AI: What a Context OS is and Why I think its the Start of a New Internet

0:01 So there's three things I want to talk about today. They are related to context engineering. but this one's going to be more of a ramble than anything else. The three things I want to talk about are as follows. The overlap between what I'm seeing and actual building out context operating systems and seeing, seeing this like approach basically like emerge independently.

0:38 and then my thoughts on what that means for future of go to market the future of knowledge work. And then I have some notes here, kind of like structure my thoughts a little bit. Why? Oh yeah, so why that is why that's happening. That is so a context operating system. I think in the simplest definition in my view is that it's all it is is a way to get the AI to behave in a way you want, in the way you wish it did.

1:20 To like imagine Jarvis from Iron man where it always knows what Tony wants, right? He can always say, yeah, I know what you're talking about. This is information, this is the work we did last week and this is how it's relevant today. it's AI that always has the context on the situation or the task or the promise that you're solving. Turns out that that's way harder than everyone thought it would be in that the language models on their own were not just we're not capable of solving that.

1:55 and I can get back to that in a second. but the actual, the implementation of what something like that looks like is effectively what I find is the most effective and elegant solution is a markdown based knowledge graph. and there's different ways to slice and dice it. But fundamentally it's like markdown mark your markdown files, that are linked to one another. So there's like a relationship between your positioning document in your ICP document.

2:31 and those are all based on research from your transcripts, which are also in the context os. and there's an information hierarchy that allows it to be navigatable by not just human beings like skimming the thing, but by AI agents that can use, they can read the file, they can do keyword search that can list the directory and that allows for agentic search, agentic electrical search basically, which is just keyword search.

3:10 and that solves a few big problems. the big problems it solves is that you're never starting from scratch. Instead of working with your AI through this problem, you know, in your chat, GPT or cloud web app, and then you get to then the context window and you go, oh, it's Stuck. I can't do anymore. I can't feed anything. And, and you have to solve that problem all over again in a new context. When in a new chat, because you are reading and writing to a local file system, you are saving information outside of the context window.

3:47 When the context window has to compact or you have to start a new one, you can go and say, hey, this is how we write copy. That copywriting document is already there, it's saved, and the agent can immediately go, okay, this is how we write copy. That is. I think that's underestimated. How much of all of everyone's problems with AI, how much? That's like 50, 55% of it, as like the root cause.

4:24 So it solves that. And, and because it solves that, instead of having to waste time constantly reteaching the AI, what to do, giving it the context over and over again, it inverts it and it becomes a compounding intelligence asset, meaning the more you use it, the smarter it gets. Once you solve, hey, this is how we want to write copy. Or, hey, this is how we want to, this is how you want to score ic or, you know, this is, this is our ICP scoring, definition.

5:04 you can use that repeatedly. Repeatedly forever. So it takes the issue of like, you know, if you copy something and you make a copy of a copy of a copy, it degrades over time. The visual analogy of this is like copying in a photo and you copy it and the resolution just drops. this is why it flips that in that you're always copying from like a living source, from a source of truth. And not only is that source of truth, not only is it a like, source of truth that solves that copying problem, it adapts and changes and is always up to date with your current context.

5:44 That's the dream state. Right? so going back to the core of this first part, all the context operating systems I've built for myself and with clients, without fail, all, all of the clients. I don't have many, by the way, because I'm being particular with who I work with. And I'm really grateful for the partners I have right now, the few of them I do, because I can learn stuff like this without fail.

6:16 They've all gone and been like, yeah, I took this system and I made five versions for myself for different topics, or I made my own personal context operating system. and I've made several other versions of it, which is amazing to see. it's really Cool. and then also having conversations with people I'm not working with directly. Like they're using analogous techniques around, using markdown in cloud code.

6:47 and git, which is the, I guess other part of the triangle there. It makes me feel. My sense is that as much as I'm like innovating, inventing and understanding, like it's me doing something right, I'm also documenting, I'm documenting this like emergent trend of technology.

7:19 and what I mean by that is like I think we take for granted that technology is this thing that's like this like fortress force of nature. It's just going to happen. That's not the case. It is like significant effort and ruthless iterations like Edisonian iteration, like the 10,000 light bulbs thing. I lost my train of thought there for a second.

7:50 But basically I think it's both. The, the, the, the core idea I'm trying to communicate here is I think there's. This is going towards something larger. It's not just, you know, it increases your iteration tempo for go to market, which is super valuable. increase your visibility which gives you better intelligence for making decisions. like not the general frames I use to kind of explain the value props for go to market, which are like very true, and highly relevant and you should care about them.

8:30 but you know, this stuff is applicable to basically all knowledge work. I could go and be like, I want to learn, an immediate one that like I do somewhat semi is I'm trying to like track my workouts or track my health. Like this is the foundation of personalized, of like ultra personalized medicine. Or you are, you know, you are a professor teaching philosophy.

9:00 You have your philosophic, like philosophical philosophy knowledge base. There you go. and it is a new way of working with computers, computing. I think that's where this is going. So that the, no, the logical thought process I see happening here is all right, you get it working for one person and it makes you work 50 times faster, 30 times faster, whatever it is.

9:34 it's a different way of working with a computer. But once you get the hang of prompting and context management and how to verge and stuff, you get the hang of okay, this is why this is valuable. I can scale my expertise, infinitely with zero marginal cost of production. The next part then is, well, how do multiple people work from the same living source of truth that one's built into?

10:07 using Git, to version control. A version control system, which is what Git is to you turn your context OS into a GitHub or just Git repository and that solves, multiple users working from the same source of truth. But there's no permissioning, there's no, you only have access to this part of the context os and there's no controllability in that.

10:38 Let's say I want to open up my context operating system section of my knowledge base. I can't control how other people's agents interface with that right now, at least if they're, if you're just going to say here's access to the GitHub repository, which then the next step I think is then, well, how do you say, hey, I want to grant programmatic access to my, this section of my context os.

11:10 Let's say it's your, you know, your copywriting principles. how do I give someone access to only this portion of it? Like you're sharing, like your Google Docs sharing your you know, your visibility settings on social media. Whatever, whatever the analogy is that makes sense for you. the way I see that working is that you have open protocol, right? So probably MCP and you have another, you know, AI in between that you tell, hey, this is how I want my context to show up in other people's systems.

11:48 This is how you interface with it. And then this is how you, this is how you interface with the context internally. And this is how you answer questions when their AI asks you. The AI is not a you, right? But when it, when it tries to, when it sends a request for information and that's how you control the quality of the context that shows up. This, that's how you control, you know, it's authentication, right? and that's permissioning, permissioning and all these things together.

12:21 It feels like it's a new layer of the Internet emerging that is bigger than my specific context OS implementations, or frankly anyone I'm talking to or working with, it feels like it is the natural evolution. And that's really crazy because going back to that idea of like, it's both a ton of intentional work by very smart, hard working people. Right? In general. Right?

12:51 Not, not me. it does seem like, well, this is the natural conclusion where we're going to get one plus one equals three. Where you have all these systems coming together to make something larger than the sum of their parts. so that was the first topic, I guess the interplay between. I feel like I'm like actively innovating which is really Fun. And I'm really grateful I get to be in this position.

13:21 And also like documenting my findings is what I'm seeing. I talked about the emergent layer of the Internet. I guess then we can talk a little bit now about the different ways this will go about technically, from like a how perspective, and why I think it'll go that way. So getting more concrete, the, the core of what I think makes all of this work and at least like what I'm betting on is simplicity wins over complexity over time in complex systems only are useful when, when you have a bunch of very small sub components coming together to get that one plus one equals three.

14:18 There's like some law or like rule of thumb or something called this where it's like you can't design a complex system and expect it to be useful or valuable from the start. It's an emergent thing. and getting into the technical parts of this I think it really comes down to how are you giving, how are you tooling the systems around and controlling the base foundation model, to solve the context, to solve the, you always have to give it the context over and over again problem, and related problems like hallucinations and shit like that.

15:07 But I'm going to focus on that right now because I think it's the most important one and I really do think if we're going to apply that principle and what I've seen work empirically, it basically comes down to a vector database and semantic embedding, versus a far more simple agentic search against markdown against markdown based, well not just markdown but agentic search against plain English encode.

15:43 So we could bucket that as like markdown as JSON, as YAML, and then your object oriented programming languages like Python and JavaScript. sir, go throw in Rust if you want, whatever. I'm not trying to, the thing I'm trying to get to is when you have a markdown based solution, your context os, it is both human readable and machine readable.

16:19 Both you, the human being and the system you're trying to have interpret that information are working from the same level versus if you go try and use a vector database and you say all right, here's all of my documents and transcripts, about this problem set. This is all of our sales transcripts and we're going to embed these, that becomes a probabilistic black box. and not only is it a probabilistic black box, Is that it's, it becomes a, it flattens the contextual relationships between ideas, people, places, things like your ontology, basically like how you group things to make sense of the world, versus if you solve it in, in a context architecture perspective where it's markdown based and you have metadata around.

17:16 This is the summary of the call. This is how this relates to these people in the system. And these are the main topics that were discussed. Let's say that's like your organization pyramid. The answer quality you get from an agent querying, it's that using something like a grep in file search. Basically it's one farmer like if you asked human being to go and search through that and summarize what it found, it's like far more, it maintains those relationships far better because it didn't flatten that information.

17:51 It's auditable because you can get the agents with a reasonable level of confidence to say I got this idea from this file that was linked to this other file. And I summarize them this way. That is an audit chain, right? That's super, super important for coming back to the next idea or to the idea we mentioned previously. mitigating hallucinations. Because my core belief is that there's no fundamental solution to hallucinations with this current architecture of AI.

18:26 it's a feature and a bug, it's just inherent. What you should solve for is making generations verifiable. So rather than saying well we can't, instead of trying to push the model towards it's never going to hallucinate ever. Which is not like, doesn't make sense technically the next best solution is saying well we can get it to make a claim that we can verify and it's either true or not, but like it's moved towards being able to be falsified, which is what that audit chain does.

19:09 just trying to collect my thoughts. Yeah, so that's the big argument from my point of view, on the value of the far more simple approach. Oh, the last one I guess is there's just less of an engineering loop in between. someone actually working against the markdown files versus someone saying well we have a bunch of files. Hey engineering team, can you go vectorize these please?

19:42 And tell us how the output changes instead of just make a change to the file, see how the agent query has been fit. It's like a one to one, right? I think, I think you should remove as much as many layers of abstraction between the user, the human being, the context and the AI agent, the coupling there needs to be as tight as possible and remove anything between that gets in the way of it, get rid of it.

20:10 It's worthless. it's not worth this. It's actively harmful. Harmful. so that's, that's my kind of take on how this is going to happen from like a technical perspective is you're going to get, you're going. Because you're going to have simple pieces of technology like that. All the things that make again a gentic agentic.

20:43 Let's go search work like grep and grep is from 1973. Goaded. Right. or like versions of it and maybe like a bit faster. But fundamentally the these, the system that is more simple and the system that is more simple is going to be more effective and it's going to win and you're going to instead, instead of tacking on these unnecessary pieces, eventually you're going to get to the point where like wow, everyone has a lot of context that they want to share around what's the way.

21:23 How can I grant someone programmatic access to my context but still control that? And I think you're going to see context sharing. You already see it now but you're going. That's the point I'm trying to get to is this is where the emergent nature of that context driven Internet, comes into play. So. I'm trying to like figure out a way to tie it all together. But yeah, I guess this is just like my, my perspective on context engineering and context operating systems and where I see the space going and I guess to make it tangible for why you should care, you should care about this because, Because you can read English.

22:28 I hope you can if you're watching this video, let's say because you can read written language. Written language is now code. You can take any thought and figure out how to structure it into this executable thing as executable program that your computer can now go do. And it can do it for a zero marginal cost of production which means you can take your expertise on given subject or even just like your point of view, and it scales infinitely.

23:07 Right? That's a ton of very powerful, that's a ton of leverage that I think we're all still trying to figure out. But the better you get at clarifying your thoughts and getting into a format, right? Taking your taste and getting into a format that can be executed by these novel programs, these AI agents, The more value you'll be Able to create, whether from an economics perspective or like, man, it's just really cool that I can learn all this stuff about my favorite autistically obsessed, subject.

23:47 I don't say that in a derogatory way, by the way. That's just like, this is the thing I care so much about. I love it. Right. And you can scale your, your understanding and perspective on that topic or like the health thing, right? That healthcare example I gave you. You can suddenly know everything about all these pieces of medication or supplements or protocols or whatever you want to say that you didn't have time to research before, but now they can be accessible to you.

24:15 I guess you could think of it as like AI is democratizing access to expertise and taste. Well, that's a new idea. I just had that one. so in a world where that is the case, what does, what does, I guess, success, if we're going to go solve for that, look like for you and I think it's be a lot weirder, than we think it will be.

24:51 I know it's a super esoteric place to to end the video. but yeah, that's the point of us. A rambling video, I guess. So sum it up context. Oss are goated. You should try and build one for yourself. I will probably. I will, I will try and make like a. Hey, this is how these work. This is how to build these things, video. And I think it's going towards something larger.

25:22 And I think that larger trend is going to be driven by simple sub components that come together to make something that is both highly effective, elegant. and greater than the sum of its parts. And winning in that situation. Looks like, Still an open question. All right, thank you so much for your time.

25:55 I I really appreciate it and I hope this rambling video was at least interesting, maybe valuable, who knows, but at least interesting. Thank you so much for your time. I hope you have a wonderful rest of your day. Bye.