Governance Is the Real Blocker

Spencer Kimball00:00

You know what I think is true right now in the world? Like every software company, Stripe, these two payments, now they're building what I'd call corporate AGI. Take a step back and you squint a little. You're like, oh my God, everyone is building to the same thing. Can we get AI to do things that were only the provenance of humans previously? And to the extent you succeed on that, there's a lot of alpha, right? It's a very ambitious, but visible horizon.

Jake Moshenko00:30

Welcome, everybody, to the Base Layer Podcast. I'm Jake, one of the co-founders and the CEO at AuthZed. And today I'm here with Spencer Kimball. I'm really excited for today's guest. Spencer co-founded and is the CEO of Cockroach Labs, the company behind Cockroach Database. I've known Spencer for more than a decade. We met when Spencer and I used to work out of Workbench, a local New York City co-working space and venture capital fund. He was just starting on CRDB, and we had recently sold our product to Quay to CoreOS. I'm not exaggerating when I say that he's one of the most prolific inventors and technologists I've had this chance to spend time around. CockroachDB is a big database built for the kind of world modern applications actually live in. Distributed, global, failure-prone, always on, and constantly scaling. At AuthZed, we use CockroachDB heavily, so this is not an abstract conversation for me. Spencer's work is part of the infrastructure we rely on every day. But CockroachDB is only one chapter in a much longer story. While he was still at Berkeley, Spencer co-created GIMP, the new image manipulation program. I think it had a different name originally, but that's what it's called now, which became one of the iconic projects in open source software. He went on to work at Google on large-scale systems, co-founded multiple companies, and has spent his career chasing hard technical problems that sit right at the boundary between theory and real-world use. That makes him an especially fun person to talk to right now. Because today, we're going to dig into the intersection of AI and authorization. As AI systems become more authentic, more personalized, and more deeply connected to company data, authorization stops being a side concern and starts becoming part of the core architecture. The memory layer is especially interesting here, where an AI system remembers who it can remember for and what that memory lives and how access to it is governed over time. CockroachDB existed before modern AI, but it's more relevant than ever. So, please welcome Spencer Kimball to the chat. Thanks, Spencer, for joining.

Spencer Kimball02:23

Thank you, Jake. That was a very kind intro. I always love to hear nice things about myself. They're only partially true. There's always another side to each of those points that maybe isn't all good. Okay.

Jake Moshenko02:38

Well, I don't want to know, because every time I open GIMP, I'm like, "This is Spencer's Jam."

Spencer Kimball02:46

Well, you know what? That's a great example, because Peter and I, that's actually my co-founder in multiple ventures. He brought me into Google. He was my roommate at Berkeley. We did GIMP together. We just kind of left GIMP to the community somewhat unceremoniously in 1997. And thank God for open source, because that thing grew into an amazing project without us. So, I think we're good at starting things, but in that case, we relied on hundreds, maybe thousands of people externally to carry the torch and make it actually the product it is today.

Jake Moshenko03:26

And I think a lot of times when that happens, when you have a visionary or visionaries such as yourselves who set the tone for a project and then turn it over. Sometimes we see it turn into this amazing thing, like what GIMP has gone on to become. And sometimes we see it kind of stagnate or turn static. So whatever you did, when you passed over the torch, you did it the right Because the community just took it and ran with it. You know, I've been following along sort of since 1999 and seen all of the iterations with like the interactive timeline, the reprocessing thing, like everything that they tried in GIMP. You know, I was right there as a spectator being like, that's really awesome. So, you know, you also have a long and storied history in open source, and I think that that's pretty amazing too. But to kick it off, I just, you know, everybody's working on AI these days. So what's the coolest thing that you've done with AI?

Spencer Kimball04:18

Well, I actually had a welcome to new family member in February. So I had a paternity leave and as CEO, I couldn't take as much time off as you know, you might want. So it took about a month. And in that month, I finally had some time and dedicated it to agentic coding.

Jake Moshenko04:39

Was that your first time trying out agentic coding?

Spencer Kimball04:45

My first real time. I sort of poked around and Claude Code a little bit, but that was where I said, you know what, I've got an idea that I really want to see happen. And I had some motivating examples, which, by the way, is like, that's the, they say that necessity is the mother of invention, right? And nothing I've built didn't come, that was any good, didn't come from a, didn't fail to come from a real problem that I personally was experiencing. I'm very motivated to solve my own problems and you know, I'm a pretty good judge of whether the solution makes sense because, well, did it, did it work for me or not is a, is a great litmus test. When you're building for someone else, that can, that can be a high latency, you know, exchange to figure out whether it's actually any good. You need a really dedicated design partner or partners for yourself. It's, you know, part of the inner loop. So that works pretty well. Uh, you know, what I decided to work on was, uh, how do you get, uh, how do you operationalize AI agents? And I wanted to do it for, uh, at least two and maybe three use cases and two of them were at cockroach labs and one of them was, uh, for a personal, uh, desire that I had. And so let me explain the, the sort of work related ones first and then I'll, I'll, um, cover the personal one. So the work related ones, there's two incredible sources of information we have that are just, um, extremely dense information wise, and they're very important to our business. On one hand, we have gong call transcripts. So, you know, think about when you have a call with, um, you know, someone on the cockroach lab side, it might be sales, it might be, um, customer success, it might be a sales engineer or a solutions architect. And on one end is a customer and on the other end is a, you know, one of these roles at cockroach and there's questions and there's answers and there's discussion and, uh, you know, these things usually last 30 minutes to an hour and there are many of them. There are so many of them. And often there is a, um, you know, they could be talking about commercials for a deal. They could be talking about a POC, uh, implementation or a, um, a workload optimization or a, um, you know, any number of things. And the, having, uh, Claude Code, giving it access to these, uh, conversations and asking it to, uh, spend some time and dig through a particular account, answer a question, it's totally possible. It works very well actually. Claude's very good at that. Uh, so, so, uh, you know, any number of these AI models and harnesses and so forth. The problem is it's very latency intensive, right? It's like, it's a lot of data and it's got to crawl through and organize it. And you might spend, uh, you know, you might spend 10 minutes or more on such a task and, um.

Jake Moshenko07:32

For just one call.

Spencer Kimball07:33

Oh, well, not one call, but if you want to, if you want to go through the history of an account answer like, you know, where's this account trending or what are their biggest problems or let's talk about what are the patterns if I'm going to do, uh, you know, sales enablement, like what are the questions? What are the objections from the customer? How do we answer them? What are the best answers? Any one of those questions, you know, you're talking about 30 minutes to an hour. And when you think about operationalizing, like what could you do with this? Okay, can we like, can we do adaptive learning for enablement? Okay, that's a cool use case. Could we actually have a sales co-pilot that as you're on a call, it's like, it's actually giving you answers and evidence and like examples from other customers? Okay, well, if that, that's always going to, we put these conversations, the transcripts into Snowflake. If it's going to Snowflake, querying that and, you know, digging through it and analyzing it, there's no real time possibility, none at all. And by the way, all the tokens and things of crawling through with a good model, you do not want to keep spending that money unless you're into token maxing still, which hopefully is dead. So that is a really cool use case. And the idea is like, okay, well, what do you, what exactly would you pull out there? And how would you organize it? Yeah, you probably want a database. Use Cockroach, of course. I'm the founder of the company. It'd be weird if I didn't. You could use any kind of database, I guess. But what's interesting is I started building that system and it uses basically every single thing that Cockroach has. You certainly wouldn't want to just put this in a vector database. Like, absolutely no way, zero chance. You need transactions. You need to be very careful with how you use transactions so they don't become a hindrance. But they actually, they're very important because you're gonna have lots of agents accessing the data. And then you have things updating the data, and you want that to be both efficient and safe. And if you can't reason about that, you just create a game of insane whack-a-mole. So like, and you want this to be resilient. You want it to be scalable. Okay, well, so now you ended up at Cockroach anyway. So it was good that I was able to start on that. Another really cool use case at Cockroach, and it's a crazy one, is Zendesk tickets. So Zendesk is the system we use, and you can use any one. That part's arbitrary. But how do customers file issues? And then how do we respond to them? What's that process? What artifacts does the customer paste in? A lot of times we can't even see the cluster. It's kind of we debug through a pinhole because it's a self-hosted deployment, and we don't have access to the cluster. And so we're playing a game of telephone kind of asking the customer, and then they go and they look for something that's high latency. And sometimes these things are, I think the mean time to resolution is like five days for some of these certain classes of these tickets. It's not pleasant, right? Oftentimes we can get things mitigated much sooner than that. But to get a full root cause analysis takes some time. And it's a lot of work, and there's a lot of customer pain on the other end. And so you'd like to understand, are there patterns in this data where if we comb through them, how quickly could we get an answer to some of these tickets that are filed? And then maybe I'll tell you a little bit about the results next, if you're interested. But the final use case, which is super motivating for me, is I have a rental property that I've had for 18 years. And it's a really cool place. I built it myself after Google had some success. I was very interested in doing it. It's really cool. What did you play yourself?

Jake Moshenko10:54

Like you were swinging a hammer?

Spencer Kimball10:55

No, no, no, no. I worked on the design of it. I went down there. Okay. It was in the Dominican Republic. And so it's a big house. And people rent it for family reunions and weddings. We host weddings. And it's a really cool place. It's called Casa Kimball. People want to check it out. And what I wanted to do there is say, I had this idea, right? If you really think about what AI portends for the world of websites, all sorts, every business, every rental property, every hotel, whatever, every single business in the world, you have to realize that there's going to be a corporate emissary, an AI representation of like the ultimate emissary of the business, right? In Cockroach's case, it might be capable of debugging your cluster, helping you implement a workload or migrate something, knows everything about, you know, cockroaches capabilities, can answer any question about the pricing. It's like your ultimate DBA and SRE and like, you know, solutions architect and, you know, salesperson and SDR. It's all those things. And you just come and you knock on the front door of that website. And instead of just static pages and tons and tons of content and maybe a little search thing and one of those crappy agents that are currently there, this thing actually is like, first of all, it's always home. And second of all, it's as good or better than any human you can have in there. And that's where things are going to go, 100 percent. And so when you even think about the Casa Kimball website, which was kind of a piece of crap until recently, we really updated it, right? I have not put this in there yet. But the goal is to have an incredibly knowledgeable emissary for the house, like a virtual concierge that actually really can answer the questions. How can you do that? It's not just crawling and organizing the website's content. Turns out we have about 20 gigabytes over 18 years of email. That's not even counting attachments. It's a lot of email. Only a fraction of it's really guest correspondence. But like, can that information be anonymized? And then patterns extracted. Every kind of question, every arrangement. Can I do this for my wedding? Can I have fire dancers? What kind of wines do you have? Can I bring my own bottles? How do I ship things down there? What's the best airport if I arrive on this time? Can I have fireworks? Can I have boating cruises and deep sea things? And what happens if someone's disabled? There's a lot of things in there. It's not like it's all the same questions. It's amazing the variety and the care that's been taken over 18 years And so can you put all of that into something that makes the right kind of sense of it such that when you there's an easy eval to this, just like there's an easy eval to the Zendesk ticket problem, which is like, let's let's not train it on everything. Let's take like questions or tickets filed or whatever. And we have an answer. We have a resolution to the ticket, right? We have an answer to an objection in the gone call. And we're just going to see like with 500, you know, kept aside examples, whether the system using the A.I. memory, we'll get into a little bit more of what that actually should mean. You know, can it answer that? And with what fidelity? And you have you have a judge that's also an A.I., you know, a better model, actually score it. And, you know, you're interested in latency minimization here. You want these things to be real time. Otherwise, like humans don't want to be on the end of a five minute weight, God forbid, right? And so the question is, can you can you do the thinking up front? Right? So you actually can you have crystallized intelligence, not like fluid. And the chatbot and hopefully you're using like you're able to actually downgrade the capability of the model. You don't want to use Fable, right? You would like to use a cerebrous open source model. And given this tool, which is 30 times faster, right? Given this tool, can it do as well as Fable without the tool? Or close to as well as Fable with the tool? Because the most of the hard thinking has already been done. All that sifting and pattern recognition and extraction and synthesis has already been done. And so it's really just doing a fast look up that's on the fast path of Cockroach, like tens of milliseconds and you get to do to do all these results. And in the the model is actually taking the queries from the customer along with the whole conversation, the standard stuff, right? And it's saying, OK, with this latest thing, what should I search for based on what I already know here? Put that into the system and then it's rag basically, right? But the question is, what's the quality of what rags returning? If it's just returning snippets of raw conversations, you are sunk. That's garbage. The whole world knows that's garbage already. That kind of rag doesn't work.

Jake Moshenko15:43

It doesn't work well, right?

Spencer Kimball15:44

It's just not good enough. It's not better than a human. And so the question is, can you get better than a human? And actually, the one that I think is the very best example here is the Zendesk ticket one. With this system, the Zendesk ticket example, 41% of all tickets from just the initial filing, the customer saying, oh, I've got a problem. My cluster's down here, the symptoms. No, like follow up. Like, well, did you check this thing? Like, what does this metric say? Or like, you know, what are your settings? I need to double check this. Or can you paste these things into the ticket commentary? Just the first initial thing, 41% of the time, it nails it. Full root cause analysis. That gives you an idea of just how much information is there to be extracted as patterns. That's incredible. 41% of the time. And this costs $1.50 using Opus. And it is more or less instantaneous. Right? It takes like less than a minute.

Jake Moshenko16:43

Do you know what the corresponding hit rate for a human is? Yeah.

Spencer Kimball16:49

Well, I mean, it's extremely low. Okay. Humans almost never get these things on the first one. Okay. I mean, it's like, you know, another interesting thing is what's the hit rate of Opus without Chita. It's actually about 10% is better than a human. We base the human one on basically whether someone responds with, like, the right answer. And that's not necessarily fair because humans might be, you know, being careful. But the reality is, you know, there's a lot of latency in that process. And so the goal here is when you show up, when a TSE even gets the page and gets to the Zendesk interface and looks at the ticket, this thing's already there waiting for them. And so it kind of like it runs this experiment where it gives just the initial ticket. And if the judge scores it as being, like, 100% correct, it's done. Otherwise, it'll go and it'll add the first comment. And it stops before the resolutions provided by a cockroach representative, right? So it's kind of like, okay, with additional little bits of signal, can it get it this time? Okay, let's put a little bit more signal and can get it this time, a little bit more. And in that process, it will get to 60% full root cause analysis and before the resolution is actually given. So those are very high numbers. 92% of the time, it has the mostly correct answer.

Jake Moshenko18:05

I think maybe we missed a bit of the story, though. So you're on pat leave, you've got 30 days to build something, and you've got these three use cases, two of them across Cockroach Labs and one of them across your rental homes website. What did you build in those 30 days? And I think you may have dropped a code name in there at some point, or a name of the project.

Spencer Kimball18:31

Yeah, sorry, I jumped ahead to all the exciting stuff.

Jake Moshenko18:33

Yeah, the outcomes are certainly exciting.

Spencer Kimball18:36

Because I'm still working on it, by the way. It wasn't just a month I was on pat leave. Now this thing has taken on a life of its own. I'm so interested in it because it has such profoundly transformative effects. So what I decided to build, and I didn't exactly know what it would be. It's kind of like SpaceX building their first star hopper. Yeah. It looks like a water tank on top of a building. They didn't know how to build a rocket exactly. They had great people. They built something they're like, hey, what do we do to improve this? That's sort of the process I engaged in. I decided ultimately to call this chitta, which is a Sanskrit word, also a Buddhist Pali term. And what it means is the storehouse of all memory and impressions that lead to action. And it just seemed like a very appropriate term for what you actually want this thing to be. It's like, how do you how do you distill knowledge in domains and then make that available to agents? Because you think about an enterprise situation, no one's going to give in the real enterprise. I'm not talking about startups or even cockroach level companies. I'm talking about like, you know, Bank of America or something like that, right? The idea that you're going to give an agent access to Snowflake in an operational context is preposterous, right? Never, never, never going to happen zero chance. We need to do is you need to create like sandboxed curated knowledge domains that like you know exactly what this thing has access to and you trust it. And and again, what's been very interesting about this is to realize that it's not a general purpose distillation of information and organization. The thing that makes it truly useful is to understand that any data source needs to be distilled with a purpose in mind if you want it to be truly effective. Like when you understand that this is that debug a complex distributed system. The the interpretation of those Zendesk tickets like further debugging is very different. Another case is very interesting is source code, but you can actually if your purpose is to add more features to a source code repository. That's actually a different set of decisions that the AI is going to make and thinking about it and crystallizing these things and organizing it. Then if you say that you want to understand the source codes that you can find or maybe even exploit vulnerabilities or you want to understand it so that you can operationally manage it. Those are like very, very different things that you're going to pay attention to just like a human would and so understanding your use case like I want to answer every customer objection versus I want to enable employees those might be somewhat similar, but maybe what's wrong with our product as far as customer conversations go. Well, that's a very different kind of distillation. So shit is the name and and it was a incredible process of evolution to kind of end up where it is today. A lot of what I'm telling you was not at all in my head. It couldn't have been and what I was able to do, though, that was so interesting and trying to trying to bring this future trying to find a responsible way to bring the future. I embraced into cockroach labs because we're of course we're very, very careful for obvious reasons with cockroach DB in particular. And this is an opportunity for me not to be careful, which has is a very incredible learning experience. You know, in this case, I was like, I'm not going to look at source code. I'm not going to look at the diffs. I'm not going to look at its tests, its effort to make sure that it's running them and adding them right. And I'm just going to go at maximum velocity and see what the AI can and can't do. And if it can't really do something that well, then I'll pay more attention to the plans and try to give it advice. And with the expectation that it will get better and then I will be needed less and less in there. And it's all just about can I direct this thing to make my vision a reality. And I'll say this, like just in the recent past, I've had to stop paying attention to plans. I haven't had to continue paying attention to plans in the way that I used to with Opus. Fable has been a massive sea change.

Jake Moshenko22:50

Yeah, that's that's a super interesting journey. Is this like, can I buy this thing from you? Like, what is it?

Spencer Kimball23:00

We're going to open source it.

Jake Moshenko23:01

It's a feature in CockroachDB. Okay.

Spencer Kimball23:02

It runs on Cockroach. But I imagine as soon as it's open source, people will port it to whatever. You can buy TiDB, Postgres, Oracle. I'd be surprised about that one, but you never know. Yeah, that open source is the key here, right? Because like this system, first of all, it's so it's so powerful as a foundational building block that I want to see these things used. I hate not having things that are not open source, even though sometimes that's just the decision we had to make eventually with Cockroach. It's like, can you actually build a business, you know, with something that's pure open sources and we tried various mechanisms in the end. It just turned out that the better you make your product open source, the more even a bank is willing not to pay you for it. And so it's like, do we do we want to keep building this as a business or not? Because if not, then, yeah, we can just make this thing, you know, so that anyone can use it and repurpose it and so forth. But if you want to have a business that doesn't have an impossible Achilles heel, then, you know, you have to make some decisions there. Chitta gives us an opportunity to actually contribute something immensely valuable to the world that actually runs on Cockroach and also will run on other databases, of course. I think Cockroach is a very responsible choice to run this on if you want to put it into large scale operational context, certainly if you're in the enterprise. And so, you know, this is a channel and it's very valuable as such, assuming people like it. I don't know the answer to that, and that's that's something that, you know, you only find out from empirical observation. So the plan is to open source this and actually open source some of these examples. The thing I've been building for this rental property, Casa Kimball, is something that I'd like to open source. I mean, Casa Kimball doesn't have this on the website yet, but it will. And what I'd like to do is, as soon as that's working, extract out the, let's call it like the, the AI powered rental property. Like what is that project that works on Chitta, Chitta works on.

Jake Moshenko25:15

The operating system for Right.

Spencer Kimball25:17

So anyone that has a rental property that could use Claude Code could build this themselves very easily. It's just like, okay, just put the house website onto this thing, style it differently. You're good to go. Right. You got to stand some things up. By the way, you can use cockroaches basic tier thing. It's, it's so cheap. Right. And then use cloud run. And by the way, that that architecture, you know what's amazing about it? You can scale this up to the size of, you know, Hyatt hotels. Like there'd be nothing to stop you. You can have as many cloud run instances with your application state. You can have, you know, cockroach, You know, a basic tier, which starts off as free for you. And they could scale to the moon, basically. This is a very, very powerful system. Ultimately, I think most house owners wouldn't wouldn't use Claude Code. But there's lots of companies that could be, you know, adopting this technology and doing it on behalf of customers. And by the way, this isn't just houses and rental properties. It's anything, any kind of business where you actually want to, You have enough information about your customer interactions and you'd like to operationalize a fast day high with low latency to be there as the like always someone's always home, right? Like that's a great concept. And then, you know, as soon as the conversation is getting started with someone that's interested, it can it can bring in, you know, an actual human into the same conversation where they get the handoff. And these are not new ideas by any means. The question is, how good is the AI? Because I don't know about your experience, But anytime I've tried an AI, God forbid on FedEx's website. Oh, my God, like it's actually it's actually a slap in the face. And not only does it not work, but it almost like invite you to believe that it might. And then it's like you ask the simplest question. Sorry, I can't help you with that. And it's not like, hey, you know, here's a human or something. It's just like, sorry, there's nothing you can do. Good luck.

Jake Moshenko27:04

I've had AIs for some of our vendors just completely fabricate things that I very much want. I'm like, how do I do this thing on your platform? And they they like make up a like go click here and click here and do this and go here and like none of those things exist. So you're totally right. It's a slap in the face. It's like how like it would be very cool if that thing did exist, but it doesn't.

Spencer Kimball27:27

Well, here's what's so interesting, right? Because the tools that you give an AI shape its thinking, fundamentally shape it. And if it knows it has Chitta and it's explained that you explained to it, Like not only what Chitta is and the tools that it's giving you, which are quite simple surface area, but also what pools does that Chitta persona, is what I call them, what pools does it have access to? And then what is the purpose of those pools? What is this for? What should you expect inside of it? And then also like that persona also says, for example, in the Zendesk diagnostic Persona. Actually, it doesn't just have the Zendesk ticket distillation and organization actually also has all of our runbooks, right? And it has all of our docs. So it's like it's it's able like what the persona also describes to the agent is like, You're you're you're going to diagnose, you know, problems, some of them very complex with CockroachDB. Most of the AIs know what that is, but it will say like you've got runbooks, You've got all the past Zendesk incidences that are totally, you know, well, that's probably the biggest source. And you've got like all the cockroach documentation. This is a good way to think about using this, you start with this, then move on to this, double check it with this. When you when you when you lay all that out, the thing stops hallucinating. It really does. And it will actually tell you when it's got low confidence and when it doesn't have the answer, because once you give it those tools, It's like, OK, I'm going to use these tools. Did this answer my question? And the answer is no, then that's it's safe to it's not it's not incentivized to to fabricate beyond that. When you just when you say, like, you know, answer me this question and it's just the model, it's going to do its best. It's not like, you know, obviously, you can shape prompts and make it behave a little bit better and so forth. But I think the gift of tools actually has some interesting follow on effects, and I think the decrease in hallucinations is one of them. Did you manage to get this tool Chitta integrated with Chitta development itself? Did you expose it as a tool to Claude Code? Yeah, I mean, actually, source code repositories is another interesting use case that I really spent a lot of time on. An interesting thing about source code repositories is that they are they're so authoritative and things are constantly turning in them. So if you don't notice the turn through the Git log diffs, right, like if you don't get rid of all of the organization, all the patterns that might have been touched, all the concepts that might have had assertions, you know, that help build them. But those assertions might have just dropped out in the latest commit that you process, and you've got to update everything for the both the additions and the subtractions and the modifications, right? So this thing has to be capable of rewriting its history to exact specifications for source code, otherwise you get this vestigial crap that never goes away and it just like misleads the model. And so you start to have the chita becomes unveiled valuable compared to the way that Claude Code shoots x rays through your code with greps and things like that, right? Like that's it's amazing what it can do there, but we've all seen how it misses the bigger picture. Sometimes it misses like the necessary thing and you can start putting these into cloud.mds and, you know, you know, kind of create something that has some of the attributes of chita, you know, and I think, by the way, that models just can get better and better in the harness and get better and better. Chita's not required for this, right? The question is, can it do that? Well, and so, yes, of course, actually, I have a chita instance that are a pool that's running, which is the chita source code. And there's a there's a GitHub action that just frozen in the latest commit and it is it's completely up to date. And there's a, you know, a requirement that I have on Claude Code to consult it when it's working on the chita source code and also to judge its utility. Like what did it help you find that you didn't have and like, you know, also what, you know, what can be improved, I haven't really hooked up an automatic attempt to improve itself just through those signals. But it does surface where chita is being useful and where it's actually falling down or giving incorrect information according to like what it follows up on. And it's, you know, it actually is, it likes to use the tool now, at least it seemingly does, who knows, maybe it's just because it wants to make wants to make me happy. But yeah, you never know with these AIs. But it, it, it always is like, wow, chita was really doing the right thing here, because it gave me this thing which I never would have seen and blah, blah, blah. And so like there are there, it's finding those things because they are visible through some of the organizational mechanisms that chita employs. And there's, there's a couple really maybe three, I think, maybe four, it kind of to sort of blend into each other. So the first stage is that you take a source memory, so this would be a gone called transcript or a Git log commit, and or a Zendesk ticket in its entirety, right, or a thread in email for a guest correspondence. Those are all some of the examples that I've really tackled recently. And in all those cases, this is the source memory, but you don't want to just break those into chunks like the old fashioned rag, right? You actually want to understand every one of these source memories as a hierarchy of meaning, which isn't that hard to do. There's like distillation instructions, and they're actually, they're tailored to each of these different use cases. And you can sort of arbitrarily tailor them, there's a plug in architecture. So if you're like, I have an inbox one, which I use to organize my personal email, it's very interesting. When you hear some of these other things, imagine it applied to your personal life. But like, I made a version of that for guest correspondence mail, which had specific instructions of what to look for and what to just completely ignore. Like, I don't want to hear about vendors and procurement crap. I want the guest conversations. That's all right. So there's like a there's a there's one that has tailored instructions. So the hierarchy of meaning, what I mean by that is like, we're having this podcast conversation. If someone's listening, they're going to have the gist they come away with, like what's the overarching narrative. And that's a that's a very conceptual and abstract representation of the larger thing, what did you take away from it. And then you have the major topics, and then you might have minor topics and you might have supporting detail, something like that happens in every meeting we have every experience we have. But something else happens. So it captures all that right and it like it vector indexes and full text indexes all those little chunks because it creates essentially a tree. But when you think about this podcast, as we're talking about AI memory, a listener might say, well, actually, I'm familiar with, you know, these other products. What's what are they talking about? That reminds me of this thing. And you kind of pull information in that exists already about a particular area of. You know, topical area and you're actually like augmenting that information based on what I'm what I'm saying, for example, here about Chitta and putting it away for future reference, you've changed that topical summary of information in your mind. And that's what it does. It organizes the pieces of this this hierarchy for a source memory, each of those pieces gets located into an ontology. It's like a taxonomy of topical meaning. And so this cross cuts all of your source memories. You might have 20,000 commits or something like that. And the commits are all broken down like, OK, this is this is about adding a feature X to, you know, Chitta, right? And then it's like we had to modify this package in this package, this package in this way. Here's the actual detail. There's actually some supporting detail, which is actually code snippets and things. All of that, you might put the gist into something about adding features. It figures it out based on like, you know, judgment and, you know, cosine similarity in the vector embedding spaces and things like that. Right.

Jake Moshenko35:31

So those that ontology is designed by the AI. Yeah. It's not something that I have to specify.

Spencer Kimball35:35

Yeah, it uses this uses this mechanism for semantic tree management and topic management called cobweb. Like I started off with my own, I just it's kind of the difference between the two is like I started off with a rule based optimizer. And eventually, like I was like, this thing is not working that well, because you get like somebody solve this. Yeah, exactly. I was like, you know, how could you could you there's got to be some research on this. Like, oh, yeah, there's tons. I was like, why didn't you tell me about that before?

Jake Moshenko36:04

But, you know, I've noticed that as well, right? Like as soon as you ask, it's it's a genius on the topic, but it's like, why didn't you interject when we were beating our heads against the wall doing it the wrong way?

Spencer Kimball36:13

Yeah, the reason I start companies in the first place is for some reason, I'm foolish enough to believe that like, like, my thoughts are original, and I can just do it myself, right? And then you get further down the road, you're like, well, this thing isn't good enough. Let me look around a little bit, by the way, in that process, you understand all the reasons that the truly thought through research is brilliant. And like, also, like, you know, I've even improved on some of it. So like, you've got to get your mind into it somehow. I think some people do that better by like reading, reading up on like thoroughly on something becoming as much of an expert as they can, and then they move to implementation and they kind of start from maybe much further up the mountain than I tend to. I just start at the bottom. I'm like, I can climb this mountain, start going up K2, you know, obviously, I fail somewhere along the way. But you know, it's not for lack of trying. So, so it does these ontologies right and then what it does is synthesize it sort of locates these sub memories into like the hierarchy of the additional hierarchy of meaning, which is the whole cross cutting topical hierarchy. And then it synthesizes and tries to make meaning at every stage in this and again it goes from a level of, you know, concreteness to a level of abstraction at the top. And then it also extracts assertions think of these is truth claims about reality in this knowledge domain and organizes those they linked to concepts and so it'll build up an incredible like base of assertions about different concepts that matter to that knowledge domain. Like in cockroach TV, I'll have a ton of stuff about different versions of cockroach what was wrong with them. What worked well, but didn't like what was not performing and it'll have all kinds of information about features of cockroach what doesn't work. You know, like you see the feature that's most mentioned in the zendesk tickets. That's our that's our biggest problem for sure right. And there are some of them. And so like it's a it's a pretty interesting there, the really cool thing, though, is, and this is it's very important to to to find the right. Ways to extract patterns in each knowledge domain patterns are very, very powerful. The right way to think about a pattern is it's what you expect to be a recurring situation. So in zendesk tickets clearly failures like failure scenarios where there's run books right there's there's sort of debugging, you know algorithms are like flow charts essentially. And so in the same thing in guest correspondence and zendesk tickets and coding it's just like the every knowledge domain has very valuable patterns that can be extracted. The way it does is it suggests recurring situations as these fingerprints, and it's it tries to find like a an English language sentence that that really actually it's language agnostic, but it'll find a sentence that that. That describes. Without adornment, the recurring situation in this domain. And then. Right. Exactly. In a vector space it finds clustering and then that's thing of that is like it's evidentiary weight behind a pattern. Some patterns have like 60 pieces of evidence and so you have to synthesize all that evidence into like, you know, essentially the patterns are so powerful. Like on the zendesk thing is that I think like 60% of the time the answer is coming from a pattern. It's not it's a little less than that in the the guest correspondence, but it's like 40% it's definitely it's the it's the plurality of the of the solution, even in the guest correspondence. So like that that's that's another whole thing. There's there's all these ways to browse it. It's cool thing. I think I'll be very interested to see what kind of reception it gets in open source could be crickets. Like a lot of people are working on this problem. Like it's just I see this as a as a as a as a necessary tool if you're going to sell the database in 2026.

Jake Moshenko40:08

Yeah, I mean, I'm 100% with you about the need to organize information. Humanity has been organizing information for as long as we've had writing right. So the idea that people come to me and they tell me well I'm just going to connect the AI up to my database database and let it run SQL queries and now I don't need HubSpot anymore or now I don't need Looker or now I don't need Snowflake or any of these things right. I'm just going to attach it to the raw sources of data and that just can't possibly work right there's just too much data. There's not there's not enough context in the world to be able to handle to be able to handle these problems that way. And even if there were it's like you say you'd burn the tokens over and over and over again coming to the same conclusions where if you had just sort of distilled put this data. I mean the way you're what you're describing it sounds to me like kind of an ETL process for data to turn it into intelligence. Exactly. And and without that I just think these things are going to be so lost and so expensive and everybody's going to you know sour on the whole concept. It just didn't seem to do what it was promised. But when you hook it up to this sort of distilled intelligence I can see how the utility would go up a ton.

Spencer Kimball41:17

You know what I think is true right now in the world like every software company stripe right these two payments now they're they're building what I'd call corporate AGI. Like that like if you if you take a step back and you squint a little you're like oh my god everyone is building to the same thing can we get AI to do things that were only the provenance of humans previously. And to the extent you succeed in that there's there's a lot of alpha right it is a it's a very ambitious but visible horizon. Right like it's to my mind like I know there's been a little bit of not a little a lot of backlash against AI recently and people love to be like oh it doesn't work the ROI is not there and blah blah blah because it's like yeah people don't want it to work on one level we don't I don't you know like. I was like oh my god this is terrible like like what what good is like all this stuff that I did my whole career and love doing you know and and was proud of. And you know I think there's answers to that question but like you can see that people want AI to fail on a level many people do. And and yet and and there's some truth to all of the criticisms of course because there's a lot of hype. And when you just try it you know it was amazing my god and then all of a sudden you're like well geez how about this oh that's that that's pretty. That's an ugly little wart that maybe is almost disqualifying and then and then people are like now that doesn't work and then that becomes like you can't let that become an entrenched perspective. God forbid in your company because the reality is this stuff is just working better and better all the time. But what Chitta is really attempting to do for for us and for me personally is is to cross that chasm really definitively cross it. Like I'll tell you this our support team they rely on Chitta now. They say that it stays on every ticket hours hours of work. Do you know expensive that is. And it's not our costs. It's our customers bleeding. I mean you don't want your database to be down. We're not supposed to go down. That's our brand promise.

Jake Moshenko43:27

I do know how expensive that is. Yeah. Yeah. This is really interesting. You talked about starting this on your 30 days of leave. And it's now taken on a life of its own. It's it sounds like it's a full fledged project within Cockroach Labs.

Spencer Kimball43:46

Is that still working on I've got like there's various interested other people that are so it's not quite true, but I'm the only like. There's no full time employee working on it yet. And that I think that you know that's a bit of an oversight is going to be corrected. But, you know, we are also at the same time working on the whole next version of like what what is Cockroach in the world of the era of AI. And that's a that's as you can imagine, an all consuming project for the company. And it's a very, very exciting evolution as well.

Jake Moshenko44:20

I'd love to get to that in just a second. When I asked this question of senior leaders, what you know, what are you doing that's the most exciting with the IRA right now. It always seems to distill back to agency and being able to take on these kinds of things ourselves right like I I've done the same thing at offset I was like I've got this thing it's it's a work and I want to fix it and let me see how far I can get. And sometimes I don't get very far and sometimes I get far enough that it turns into like an actual initiative at the company. And so I think that you're the story you told is also sort of following that same path. Right as CEO of cockroach labs, how would you have accomplished this before you had these tools available.

Spencer Kimball45:07

Well, I would have I would have come with my hat in my hand and beg the organization, you know, particularly R&D, you know, primarily R&D to try to make some room. Well, what is what do you want us to not do instead, you know, that's always the answer. And when you try to hire people, well, okay, that's going to take months and then they're going to ramp and like budgets and this and that. And it's like nothing speaks louder than something that actually you can deliver that actually people can be like, oh, wow, that's cool. Yes, of course, we should do that. Not like, oh, guys, please, can we find some room to do this idea that I have that might not work. That that was the old world right now. It's like, you know, the I would don't know if I would have actually gotten there if it wasn't for this pat leave. But you know that. I can't tell you how happy I've been not to be coding again, because that's not what I'm doing exactly, but to be creating again. You can argue that, you know, hey, as CEOs, we're creating, you know, the future of our companies. We're shepherding it or something like that, but it's definitely not the same thing. I'm sure you feel it, too.

Jake Moshenko46:14

Well, that's actually why I asked early on whether you were swinging a hammer, because that is kind of one of my dreams, right, is to actually go and build something that has like a lasting presence on my life with my own bare hands. But yeah, I think, you know, our archetype of let's build the things that we need because they're missing. Right, and we can do it in a tight loop. We could do it in a tight loop because we are the consumer for these things. Right. And then now we have the tools to go and do them without at least to get to prototype phase, right? Like we have the tools to go and do these things without, like you say, engaging a lot of people or like having to explain to the board why I'm working on this, right? Like why I have 10 people working on this. Like that's that doesn't sound like cockroach TV. What are you doing? Right? Everything is just a lot. We have a lot more agency now. And that's been a really exciting thing for me.

Spencer Kimball47:08

And, you know, what you want to do is find ways to promulgate that thinking throughout the organization because we have a lot of really good people. And one of the ways we did it and it was an employee that left us recently. He's great. His name is Jordan Lewis. He actually started a new company. You might know Jordan. Yeah. He's been around the cockroach for a long time before he's gone to do his own thing. He bequeathed to us in February a system that we call Mica. And Mica is pretty, pretty wild. It's basically think of it as like Claude co-work. It's essentially like that. But the difference that is so crucial is that when you when you run Mica, first of all, it hosts applications that it builds. It uses cockroach serverless databases. There's like thousands of them now thousands of applications that have been built in two months. There were a thousand applications. There's thousands now and you would not believe the quality of these things. They blow away the long tail of SAS. So that SAS apocalypse. It is real to an extent, certainly the low order SAS things that it just like there's no there's no point because you don't the responsibility of taking it on yourself and just inventing exactly what you need is not so great. That you know, like I don't think Salesforce is going away. I'm not like one of those people because there's just so much responsibility that you need to invest in someone and a tool like that and taking on yourself is just not the right trade off. But like, you know, we used to buy these, you know, 10 seats of this marketing thing or this thing for HR or whatever we had like hundreds of those sadly right now we're just building them ourselves. My chief people officer I've been working with forever was at Google with me. She built some of the most I look at these things. I'm like, this is this is amazing. And it's like this very thoughtful app that's perfect for what we need. And it's literally for like it'll be for like just one use case like let's evaluate, you know, people on this new set of metrics around like AI usage, let's say, and just see where we need more help and more enablement. And it like it was done in like a like a partial weekend and it's so good. Anyway, like Jordan Jordan just built this thing honestly in a week and he released it. And what really makes it different from cloud co work is that when someone signs up not only get all that hosting and everything that's integrated, but it has access to their the same sources that their credentials have. Now, again, like in the enterprise, I don't think that would fly at a fast moving or supposed to be fast moving company at our scale. Right, like this is this is a game changer because you can ask it any question that you yourself if you knew all the schemas and snowflake and all the different data that's available, which nobody does. Right. But if you knew all that, and you were really smart, like as smart as Opus or Fable, you could do this yourself, but you know almost nobody can and if they could, they don't have the time. And this thing, you know, may go off for 20, 30 minutes and it will help anyone like literally anyone in the company build a, you know, an app that people are very proud of. So it's incredible that, you know, you see an engineering leader like Jordan is able to impart this to the organization and scale the AI capability to everyone. And like, you know, that's really, again, when you hear some of the like, I don't know, the doomsaying or the doubts about AI actually having, you know, some salutary impact. I'll just say like the quality of what we're producing and the velocity is light years ahead of what we used to have to like procure. Right. And like partial solutions that nobody really liked that we're draining cash on.

Jake Moshenko50:53

Yeah. Is that something that you've open sourced or made available as a product?

Spencer Kimball50:58

We have not open sourced that. I think we should honestly. I don't know what the thinking is there, but Jordan.

Jake Moshenko51:03

It sounds like a dream from a go to market top of funnel perspective, right? Thousands of cockroach basic databases. That's

Spencer Kimball51:10

I agree with you. Here's a thousand new leads. Go. That sounds amazing. Yeah, probably, probably should do that. That's a good idea, Jake. Thank you. Yeah. It should have been obvious to me.

Jake Moshenko51:23

Well, 51 minutes later, we've just about answered the first question. One of the things that I always find interesting is as people are bringing in data sources and doing this sort of like ETL for data to intelligence process, those data sources have their own governance around them. And the pools that you're generating should theoretically inherit that governance. Are you thinking about that problem at all? How are you addressing that?

Spencer Kimball51:54

Yeah, I mean, like that's something I'd like to work with offset on. And, you know, that feels very appropriate to me in order for this to get really broad usage. Those problems have to be solved. And the only way around that is, you know, not not necessarily the. Dispositive or anything like that, but, you know, you build a you build a pool specifically for one type of agent. And but then, you know, you may want to. You know, I think in the end, you have to build it not just for one type of agent, but then you need a pool for each account separately, you can't have the details leak. And it's so important to get that right, because governance is like the the ultimate blocker to AI adoption. And you have to if you don't solve that problem, you don't give people the right controls, then the adoption is never going to follow, certainly in the enterprise where like the long term value ultimately accrues.

Jake Moshenko52:58

Yeah, these things, I think, often get stuck as POCs if you don't address the governance issue, because it's going to demo amazingly. Right bring in all this data, I, as the CEO can bring in data from anywhere. And so my agent that I build with this process is the smartest agent, but when somebody else tries it with just the data they have access to maybe it doesn't work as well or. You can't launch it because it brought in things that only Jake should know so those are the kinds of problems were thinking about. At outset, so I was just wondering, the minute you open source it, I'll fire up Claude Code and see what I can do.

Spencer Kimball53:33

I think I think we're going to do it in a month. So it's imminent. Perfect. Cool.

Jake Moshenko53:40

I guess we'll kind of fast forward a little bit. If you were going to start cockroach today in the AI era, would you change anything about the pitch or the architecture?

Spencer Kimball53:49

Oh, yeah, that's exactly the whole thing that the bigger thing that, you know, I'm working on. I got into my my science project with you a little bit too too long. Sorry about that. But, you know, for for cockroach. Yes, it's a new era. Right. And I think the core central thesis that we are organizing our business around and it's existential is that there's so much scale coming. Think about agents. They're busy. They never sleep. There's no limit to their proliferation. The right way to think of it is like, you know, human traffic has kind of gone in a signal. Right. Right. We had people with desktops and laptops and then mobile devices and then everyone in the world was getting on multiple mobile devices. And then it's like, what's your saturation point? There's still growth because use cases become more complex. It's just like we're talking about. Right. Go ahead.

Jake Moshenko54:40

Can I pause you for just a second? Yeah. Yeah. We'll tell the company. And I think it's a Yann LeCun quote, which is nature of horror is a raw exponential. All the world's a sigmoid. Yeah. Okay.

Spencer Kimball54:52

This is true. There is a limit to even agentic because it's got to be some level tied to human desires. Right. Right. Are there going to be 10 billion agents? Yes. Are there going to be 100 billion? Yes. Are there going to be a trillion? Yes. Are there going to be 10 trillion? A hundred trillion? Maybe eventually? I don't know. Yeah. There's probably a sigmoid.

Jake Moshenko55:12

As soon as they have their own desires. Yes. Oh, God.

Spencer Kimball55:15

If they start creating agents for agents, which will happen, then maybe God only knows what happens, right? Yeah. That is a very interesting thought experiment. How quickly are those things going to happen? Because the world right now runs a monolithic infrastructure. Like 90 something percent of all databases are like Postgres and MySQL and Oracle. Like, are those going to survive a 10x scale up? Well, the question is, how much headroom do you have in your 64, you know, VCPU monolithic database? Or, you know, there'll be more VCPUs in course, you know? Okay, that increases. It's been 10x every 10 years in the past. Now it's going to be 10x in three years. And then it's going to be the next 10x in two years or something like that, right? So, okay, what does a 10x mean? What does a 100x mean? Oh, my God. What we've realized, and this is the central thesis. The cheapest or most cost efficient database at scale is going to win the market. Full stop, right? Yeah. You cannot be an expensive database in this world. We are an expensive database. Like, okay, we got to change that, right? That's what we're doing. Like, the beauty of Cockroach in a distributed architecture is there's like, there seems to be a lot of room at the bottom, like a lot. There's so many opportunities for incredible cost savings. And, you know, we've optimized Cockroach. It's twice as fast now as it used to be last year. That's great, right? But there's limits to how far you can go in the micro optimization. Much more interesting is, like, how do you move from this plague on databases and storage and so forth, which is, like, you have to provision for peak capacity. Mm-hmm. And as a consequence out there, the dirty secret of operational databases is they're like five to 15% utilization at baseline. So you have to plan for your biggest peak. And if you don't, then you get an outage, right? And then you need some buffer on that peak. So people are running at these, like, very safe levels. But, like, you know, you've got things like virtualization and auto scaling. And, you know, there's an ability to decouple compute and storage. None of these are new ideas at all. But, like, we've been having them baking for a long time. And now it's bringing it all into a repositioned perspective. We call it the agentic database cloud. But this idea that, you know, in the end, you really get the economics, extremely superior economics. When you own the cloud and the cloud is able to impact things with virtualization and you're talking about use cases that run the spectrum, you're going to have literally thousands, tens of thousands, hundreds of thousands of agentic workloads. Either fractional, tiny little tools that were built by AI or agents themselves that store metadata coordination between agents. There's a huge distributed systems problem. Just use cockroach clusters. Cockroach solves all those problems, trivially, actually, right? Then you have, like, all of your, like, smaller to medium size to largish to petabyte scale clusters. Put all of that into your own database cloud when you own it and you can bin pack and you can auto scale and you're not paying for, like, these peaks anymore. Then you can actually realize a 10 to 20 X. That's that's kind of our near term horizon cost improvement. I mean, we want to make cockroach much cheaper than RDS with all of the superior resilience. And that's really what we're aiming for right now.

Jake Moshenko58:34

So help me make that real. Does that mean you're building your own cloud? Does that mean I'm building my own cloud and running the soft, like cockroach software on it? How am I making this happen?

Spencer Kimball58:44

So we call it the agentic database cloud, only meaning that, like, it's the fleet of databases. And, you know, you're looking not at workloads in isolation, but at the, you know. Ah, I see. I see. You would almost certainly, I mean, you can run it anywhere.

Jake Moshenko59:02

So you're just saying the cost gets amortized over tons and tons more jobs. Right.

Spencer Kimball59:10

You're like, you're really utilizing things. That's why I say there's a lot of space at the bottom. I can you get cockroaches superior things like elastic scalability from a tiny thing to a massive thing and all the different sizes in between. And the resilience applied across your whole estate and that you would you would probably run that, you know, on AWS or GCP or Azure across them or like hybrid or you're going to run in the Neo clouds. Where's the inference cheapest part of this idea is like it's not so much. It's like giving people ownership, which gives you freedom, right? It's the superior economics that's the absolute casus belli. But it's also the idea that you need to go where where where your workloads make sense. And you know, with like cloud 2.0 rapidly developing like, OK, where are the GPUs like where's where am I going to get the low latency wafer scale engines, whatever it is right that you need for your use cases. It's not just like, oh, well, everything's in AWS, right? That is that's history, right? You sure need AWS. You need to run there. You need to run other things. You increasingly need to run across clouds like the this this idea of like getting off your legacy sprawl and getting off the idea of renting the cloud just from like one provider. But you actually own the cloud. And by the way, your agents run on it. Our agents run it for you. That's kind of that's the only thing that makes sense. You want you want to you want to handle 10x the database scale. Well, you want 10x your DBAs. It's like, no, I don't think so. Right. You don't want 10x the cost on anything. God forbid the human labor. You want to give all your people leverage. Right. So like this is a very comprehensive repositioning that actually, you know, it packages a lot of the innovation that we've been working on. Not not quite so clearly for this purpose, but in the end, we've done the right things. And we're doing more of them with a much more intentional effort towards solving this this huge problem that's going to afflict every company.

Jake Moshenko1:01:03

Yeah, I thought we were going to get a you heard it here first folks cockroach is going toward disaggregated architecture or something along those lines.

Spencer Kimball1:01:11

Well, we are. Okay. That is key. I said, that's a big part of the story. You know, that's how auto scaling works, right? You want to, you want to take a, let's say a database to state that has, you know, 1000 nodes and or maybe a workload that has 20 nodes in it or something like that, a very, very important big workload, and you need to double it. Like that can be done in minutes. Yeah, data movement, no cost. You know, that's that's the promise of these things. And it also allows you to do crazy, amazing things like offload compute. So you can do like your backfills and your schema changes without. Right now, we're very, I think we're good, considering that that's mixed with foreground traffic at making these things very isolated, non impactful. But wow, you know, if you can just move that on the spot instances, then you really get a you get the best of all worlds. It's pretty sweet. And, you know, I think there's, there's really interesting opportunities when in this new capability, because think of what we're one of the things we're adding we call plenum, which is the storage substrate for this, this, this new agentic database cloud concept that that storage system, it it really is able to collapse costs by doing a ton of sharing and it creates the ability for the auto scaling, but also the ability to Let's see how the right way to describe it. Essentially, it's going to since we control it so deeply, it's not just EBS that we're kind of locked into in various ways with like, you know, without an ability to go very deep into the storage layer and tie that deep into the cockroach layer, we're actually able to do things like imagine this fork your entire estate. In other words, you could you could try an estate level policy change. And like every single database you run could be trying like a new version of it or some you can control exactly which ones and in which order and like, this is all very, you know, interesting stuff that this isn't gonna be in the first version, but like we have now the potential to do these things just like, you know, really. Sort of let's call it like paradigm shifting capabilities in the database, especially when you start thinking about companies that have lots of databases. And that's the interesting thing, right? Every companies have lots of databases. Yep. Just think of the mic example I told you about 1000 1000 applications developed in two months by 500 people. That's, that's, that's the sign of the future.

Jake Moshenko1:03:42

I had, I had Zac Smith, the CEO of Datum, a little bit ago, and he was making a lot of the same noises around you're going to have multiple clouds and you're going to be stitching compute together and you're going to have workloads running everywhere. And he was painting this picture kind of like what you're saying about owning that whole infrastructure, owning that whole substrate and turning it into one cohesive thing for your business. So interesting to hear.

Spencer Kimball1:04:07

That would be a horrible thought, honestly, Jake. And you know it too, right? Because you guys run cockroach. It's like, whoo, like, you know, the people you need and the like the expertise you need. Like this, this is not feasible. No, that's just not that's not maybe maybe JP Morgan could do it or something like that. Yeah. Right. But like, you know, one of the interesting experiences I've had, and what's led us to this idea of like agents can help operate all of this stuff. For users is which I didn't waste my time trying to become, you know, an expert at running cockroach myself. I just asked Opus to do it for me. And I was really impressed and I continue to be at how good Opus is now when you give Opus access to Chitta that has all this other information about cockroach. It becomes truly superhuman. And like at that point, what you realize is that humans deserve to be elevated into the position of policy and governance with leverage. And what this AI provides is the grunt work and grunt work you could never historically have gotten human to do ever. Right. Like humans typically will respond to emergencies and alerts that are very important. Right. But like, what about all those little niggling things that might indicate if you pulled on the thread like there's a there's something nasty under this rock. Right. Right. And we should fix it. I think AI does all that. You know, it goes through log files. And, you know, but the question is, can you get the fast, cheap models to do these things? It's all a hierarchy. But my experience with Chitta was just that, my God, I'm glad I didn't try to become an expert on cockroach because this thing in Ocus outclassed me much less fable. Yeah. Right.

Jake Moshenko1:05:40

Would have been wasted effort. Yeah, totally. All right. We are way over time. I just have one last question for you. Do you consider yourself an AI Doomer? Is AI going to take over the world and end humanity and take all the jobs? Or are you more of an optimist, a futurist? I'm an optimist, for sure.

Spencer Kimball1:05:59

I mean, this is like the ultimate weapon. It's also the ultimate tool. And, you know, we have figured out how to not blow up the world with nuclear weapons, thankfully. I think we may get burned a bit by AI, but in the end, it's inevitable. And, you know, I I look at the immense satisfaction and pleasure of giving birth to my having. Yeah, maybe the right way to say it is like AI has has brought into existence the creative impulse that previously would have been unable to act on. And that is a beautiful thing. And that's what everyone should truly realize that their personal future is with AI. Like, how do you turn your ideas into reality? And we're going to have such a boom of entrepreneurship and human capacity and flourishing some I'm actually optimistic. But don't get me wrong. We're going to get burned. We're definitely going to be already getting burned. Right. Like the short term pain is of the you know, the transition is already biting hard and it's causing huge amounts of anxiety. Yeah. And, you know, it's inevitable. And I don't we can't turn back. And I don't think we should either, even if we could. It's a moot point, honestly. But I think that, you know, I've I've gained as I've gotten older. I don't know if you see my gray hair. I've definitely gotten a little bit more of a sense that there's there's there's there's a deeper symmetry in reality. You know, we're not we're not just like always on the edge of extinction. That's that's that's a perspective. I just don't buy anymore. I've been a lot of doomers throughout every age of humanity. And you may you may be tempted to say this one's different. You know, I have at times I have my own doubts, but I do have a growing sense that there's there's meaning to reality. We're not just, you know, you know, big mistakes emergent properties, something deeper than that, and I think that we certainly can screw up, but I'm optimistic.

Jake Moshenko1:08:08

I think, you know, I have kids and I think my kids and your kids are of a similar generation or a similar cohort. So one way I like to frame the question is, how would you paint a picture of the future for my kids, but how are you painting a picture of the future for your own kids.

Spencer Kimball1:08:23

Yeah, I mean, I want to I want them to realize that their their dreams and ambitions are theirs to make real even at the age of a nine year old, you know, I've nine all the way to, you know, you're not a half a year. And, you know, my nine year old is already starting to realize this and she's built like some really interesting things I helped her build a little business. And she uses AI all the time, you know, to satisfy her curiosity or to like, you know, suggest things come up with ideas. And, you know, I think that everyone should be an entrepreneur. Everyone should be a dreamer. You know, that's that's what I want to encourage. It's not like, hey, you're going to be in some nine to five job working in a cubicle and you're going to like it. You know, think about office space, you know, back in back in the day when we watched that movie. And it was so funny because like so many people had that job and so many people had that boss or whatever. And like, it just seemed like that was right.

Jake Moshenko1:09:25

I quoted that movie already today. So you did.

Spencer Kimball1:09:28

Yeah, we're on the same wavelength, but that's not the future. That's beautiful.

Jake Moshenko1:09:34

Um, well, if people want to hear more of what you have to say, like, where do you like to hang out or where can they follow you?

Spencer Kimball1:09:40

LinkedIn is kind of the social I have. I need to. I'm going to I'm going to start talking about more of this stuff on on X, too. I think that's probably one of the better forums these days. Okay. All right.

Jake Moshenko1:09:54

Well, thank you so much for coming in and chatting with me today. And I feel like I learned a lot. And I hope our listeners do, too. Thanks, Jake.

Spencer Kimball1:10:02

Thanks for having me. It was a pleasure. Yeah.

Jake Moshenko1:10:05

Have a good one.