Canonical Knowledge is Key

All right, well, Joe, it's great to see you again. We're really excited to have you on the show today. When we were talking last, we were kind of discussing canonical knowledge and this idea that organizations if they really want to take agentic AI seriously, one of the first steps is to create a source of truth for their organization. We've had conversations with Jeff McMillan and David Wu at Morgan Stanley, and they actually worked with Sam Altman and people at OpenAI pre-chat GPT on kind of building that source of truth for their advisors. And they, you know, they discovered that it was this really laborious process and that it was a lot more than just dumping data and feeding it to an LLM. Like they had to go through and figure out systems for assigning like a time to live to documents, making sure that every piece of information had a human attached to it that could provide oversight. So, you know, it's a ton of work, but it really is that I think one of those critical first steps. So maybe we could start there. I know you have lot of thoughts about canonical knowledge and then you had mentioned the idea of a semantic layer being a key piece. So we could get into that a bit as well.

So I think that what you're describing is kind of a remarkable change of events over the way data worked 10 years ago. So 10 years ago, the whole idea was that you structured data specifically for consumption. You built a data warehouse. You kind of really carefully organized this data. And you knew what it meant. And life was kind of easy and straightforward when you were basically going through that. was political, your notion of agreeing on truth.

was based on kind of a formalized business requirement. But now what you just described is like, what's true on the internet? Like what could be a more profound question than that? Like how do you actually have, you know, describe some kind of crazy source of truth? So what has been the irony of that is the thing that you just described that is so very difficult resulted in LLMs like ChatGPT and Gemini that are actually really good at siphoning through public content and unstructured data.

an inferring truth. But they did that. And they kind of forgot the first part. They kind of forgot like, by the way, there actually is a real life calculated value for revenue, a calculated right and wrong answer, right? So those things aren't inferred. And so if you kind of think about it, LLMs are really just predictive models that anticipate what you probably would mean, what you probably would interpret. And a lot of classic governance is actually, that's not

That's not true. Your opinion of this is actually not particularly relevant. We have local standards, right? And so I think that what's happening right now is the convergence of these things. Like, how do you actually start to take advantage of the way that LLMs have started to understand that which is out there and then combine it with that which you spent all this time and energy to define and consume and create standards around? And so that's been, I think, the big challenge because one of my colleagues said to me one time, you

LLMs are pretty good at words, but they suck at math. They suck at math. that's our...

Yeah. Well, depends. Yeah. Probabilities, no, but yes. Deterministic math for sure. yeah, I've always, I don't know why in my engineering brain, you know, that side of it, I'm like always thinking about implicit and explicit knowledge and information. And, and in some way, like if I took an LLM and I tried to make it a thing, if I like tried to describe Like in an overall system, what is this? What is this piece? Like if I was looking at a car engine and I looked at a, you know, the, the fuel injectors, I'd be like, well, what's this thing, right? What does this thing do? And so if a kid looked at me and like pointed at an LLM as a part of an agentic runtime system and said like, what does this, what does this piece do? I'd be like, well, it, it's really good at converting implicit to explicit. knowledge and explicit to implicit and sometimes explicit to explicit or implicit to implicit. But generally speaking, people are really impressed when it converts from implicit to explicit. That seems to be when they were like, wow, how did it know? And that could be a sentence to a complete application. and, and it's, but when you think about that and you're like, okay, so it's going to go from implicit to explicit, but there are multiple ways. There's no right answer. When you're going from something like, what did you imply when you said that? Well, you could have implied 10 different things and all of those things could be correct. So I don't, it's like this idea that we could correct the thing makes no sense because it's literally 10 right answers. There's no one right answer. There's no way to say, we can make these things accurate because accuracy is just in the eye of the beholder, right? Only the context.

So you know, like for generations, we've known inherently that there's kind of a difference between B2C businesses and B2B businesses, right? They just do sell differently, you act differently. And what you described is very much a B2C kind of a construct. The world is out there. We all have different opinions. We all have different interpretations of the world around us. And that's reasonable and that's logical. A lot of times what happens in a B2B context is, you know what revenue was last year?

$1.625651 billion, period, exactly. But I don't want it to be true. have different opinion. Your opinion is not welcome here. So as you start to think about how do you bind that, if you think about this implicit, people, the beauty of Geni is what you just described. How do you infer? How do you make sense of an overwhelming amount of But what's your launching pad? So our hypothesis at Workday has always been like, OK.

I've got a problem statement. I had X problem with my business. This product isn't doing very well. Or I had attrition in this group for these employee groups, right? The first step is, do you have your facts right? So I need to know that this problem actually existed. This is actually what revenue was. This is what churn actually was. And then if you can somehow create an inference system that starts to label on top of that to say, how do you interpret that? Then interpretation is wildly powerful.

We're creating some of the things that humans do on top of that. part of the thing that is a complete drain in an organization is when 10 people walk into a meeting and everyone's got their own set of facts. You really just agree on what to do to solve a problem. But imagine that you could deploy agents to use your words to create these explicit things from the implicit things. That's where the power is. You can set these things to make you smarter.

Mm-hmm. You Yeah, so I think it's an alignment question. It's not a right-wrong question, it's an alignment. What are your values? What do you want it to mean? And let's make sure that that's what it means to other people that agree with you.

Well, in that case, isn't that governance from 1990? Not much has changed in a classic governance context from 1990. And then what happened is when we, the process that you described around how LLMs were largely trained, we didn't use any of that. We did nothing to reestablish those governance principles. It was a different muscle that we were doing. So now I think what you need to do is to put both back together.

We can't make much progress as a company if we can't agree on how much money we made last quarter by product, by geography. We can't decide where we're going to hire, where we're going to invest if we let people create their own realities. So the answer to that is we've been trying to say, I think people are clear that there's protocols that we're establishing, MCP, and different ways that things can talk to each other. But what is the actual way that you can separate

the implicit and explicit workloads. Turns out implicit workloads are kind of easy nowadays because you just ask me a question from some unstructured data and I'll generate an opinion on anything you want.

Whereas on the explicit things, we are seeing the dawn of the importance of semantic layers that can wrap around that data that really can speak to machines in the same way that they spoke to people. So one of the ways I kind of try to tell the story is if you were a business person and you said to your BI analyst, hey, you know what? I'm hoping to get a list of the swaps we had last quarter.

What's that? What's a swap? What do you mean? What do you want? Like, so there's this there's this kind of idea of inferring what you mean. I don't know about you, but as chat GBT once ever asked you, what do you mean? Oh, no, no, no, no, they're off. They're off and running and doing this stuff, right? So this idea of trying to simulate that idea of

Yes. Unless you ask it to ask you, what do you mean?

Exactly.

Yeah.

Right. And so the idea here would be like, how do we get the prompt to ask you back and clarify what do you mean? And then how do you actually structure the question against a known corpus of data with known definitions, right? And that's not instead of, that's in complement to. Sometimes there is a right and a wrong. Sometimes there's just a prediction and those two have to live equally side by side in our brain.

Mm-hmm. Yes. You got it. Yes. Yeah. Yeah. Yeah. But we're also, I think most people are lost in the space even as to how, like they're like chasing the wrong sort of the wrong branch of that tree because I always use the weather as a great example. I'm going to just, I'm going to jump in all the weather data from all of my, it turns out that 50 % of my customer service calls begin with how's the weather out there? Right? So I'm to dump all that in and then I'm going to feed it into an LLM and then someone's going to call in and it's going to say, how's the weather? They're going to ask, how's the weather? And it's going to go sunny every day because most of the time that was the, it's nothing to do with what the weather is. Right? It's just that I fed it all of these sentences that said sunny and nice. And then someone's like, well, It's so stupid. It doesn't know what the actual weather is. And you're like, well, it's, it's not an intelligence in that sense. It's, it, it's not, it's not logical. It's just an interface guys. This isn't, this isn't data. This is an interface and an interface to what an interface to other systems that are deterministic. And once we realize like the LLM is not the system, it's just the interface. to the system, then we start to get to understand that we're way overemphasizing these things as being the center of the solution instead of the new interface, which is extraordinary to our old systems.

So you are starting to see this emergence of the importance of context. And effectively, what you're describing is just that. does the LLM, which is just the front end, understand the context of your question to know how appropriately to route it? I am not asking. There's some actually, there's a story out there on the internet that if you ask all the three big LLMs, give me a random number, they'll all say 57.

And it's the same logic, right? It's not to do, it's a predictive model about what we would anticipate someone would say under these circumstances. And so what we've been trying to do is to say, ask someone else a question. What is the nature of that question? Is it a deterministic question with a real right, right, wrong answer? We route it this way. Is it an inference question? Is it like, all of these different things require different strategies? Cause you're right. We need the front end. The LLM can just handle those things. But if we just leave it by itself, it will just

predict what you probably wanted to hear as opposed to give you the answer that you

Yeah, it's like building a front end UI without a backend. You're like, your front end without a backend is, is just an interface. And we would never ask the question. Like we'd never, we'd never expect it to know anything.

Yep, that's exactly right.

If it didn't have a backend, it's kind of silly if you think about it. Why do we even expect it to even know this stuff? It doesn't know the weather.

Well, so the way we've attacked that is we think about this idea of there are notions that been around for a few years around the construction of data products, thinking about data inventory with a product owner, with a backlog, with features and attributes, and really things that you commit to. And that turned out to be a really wildly successful for building out this back end, as you suggest. It's like, this is my data. It's right or wrong.

And then you complement that with governance tools. And all of sudden people can say, that is my standard definition. This is where I find the data. And what hit us over the head was just how human centric that approach it was. Because as soon as we set an agent loose on it, it was completely confused. And what was missing is the emergence of this translation layer that has to sit between the data and the agent, which is largely structured in YAML files.

And by the way, each of the different AI tools has a different syntax for the YAML. So the of coordination of that semantic layer that translates your back end source of truth to the machine's ability to understand it is actually quite tricky and really resource intensive. So Josh, you were saying like, we needed to kind of go through all these websites and assert truth. That's great. And now every single company is trying to face that same very reality with its own data. So how do I?

end up the inside and the outside because I need to go through the same activity of trying to figure out how I define truth and then how I define it in semantics that the AIs can understand.

Yeah. And we've worked with NASA and Roger over there who pretty much, you know, and yeah, he invented the NASA knowledge management system, which is a very unique approach. And we modeled our system after that. And it's just fundamentally an understanding that knowledge has to be an abstraction, like you said, you know, knowledge is not

Roger Forsgren, yeah.

the documents that you have in your document repository. is not the data you have in your database. Knowledge is the abstraction layer on top of that. And it has to be canonically correct. And to be canonically correct, means like some human, human in a loop has to sign off on that data. They have to say, this is truth. I am appointed within my organization to to determine truth within this domain of knowledge. So I am the person or persons team that is assigned the truth knower and decider. We call it alignment, right? It's who gets to decide how to align LLMs, right? Well, the LLM provider does. They decide alignment at a hierarchal. top level. Now companies need to align within their organization. We call it OAGI, Organizational AGI. And it's, now you need to create your own alignment system. Because alignment at the organizational level needs to be far more aggressive than alignment at the top LLM level, right? Because you're talking about a much higher opinionated structure. And then the question is like, who can align that? Can some vendor align that for you? Like, absolutely not. We're talking about your values, you know, your, your core, your knowledge, the things that are in people's heads. and then you've got to, you've got to create a layer of that knowledge that is truth. And then, and then you can start using agents. Now you're ready to start using agents.

Aha. So one of the things that I, that as you tell that story, sometimes I think about in the AI age, how much of that story that you just told was true 10 years ago. And I would say almost every word, almost every

Absolutely. Here's the difference. I'm glad you said that. Here's the difference. But every company, I'll call it lazy. It's not a good word to use for it, but every company was too lazy to create and care about data. your career is a career of trying to get people to care about things that they should very, very, very much care about and could give a flying F about. And

person real. so sad and so true. And this is why CDO is only the average lifespan is like two and a half years, right? No, you're absolutely right. So I think that what you're just talking about is in a world where this has always been true, imagine the wind that this puts in your sails as a data professional to go, thank God people finally care about this. So the thing is, if you are in most organizations, there is a certain kind of bureaucracy associated with data management that is really annoying.

Right. my god, I know. Yes. Mm-hmm.

And so what you do instead is you create your own localized, this is my posse of people, and they create my charts and my graphs and my insights and they give me what I need whenever I want. And that works so long as I don't, and then once in a while I bump into some other different variation of truth and I fight with them for an hour and that happens from time to time. And this idea of the relative importance of truth, it's not what keeps people up at night. But now,

you have CEOs that are fundamentally trying to say, I actually don't really want to have hundreds of BI developers. I want to have agents that are going to take over some of these functions, right? And if you can start to make the case that's like, the fundamental foundation for agents is data. All of a sudden, what was once kind of an unnecessary bureaucracy is now front and center in the agentic. Right? It's not any different. It's just psychologically different. I love it. What a great time to be alive.

Yes. Yeah. Yeah. I know, especially for data person. I think the irony of being in data is that your job is to measure things, particularly like ROIs. And then some CEO comes to the team, their data team, and says, what's the ROI of canonical data? And they're like, that's very hard to measure. So I'm not going to invest in it.

So it's good that you said that. I'm so glad that you said that, because I'll be participating in some CDO boards and different things like that. And there are often these discussions around, like, what's your ROI of data? And it's like those old MasterCard commercials. It's priceless. So I've basically given up that narrative. There is no ROI for data itself. And the metaphor that I say is like, so imagine you go to build a house, and you want to talk

your builder about granite countertops and about marbled bathrooms, and they go, great, I'm going to spend a month on the foundation. You can't say, I don't want it. Right? It's just part of the deal. And so what I've tried to do is embed the data part as the bill of sales, the embedded in terms of what is required as part of the cost of delivering an AI solution. And I think no one ever says like, it should be free. It's not as a cost. So it's a cost, but here the value is so much more explicit, right? Like I'm delivering that agent and that agent, I'm going to deliver these kinds of capabilities. And these people are so desperate for them that the ROI, it's not my ROI. My joke is always like, I'm like a Bette Midler song. I'm just the wind beneath your wings, So that's what we're really trying to do is we're trying to not so much quantify the value of a canonical model, but really reinforce the importance of it, the essential need for it.

as we start to deploy these agents at speed.

Yeah. Yeah. And I think, I think like history, and this is something Roger and I have discussed many times. History tells you that you'll never get people excited about spending money on data. and, and that now is okay because what we've been able to do now is automate a lot of the stuff that would have been manual to manage canonical data. We can, now that we can automate most of that, there is like hope. You know, now, now you like your point, A, you have to care about this, but, B, it's not as big a lift as it used to be because a lot of the mundane tasks of, of, know, managing truth can now be done. And it's this interesting thing though, this like this, this loop, because you're like, well, the first things you should tackle aren't the use cases. The first thing you could tackle, they're use cases of managing knowledge, right? And then they're like, but what jobs can we eliminate? We're like, well, you're not doing it, but you should be doing it. And so you get into this like, but how will we save money? And you're like, well, you'll save money later by doing the thing you should have been doing that you can now do much more effectively and efficiently using AI. And it gets into this like, but I... But I I want I want use cases. I want use cases. And I think that's a a tough spot for companies to be in because I don't think there's a path from use cases to systemic change. think, I think you have to, you have to work through knowledge. It's like, it's like teaching someone in school, not teaching them how to read and write, but just teaching them how to answer a phone. Like thinking that we would teach humans by giving them use cases, you know, and just let them like figure out how to read because we give them, you know, examples of, you know, here, you know, here's a label from a piece of clothes. Here's another label from a piece of clothes. Now figure out, you know, that this is shirt and this is pants.

Yeah, I think you're on to something, but I would argue that the thing that separates the most successful people in data from those that come in, talk a lot about taxonomies and autonomies and lose their job a year later, is their ability to connect the dots, right? There's no way that I can own, manage, or inventory all the possible use cases of data. But the strategy for anybody who's in the chief data officer position is to figure out.

who are the people that will value most in their use cases? And now the answer is principally AI ones with the data. And then you essentially contract to them to give them what they need. And much in the way you build a house, you're signaling to them, I know that you think I should just be able to pound stuff together in an Excel spreadsheet. But what we're going to do is you are going to be here a month from now, and you're going to say, I need to change it, and it's going to take me a month to build and two months to change that and three years to change that. Can you just give me some rope?

And for people that have come from companies, they will give you that rope. But you can't do that universally. So there's kind of a paradox. Because what you're saying is true. The first thing to do is establish the foundation. And the foundation has no value. Well, how am going to do that? the first thing is not that. The first thing is identification of a use case that allows you to flex that foundation more broadly. So when I first got to Workday, finance is always a really

compelling use case because people will fight over numbers and they will, you there's things in the accounting that you want to reconcile. Truth matters. Great. Also we had, you know, in software industry, there's always friction that exists between marketing that thinks it's doing a terrific job generating leads and sales that thinks that marketing is doing a terrible job. So we picked these use cases specifically to help executives reconcile problems.

and explain to people that we were going to slowly build up the foundation underneath them. And now we're 18 months later. And in some respects, you could argue you're distracting them. You're distracting them with use cases while they give you liberty to pursue the foundation. But I'm not going make any friends in the executive suite by talking about canonical models and ontologies, even though I know the most important thing. What I do say is people get the concept of garbage in, garbage out. You know that AI relies on

data and I say like just so you know that's hard I got it and they'll give you a little bit of a leash to do that but if you can't really articulate who's gonna use it I think there's a reasonable question that's what the heck are you doing right

Yeah, yeah, I agree.

Yeah. Well, I think who's going to use it determines how you build the data model too, right? if it's interesting, cause like software, this might be a simplification, but it seemed like software in the past was a lot about telling machines exactly what to do. And now it's sort of morphed to where a lot of it is about telling them what not to do. Like if you, if you ask an LLM a vague question, like if you said like, what's IRA, you know, it, it'll say, well, it could be this.

Yes.

savings thing, it could be a person's name, like it'll come up with all these different things that it could be. And that data and that context as to like who's using it allows it to say, well, this person's a financial client, they're asking about this and I have this information about them, they're probably asking about their IRA and it can narrow all that and make it useful.

Yeah. Yeah. Yeah. We're coming back to alignment. You can't align something if you don't know who you're aligning it for.

That's exactly right. Josh, I really love your point because effectively, you know, broadly speaking, if you talk to data nerds like us, they would say the canonical model doesn't really have boundaries. It just kind of expands forever to kind of take on whatever you do as a company. But what you're describing is super important, which is you got to start somewhere. And so if you're going to start with somebody who's going to be like, run wealth management. I want somebody to go to my clients. Great. What is the vocabulary of their everyday life? What is it? What is it that they need? And then you can start to build out a corner of your canonical model.

and then start to append to it as you move through. I think that's super important.

Yeah, you built stuff, Josh, recently. You showed me something. Could you show it?

yeah, yeah, I can give you a look. This might be useful to this conversation. Let me just screen share.

This is using an agent runtime, our agent runtime.

Well, so for our book, we created sort of an AI agent that's trained on the knowledge in the book. But, one of the things that was funny to us even about writing a book in the first place is like, or especially a book about AI and agentic AI is like, you know, this is on printed paper, like the first edition of our book came out before chat GPT. So but now we have this podcast. And so what I really wanted to be able to do was because the the book agent is

I agree.

is sort of an expert on agentic AI by virtue of, of having all the information in our book. Now I want to start adding the information that we're gathering as we have conversations with, with all these experts. And so I have this knowledge model here and what you're seeing, the color coding kind of refers to different episodes of this season of the podcast. And so I took transcripts of each episode and then generated summaries and then went through manually and made sure it had the right information and was calling out the stuff that I thought was really impactful from each episode and then fed that into this knowledge model and then it creates all these tags. I'll zoom in. is actually an area where I've been

And this is what you're talking about, like boundaries of knowledge. Like, so if you had a meeting and you took a transcript from that meeting, like this call, we wouldn't talk about where we went on vacancy. Like none of that would get sucked in, but only the pieces that are relevant to the boundary of this expert, this agentic model, right? But it's always learning. It's not like a static, this is

an episode. is a this is now becoming an expert in a specific area like OGI. Sorry, Josh didn't.

Yeah, no, that's all good. And like this area here is actually, this is where I've been grooming recently. So you can actually go through and open these up and verify the information and connect it to different tags. But this is from our conversation with Steven wit, who wrote, the thinking machine, which is about Nvidia and Jensen Huang. And you can see like some of the ideas from this episode go to touch things that Brian cotton Zara, who, works at Nvidia. told us when we talked to him. And then there's even ideas that branch over here to things that we learned when we spoke with Karen Howe. And as I feed it more information, it grows in size. So it becomes challenging to you really, for me anyway, I've just kind of, dive into specific areas and I start checking these tags and I start checking the information in the notes. And then I'm thinking about how does it connect? Is it missing connections? Do other connections in my mind that I think like, this person in this episode, I remember them saying this thing, let's figure out how to map that.

I think it's brilliant. I described this to my team like the International Space Station. Every single module does something by itself, but somehow they connect the broader ecosystem to make the broader ecosystem better, right? And there's connectivity and you need to figure out how they're connected. What you've done here, which I think is brilliant is there's some kind of like purpose to why I launched that specific module. You're just like, I don't know, that module just happened to appear. How is it connected to everything else? I think it's brilliant. Yeah.

Yeah, and it's self-learning. it's like listening to his email box. It's listening to Slack channels. It's constantly being fed. And then what I think gets really interesting from a canonical knowledge standpoint is in his IDW, it's pushing the new knowledge. It's learning to him on a daily basis. So it's keeping, it's flipped from the old days of you need a team to keep your canonical knowledge up to date. To now, the canonical knowledge AI system is keeping Josh up to date. now we see, okay, organizations in an agentic world can now, it's just the push of a button, just decide, have all these experts and on the far right there, it's MCP. So now they can talk to each other, right? And share knowledge. It just gets...

Well,

It's such a great time to be in data, isn't it?

can we come full circle to your comment around the implicit versus explicit here? Because I actually think that has a lot of meaning in this context, right? So a lot of what's happening in the context of Josh's graph here is more of the derivation. There are implicit things that are happening, and you're trying to make that implicit explicit, right? Now imagine that I said to you,

Yeah, yeah, I love that.

Workday's revenue last year was $9.8 billion. And that is not probably something that we should triangulate with other companies. That is something that is, in fact, true and verifiable. So now we come full circle to the whole part about how do we tell the difference? How do you actually get data nerds that are traditionally about right and wrong?

huh. Yeah. Right. Mm-hmm.

and say, hold on, hold onto that with every fiber of your being. We can only have one revenue number. We can only have one employee count. There is truth and there's value into the organization. The lack of friction that that creates, that is so valuable to just have the answers at your disposal and deploy analytics to fix it. Right. And to have these kinds of things color it, surround it, inform it. Right. I think that's where this power is because. So what I'm kind of thinking about right now is your

Mm-hmm. Mm-hmm. Yeah. Yep. Absolutely.

collecting semantics and you're using LLMs to bind them together. Very, very powerful. And then there's an inside out bit, is just like, how do I actually make sure that words are governed and have meaning and complement that? And how would we insert those same concepts into this very same graph? That's, think, the struggle of our generation right there. So ready.

Mm-hmm. Yes. So, so let me try to blow your mind here. Josh, if you just clicked on one of those notes, right? It was very early on a couple of years ago, we figured out that as we were, as we were doing this and we were saying, okay, so, so this is an idea, right? This is a canonical idea. And therefore if some, if we, if, this idea comes up in a conversation somewhere else,

sure.

It won't exist as a duplicate note. It will have to reconcile with this note, that idea. And LLMs can determine that this is the same idea, you know, because we're not looking at bits and bytes. We're looking at implicit data. So we can now say, okay, this is the same idea. Some human has to reconcile these two things and decide which alignment we want, right? So this is aligning it. But we all of a sudden realized, what if we put code in here? What if we took out the language and put code? Could it build applications on the fly canonically? And isn't our business processes in code already? A lot of them that are canonically correct because we manage code more canonically accurate than we manage our documents. And so we played with that. We actually put code in here and realized that it can execute and create applications on the fly. execute them on the fly, ephemerally. And you lose this whole concept of pre-created apps, right? And now you have canonical knowledge stored as code because code is more capable at being precise than language. And it's crazy how well it works. And then you mix the two. You have language and code together like a Jupiter notebook, essentially.

And then you get this like crazy power where you can ask the weather.

It's interesting that you are on to this point around the power of code. And it's so easy to read code. It's so easy to infer what's happening. And it's very causal. can look at the target. can look at the source. I can look at the transformation. I can look at exactly what the rules are of the governance.

Mm-hmm. Mm-hmm.

Here is the challenge that I have found. I have encountered a lot of vendors that have said, you know what we can do? We can create this knowledge graph that you have here by looking inside of your Power BI dashboards or by looking inside of your Tableau logic. And then we'll be able to infer what happened. The problem is that there is an expectation that that is, in fact, true and accurate. And what you find is, once you go over there, you found that there's 35 different variations of the same thing.

the pers there's precision, you know that there's precision, but not accuracy, if you know what mean? Like, there's, yeah, there's, there's just too many variations. And then you find yourself back to the drawing board.

Yes, 100%.

It's, yeah, it's my address, just the wrong one. Yeah, and I think, yeah, because we had to focus, the trick wasn't reconciling conflicts. The trick we had to solve was identifying ideas that were the same. That was the hardest component is saying, wait, this idea is canonically

It's exactly right.

similar, whether it's and going beyond semantics, right? You couldn't purely use semantics to figure that out because you like the three little pigs, right? You can write that whole story, feed it into, you know, a rag system or something, and then say, tell me a good kid story. And it'll be like, I don't know one because the word kids and the word story isn't in the story. And so

this idea of being able to use an LLM to explore a graph. And one of the things we were inspired by, I don't know if you're familiar with it, but Zettelkasten, which is like an old German. It's like whenever you see those crime shows and they have like the strings and the cards up and they're trying to solve the crime, that is based on Zettelkasten. And it's the whole idea that like you can't have two people in there that are redundant. The whole point is to understand the connections so you can solve the crime, right? And so it's a whole methodology that turns out LLMs are really good at applying.

connections. This is really powerful. So I think that I used to work with a guy, he's probably the smartest guy I ever worked with, and he had this great expression. He'd say, when intelligent people disagree, when presented the same facts and the same assumptions, something is wrong. we have this notion that people just kind of disagree, but if you give people same facts, same assumptions, they'll usually agree. When they don't, you start arguing about facts. And what you need to do is actually focus on the assumptions.

So what is the goal here?

What are we trying to do? And I think what you just said is actually a really powerful part about trying to frame the question. Like how do we actually even understand what it is that we're talking about? And then what we need to do is to somehow include the assumptions and facts behind the scene. So what we have here is something that just basically says people were talking about things and they were kind of talking about the same things.

Now, if you had a problem that you let's say that you had, I saw that GPU was one of your concerns, right? So let's say that you were like the price, I'd like to know the price of a GPU, right? Well, it's a price differential between a standard chip and a GPU for instance, right? That might not necessarily be here, but you could insert it. It would first of all say, I would need to understand what a GPU is. Then I'd actually insert a new canonical term called the price.

then I would give a search strategy for that price and it's something that I'm feeding into it, right? So I think that that's the part that I think is the how do you marry up the facts bits? These are facts. There's the concept of how we frame the argument and then there's the way how do we actually make sure that we're talking about the right facts and assumptions, you know? That's...

Mm-hmm. Yes. Yeah. And facts deserve code. mean, that's, and so I think that's where notebooks come in. Like where you mix code in with the language, where you have a hybrid of this knowledge is code and human readable language. So I always say machine readable and human readable hybrids and things like facts are best in those structured ways. And then the implicit stuff.

Facts deserve code. That's right. after it.

you put into the end. this, a good friend of mine is the creator of Jupiter notebooks. And so I came into this with a bias towards like, what if every note was a Jupiter notebook, right? And it was hybrid. And that's kind of how it led me to this place of canonical knowledge being Jupiter notebooks versus, you know, just pure language. But I, yeah, I think think what you're saying is interesting that the other layer that we found extremely valuable to LLMs is each one of those cards and the graph being temporal, meaning you have the full history of every change as an idea evolved over time. And so when you're asking the system and it's finding this canonical note, it also has the history in context of how that note got to where it is, where it was sourced and the changes along the way. And I think that's what you were kind of talking about. You need the context behind that, not just the fact. Because the snapshot in time is far less interesting than many snapshots in time because now you get to forecasting, right? Now that you have history, you can say, where's this idea going?

All right, so let's take a use case here, because I really like where we're headed. So a great example of things that kind of change over time in a software context are sales territories. You're constantly trying to tweak and optimize. What if I break up Germany between financial services companies and insurance companies? So one of the things that happens is it becomes difficult to track performance over time, because the territories themselves are kind of changing in the background, right? But that's a fact.

No. Cool. Mm-hmm. Right.

Those shouldn't be something that are open for interpretation. How those territories were structured and what accounts were bound to those territories is a fact. It's got to be baked in the assumption. Now, what we start to see is, imagine that there's a concept which is about quote attainment or customer success. If you can find a way to bind these concepts, hey, what have been historical things that have grown my territory?

There's a factual component that you're introducing to your point code. When I say data product, I'm really talking about code that generates a specific, generally either physical or logical construct that is very real and driven by code. Then you mix it with these other things, which are introduced by the canonical representation. And you need both. What is good? Tell me what good performance is. And you can be open to the AI saying, well, if you consider it by

By revenue, looks like this. you consider it by CSAT, it looks like this. So that's kind of the nirvana, right? Combine the code-based bits that are what you will know, that's the facts, with all the middle assumptions and ways of thinking about a problem that you can start to layer on top of it.

Yeah. Yeah. So let me bring this home. Especially for what you said earlier, which is I was talking to Patagonia CFO and we were going through this, was, know, similar things, but and, you know, I always use the example of it was IDEO like I three years ago created a lot bot. where you could find out if there's a space open at your office, a parking space, because they had fewer spaces than cars. And they put up these cameras they called cargo oils, and each cargo oil could talk, each camera could talk. And then there was a central interaction point with canonical knowledge that that could talk, right? So it wasn't that each system

was communicating via API or it was language, right? Hey, is there a space available in your part of the parking lot? No, how about yours? So these things are all talking in natural language to each other. And we were just riffing on the show and we were like, okay, like what if it starts booking a desk for you, right? Like, you're coming to the office. I'll go ahead and reserve a desk. And hey, like what about it knowing when other people and your team are going to be in the office and like suggesting, hey, you should come on Wednesday because and then it can get to like, do you want me to just coordinate your team coming to the office on the same day so that we can make it happen? And then it's about what projects you're going to work on. We should get agendas. We should talk about what you guys are going to work on. And the next thing you know, we were like, it's running the company, right? And we were going like, but it's still called lot bot. That was our joke. Like it's still going to a lot bot, but it's running the company. And it's to your point, like you can start small and evolve into this. You don't have to have all of your company's knowledge to get used out of it, but you start with that use case and the canonical knowledge necessary for it. And I did a similar thing here where I was like, imagine that every Patagonia, I also did it for a company that does commercial washing machine. So same idea. What if every Patagonia product had an expert like this on that product that could talk to other products that could manage its own. So you start really simply, you can answer questions for customers about the product. It could help you with repairs, but then it eventually can like help you, help you find ways to repair it. can, then it can work, you know, the entire supply chain of, of, could start handling inventory. could, and like eventually each, you have this, this, these, these small intelligent product expert agents that are just working in a decentralized way, but ultimately to the, to the point where it's practically running everything versus some centralized system that you start building that's eventually gonna run everything. I don't know if I'm making any sense, but.

I love it. Yeah, no, it makes total sense. think that the idea, there's two important themes of what you're bringing up. Number one is the thing that I have been scared about from the dawn of agents is how easy they are to build and how quickly decentralized they can become. And so you can imagine a world in which there's hundreds of thousands of agents just running amok, each with their own kind of interpretations, semantic models, and you, it good in the moment, but it's no good.

So this idea of start small, grow into something big, you're really trying to tie architecturally people back to the mothership and it can grow, right? So you're really anticipating the sprawl. And so this idea of kind of how you think about the growth of the canonical model on a case by case basis, so much grief later on as you try to, you don't have to unpack all that stuff. So now you can tell the agents to go to the same canonical model to get the information.

Is it?

The key thing here is that you need to be in a position, this is a second point, there will always be, when you get to the next use case, some lack of semantic construct or fact base that's missing to be able to do that. So let's take your lot bot example, and you're like, I can get you a desk. Cool. The first question is, what's a desk? And as simple as that is, you need to then kind of say, so the concept of a desk can be understood. But now,

Yes. Right. Yes.

Yeah.

how many desks are available is actually now you're going to tie the concept to a physical data store. And then we get into the code bit, right? So now you can wrap all of these kind of AI constructs, but you at least kind of onboard these new data sets and these new interpretations of the sets a little bit at a time so we can do more. And I'm finding that that's the skill that we need to start to introduce. Like governance people are not really great coders.

Coders don't really understand the power of semantics. So the people component of standing that up, like, great, we're going to go do this. How do I express the thing you just described in a knowledge graph? How do I help it connect to a data store? How do I build the semantics around the data and have that be a service that can grow over time? This is super interesting.

You huh. Yeah. Yeah. And I feel like, I mean, it's no secret. I'm a huge proponent of agent runtimes. mean, I've been working on them for 10 years and I think that's the missing piece. It's exactly, said a third grader could build an agent. Building agents is just yawn. Anyone can do that. No one should be buying agents. That makes no sense. They're too easy to build. But an environment they can run in is what you're talking about. That's hard. That's important. That's critical.

Yeah, let me kind of tweak what you said. I think that what we want out of the agents that we buy are really deep and powerful and bring with them some subject matter expertise in what they do. So, I mean, I worked for Workday and what you'd expect from an agent that's delivered for Workday is something that can really completely juice your recruitment process, completely reduce your review process, all these different things, right?

Yeah. Yeah

So I think that we're going to look to buy agents that are really deep. What we need to build though is like, how do you actually put those together and orchestrate things? Because most business processes do not respect the four walls of an application. So it's great that Zoom can record your conversations. That's terrific, right? But that's not going to actually then, how do you connect that to the calendar, to the, it's kind of Josh, what you just built, right?

Exactly.

So I think those are the agents that are, those are trivially easy to build. We should be building those. And now the question is you have to build them smartly so they don't spiral out of control. So there are gazillions of them all over the.

Yeah. Well, and I'll challenge you back because what we've built, you know, is a runtime and runtimes acknowledge the fact that agents, so LLMs can write their own prompts, right? So you don't need, people are worse at writing prompts than LLMs are. This is an illusion we have that we're better. We are not, we are worse. And if they can write their own prompts, they can tune their own prompts and they can auto tune their own prompts. And they're faster at getting feedback and applying that feedback to themselves and improving. And the tools they use, guess what? They can write those too, and they can fix those. So you end up with a system that's very circular in nature. You, you, it's, it needs to be adaptive. There is no like, I bought this thing and it is static and And, I get upgrades. is this thing is evolving and learning and as changes to my system and my canonical knowledge are changed, it's adapting to those systems so that, you know, two days after I implemented it, it is, it looks nothing like any other company. It's, it's its own, because it has to live in the context of all of my other agents and it has to adapt to my environment. It can't be static, right? Because I have a dynamic environment that it needs to keep moving in. But I do agree with you that there will be almost like drugs, right? Like, like efficacy rates and we'll be like, okay, this, like, let's say onboarding employee and them being happy. They'll be like, okay, this one, this one has had enough data. We fed it enough that it's done at enough times that it is optimized and you're better off using this as your launching off point. Something that's already intelligent, just kind of like we've educated this person, you're better off taking an educated person than having to teach them how to read and write. And then using that as a launching pad for it adapting to your system versus starting from scratch. And I think that's definitely true.

So I think you're onto something that's important about, it's not code, it's something that's living and breathing and it evolves over time and it evolves over time based on the data to which it's exposed. That is unambiguously true. I think that what we are up against is knowing that that is the case, how do we put up guardrails? How do we monitor things around bias? And who do we hold accountable for those things?

Yeah. Yeah.

And so I think that where we're headed is, in the next, we're already starting to do it with our vendors. We're starting to hold them accountable for AI constructs. We're buying this from you because we know that you'll manage code. But we're holding you accountable for way more than that. We're holding you accountable for restrictions and data assets, guardrails to certain kind of privacy restrictions, ethics, AI, all these kind of different things. And so that's a lot to manage across a lot of different things. And so I just think that that

If you have systems and people that are dedicated to these domains, it's just easier to manage that way. Like, it's not a question of whether you can build it or not. Of course you can. It's just a question of all the things that you have to wrap around building it.

Yeah. Yeah, and this is a journey, you know, from where people are to where they need to go. And that journey requires a sherpa for sure. You don't want to be inventing this stuff on the way. You need somebody that's helping guide you that's been there a number of times. Yeah, I totally agree. And it's just interesting to see where we're going because

That's right.

this like self adaptive system. That's why we call the book invisible machines because at some point there is no there is no interface. We don't it's we just talk and that's it and and they just and they just build right.

Yeah, that's right.

Yeah. And in some And the company is now lot bot. But, for them, that experience might not have really changed that much. They're still communicating with the same interface. there's, there's this strange like narrow, like this, this complexity on the backend that's sort of inverse to the amount of simplicity that gets presented on the front end.

Yeah.

There's no doubt. It's like a person. They learn, they grow. They're different the next time you talk to them. So there was one story that you were, as you were telling that story, I was reminded of a, you know, there's always kind of like data science folklore out there. So a friend of mine told me this story. I don't even know if it's true, but I'd like it to be true. So there's a telephone company, a cell phone company, and they're having a certain attrition problem. And so they call these data scientists in to do the evaluation.

of what's going on in the system. And this data scientist spends two weeks and he goes, I got it, Eureka, I got it, get everyone together, I got the answer. And so he gets the whole bunch of executives together and he says, every single time that that flag turns from N to Y, 75 % of your customers are gone within the first three months. Boom, mic drop. And a sheepish hand goes up in the back of the room and it says, that's the deceased indicator.

Hahaha!

And so I feel like this is the...

So we're getting into longevity drugs.

Right.

So I guess my point in telling that story is, as you start to think about all the gotchas around the corner of stuff like that, you're looking for things. The machines can go off and find all these things, right? But there's so much historical institutional knowledge about what's right, what's real, what's fake. And I think that this transition that you're talking about, I think we have to rely on vendors that have subject and domain expertise around these particular areas.

Otherwise, we can get there. It just takes forever because we're going to jump into all these flags that the machines encounter that they feel are so important that turn out to be not. So that really gets back full circle to your canonical model. How do we start to take that which is there? We should expect that our business partners and that our software providers express their own relative canonical models for the business that they do within the four walls of their application.

Yeah.

And then we have to figure out a way for that to all be expressed in the kind of model that Josh showed to just save us the time of binding it together. Take care of your business. Help me bind it all together. Because that's where the magic helps.

Well, and like the so much of it is, is human too. That's what's so interesting is like, we get caught up thinking about process automation and there's obviously a lot of exciting things that can happen there, but almost all of it. If it, if it's a, if the first step is kind of establishing canonical knowledge, it's so many human hands. Like it's, it really is an augmentation effort. It has far less to do with automation. It's more about finding ways to. safely and responsibly augment human abilities and that and that means you need a lot of people in the mix.

That's such a great point, the idea of augmentation. So I don't know if you're familiar with the The New Breed that was written by Kate Darling. She's at MIT. And she talks about how do people think about what is an AI? What's a robot? What's it to do? And people are like, it's like a replacement person. she's like, actually, history kind of teaches us that it's kind of closer to a pet. Right?

Yeah. Yeah.

We had to make these distinctions of like an oxen helps us work, a dog helps us play, right? And you have these things that are really about enhancing the own personal experience, but also like kind of bringing out time for us to be, I guess, explicitly more human, right? So I think that this is kind of part of the experience. Like, how do you, if you think about them as pets, they're not quite as scary. You send them off to do things.

They enrich your own knowledge of your data, of your business, of your life, right? And they can do that at such massive scales. Like Rob, that's not to undermine your point around like, my God, can they figure things out a gazillion times faster than we do? Writing prompts is the tip of the iceberg in terms of what they can faster. But ultimately, kind of how we think about our company performance, how we think about our relationships with other people, that's what we hold onto.

am, you know, I, I always say, I hate that we anthropomorphize this stuff and, then I keep doing it because it's such an effective way to try to dumb it down and try to explain it. It's so efficient, but then I, then I put us, know, so I, I, I'm a victim of doing it. I, I, I'm a perpetrator of continuing to talk in this way, but then it's, it's, it really puts us in a box with this technology because We start to think that it's not at service to humans and we start to put it on our level and we go like, wait, that makes no sense. Why would we, why would we ever build machines that are at our level? We only build machines as tools and tools are there to help us execute our objectives. and, and of course we're not going to, they're not going to live among us. It makes no sense. They're going to be at our service. It's almost silly to say it. You're almost like, why are we even having to say this? Of course.

Yeah. Well, that's why. Yeah, I think that's what I found so compelling about her book, which like the anthropomorphization that you're describing was like, well, I need something else instead. And she's like, if you can think of an AI or a robot as being like an animal, we need these things for very different reasons. But no one ever goes, my God, my wife loves the dog. I guess I'm out of a job. We've put them into these categories where they fulfill certain needs. It's up to us to figure out what those needs are.

getting back full circle, what is the business that we're trying to do? What are the problems we're trying to solve? Those are the things that really kind of are of interest to us, right?

Yeah, I wonder about that. Like part of me thinks like she's on the right track. Like she's bringing people down. Like, okay, it's not a person. It's step, baby steps to like, okay, but where's the destination? Like it's just a smarter machine. And we've been building machines for a long time. So I've said this from the beginning, a skillsaw that knows the difference between cutting through a piece of wood and a finger is just a smarter skillsaw. It's not a human. A stove that can see a pot boiling over and turn the gas and turn the fire off is not a cook or a chef. It is a stove that is smarter. A washing machine that, you know, can recognize materials that shouldn't be washed in hot water and turn on the cold water is not a human washer, washing person. It is a machine that is smarter and And these are just smarter machines and it's on a continuum of making machines smarter that we've already been on and that dumb machines are dangerous machines. They kill people already. Smarter machines save fingers. They save houses from burning down and they don't shrink your clothes.

So I love that. I love that story. I love the way you told that story. I'm going to get back to the kind of B2C versus B2B nature of it. I think that most people can conceptualize smarter machines in their lives. The smarter machines in their workplace feel threatening. And I think that's kind of a real fundamental definition. I think the way to think about that is if I just tweak your word machine and say process,

Yeah, like, haven't we been working on systems and processes that are about automating the company experience for as long as there have been companies, right? So I think that that's going to be like, how do we start to get people comfortable in a professional context with the same way that they get comfortable in a personal context? So take your washing machine example. Imagine that you could just like dump all your laundry into a gigantic pile, three feet tall, and it would just be done four hours later. No one would say,

my God, what am I going to do now that I don't have to do laundry? Right? But I think because we don't pay people for that, right? And so what's in front of us is really trying to figure out how does the marketplace change to take advantage of these radical exponential increases in productivity and how do we create the jobs that are effectively more human, right? More decision-making oriented, more in touch with things. And I think that's

It's taking my job away. Yeah. Yeah. Yeah.

you're on to a really good way of thinking about it that it's still scary when it's a professional context.

So I have a friend, one of his like claim to fame is working with Starbucks and convincing them to install cappuccino machines that were lower in height so that the barista could see and look into the eyes of the person they were serving. And I can't help but just tear that all apart and say, okay, the beginning of Starbucks is the automated home barista machines. Starbucks doesn't exist. We created a way to create cappuccinos at home. And at that point, you went to Denny's to get a cup of coffee. The barista was called a waitress or waiter at that time. And so we invent this machine that can give you cappuccinos at home. And the next thing we know, there's coffee and that, and that to improve the experience for them to, to make an improvement, was like, let's like lower this. Right. So that because, because of course I came there to have a human make me a cup of coffee, not to be like, why most of the stuff we buy in our minds, we need. And most of the jobs we do, we think are productive. I was in the film industry for 10 years. can promise you that winning meant I wasted two hours of most people in the world's time. Like the more people I could help to kill two hours of their life sitting was winning, right? Is that productivity? So there's this underlying first principle of like, as most of the things we make and do really needed. And if not, then we'll just invent more things. we're just going to invent these new jobs because we need them.

You've, um, you've just created a, um, a framework for me that I'm going to say out loud for the first time that I'm not sure I've thought through, but you've made a distinction between the things that we need and other things that are really just kind of craving human contact. And, um, so I feel like perhaps a metaphor is the ATM, right? Like we don't want to talk to a teller. We want money. Right? there's a very, and maybe I'd have to kind of tease this apart a little bit, but maybe the construct that you're creating is whenever we want a task to be performed, whenever that task and the efficiency of that task is sacrosanct, robotic automation AI is always going to be a great answer. But we have to probably take a step back and go, what percentage of the things that we actually want are that? Which are category A and category B? We don't want to have a coffee. We want to have a person make us a coffee. There's a human kind of a connection. There's a kind of a smell. There are now machines at the San Francisco airport that are called Cafe X and they have like the robots that kind of make your coffee and stuff. There's never a line. And I hadn't thought much about it until you said that, right? Maybe that's a social construct, right? We don't want to go to coffee. We want to sit in a place that smells a certain way, that's got a certain aesthetic and a certain kind of music.

You

this thing that feels somehow. Yeah.

Yes, we're so freaking lonely that we'll just take a five second gesture from a barista that's cranking out coffee that says, hey. We're so starved for connection that we'll pay eight bucks just for a hey.

Maybe we're onto kind of the future of work, right? So as AI starts to take on these things, the future of work is going to be dominated by things that really reinforce our need for these human connections. So clearly, there is a bunch of people that could go to Nordstrom and have somebody fawn over them for half an hour, but choose to go to Amazon instead because they don't have to talk to people, right? And there are people that want the human connection. So we're to have to craft these very human connections.

Yeah, I think we are. Yeah.

Because I still feel like sometimes when you call customer service, you're not talking to a person and it's annoying, right? And you can get, they can answer the specific questions, but it's really irritating. What causes that? Like, I think the answer based on listening to you say is I want someone to understand that I'm pissed and say, I'm sorry. That's what I really want right now, right? The machines aren't doing that for me. So you can do the task, but you're not actually creating the human experience, right? It's a very interesting way.

Mm-hmm. Yeah! Yeah. Yeah. And so what if every job is just customer facing? There's no back office and we all, and, and, and the quality of that connection now lowering, like everyone's got to lower the cappuccino machine. Everyone's slow down a little bit, have a little bit more interaction. Like, like, like it may be, maybe all jobs are just human facing jobs and the quality of the connection is what differentiates one company from the next company. And

Yeah.

Machines doing shit in the back room who the hell cares because

funny that you say that is isn't that true already? Isn't that kind of the connection that you make really working? And you can call that brand like why does someone's brand connect more? Why does someone's customer service score connect more? You know, so it's very rarely the the service itself that's defining your satisfaction. It's it's some interaction that you've had. It's very, it's very.

Yeah! Yeah. Why were you just on the call with customers? Like, why didn't you just send a chatbot?

Yep. No, that's

right. They want to know, people want to know what you are, that you feel what they need, as opposed to have a predictive model that anticipated what they were going to need. That's, think, maybe you've touched upon the very thing that it means to be human. I need to be understood. I do not feel

Mm-hmm. Yeah. and connection and the transition, I don't want to lighten it. Like it's going to be tough. There's going to be people with back office jobs that sit down at desks that have to figure out what the new, their new meaning is and their new value. But, it's going to be fast, but I do think, I think we're just desperate to connect. We just are and we'll invent all kinds of excuses to do it.

Well, I think if there's a world where, where, work is more service oriented, like serving other people that could actually end up being a somewhat beautiful thing, right? Because it feels good to help other people. If, your job is now more centered around helping other people and, and even starts to branch out from like, yeah, I work for this company, but I also work for this food bank. Like I, I, I realized that I'm addicted to connection, so I'm looking for all these opportunities to connect. And I'm in a world where. you can make a living by connecting with people. that outcome isn't terrible.

Yeah.

Sounds like a hill for introverts.

Yeah. I know. Yes, yes. But there's LLMs for them.

Yeah, that could be a little rough. There's service work for introverts too.

Well, think I think that I think that the idea of connection, connection doesn't have to be defined by an explicit extroverted thing, right? It's it's about being seen, felt, understood, connected in a way that feels right. And I also think that like, you know, what what is the future of work? If you automate away 50 % of work, do you need a 40 hour per week work week? Like the whole notion of what it looks like? I like Josh, what you said around like,

this gives you time to start thinking about food banks and volunteer service and different things like that. I just think society in general, this might be a little utopian, right? But if back office work goes away and businesses are formed around connections and service, then that should in fact free us up to do other things that hopefully reduce the temperature of some of the interactions that we have because we're all so stressed out.

Yeah. Yes Yeah, I agree with that. think it's a fascinating idea to think that, you know, we co-create together. know, a conversation is two people creating art. I know that seems very esoteric and, you know, but in a way, you know, what is an abstract painting, but two people creating art together, one the observer, one the painter. And maybe I've always contemplate the idea that co-creation is connection. Like if you were like, what is the substance of quality connection? It's creating something with somebody. And if that's just us iterating on each other's ideas right now, that's us creating. And the more creative you feel you are, the less you're just regurgitating old words, the more you enjoy the things, right? You're like, wow, that was really great.

It is remarkable to

Yeah. Well, I agree with you that this podcast is art. Yeah.

think that like. I know in the history of the world, no, this conversation has never happened before. That's what you're talking about. There's something kind of profound about that. That every time you have a conversation with somebody, chances are that conversation has literally never happened before. So how do you, coming back full circle to the canonical model, how do the things that we create in these sessions start to become part of the broader canonical model, if you will? That they're just easy to find, they're...

You Right. Yeah.

that we can understand what's going on in different parts of the data space, the world, and they'll just kind of become part of the conversation that we can have. I think that's pretty.

Yeah, I agree. how can it enable humans, not replace humans, but enable humans to have more quality connections? How can it connect people with customers? How can it improve the quality of those connections? think that's it. Ultimately, we'll look back and realize that it's a technology that improved our connection with each other. And that's the service that it's going to provide. And that may just be going to that coffee machine so you can get back to the conversation you're having with somebody. So anyway, this was awesome.

That's it, right? Sometimes I want a coffee, sometimes I want an experience. And we don't right now have a model for telling the two apart.

Yeah.

Exactly.

I agree. Well, this was great. Thanks for coming on. You're definitely sitting at the center of this. know, workday is the workday and the workday is changing.

Hahaha Yeah, really appreciate it, Joe. You

Yeah, we spend a lot of time thinking about this, right? Like the, is the future of work? What is the future of the worker? What is the thing that helps the agent to compliment and enhance that worker experience? Yeah, so we're spending a lot of time on this, not just in terms of our internal life, but in terms of how we go to market as well.

Yeah. Yeah, it's a cool vantage point you have. This is awesome.

Yeah. Yeah, thanks Joe.

All right. Well, thanks, guys. It was good talking with you.