{
  "video_id": "okSgumIGLBY",
  "channel_slug": "deeplearningai",
  "channel_handle": "DeepLearningAI",
  "title": "AI Dev 26 x SF | Carter Rabasa: File Systems Are the New Primitive for AI Agents",
  "duration_seconds": 1960,
  "url": "https://www.youtube.com/watch?v=okSgumIGLBY",
  "upload_date": "20260521",
  "transcript": "What's going on everybody? Uh, super\nstoked to be here. This is my first uh\nfirst AI dev and uh yeah, I'm really\nhappy to talk to you about this kind of\na fun idea. Um, interesting idea I think\nthat file systems are the new primitive\nfor AI agents. Um, first a very quick\nintroduction to myself. My name is\nCarter Robasa. I lead developer\nrelations at Box. Um I have been I\nstarted my career in software back in\n2000. So I've been building software for\nover 26 years. Um and I've been\nspecifically building developer tools\nsince 2012. So I joined Twilio in 2012\nand I've been working at developer\ntooling companies uh ever since and uh\nnow I'm here at Box. uh when I'm not\nworking uh on you know file systems and\nagents uh I'm an organizer or a\nconference for web and AI developers\ncalled CascadiaJS\nuh our event our event's actually coming\nup uh June 1 and two check us out uh and\nthat event is taking place up in Seattle\nWashington which is where I live uh with\nmy wife and uh two children so uh quick\nshow of hands uh who here uh builds on\nBox right now today\nall Right. That's fantastic. I'm in the\nright place. Um, so what what is Box?\nRight. So Box is a company historically\nthat helps companies and enterprises\nmanage their unstructured data. The kind\nof data that was like the lifeblood of\ndoing business. You know, your PDFs,\nyour Word documents, your Excel\nspreadsheets. Um, and now of course like\nuh this this this these workloads are\nmoving um into the world of agents. Um,\none thing that's kind of interesting is\nthat 90% of like all of the data that\nkind of it floats around enterprises is\nunstructured data. So like a as a sort\nof a a a full stack developer my entire\ncareer like this was shocking for me,\nright? Like I'm used to building\nsoftware using things like relational\ndatabases, NoSQL databases. Um you know\nmy my my career has been mired in\napparently like 10% of how work gets\ndone. Um and it's really in in a box we\nare thinking very very deeply about how\ndo you unlock the value of all of that\nunstructured data across companies and\nacross enterprises.\nAnd lately a lot of people have been\nthinking very hard about this sort of\nquestion and problem around long-term\nmemory for agents. Like what does this\nmean and where does this come from? So I\nthink all of us I'm sure every single\nperson in this room has been at a bare\nminimum playing with agents, right? I\nmean when I and I I just mean chat GPT\nor clawed code, right? any kind of LLM\nwired up to any to even a minimal\nharness that that allows you to do work.\nAnd the immediate problem that people\nstart to run into, and this has been\ngoing on for a while, is that agents\njust keep forgetting things, right? They\nlose context, they forget things, and\nthat makes it very difficult to actually\ndo work or get things done. So anyway,\nso there's sort of this question about\nuh how do we give our agents long-term\nmemory? Um, let's talk a little bit\nabout what are the characteristics of a\nlong-term memory solution for agents\nthat I think we can all agree on. Right?\nSo, one characteristic\nit needs to retain facts across\nsessions, right? This is super\nintuitive. Like I, you know, I connect\nto Claude code multiple times a day,\nmultiple days a week, right? like I want\nmy I want claude code to remember the\nthings that it learns or that I teach it\nacross these sessions. Um two, it needs\nto be able to retrieve these facts\nselectively. So this is actually fairly\nnuanced. Um if the data set is\nsufficiently small, don't worry about\nretrieving things selectively. Just\nretrieve everything and put it in the\ncontext window. that obviously falls\ndown the more data you start to\naccumulate, the more memory you start to\naccumulate. So, it's critical that\nagents be able to retrieve certain facts\nin a very selective manner. And we'll\nkind of go into what that means. Um, the\nagent needs to be able to update and\ncorrect this data over time. Once again,\nthis is fairly intuitive, but it's it\nrules out like a lot of, you know,\nreadonly solutions, right? um the agent\nneeds to be able to intuitively update\nthis information in a way that is not\ntoo difficult. Um and then the fourth\nthing which uh I think is kind of\ninteresting to think about is that this\nmemory needs to be inspectable by\noperators, right? It needs to be\nreadable by people. Um it can't be a\ncompletely black box. The reason for I\nthink there are many many reasons for\nthis but this the most intuitive reason\nis that LLMs are non-deterministic.\nUm and if you have a non-deterministic\nagent and you have a black box of\nmemory, it makes it really difficult to\nkind of understand and introspect uh\nwhat what's going on with your agent. So\ncool. So like let's let's operate from\nthe perspective that these are\nreasonable requirements for long-term\nmemory. So great. So, let's start to\nthink about like what could we what\ncould fill in the gap, right? Like what\ncould serve as the magical and amazing\nsolution for long-term memory? Well,\nI've got a bunch of options here. It\ncould be relational databases. I've\nbuilt my entire career on working at\nleast initially working with relational\ndatabases, Oracle, MySQL, Postgress. Um,\nthat's that's possible. Um I I think\nthat relational databases uh fall down\non some of the characteristics that\nwe're that we're talking about. Um not\nvery human inspectable being sort of the\nsimplest. Uh NoSQL databases, you know,\nsimilar vector databases. Um vector\ndatabases are amazing. I love vector\ndatabases. Uh I used to work for a\ncompany that shipped a vector database.\nUm they they definitely uh fall down on\nsome of the characteristics that I\nmentioned. They're fantastic for\nsimilarity search. It just turns out\nthat that's kind of where they're\nnarrowly useful. Um, things like graph\ndatabases, um, or data APIs, right?\nLike, forget about off-the-shelf\ndatabases. Like, why not design the\nperfect bespoke\nlong-term memory tool for agents, right?\nUm, well, I think that there's a really\ngood reason to be skeptical of all of\nthese approaches for LLMs,\nand it's because LLMs already know how\nto do these things perfectly. In fact,\nraise your hand if you know what these\ncommands are. Every single person raised\ntheir hand. Why did that happen? Because\nthis is Unix or maybe Linux or, you\nknow, whatever. It's a star nyx. These\nare the fundamental building blocks for\nUnix file systems. All of you or most of\nyou as you learned to program, learned\nto operate computers, you learned how to\nmanipulate file systems using all of\nthese commands. And the amazing thing is\nthat every single LLM in the world has\nbeen trained on these commands and the\nsemantics of these commands and how to\ncombine these commands. There are like\nuncountable billions of tokens that have\nbeen spent training LLM on what these\nthings are and how to use them. It's\nkind of mind-blowing when you think\nabout it.\nSo long before I was born, uh, a couple\nof brilliant people, Ken Thompson and\nDennis Richie, they invented this thing\ncalled Unix. And it was sort of, it came\nout of the wreckage of, uh, a more\ncomplicated multi-user computing system\ncalled Multix.\nAnd Multix failed for a number of\nreasons,\nbut the number one reason that it failed\nand the number one reason that Unix\nsucceeded was like the Unix philosophy.\nAnd the Unix philosophy is that you can\nbuild powerful complex systems that do\nthat do everything that we you know that\nit powers my phone and empowers our\nLLMs. You can build all of these things\non top of incredibly simple atomic\nbuilding blocks and that is how you\nengineer complex powerful repeatable\nsystems. Um, and one of the things that\nthey invented, which at the time was\nmind-blowing, was that everything in the\nkernel and everything in the in the\ncomputer could actually be represented\nby this thing called a file descriptor.\nAnd if you don't know what that means,\nit means that the files in the folders\nthat you're familiar with are\nrepresented by file descriptors, but so\nis everything else. So are the current\nrunning processes. So are so is a disio\nwith devices. It's all represented by\nthe same thing called a file descriptor.\nAnd those file descriptors because they\nare uniform across all of these\nheterogeneous resources. They can all be\nmanipulated using these exact same\ncommands.\nAnd this was outrageously brilliant. Um\nand uh and it and it basically just sort\nof invented the idea that simplicity and\nencapsulation is not the enemy of\ncomplexity, right? Like you can build\nradically complex things on top of these\nsimple interfaces. And the irony is that\nwe 50 years later, we keep trying to\nlike run away from these principles and\nreinvent new bespoke complicated ways to\nstore and retrieve information that\nmaybe make sense to us as human beings,\nbut are basically foreign to LLMs. And\nyou have to spend disproportionate time\ntraining and teaching the LLM how to use\nyour bespoke API or how to query your\nbespoke relational database schema.\nLLMs have 50 years of training on the\nsemantics of file systems and it's\nunbelievable like how far you can get\nwith these semantics.\nAnd honestly, once you think once you\nrealize this,\nit's like it's like you're wearing these\nmagical sunglasses and all of a sudden\nlike everywhere you look, you just see\nfiles everywhere. Um, it's it it it's\nremarkable and and I'll be very honest\nwith you, this flies in the face of how\nI have been building software for my\nentire career. Like it's really\nmindblowing\num to to understand that LLMs are\nabsolute black belt\nkung fu masters at manipulating file\nsystems and working with files in ways\num that you can really only start to\ngrasp the surface of. And so and you you\njust start to see it everywhere. So this\nkind of came on the scene with\nclaude.md, right? Claude do claude code\ncomes out almost immediately.\nUh Boris and the people who work on\nclaude code started using a simple\nmarkdown file to help provide Claude\nwith more information on how to behave\nand and how to react and and to give it\nskills, right?\nThis became generalized into something\nthat is now sort of standardized into\nthe concept of an agent skill. All\nagents know what agent skills are. It's\nbeen standardized into effectively a\ndirectory, a skill.mmd file and a bunch\nof subdirectories.\nAnd why is this happening? Like once\nagain this is happening because LLMs\ninstantaneously understand what this is.\nThere is no training required. Think of\nevery GitHub repo that they've ever\nconsumed. Every Linux man page. They\nabsolutely instinctively natively\nunderstand what directory structures\nare, what files are, where and where to\nlook for things.\nLangchain has a deep agents SDK and\nframework. You can use this to build\nincredibly powerful agents that can do\ndeep deep thinking and deep work. And\nthey have a pluggable backend system\nthat you can use to give those deep\nagents all of the context and all the\ninformation that they need. And this\npluggable backend uses file system\nsemantics. write file, edit file, glob,\ngp. And once again, like this is no\naccident. Um, Harrison was just on this\nstage like two hours ago, right? Like\nthe engineers at Langchain are not using\ntheir crazy big brains to invent\nbespoke,\nsuper powerful\nAPIs for LLMs. They are leveraging what\nLLM already know and understand.\nThe open AI agents SDK\nthey have they allow enable you to\ndefine workspaces for your sandbox a\nsandbox agents to operate in.\nRecently they came out with a new\nversion of their SDK that supports\nremote file system mounts. So this is a\nscreenshot of the documentation for the\nbox mount, but there are many other\ncloud-based file systems that they\nsupport, but they also support box.\nThis enables the open AI agent SDK to\nmount a Box drive or Box folder as in a\nway where the agent feels like it's\noperating with a completely local file\nsystem. And I'm going to show you\nexactly uh what that looks like. So\nhere's my box. This is what the box user\ninterface looks like. I've created a\nfolder called team management and it has\ntwo files in it and they're just\nmarkdown files. Let me show you what\nthese files look like. So here's\nteam.mmd. It's just a t and by the way\nthis is all markdown. This is just sort\nof like the pretty view for it. Here's a\ntable of me and the members of my team.\nWhat's up Andrew? Um hello Olga in\nPoland. So there's three of us on the\nDevril team and this is us and here's a\nlist of some you know sort of fake sort\nof real tasks uh that have been assigned\nto the team. Let me go over to\ntasks.mmd.\nSo here is a different view of the exact\nsame data um organized instead by the\nstate of the task. Um and this\nultimately represents like the the the\nthe agentic application that I want to\nuse to manage my team. Now, you might be\nthinking like this is an incredibly\nsimplistic use case, but that's entirely\nintentional. I want you to understand\nhow this very simple use case works so\nthat you can start to generalize these\nsimple principles and build more complex\nsystems. So, now let me show you what my\napp looks like. So, it's just a Python\napp.\nUh almost all of the code in this and\nthis is all on on GitHub and you'll I'll\nshare the link with you shortly. Um this\nuh all almost all of the code in the\nCLI.py Pi is just Bootstrap for the CLI.\nMy agent is defined in like less than 20\nlines of code. Mostly just a prompt.\nYou're a helpful sandbox agent. Here's\nwhere you can find files. Feel free to\nread and write. Um, please inspect the\nfile system. Don't make assumptions. Be\nconcise. Okay. These are the\ninstructions for my agent. Um, a couple\nthings I want to dive into. Uh, OpenAI's\nagent SDK uh for sandbox agents. It\nallows you to be really really crisp and\nclear about what capabilities the\nsandbox agent has. Here I'm giving it\naccess to the file system. I'm given\naccess to the shell and I'm giving it\naccess to a skill that I created. Let's\nlook at that skill.\nSuper simple.\n25 lines of markdown the skill and\nbasically saying hey you're you're a\npersonal chief of staff for this team.\nHelp this team manage their tasks and\nany other you know team related things,\nright? So, a very very simple, concise\ntask, a skill. The task references a\ncouple sub subreferences. So, I can go\nahead and show you what those look like.\nUh, here's here's the here's some some\nuh some markdown that explains how to\nwork with the team.mmd file. Here's some\nmarkdown that explains how to work with\ntasks. And here's uh here's some\nmarkdown that that basically explains to\nthe agent explicitly that there is data\nin both of these files that is\ndenormalized. And I wanted to make sure\nthat these tasks that are replicated in\ntwo different files are synchronized\nacross the files. That's pretty\ninteresting. So, let's uh let's go ahead\nand do something fun. I'm going to go\nahead and mark in this web UI in this\nhuman readable UI. I'm going to mark\nthis task is complete and I'm going to\ngo ahead and save this file. So once\nagain, like this if if you're sort of\nthinking about agents that you might\nwant to build, this is a fantastic\nexample of human in the loop. Imagine\ngiving humans access to this box account\nand letting them manipulate these files\nthrough this human readable interface.\nBut now I'm going to go over here to my\nand I'm going to fire up my my uh my\nCLI. So I'm going to fire up Box Sandbox\nCLI\nand this is going to bootstrap my agent.\nGreat. So, here I am. So, now I'm just\ngonna I don't know just gonna ask us\nsome questions. I'm going to show I'm\nsay uh uh show me a status report of the\ntasks.\nAll right. So, it's, you know, it's\nthinking. Um hopefully it's going to uh\nknow that it needs to load the skill\nthat I gave it. Um let's uh let's find\nout. Live demos are always fun.\nAll right. Awesome. It's going to use\nthe team management workflow. It's going\nto start call. By the way, this is\noverly ver I've turned verbosity on so\nthat for debugging purposes so you can\nkind of introspect what the agent is\ndoing in a real userfacing scenario, you\nwould turn this off and have much more\nconcise responses. But you'll notice it\nhas detected that tasks.md is newer. Um,\nand it's going to synchronize both of\nthese files. It's going ahead to do that\nwork.\nAll right, perfect. So now I'm going to\ngo back over here\nand let's see what has happened.\nOkay, so we've got tasks.md\nis updated. Let's go to team.md.\nAnd there we go. So the agent saw that\nI'd updated tasks.md and it knew that it\nneeded to synchronize that with team.md\nand it went ahead and did that. So\nawesome. So we've got these two files\ncompletely synchronized. And what's\nreally interesting about this is that\nlike I did not have to connect this\nagent to an API. I didn't have to\nconnect it to MCP. I didn't have to give\nit access to a CLI. It simply used basic\nfile system semantics to like understand\nwhat the state of these things were and\nto update them appropriately.\nAwesome. So let's see. And I mentioned\nthat uh you want to make sure uh like\none of the benefits of using a\ncloud-based file system for something\nlike this is that you can you have an\naudit trail and you have inspectability\nof exactly everything that the agent is\ndoing. Um and of course this is powered\nby an API. Let's go ahead and go back to\nthe agent and let's get a list of all of\nthe events that have taken place.\nAwesome. So this is this is just a\nrepresentation of basically the audit\nlog for everything that has happened on\nthis account and you can tell the\ndifference between what uh what a human\nis doing here in this case versus what\nthe agent or the automation user is\ndoing in this case. So you can go ahead\nand uh and get the source code for this\ndemo here on GitHub and you can sign up\nfor a free Box account, Box developer\naccount at box.dev.\nUm, this is an incredibly simple uh\nthing to get set up. All you need is a\nfree box account. Uh, and you just need\nto have Docker installed. Uh, and you\nneed and of course you'll install the\nOpenAI agents SDK and you can get up and\nrunning. And like I said, this is a\nrelatively simplistic example, but you\ncan start to imagine the more\ncomplicated things that you can build.\nAnother thing uh that I want to make\nsure that we're sort of clear on is that\nfiles are just the interface. I\nmentioned this before um but like things\nlike proc and dev null are not actually\nfiles on a hard drive. Like they're file\ndescriptors but they're something that\nlike the agent can access. In fact,\nlet's look at that right now. This is\njust Ubuntu. So this is a digital ocean\ndroplet. I want to show you exactly what\nthis looks like because honestly like\nit's it's completely fascinating.\nSo I'm going to go ahead and do an ls of\nslashproc.\nThese are not files. These are actual\nprocesses running on this computer.\nRight? Look at this. I can do a cat of\nany one of these processes. How about\nthis one?\nSlashproc.\nOops. Oh, here we go. It's a directory.\nAwesome.\nThere's a bunch of stuff in there. So,\nlet's uh let's see. Um\nlet me let me cat CPU info.\nAwesome. So, this is a bunch of in\ninformation about the CPU on this VM in\nthe cloud that I'm getting out of the\nsystem using these file system\nsemantics. And I share this with you\nbecause I'm not asking anyone here to\nthrow away their databases or to throw\naway all their their APIs. That wouldn't\nbe productive and it doesn't make any\nsense. I guess what I'm suggesting is\nthat as you're building agents, be\nthoughtful about how you can build shims\nbetween the agent and that data. Um, I\nwas just talking to to someone earlier\ntoday about how the the Chroma team uh\nworked with Mitlfi to build a Chroma FS\nshim that allows Mitlifi's systems to\naccess data in Chroma using file system\nsemantics. Like it's just it's it's\namazing. Um, quick question for the\ncrowd. Does anybody know what this is?\nPlease raise your hand. Going once. What\nis it? What is this, sir?\nWhat? Absolutely. Wow, man. Five\nimaginary points for you. Um, it is an\noyster shucker knife. Um, I'm from the\nPacific Northwest. I We have amazing\noysters up there. Why am I showing you\nthis picture? This tool, this is an\namazing tool. It's only good for one\nthing. It can only shuck oysters. And in\nfact, if you try to use any other kind\nof knife to shuck an oyster, you're\ngoing to cut yourself. And you cannot\nuse this for anything else. It's a\ncompletely bespoke thing. Why am I\nshowing this to you? Because I feel like\nthis is Mike entire career, right? My\nentire career is some client or customer\ntelling me about this bespoke\napplication that they wanted to build\nand me as a software engineer thinking\nabout, okay, uh, what bespoke database\nschema will I design to support this\napplication that needs to be built? What\nAPIs, what completely bespoke REST APIs\nand endpoints am I going to build to\nmake this interface and this experience\npossible? I mean, I I we would this is\nthis is what software engineering was.\nYou would spend days or weeks or months\nagonizing over these details in order to\nbuild these experiences.\nBut what does that mean in a world where\nevery interface is collapsing down into\na box that just says like, \"Hey, what do\nyou want? What are you looking for? What\ntask can I complete?\" In a world where\nusers are trained to deal with these\nkinds of interfaces,\nwhat does this mean to you as the\ndeveloper that has to actually build the\nagent to perform these tasks and get\nthis data?\nThis tweet has 2.5 million views. Um,\nmost of it is because it's Andre\nCarpathy, like a super brilliant\nthinker. Um, but this tweet is basically\nabout the idea of an LLM operating\nsystem. Um, so you know, you go in the\npast and like what is an operating\nsystem? Oh, it's like some CPU. It's a\nCPU and some RAM and some disc drives\nand some IO, right? Now, today the LLM\noperating system, you know, uh, the\ncontext window is your RAM. Um, and\ntools are your system calls. And like\nthere's just, you know, all these\nmetaphors from the past are getting\napplied to where we are today. And one,\nand I don't think this is original\nthinking on my part. Maybe this is what\nhe was thinking. I I don't think this is\nonly because it's elegant. It's because\nLLMs understand what computers are and\nthey they understand these metaphors and\nthey understand these systems. once\nagain because they've been trained on 50\nyears worth of documents about what\nthese things are and how they work.\nSo, let's just put all of this stuff\ntogether um and and and kind of wrap it\nup.\nLLMs are already pre-trained for file\nsystems. every single thing that you can\ndo to give an LLM access to a file\nsystem or even something that is just\nhidden behind a file system interface is\ngoing to be incredibly productive for\nyou. Like I I I know I'm sort of\nhammering on this issue and I also know\nI'm biased because I work at Box, but\nlike it's really really true. Like you\nwill be these systems are so highly\ntuned for this stuff. It's worth it to\ndo things in a slightly inefficient way,\neven if these metaphors aren't perfect\nbecause LLMs are so good at it. It's\nkind of like I remember I I'm old enough\nto remember when people would write\nassembly code by scratch, by hand, and\nthey would argue that that was the right\nthing to do because it was like the most\nefficient and it squeezed the most you\ncould out of your limited CPU and\nlimited RAM. And I respected that. You\nknow, people built video games with raw\nassembly because they needed the\nperformance. Okay, we don't live in that\nworld. And but by the way, this is over.\nLike people moved on to higher level\nlanguages because the tradeoff was not\nreasonable. Like it was more it was\nbetter to use less efficient code\nbecause the developer could be more\nproductive with it. And so no one writes\nassembly code anymore. CPU has never\nbeen more abundant. disc drive has never\nbeen more abundant. Like these things\nare totally abundant. They're total\ncommodities. So being hyperefficient\nwith your custom APIs is is simply not\nnot the right not the right way to think\nabout solving these engineering\nproblems. Two, humans are trained on\nfile systems. Forget about everything\nI've said. If you are building a system\nthat has to have humans in the loop,\nokay, what interface are you giving\nthem? H how are they in the loop? Oh,\nare you gonna I don't know, you going to\nbuild a Slackbot? Are you going to I\ndon't know, build some bespoke web\ninterface or whatever? Like, yeah, sure.\nOkay. But like once again, like you're\ngoing to have to train them, teach them.\nYou're going to actually gonna you're\ngoing to have to build that that web\napp, right? Um humans are trained on\nfile systems. Like if you can simply\nleverage a file system that the agent\nand the human shares, it's it's an\ninterface that they both understand and\nit's less stuff that you have to build.\nAs I said before, file systems aren't\njust cool in a vacuum. They fit into the\nbigger picture of the LLM operating\nsystem and they fit into the idea of\ncomputer use. So, I don't know if anyone\nnoticed, but Codeex shipped a big update\nmaybe two weeks ago or three weeks ago\nthat built that has that included\nbuilt-in computer use. Now, think about\nwhat that means. So, Codex can operate\nmy Mac. That's cool. But do do I really\nbelieve that Codeex can manipulate\ncompletely bespoke custom applications\nthat I've installed on my Mac? That's a\nlittle bit less likely, right? like h\nhow would it how would it know like how\nto run a certain program and what to\nclick on and how to how to use it,\nright? However, it knows exactly how to\nuse the finder. It knows exactly how to\nmove files around on my Mac. So, like if\nyou're bought into the idea of agents\ngetting a lot of value out of like\noverall computer use, file systems are a\npart of that.\nNow a lot of what we're what we see you\nknow with agent skills and claw.md you\nknow those are all relatively local\nthings but you get a lot of mileage out\nof using cloud file systems right they\nsupport durability um they're multi-\nthey're multi-user they have versioning\nyou know for people that require\ngovernance and enterprise controls and\npermission systems these these things\nare all completely built in right so you\nknow depending on your use case using a\nfile system that's in the cloud can be\nincredibly helpful.\nAnd then ultimately, I guess what I\nwanted to leave you all with and like\nwhat I've just personally come to learn\num just sort of building a gentic like\nyou know applications on an almost daily\nbasis is that the winning engineering\nengineering principle is [snorts] to\nride the model's priors, right? Like\nthink of it like this. Every single time\nyou're spending one minute trying to\nretrain your agent at inference time,\ntake it just pause and ask yourself, why\nam I doing this? Like, should I be doing\nthis? Is there some sort of existing\nparadigm or interface or methodology\nthat the LLM is already\ntrained on that I as like the software\nengineer and the AI engineer\nthat I can just leverage out of the box.\nIt's an interesting thought experiment.\nIt has taken me a long time to untrain\nmy brain to look at every engineering\nproblem through the lens of like what\nbespoke schemas or APIs do I need to\nbuild to unlock these things. Um so I\nthink it's I think it's interesting. Um,\nso you know what, 50 years ago, uh, the\ninventors of Unix came up with this\nincredibly simple idea that like\neverything that a computer cares about\ncan be represented as a file.\nAnd I do not exaggerate. Every single\nthing that you all use today up to and\nincluding LLMs and chat GPT are all\nbuilt on this foundation. Every single\nthing. So this simple idea has scaled up\nto the largest and most complex systems\nthat human beings have ever built. And\nthe only question that we have to ask\nourselves is like why do we keep\nforgetting this? Why do we keep\nforgetting this? Um so yeah, I would\nencourage you um not to not to forget\nabout this to think a little bit\ndifferently about the things that you're\nbuilding.",
  "transcript_chars": 27830,
  "ingested_at": "2026-05-21T18:59:32.666437+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 93,
    "like_count": 2,
    "channel_id": "UCcIXc5mJsHVYTZR1maL5l9w"
  }
}