's, man, memory lane has been engaged here and some of those memories I had blotted out and now they're back in primetime. So Anthropic figured out that They were pushing a lot of code. So they went from about 2,500 merged code changes a day in May 2026 to more than 30,000 a day by September. And what they figured out is Git's offline, every clone is complete design couldn't keep up because their main branch history alone grew to 100 gigabytes and CI is running roughly 25 million checkouts a day. It's basically what I do. So I guess I'm on par with a frontier model.

12:47Matt Coles 600 petabytes of traffic per day is their estimate.

12:52Chris Romeo That's a lot of petabytes.

12:54Izar Tarandach That's a lot of petabytes.

12:55Matt Coles That's a lot. That's a lot of bytes.

12:58Chris Romeo So what they did though is they built this thing called Scion. They, and they really were, were happy to tote that it was done by a 7-person team. Of course, they do have unlimited access to tokens.

13:11Matt Coles Yeah.

13:11Chris Romeo Which, I mean, the three of us could build something pretty incredible if we had unlimited access to tokens where we could just say, go spend, you know, go have 7 groups, teams of agents of 100 each, go try to solve this problem. And then you decide which one's the best one and bring it back to us. Like if unlimited tokens would be a, a fun opportunity.

13:35Matt Coles Mm-hmm.

13:39Chris Romeo Interesting. So yeah, go ahead.

13:41Matt Coles No, go ahead. You first. Cause I'm just looking at what Anthropic actually did here.

13:47Chris Romeo Well, now I'm, now I want to hear what you're, it sounds like you have an analysis of their technical

13:52Matt Coles approach that I was going to say, why didn't they just use Perforce or SVN or one of those other systems that we know works without doing the replication that that Git does.

14:05Chris Romeo I know we were talking about this before we started. What— I just can't remember. Perforce. I remember Perforce. I've used it. I just don't remember. I was always just a user of source code control systems. Like it was one of those things where it was like, what are the least amount of things that I have to know about this to be able to use it so that I don't get fired?

14:29Matt Coles And eventually somebody figured out how to make all the commands equivalent, so you could just use SVN, like you had a, you had an SVN to Git thing, so you could use Subversion commands in your Git repository at one point. Right. And so you had compatibility across the, you know, almost across the board, but some things were just weird.

14:48Chris Romeo You mentioned Perforce, which I don't think exists anymore.

14:51Matt Coles Perforce. It might, I don't know if it's in use. ClearCase was another one that was pretty popular.

14:58Izar Tarandach ClearCase was bad.

14:59Chris Romeo ClearCase.

15:00Matt Coles Yeah.

15:01Chris Romeo So glad you have a good memory for stupid control systems.

15:05Matt Coles For crappy things that I really shouldn't be remembering.

15:08Izar Tarandach He has a good memory for things I'm trying to forget.

15:11Chris Romeo Yeah, that's right.

15:13Matt Coles But it's interesting. I, I, so I was, I was being a little facetious about that, but then I looked and actually, you know, the very first, some of the very first information that Anthropic provides in their article here is that they used they use as reference for this project, two things around managing large or huge mono repos from Google and from Meta. And if you look at Piper, which is Google's thing, they started on Portforce and then they went to this thing called Piper,

15:46Chris Romeo which is

15:46Matt Coles homegrown, which is homegrown in 20— in 2012. They built it.

15:52Chris Romeo Oh, wow.

15:53Matt Coles Right. So this is not a new system. And it was managing something like 86 terabytes of data. And then Sapling by Meta, right? So Sapling looks a little bit more recent, at least on the, the fancy webpage you get to when you click the link.

16:15Chris Romeo Yeah. Well, compared to Scion. Scion is the cutting edge at the moment.

16:22Matt Coles Right.

16:23Chris Romeo So, but it's all about in-memory.

16:24Matt Coles They took these older, they, they took these older approaches to source control and reinvented it basically.

16:31Chris Romeo Well, and they, they, you have to be connected to use it. There's no offline mode for what Anthropic came up with. Right. But in an agentic world, there is no offline because if you go offline, the agent can't talk to the model and it's no longer working unless you have a local model. Well,

16:48Matt Coles if you do an enterprise deployment, you're probably at least locally online, meaning if you're, if you lose access to the internet, you're probably not dead in the water. So I guess it depends on how connected you have to actually be. You have to be connected to your model server and to your Git, into your source repository, which is like, which is like Perforce or SVN or those things where you were a client to a remote system, right?

17:16Chris Romeo It seems like you would have to be connected though.

17:19Matt Coles Yeah. But again, how connected? You have to be connected. If you're connected to your LLM, you can be connected to your—

17:24Chris Romeo Yeah, I think it's, well, I don't know though, because it's a deterministic software solution. It's not just, it's not like they just created a model to do the search, to handle source code and like you just ask it, can I have these files? It's like, sure, here you go.

17:42Matt Coles Well, actually, if you think about it, wouldn't it be a natural fit if that's what they actually did? And actually, if it's even possible, think of, think of LLM Wiki, right? So you talk to your model to manage a bunch of files. That's a knowledge base with a knowledge graph, right? Or OpenWiki or whatever the system is. So why couldn't you do that? And if you did, you could potentially build a knowledge base alongside your code repository and have explainability and other things like curated by the L— by the team of agents that are managing librarian system.

18:20Chris Romeo Man, once again, Matt invents another thing on the AI security table. So if you just listen to this, that is patent pending, trademark pending.

18:30Matt Coles Uh, although if Andrej Kapasi is out there, call me.

18:33Chris Romeo All right. I'm trying to analyze the technical architecture.

18:37Matt Coles That seems like it would be a natural fit and it would suggest you're always online, right? Your client. Oh, and I guess the other thing that, that Anthropics basically stated here was, so Git, you do all your work locally and then you push it to the remote system, right? That's where a pull request happens. And so instead, if all of the work is done there, there's no server, there's no client-to-server transfer. Right. Merge requests are done. If there's a concept of merge requests, code branching and things happen on the server side, right?

19:12Chris Romeo There is a concept of merge requests and what they've done. It doesn't look like it's model driven. It looks like it's like they've created a deterministic set of tooling.

19:26Izar Tarandach Well, it has to be deterministic, right? Because you want to know that the code that you think is there is there, but They define it as a write-ahead log.

19:35Matt Coles Well, it does say in server, the server answers questions about file history. So there's probably an LLM attached to it.

19:42Izar Tarandach No, they say that there's an API for it. So where there's an API, there's an MCP, but it is write-ahead log. Where have we seen this thing before? Write-ahead logs besides big databases.

19:57Matt Coles Remind me of the feature of a write-ahead log.

20:00Izar Tarandach It's a write-ahead log and change sets. This is starting to look very much like a ledger.

20:06Matt Coles So that's like a normal source repository, right? I mean, before Git came along.

20:11Izar Tarandach Yeah, but what is this?

20:16Matt Coles SVN? Was it SVN or was it— there was another one. Yeah, it was SVN. Everything was a write-commit to a, to a database. And so you had, you had a transaction log, right?

20:26Izar Tarandach And you could even rebuild the

20:28Matt Coles Yeah, you could

20:29Izar Tarandach reproduce from the log.

20:32Matt Coles Rolling back a change was rolling back the transaction.

20:36Izar Tarandach By exposing source control queries through an API rather than requiring a full local checkout, Scion allows automated jobs, including cloud-based coding agents, to inspect repository state, history, and metadata without the overhead of cloning or syncing a working copy.

20:50Chris Romeo That's open. That's magic right there.

20:53Izar Tarandach So it's centralized.

20:55Matt Coles Yeah, it's a centralized— yeah, it's a centralized system. And what's interesting is the client maintains a bounded local object cache, lays out the working copy files, working copy files on disk, monitors for changes, and basically interfaces with the server. So it's a— obviously it's a client-server system. They've already explained that. It works through APIs. So the client's doing all of that work through APIs on the server.

21:21Chris Romeo Brilliant.

21:23Matt Coles Brilliant.

21:24Izar Tarandach Reinvented the wheel. Why does it sound like, yeah, why does it sound like everything again, another case of everything that's old is new again?

21:30Matt Coles Yeah. We had to reinvent because it's AI, but we use it to be able to—

21:34Chris Romeo But it operates at a thousand times faster though.

21:37Matt Coles Sure.

21:37Chris Romeo The challenge was they were getting buried in the amount of effort it took for agents to clone, maintain, and then merge. Conflicts and everything across various branches.

21:53Izar Tarandach And so, but they're saying Scion is by no means a suitable replacement for Git for general use. So they are answering a, but yeah, it's a very,

22:03Matt Coles yeah, but why not?

22:04Chris Romeo Because, well, because they're using it for agentic development and that's because that's what they do. They don't do human development anymore.

22:10Matt Coles But why couldn't you, why couldn't it be a general replacement just because it's a distributed file system with, with. It's so the entire repository, the entire repository is distributed, right? Which means there's multiple, which means there's a multiple server system in the backend.

22:26Chris Romeo Well, I think the reason they wrote that is because they don't want to open source it.

22:30Izar Tarandach Yeah.

22:30Matt Coles I thought it was open sourced.

22:32Izar Tarandach It is.

22:32Chris Romeo I don't know. I don't know. It's closed source. They're just, they've been writing these articles. They just had a great one on code review that we'll have to examine separately in a different episode. But they, these articles announce something really exciting, but then also bolster Claude's position. 'Cause like if you read this one, there is an underlying kind of explanation for how they used Claude to build the whole thing. They could never write something like this from scratch. They did it with 7 people. It's an advertisement for how good Claude is.

23:11Izar Tarandach But is this really exciting?

23:15Matt Coles It's, it's innovative. It's innovative in its use case, but not necessarily in its implementation.

23:21Izar Tarandach Yeah, but is it exciting? Are we seeing riots in the streets of people clamoring, give me a new source management?

23:28Chris Romeo My agents are. They're telling me all the time, this sucks. Why do you make me use this Git thing? It's terrible. This is a problem that you have to be. This is not a problem that everybody has. This is not a problem that everybody will ever have.

23:42Izar Tarandach No.

23:43Chris Romeo This is hyperscale, edge-of-your-seat, frontier-style company problem.

23:52Matt Coles Mm-hmm.

23:53Chris Romeo Today.

23:53Izar Tarandach But it is, but it is in a way, perhaps this is not the most extreme case, but it is a case of rebuilding infrastructure to suit agents.

24:05Chris Romeo Yes. Right.

24:06Matt Coles But it's utilizing techniques and technologies that are not new, not agent-invented.

24:15Izar Tarandach Right. My point is that—

24:16Matt Coles So we had it, we were right all along.

24:19Izar Tarandach My point is that if you start optimizing things like this for agents, at some point you are going to get into a place where it's going to be much more comfortable for an agent to work in things than for humans to step in and look at what the agent is doing. And then we close the door and everybody goes to the beach to drink Mai Tais or something. I don't know. But the point is that we are starting to optimize things to, to, for agents. And at some point that may come back. Did I just say that? Then it may come back and bite us in the, uh, behind.

24:52Chris Romeo But this is just the beginning. This is the beginning of optimizing everything for agents.

24:59Matt Coles Okay. So funny thing. Their client name is SVN.

25:04Izar Tarandach SVN?

25:05Matt Coles SVN, not SVN. SVN.

25:07Izar Tarandach I would be, I would be writing an alias so fast. Not even funny. It took me years to stop typing SVN.

25:19Matt Coles It does claim that it's less CPU intensive than Git servers.

25:23Izar Tarandach I'm sure it is. Cause it's from what I understand, and that's not a lot. It's just writing ahead and hoping that something fixes the conflicts down the road.

25:34Matt Coles The Cyan daemon, the Cyan cache, and the Cyan server all respond to identical RPCs for this protocol.

25:43Izar Tarandach Or to tell you the truth, because it does have an API and it is API-based, it could be a way to write code, already send it to review, and only after the review, the thing gets committed or something.

25:57Matt Coles Well, all the changes happen on the server and it uses RPC calls. Where have we heard that before? Everywhere. Well, NFS, MCP.

26:09Izar Tarandach MCP.

26:11Matt Coles MCP is RPC underneath, if I remember correctly.

26:15Izar Tarandach It can, it may, it may use SDIO or whatever, but it can.

26:20Matt Coles But it's basically RPC. RPC is not a new concept. It works well. One thing I do know that's missing from this conversation, and given the name of our podcast, I don't see a mention of security anywhere in this article. Now, on the plus side, it's a server-based defined APIs. That's, that's a well-known security problem. Right? How to protect that?

26:50Izar Tarandach I'm sure that there are tokens everywhere and token exchanges and some cabbers somewhere and somebody passing a hash.

27:00Chris Romeo They do mention security.

27:02Matt Coles Oh, they do mention security.

27:03Izar Tarandach They say we don't have any.

27:05Chris Romeo Not that I detected it myself, but a friend of mine was just telling me about it. They discussed access control. They built what they called a rich source control IAM system. Supports access control at the repo level and permissions on individual directories. So the most sensitive code could be locked down without splitting a monorepo.

27:25Matt Coles We, uh, we laid the groundwork for new forms of monorepo security. Yep.

27:30Izar Tarandach But when you, when you said RPC, you turned on a bulb for me. NFS. This doesn't seem to be very different from a file system for source code.

27:41Matt Coles The only difference is there's a server managing the file system as opposed to the client.

27:46Izar Tarandach It's a distributed file system.

27:48Matt Coles Right. Pretty much. Yeah. And we know how to secure that. That's like a, not a new thing. What I'm, I guess the other part that I'm sort of really touching, I guess what I'm really focusing on as I read this, everything is a monorepo.

28:06Izar Tarandach Yeah.

28:07Matt Coles Why?

28:09Izar Tarandach Why not?

28:09Chris Romeo Because that's where everybody lives now. Yeah. They're back to monorepos are cool.

28:14Matt Coles But monorepos is just a behemoth of data, right? I mean, if everything is in one repo and you don't consume other things, right? I mean, dependencies live in other repos.

28:28Izar Tarandach Look, as anybody who ever checked out a repo that has a lot of other repo dependencies and just watches the thing check out other things by itself, right?

28:40Matt Coles And you go, wait, I didn't ask for 8 billion gigabytes of data.

28:45Izar Tarandach The point is, from an architectural point of view, talking monorepo or microservices or whatever, that makes absolute sense. But to go back to the repo itself and say, oh, it's a monorepo, it's big. Okay, so it's 1,000 microservices and you need all of them for your local tests. That's not big. That's not Whatever it is. So I don't know, just that's the first thing I would

29:11Matt Coles say. So model repo makes it a bounded problem. By calling it, by calling and enforcing it as a model repo, you have a bounds to the size and the complexity problem.

29:23Izar Tarandach But it's theoretic because at the end of the day, you need the whole thing to get it running, right?

29:29Chris Romeo Yeah. And they've created this in such a way that it it, it doesn't really matter how big the repo is, what's in the repo, whether you had 75 repos, it's still like, I like where you guys were landing on this. It's just a, it's just a virtual file system that operates at 1,000 miles an hour and you can, you can swap out individual tiny pieces and it tracks it all so that it doesn't destroy the whole.

29:58Matt Coles Right. And maybe you get—

30:00Chris Romeo Integrity of the source code.

30:01Matt Coles And maybe you get some benefits along with it, like explainability or, or traceability and whatnot. Yeah. This seems like, this seems like a win. There's, there's, this is good stuff.

30:12Chris Romeo I wanna throw out something here that, that the article frames as a negative, 'cause I think this is worth exploring a little bit in this newfangled age of AI. So they did an adversarial review against Scion using Claude, and the output of that was overly cautious about trusting another internal service. And so then Claude added excessive validation. That's their assessment, excessive validation in response to something from this adversarial review. And then they, they claim they steered Claude back. To reality to not over-engineer some type of response to a, a potential weak area, you know, internal service validation. That's happened to me too. What do you, what's your take?

31:07Izar Tarandach It has happened to me too.

31:08Chris Romeo Yeah.

31:09Izar Tarandach But I was Claude. I was Claude in that, that scenario.

31:12Matt Coles People had to, there's a t-shirt idea for you.

31:15Izar Tarandach Yeah. It was like, yeah, Claude been there.

31:19Chris Romeo Okay, so people were trying to argue that, yeah, I get it now. Okay. We've probably all been there or there. Well, you're just being excessive. That's why they pay us here is it's not to be, you know,

31:38Izar Tarandach that, that's why we came to and came up with a reasonable security, right?

31:41Chris Romeo Yeah.

31:43Izar Tarandach Remember that? Remember?

31:45Chris Romeo Remember when?

31:46Izar Tarandach Remember when?

31:47Chris Romeo It's a country song's about to start here.

31:50Izar Tarandach Pepperidge Farm, remember

31:51Chris Romeo Pepperidge Farm?

31:53Izar Tarandach Isn't that how it goes?

31:56Chris Romeo Pepperidge Farm. How can you, how can you prove to the world you're old without saying you're old? Make a reference to Pepperidge Farm. Do they even make Pepperidge Farm? People on this that are listening to this under the age of 50 are like, what are these? People talking about? It was a cookie brand, people.

32:17Izar Tarandach I think— oh, wait, bird snappers.

32:18Chris Romeo It was a cookie brand, right?

32:20Matt Coles Oh yeah.

32:20Izar Tarandach Yeah.

32:20Matt Coles Full of high fructose corn syrup and—

32:22Chris Romeo does it still exist?

32:23Matt Coles Pepperidge Farm? Yeah.

32:25Chris Romeo Yeah.

32:26Izar Tarandach I don't know.

32:26Matt Coles How else do they remember?

32:27Izar Tarandach They remember.

32:28Chris Romeo Yeah. No, we remember. That's 'cause we're old. Like Matt, you went through and traced every source code control system since 1974 from memory.

32:38Matt Coles From memory.

32:38Chris Romeo So this is a, this is a group that's a little bit different than the rest of the world.

32:44Matt Coles Yeah, we have long-term contextual storage memory, but the Pepperidge Farm, of course Pepperidge Farm exists. I walk by the display case every time I'm in the supermarket.

32:56Chris Romeo You pay much more attention to the world around you than I do.

32:58Izar Tarandach You still go to the supermarket?

33:01Matt Coles What do you do for food?

33:02Chris Romeo How do you eat?

33:05Izar Tarandach Is that everybody ordering all the time?

33:07Matt Coles Is it just me?

33:07Chris Romeo It's just you.

33:08Matt Coles It's just you.

33:09Chris Romeo Damn it. Man, bourgeoisie over here.

33:12Matt Coles Wow.

33:13Izar Tarandach No, I call it laziness. Who has the time to cook?

33:16Chris Romeo It's just, it's just the busyness of life.

33:18Matt Coles So do you ask Claude to order you food?

33:21Chris Romeo Oh, there's an idea.

33:23Izar Tarandach Actually, just ask my son.

33:25Matt Coles Does, does, does pro— does, does, uh, protein— does calorie-dense food show up at your house? Because Claude—

33:32Izar Tarandach Silent Green, man. Silent Green.

33:35Matt Coles So, so I was just looking to digging at the end of the article in the footnotes. And it mentions something called Mononoke from Meta, which is one of the source repositories that they use as input, including a data model. So Mononoke is written in Rust and it has a data model that they basically, I don't know if they lifted it, but they used it as heavily influencing the development of this repository system that they built. So it's interesting to see they, they're not inventing something new. They're inventing something that is new for them and that is useful for their use case and probably for others based on technologies that have been seen. And of course we know this is how LLMs work, right? Training data.

34:30Chris Romeo This is where LLMs excel. Is when you can point it at something else that's open source and say, make me one of these. Make me one of these, but in green. Yes. But just, yeah, make it a little different, little lighter green color to match my brand.

34:48Matt Coles I know we will be talking at some point about also pointing it at something that's not open source and saying, make me one of these.

34:56Chris Romeo That's possible too.

34:58Izar Tarandach That's how you get, you get freedom, isn't it?

35:01Chris Romeo What? It's been in the news the last couple of weeks, but you guys have seen the Adobe and Microsoft knockoff projects, right? Yeah. I haven't tried 'em yet. I need to just, I'm really curious on the Adobe side because, you know, I was a user of Photoshop, but not very good at it for a long time. Like used it for very special purpose use cases that I would now just ask a model to do for me.

35:29Matt Coles There

35:30Chris Romeo are

35:30Matt Coles open source variations on many of those tools already,

35:33Chris Romeo but there were, but there were, but this time somebody pointed an LLM at it and said, make a copy of this.

35:43Matt Coles Right.

35:44Chris Romeo And from what I understand on the Adobe side, they actually reverse engineered binaries and that's at least what I heard reported.

35:51Matt Coles Yeah, we need to talk, we need to talk about that in more in depth because there's a whole host of interesting Challenges, I imagine with that, like if there's any lawyers listening.

36:00Izar Tarandach No, let's not do the lawyer thing.

36:02Matt Coles We don't need to do the lawyer thing, but I'm sure that somebody is.

36:06Chris Romeo Yeah. I don't know that we have a lot of lawyers in our audience, but maybe,

36:10Matt Coles maybe there's a, maybe there's an AI legal table out there somewhere.

36:14Chris Romeo That could be another, could be another podcast for us. Yeah. We got a lot of opinions.

36:19Matt Coles We do.

36:20Chris Romeo We don't have a lot of facts to back it up, but we got a lot.

36:23Matt Coles We stayed, we may have stayed at a Holiday Inn last night, but Express.

36:27Chris Romeo Oh, it's not just a regular Holiday Inn. Come on.

36:29Matt Coles That's right.

36:30Chris Romeo Another dated reference. But yeah, I think this is this, this whole creating software. Do you guys see the Beehive one?

36:40Matt Coles No.

36:40Chris Romeo From the last couple of days. So the founder of Beehive, it's like a newsletter. I use it to send the Reasonable AppSec newsletter. It's a pretty popular newsletter hosting site and they're trying to branch into other things.

36:54Matt Coles Constant Contact used to be, or—

36:57Chris Romeo Yeah, but imagine if it didn't suck. I used Constant Contact and they can be mad at me if they want. Make your software better. You should have made it better a long time ago. I don't even know how it is now. I haven't looked at it in a long time, but it used to be terrible. Okay. So to not have offended a good chunk of the tech population in the Northeast, let me try and do that to other somewhere else. So. What the, the CEO, founder, CEO of Beehive, I don't even know how big of a company they are at this point, like how much revenue or whatever, but they're a good size company, comes out and says, basically, you can't copy what we did. And then he tweets, it's like a, a large article tweet of all of their feature set, like broken down into 2 and 3 sentences at a time of everything that they, their platform does. Line by line. And so literally 12 or 20, 24 hours later, somebody responds to it and says, hey, I built it and I'm now charging half the price. And then he had to backpedal and say, well, it's, it's all about deliv— people were like, it's all about email deliverability. And that's what, that's, they, Beehive gets you 99.4% email deliverability or whatever. But I'm just sitting there thinking to myself, there's a lot of email providers in the world that can warm up an email path for me in a couple of days. But it's not, Beehive doesn't have that counter. But it was like, I don't know what the Beehive CEO was thinking was going to happen as a result of like, the argument he, I think he was trying to make was like, oh, you can't copy, you know, this whole idea of AI being able to copy people. Isn't, isn't real, isn't possible. And then somebody does it in 24 hours, comes back with, and then starts offering it on the thread at half the price of what Beehive sells for.

38:56Matt Coles Apparently some, apparently somebody watched and believed the LifeLock commercial.

39:03Izar Tarandach Yeah.

39:03Chris Romeo That LifeLock guy got his identity stolen like 9 times while he was the CEO of LifeLock. So it didn't work, and we should have known.

39:17Matt Coles Dude,

39:17Chris Romeo that's a whole other episode about how people are being ripped off in the world of technology right now. Like, did you know, did you know Norton Antivirus is still a thing?

39:32Izar Tarandach Yes,

39:32Matt Coles yes,

39:33Chris Romeo that gets sold— I'm gonna say it— predatorily.

39:37Izar Tarandach Yeah,

39:37Chris Romeo to old people. And you know what, they could come for, come for me, bring it. Challenge me. I'll, I'll go in the public square and explain why you're predatory and you won't like the outcome of it because I'm not gonna be nice. Like they, it's owned by, it's not the original.

39:54Matt Coles No, it's not. No, it's owned by a VC now, isn't it?

39:57Chris Romeo Or PE or something like that. Yeah. But I just experienced it through the lens of somebody else's computer and I was like, what the What is this? Like, why are they trying to sell it? And it's like a fearmongering sales pattern. And I'm like, of course this is going to work on people who don't have a lot of experience in tech. They're going to feel like their arm's going to be twisted. I have to do this or my computer's going to fall apart or something.

40:29Matt Coles Well, and AV solutions now, I think AV has been been done to death. And the value add that they, that they get you with is, oh, we searched your, your, your, your name from the dark web, or, you know, we look for these other threats against you that are nothing to do with malware.

40:45Chris Romeo Yeah.

40:46Matt Coles Right. And so they're branching out into other areas because naturally they're tied into all of your data, all of your systems, all your personal information, you know, et cetera.

40:56Chris Romeo And they're that connected on the dark web. I'll just, I'll just poke around. I mean, I remember when the dark web was first becoming a thing and like all of us, I did some exploration and I'm like, man, people talk about this like it's Google for bad stuff. Like, no, it's very hard to find things. Like you really do have to have a lot of technical expertise to move around and to find things.

41:21Matt Coles The dark web is basically the old internet before you had directory lookup services. Search engine indexing and a host of other things, right? That's the whole point about it being a dark web. It's all the websites that didn't get indexed.

41:35Chris Romeo Yeah. And it's all over Tor, right? So it's in its own,

41:41Matt Coles yep. Or it has, or it has space, WireGuard front end and things. It's like, it's got, it's got defenses so it doesn't get indexed. Yeah. And therefore it's not searchable. You have to sort of, it's sort of like with the island in the Pirates of the Caribbean, you can only get to it if you've already been there. You can only find it if you've only been there. That kind of thing.

42:01Chris Romeo That's a good island. Yeah. All right. Well, I think that's it for this episode. We made it through one thing. We'll save self-replication for the future. Uh, we are trying to make—

42:11Matt Coles and we can talk about clone and replace as, as a topic.

42:15Chris Romeo Yeah. We're trying to make multiple versions of ISAR, so we'll save replication for next time.

42:21Izar Tarandach Just promise me that you're going to add some patches.

42:23Matt Coles Yeah. If, if we're gonna do this, we have to make it better.

42:25Chris Romeo Did you ever see the, The movie with Michael Keaton where he's the, where they somehow he gets replicated and there's different versions of him. You should check it out.

42:35Matt Coles I've seen that.

42:36Chris Romeo It's pretty, it's especially in this age of thinking about replication becoming more of a potential thing. Like they have various IQ levels of the replicated versions and it's, it's just, it's pretty funny. I don't know.

42:51Matt Coles The real Michael Keaton, please stand up.

42:52Chris Romeo That's right. I think Marshall Mathers has a song about that.

42:56Matt Coles Eminem.

42:58Chris Romeo Well, I'm on a first name

42:59Izar Tarandach basis. I'll have one. Thank you.

43:00Matt Coles You're on a first name basis.

43:02Chris Romeo Yeah. Call him by his stage name. Come on. We're hanging out. Come on. All right. Well, it's been real. It's been fun.

43:10Matt Coles It has been real actually, hasn't it?

43:12Chris Romeo Yeah. Other than we not, we're not sure if a, if Isar is a model or not at this point, but If

43:19Matt Coles you are, you could write an article about it and be famous.

43:24Izar Tarandach Dude, just walk around me in any security conference and the only conclusion you can have is that I am indeed a model.

43:33Matt Coles Hand model, foot model, what type of model are you?

43:36Izar Tarandach Yeah, let's go with hands.

43:39Chris Romeo The world's first application security model. Supermodel. I am the security supermodel.

43:47Izar Tarandach I am the model that your threat warned you about.

43:52Chris Romeo ASSM. That's a new— don't put that on a t-shirt, please.

43:57Matt Coles Please.

43:58Izar Tarandach I'm out of the business of t-shirts.

44:01Matt Coles We're going to need to put that on a t-shirt.

44:01Chris Romeo Are you out of the t-shirt business?

44:03Izar Tarandach I'm out of the t-shirt business.

44:04Chris Romeo Dude, why? Stop burying the lead here. We're 46 minutes in and you're out of the t-shirt business?

44:09Izar Tarandach I'm off. I'm off.

44:11Chris Romeo He's out.

44:11Izar Tarandach The world is not ready for my t-shirts.

44:14Matt Coles We're going to need, uh, we're going to need the paparazzi at the next, uh, OWASP conference for, for Isar.

44:18Chris Romeo We're going to have to hire some people if we want that to happen.

44:22Izar Tarandach Which, by the way, people, is in the first week of November in San Francisco. If you haven't gotten your ticket yet, do. It's going to be fun. We're going to have an OWASP by Design minicon and yeah.

44:36Matt Coles Party time.

44:37Speaker D Thanks for joining us at the AI Security Table. Subscribe, rate, review, and follow wherever you listen. New episodes every week. AI Security on the Table.

Transcript supplied by AssemblyAI. It may contain transcription errors.