Age of AI: The landscape of AI agents in 2025 (and one platform to build them all) — Eduardo Gonzalez

All right. Hello, everybody. Welcome back or welcome to the age of AI podcast where my goal is to keep it the realest AI podcast in the world, right? On this show, I talked to CEOs of interesting companies in the world who are working in, you know, AI and really working in AI in the machine learning field. And I, this is actually season two of this podcast season one was back in 2021, 2022 and back then most of the attention was on Web 3, you know, that was a crypto and Web 3 was the all the rage and there was really hype about AI. And now when I'm starting season two, there's too much hype about AI. So the goal here is that, you know, most podcasts about AI are either super technical and they're for a very engineering focused audience and they're hard to follow for me at least. Because they're talking about some research paper that came out or it's usually by some management consultant and it's a lot of buzzword bingo right and a lot of window dressing about. And that's also hard to follow for me personally at least so we try to find the middle ground where we try to educate ourselves about what's really going on and try to understand the latest technology. But in the context of, you know, the whole industry in which so yeah, that's the that's the podcast for those who are new and now my guest today is Eduardo and he's the CEO of express AI and express AI what they do is they help they have a platform to help build AI agents and we're going to talk about what AI agents really are and what they're not. But let's let's start with let's start with some value upfront so Eduardo yeah, I'm going to I'm going to start with the big question right upfront, which is the industry landscape, right, which is the current offerings of what we consider an AI agent. I guess we could start by talking about the when people say AI agent right every like open for open AI to Google everybody and their mom is kind of building an AI agent right now. How would you categorize the different kinds of things that people call AI agents. Like what's and what are the differences you know let's start with that yeah that's a really good question because you know like there's no like scientific yeah kind of definition of these things and people are throwing them throwing the term around to suit what they're doing now. Yeah, for me I subscribe to the original kind of like reinforcement learning definition of an agent what guys say agent right now in reinforcement learning that's like this idea of having a program that can interact in an environment like a video game. You learn about the environment and do better you know usually with a reward function or something like that yeah and the idea is that you have a software program that has agency over the environment that it's living in. And you can think of a robot as you know a robot in the real world is an agent of sorts so your room is like an agent it's cleaning and we're cleaning until it doesn't see any more dust if it's an advanced one or it runs out of batteries and has to go back. Not everybody you know is thinking of these agents in this sense one pattern that I see a lot is like they're like chat bots but with I guess in to their so they're like chat bots but yeah they use lm's to produce the response no I don't see how that is similar to the reinforcement learning of agent but you could argue that it has agency in its response. So in old chat bot platforms you had like a classifier that would be like oh the person is asking about x so I need to respond why and that response is usually really hard coded or it might look up some information in a database but you know like it doesn't change what it says what it says based off of what you said or anything like that. So I say there's definitely a lot of chat bots that are called agents you kind of see them as like customer support chat bots and things like that. And yeah those you know I'll give it to them they have their agency in their responses so that that's good but what is more interesting to me is the agents that can actually interact with and do work in the world. And you know I feel like that's where we're trying to go with with express air is one of our main goals is to kind of create agents that could you know be like a virtual co worker of sorts but I don't want to go into that particular word right like I think it was I think Google was one of the first companies to kind of really announce that they had these virtual co workers. And I think that's it's a little early for for that and I don't want to necessarily think of the AIs that I'm interacting with as co workers per say right like the the people I work with are especially type of thing and the AIs that assist me you know I call them agents their assistance right and they're different right. Yeah so that's how I can see it okay so to kind of like paraphrase what you said. The first thing you said is that an agent you know is defined by having agency right in the world that it operates in right that's what makes an agent makes sense that's very. That's pretty straightforward and how we classify agents in the AI world right now is it seems like it's based on the complexity of the world in which the agent operates is basically how you classify the agent so there's you know I mean they all operate in a very predefined in a very defined word for the most part right. I mean there was a boundary on the world right either it's a chat with a person and maybe the agent can produce images it can maybe up like interact with the internet right in some way and its actions are you know bounded in some form or it can work within a video game right which is still a computer program which is still a defined boundary. And how your classify agents is not as much as like how they whether they have agency or not because it seems like they all have agency within the world the operate it's more about what kind of word they work in. Is that is that accurate yeah I would say like the key is so you can think of in this like robot model agent has sensors right so for most of them it's like a chat window. And the text that's coming in and they have actuators which they can use to affect that environment and usually you know for for the simplest chat pots you know that's they can interact by putting more text into that. Yeah but you could create more useful agents by giving them more tools like they have the ability to for example write a program or execute something or send an email those kind of things and I think. Like most things that call themselves agents like on AWS and things like that the main idea is that this is an LLM with the ability to call tools in the external world that makes it an agent. Yeah and in that sense right the tool calls those API calls or just the ability to create and execute programs yeah those are the actuators that that agent has on the environment that it's running. Mix and okay so now let's talk about the names right and let's talk about who are the actual players and how we classify those in the context of what we just discussed so there's I mean. You're probably much more on the beat about the latest news than I am so. I mean let's talk about what the big players are doing and how their work falls into the classification that we just described and how and how you would also classify the small players that they're you know the how you how you personally. Talk of the industry like how does that work yeah so essentially so I would say like it's interesting to to start with opening because I think opening I really took charge with this setup when they created what they call GPT's now they wanted to avoid the term agents for various reasons I think they're they have that legitimate like you know that's a reinforcement learning thing they know what they're talking about. But yeah the GPT's that they introduced had the ability to you know look up information in some documents that you can provide and you can give it access to the web or programming tools and then yeah and then you had like these assistants that could be you know styled as various useful things. That kind of really kicked things off the other one of the big providers I think is anthropic because they added the model context protocol to their application and computer use and what that means is that now you can kind of launch anthropic on your computer and it can interact with your computer so that your own computer becomes the environment so if you tell it oh please make a text file with this information in it. It will actually create that file on your desktop and if you you know have a computer use enabled which I'm not sure if they've I don't I think it's still like a you have to kind of build it yourself type of thing. You could actually have it you know like launch applications and things of that nature so that that's all good and I would say that's the you know the personal assistant like the kind of thing in your computer that apple alluded to way back in the day. With you know what I'm talking about not way earlier than series so there's this really apple ad from like the early 90s I think called the apple navigator completely fake type of thing wasn't an actual product or anything but essentially there was this professor who had like an agent in his Mac or whatever he was talking to that would help him with his lecture notes and things like that. And I would say that's what open AI anthropic and a lot of these companies are kind of working towards now. So that becomes like this personal assistant type of agent then there's also the kind of industrial versions of that and I think AWS and tools like n 8n and others are working towards that where for AWS in particular they have their agent system. Which allows you to hook up LLM from their bedrock kind of APIs to like Lambda functions or APIs so it's a way for a programmer to create an agent that can essentially interact with a user through plain text or interact with a document through through text or images and then call APIs based off of that. And you can make an arbitrary complex workflows of that and and that's good for the programmers but you know what I see is missing from that is you know like if you're not an engineer you're not going to be working with these kind of agents right like with the AWS one in particular because it needs you need to know what API is you know what Lambda functions are you need to know how to deploy code in these kind of things. And where express AI is and where companies like n 8n and another you know people I consider our competitors. We try to make it easy for a person to make their own personal agent workflows or you know programmers to make kind of commercial workloads. So for example let's say you're a bank and you need to verify the documents that you know like the address of a person and you want to do that through having them submit the you know a phone bill or electricity bill or so on and so forth so you need to kind of take this picture and then send it to the application. Well you can have an agent actually go and look at that use its multimodal capabilities to extract the information from the image make sure that it's actually an image of this right not a selfie picture or something like that. And then match that with the information that the user provided themselves but also match the you know like the image of this bill with you know is that a real company or is this fake right like this look like somebody. It took a highlight not a highlight a bit of a racer and fix some names right like you can ask the agent you know does this have any evidence of this right yeah and so you can make this kind of programmatic workflow that you deploy as part of the signup process of your application. Do that with no code with a lot of these tools or with you know tools like chain DSP DSP why I think is what they call it or opening has this thing called swarm and I think the other one that's cool is crew AI as well as our own frameworks. Yeah that allows programmers to do this but the cool thing about the gooey ones like an agent express AI and others is they allow people to do drag and drop things. So kind of like how Zapier makes it easy to make automation of some sort well with you know my tools and others you can create an agent that has access to the tools and you're not just making a single workflow right I think that's one of the major differences like. Software has one function right ideal like you kind of think of it that way yeah people don't have one function you don't hire a blue button to red button HTML code right like that's not what you hired that person do you hired them to a role and what agents and you know the bigger picture of AI is is you're able to take a role and deliver that to that role using software so if I have. An agent that has the ability to access my calendar and my email well then I can have that agent fulfill the role of a personal assistant right is not that I want it to every day you know like just because I ask it every day when I get a transcript from the daily. Stand up meeting I want you to summarize that and put it on slack that's one thing I asked my personal assistant to do or my AI personal assistant but I didn't explicitly program that single work flowing so that's where agents kind of differ from Zapier is that you can hook them up with some tools and then as you need you can have them do. Things that put those tools together on a as needed basis so they fulfill a role rather than a specific task if that makes sense interesting interesting so to summarize the way I understand it you said when it comes to the big players right they're mostly making. One agent that you can kind of like ask question directly so they so they make an agent in a box basically for you right so they ship an agent that you install and then it's ready to go you tell it to do whatever it will figure out how to do it on so right there's no customization essentially is that accurate so far. Yeah I would say open AI and anthropic yeah for example yeah open AI and anthropic they're like they're they're they're making their their like they're making open eyes chat GPT and anthropics clawed be an agent and so what the agent you know can do in the tools that it has is kind of hard coded yeah now anthropic is doing interesting things with that model context protocol that I talked about which allows third party developers to hook into it but you know like. The way cloud will respond will always be like cloud and when it comes to the other side of things I think you have more capabilities to customize how that agent works and whether you have one or 10 different agents working together. Okay alright yeah so either as an agent in a box mostly uncustomizable or you have you know platforms where they basically let you build your own agent and you know it's like Lego engineering let's call it that right where you just play with whatever pieces you need and it's more of a creative platform let's say. To do stuff on your own right and within that there's two categories there's the category for the hardcore engineers right who really want fine tune control and I mean and then there's the you know more user friendly let's call it that right more the more user friendly approach which is. The interface of how you interact with the platform that helps you build the agents right you could do it line by line of code or you can just drag things around and you know do it that way and abstract some of the you know underlying functionality right yeah so is that that capture all the categories so far. Yeah and I think an important thing to keep in mind when it comes to like giving the agents tools is it's as much as of giving it guardrails as it is about giving it you know like the ability to do things so you know like for example one I have the ability to have the agent you know work with Google docs and you know access my own Google drive but I give it access only to a single specific folder in my Google drive not the entire thing. That way I know that it can't you know that the blast radius of its of any issues is limited to a certain folder and I think that's actually a super important. We'll talk about safety you know in in a bit so now let's talk about you know the platform itself and at the risk of you know maybe educating some potential potential competitors actually want to get into the tech like the system engineering of how a platform like express AI itself works right. But let's first talk about the systems you know like the big moving parts of an AI agent the kinds of which you can build on express AI on other platforms right regardless of platform or even including let's throw in AWS you know in there right what are the big moving parts of an AI agent. Yeah so essentially so one of the the biggest kind of main focus for a lot of people is the LLM itself right so yeah LLM is essentially a very magic function that allows it to take in text and it can do things intelligently. But you know there's limitations to that and one of them is like the memory right so the next the other thing that is really important to any LLM is the memory system a lot of people are using vector databases for for the memory system of their agents but you don't necessarily have to. But yeah you can think of there's an LLM there's the databases and then there's the kind of like integrations right so you need to have a way for the LLM to integrate with the tools those three things make up like an LLM based agent. Yeah so just what what Eduardo said just now is that when he talks about memory you know when you use when you haven't when you have an LLM that just a large language model right. It doesn't actually have any memory so if you give it one input it will give you one output and then if you give it an input again it will forget what it talked about but when you use an LLM on. You know chat.openair.com or Google Gemini they actually have a memory component so when he's when you use those you know tools to like those chat bots they're actually doing a lot of this work in the background like abstracted away under the hood where if you send in a query they will save that query and those parameters in memory and then the next time the next like the next thing you say in the chat. It will not only go like the LLM will not only get your new message but it will also get that memory so the LLM will make a more informed decision but if you just have an LLM like as in like a you know machine learning model that doesn't have the memory so that's what Eduardo is talking about where if you're building an LL and AI agent yourself right you need the LLM you have the LLM and then you have the memory component which is but anyway sorry go ahead. Go ahead yeah yeah I think it's important to so there's this really good movie called her that has a very specific moment so it shouldn't be spoiler but if you haven't watched this is worth watching it's an interesting one where the I guess maybe it could be a spoiler but you know the the. It is an old movie so if you haven't seen it I'm just kind of worth it but so the main protagonist asks the AI you know how many people are you talking to right now and the AI responds that she's actually talking to a bunch of people at this very moment right yeah and I think a lot of people feel like chat GPT or anthropic are built in that way but they are not. Right so chat GPT and anthropic is not aware of anything actually right like yeah it is a complete. Yeah like it's it's it's like a it's a bunch of numbers right that are completely fixed and your words translated into numbers that gets multiplied and then you know at the end they add that with a little bit of randomness to get the thing so you know the one thing about having the agent interact with the environment is that it needs to be kind of aware of the environment itself right so that memory adds that ability for it to kind of add. You know that I did something and something happened and if that wasn't a good thing maybe I should try something else like that kind of context is important and the way that gets added to LMS is to kind of add more text into the input. And you know eventually you're going to run out of how much input you can you can give the LMS will get slower and slower the more input you give it so it's important to give it the right context you kind of want to merge the memories into what is relevant right now for the next action that it's about to do and humans do this as well right like our memory we don't remember you know the. You know childhood things of like playing basketball or something like that when we're you know like doing like taking an exam yeah yeah taking exam you know like that there's context that kind of comes into. Consciousness as we do things so if i'm about to do something that i've practiced a lot in right it might be a lot more internal I don't need to do it but if i'm doing something new. Like for example I got a keyboard recently right it's a crazy small split build your own keyboard i need to remember the keys because it doesn't have as many keys as a normal keyboard and so you know my memories of the dollar sign i put it to this key kind of comes up to me when i'm when i need to press that right so. Yeah an agent system needs a memory to be able to to give the element the context that needs to act you know in a intelligent way to the environment yeah like memory is an essential part of having and a coherent conversation with anybody right or any coherent any coherent intelligence I guess has to have some sort of memory you know so that it's not like you know forgetting what it just did. Like five seconds ago basically right yeah exactly. One like practical application of this is so if I ask GPT like chat GPT what is express AI it gives me the wrong answer right but when I ask my agent what is express AI even though it might be powered by GPT 40. It gives it the memory of oh this person asked this question before and this conversation right and then from that conversation context it can give me the correct answer even though it's default answer is oh express AI is you know like consultancy where we give you the best AI talent right like that's what it thought but it doesn't know that you know like our product at all until I talked about it. And so yeah memories important part because if you don't have that you know you're kind of like everybody gets the same chat GPT essentially but your agents will be special because of the memories it has. I see I see so let's I guess that's also one you know one way to go back to the initial discussion about how you categorize agents which is if you have you know how we talk about agents in this today's conversation. With express AI our agents that actually keep getting better at the tasks you give them and that's I mean maybe they're not getting better at the intelligence LLM level you know although that's possible with fine tuning and stuff but more at the level of having just a memory of what it's already done and where it all where it already has a corpus of new knowledge that it can you know borrow from every time it's trying to do something new. Yeah okay right so so to go back to the you know big moving parts of an of an agent you have the LLM and the memory right and they kind of work together to produce intelligent responses I mean it makes sense because if you're if you're an AI agent reading a document and you have to do something in the document after reading it you have to have some kind of memory right otherwise you just like. You know regurgitating bullshit basically after you know constantly while losing track track of what you were doing right which was actually the case for I think AI agents or other lambs or whatever I guess three years ago you know before chat GPT came out. Yeah even today like you know LLM's are they they are professional bull shooters is how you should think about right like everything that they're saying is a lie it just so happens that they have fine tuned it or given it enough parameters that the probability of that lie being false is as low as they could make it at the time right so. Yeah I illustrated this with you know like back in the day you know like really early in one of the first versions of express AI I created an agent where I gave it like this prompt of you you are the agent of at wardigans all CEO of express AI right and then right after I asked it you know who is the CEO of express a right even though it had that information in the context. It wasn't fine to tune very very well I would say like the early versions and essentially once it said like the CEO of express AI is and you would expect right since it's had at wardigans all is in the context it would say at wardigans all is but it didn't it for some reason the sampler chose doctor right and because now had doctor in the stream and it didn't say doctor at wardigans all is in the previous part of the prompt. It made up another word and because of you know racial biases or whatever that were present in its training data it said like doctor Andrew something right instead of it wardigans all is because you know Eduardo the the token was given a certain probability and the token for like you know John was given another probability and the the LM is not in charge of which one gets picked one will get picked. And then after it's picked it has to go with it the L one doesn't have the ability to back space at least not today so and yeah and for those who might have you know no idea how like what an LLM actually is or or other what a machine learning model actually is actually will link to an article that I wrote back in 2017 I think. It explains in depth from from very non from a very non technical perspective how a machine learning model a deep neural network actually works right and you'll understand what a model is and then what a machine learning model is and then what a neural network is in a very intuitive fashion so I'll link to that. So in the in the podcast description yeah just for reference okay so we have the Mary we have the memory in the LLM and then we have how do you connect the what are the other moving parts like how do you connect the LLM's text responses to what are happening in my character like things are happening like API calls are being made not just being written but being made like how does that blumbeng can you give us an overview of how that works yeah so essentially what you what you can do with an LLM is either that like interpret the output to have it call out to something right so if it's just replying and you just reply it but if it's using a tool using a special like token or perhaps you can have it wrap it in some Jason or actually produce some code and then interpret that code. You can like get it to interact with the environment by kind of giving it some special tokens so in the earliest versions of express AI before we had like function calling the pattern was start a new line right tool in all caps and a colon and then the name of the tool and then the parameters for it right and that worked incredibly well. Once once the LLM so let's see the LLM gets a query right gets an input text string text string and then let's say it's supposed to call the weather API to which sends a you know a get request to the weather API to get the weather of a certain city at a certain time let's say that's the that's the what we're looking for and then once the data is fetched that data gets displayed or some decision gets made after that based on that data right of the latest of the weather right let's say it you know you a cab or it tells you to take a bicycle whatever I don't care I don't know right so let's say that's the that's an example of a use game right you you have an agent that you've programmed to fetch to get the time maybe of a certain time of day from the system or somehow and then it fetches the weather through a weather API which is another external service and then it sends that you know it gets it uses that information to send another response to a device let's say your phone or whatever through another service right so this is the example problem that we're going to we're going to use to let you know put bounce on this conversation let's say for right yeah we know now so far that there's an LLM with a memory right where it can store the information that we're giving it or that it's finding you know whatever like so that's that's that part there that's the brain of the LLM how does the rest of this stuff work what are the other moving parts that need to go like from what I understand I'm not an expert in LLM but like what is the output of an LLM I believe the output of an LLM is simply text like pure and simple text right or like I mean you could create an image or whatever but it's definitely not a function call yeah right because that's something like you need to compile like you need something to make that thing happen right the actuator yeah so how does it so how does it work besides the LLM and the brain and the and the memory yeah yeah so essentially so programming languages are also text right and since yeah for an LLM can can output text earlier so it's output for you it's output an API call function yeah exactly so what what you do is you you set up the system so either you you find to your LLM specifically to use the system but you don't really need to because you could just tell it to use a tool do this right and then you have your program take the output of the LLM the output of the LLM and see if the LLM has decided to use the tool or not. You don't have to do this, right? There's another way. So there's an external intelligence. So there is an external intelligence that's hard coded in some way. There's a human, hard coded intelligence as well that checks and operates upon the output of the LLM is what you're saying. Yeah, yeah, exactly. You have the program. Yeah, exactly. Yeah, you have the LLM. It produces some output. Then you have a program that takes that output and then scans it to see, is this trying to use a tool? Because if it is, then we do something. And if it isn't, then we do something else, type of thing. So a very simple thing. You can also set it up so that you have like one LLM is specifically, should I use tools? That's the only thing that's asked. So it has the question and it's like, should I use tools? Yes or no? If it is yes, then something else happens where we go to the call, what is the tool we should use? Kind of thing, but you don't have to do it that way. Like you can actually make this system in any myriad of ways, right? Yeah, yeah, yeah. And so what OpenAI and others have implemented in their API is that you can give the, excuse me, the LLM a set of tools along with the user's query. So if the user is saying something like, you know, book me a hotel near the event tomorrow, well, it has now a bunch of questions of, okay, what is the event tomorrow? What is the best hotel in these things? And it can see from, you know, there's the text that the user has sent. It will also receive text of, these are the functions available to you to help the user, right? And if you have look up whether, look up the current date, get, you know, look up hotels, get calendar items, right? Then it will be like, okay, well, first, I need to figure out what date is to be able to use the calendar. So ask what time is it? Because otherwise, the chat GPT does not know what date it is, right? It gets the time and date in its input stream. And now it goes, okay, now that I know that, now I can look up the calendar events for that day. And then it will call the function to look up the calendar events, get that text response, which might be like JSON or XML, you know, something, you know, like text-based that it can structure data that it can understand and pick apart. And then from that day, it knows, oh, it's, it has a location of maybe Tokyo, for example. So now it will use like the other tool to search for the thing. And yeah, and that's kind of how you go from just an LM that can just talk to now an LM that can use these tools. And the, you know, memory isn't necessarily required for that particular instance, but it helps a lot if it knows that, oh, when it's looking for a hotel, there's a million different hotels, right? But if it can see from the context that I really like, like the Villa Fontaine hotel chain, because I use them a lot. And it sees that in its memory, then it'll be like, oh, there's a Villa Fontaine and Shinbashi. Do you want to try that one, right? And it can, it can respond to me intelligently rather than just putting, putting me in like some capsule hotel, it just happens to be the cheapest, you know, like somewhere in Tokyo. So, so that, that's how you, you go from LM to now, this is an agent that can actually interact with the world, especially if it can make that booking. Yeah, yeah, yeah, yeah, make sense, make sense. Yeah. So, I mean, I guess it's, you know, to summarize, you have, of course, the LLM or a series of LLMs, right? You basically have a branching, you know, like a flowchart of intelligence that's in multiple boxes all over, which connects to each other and like, you know, it's, and it's, and there's, then there's external tools that it can use, right? So, and it's a mix of LLMs and a mix of like hard coded programs that act as checkpoints, right? Or branching functions or, you know, whatever we call it, right? And that, you know, kind of like brings up a side question, which is like, how big in terms of the file size or like the, if there is a file size, let's say, right? How big are these agents right now as of 2025? Like, for most like commercial type applications, can you give us some examples of like, what kind of, you know, sizes are we looking at? Because, you know, an LLM itself is a huge, you know, like billions of parameters, multiple like gigs of memory, right? Give us some context, you know, on that. Yeah, so, so I would say the, the interesting thing is around, like, you can make a chat bot LLM with like one billion parameters or 1.5 billion parameters, very small, can actually, you know, understand languages, you know, it can talk to you in multiple languages, for example. Yeah, it turns out that that kind of parameter count is not that smart, right? Like, yeah, sure. And you'll be actually surprised with how dumb these things actually are. But around 7 billion parameter models, you're, you're getting something that will understand that it can call functions and, you know, that it has enough reasoning capabilities to be like, okay, so if I need to book the hotel, I need to first figure out what date is today, right? Like, that's an idea that needs to happen internally in the LLM for it to be able to make that call. Or it just gets that kind of memory and then it doesn't need to figure it out, right? Which is something you could totally do if you're, you know, if you do enough onboarding, as we'll say. Yep. But essentially, you should, so a 7 billion parameter model, the 7 billion is a big word. How much is that really in space? So when we say 7 billion parameters, we're talking about 7 billion numbers. And a number is by default, you know, like four bytes. So that's what a 32 bit number would be. It could be more, but and those are usually compressed into like 16 bits of two bytes or even one byte. So if you compress a 7 billion parameter model into one byte, right? You get a 7 gigabyte LLM. And that's as small as you can go. Yeah, that's the smallest you can go. I mean, you could probably get some very simpler agents with a smaller model. But I would say that that right now is a kind of sweet spot of capabilities given the 47 billion parameter model, the smallest space you could, you know, reasonably get would be 7 gigabytes of, you know, space, right? Yeah, you can, you can maybe get it down to four or something like that. But yeah, but you know, that's what you should do. That kind of like bound like, you know, a ballpark, let's say, right? Okay. Yeah. And then it comes to the memory. And the memory is, is, you know, it's up to what you've done with it, right? So like it starts at zero. And then as you give it experiences, it starts adding. But it pales in comparison to the things. So like a little history of expressly, we started with a vector database. And one of the things we did with our vector databases made it really high performance. You could do, you know, queries in micro seconds of, you know, 20 gigabyte vector spaces and things like that. Turns out the vector database is really, really good as memory systems for agents. But the performance is completely unnecessary because it never gets that big, right? Like 20, like I've been using an agent for a year and a half right now. It is barely even a megabyte worth of data that we have. You know, it's like super, super small and yeah, it doesn't really, you know, it doesn't need to use that memory every single time. So it's, it's the way we program is active in the memory. So when I, you know, when it sees it made a mistake and needs to remember this for later, or actually remember it, but otherwise it doesn't. So you don't really need that much space, you know, 100 megabytes, 200 megabytes worth of memories is really all you need. And then the programming and stuff around it is, you know, you need less than a gig worth of like real libraries to execute the models and, you know, run the interpreters and all that kind of stuff. So with like 10 gigabytes of space, you actually have everything you need for, for an agent, but the vast majority of it is just the LLM because that really, that provides the magic sauce. That's how it knows what it needs to do next. And 10 and let's say seven gigs would be an example of like chat GPT, the chat GPT model or like, like for the latest model that are coming out, you know, like a race bigger, like the bigger the better. Where do those models fit in, just for context? Yeah. So, so chat GPT or open AI hasn't disclosed how big GPT 4 is, nor deep sea, but deep sea has created a model that, you know, they say rivals the '01. And it's a 405 billion parameters and that's that is a big 405 gigs, let's say that's yeah, it's like 400 gigabytes. Now, so to give you context on the capabilities of that, right? If you ask deep sea, how many Rs are there in strawberry? It will not get that answer correct. It doesn't always get correct, right? It will think about it. Like the interesting thing about deep sea in particular is it forces itself to think about the problem and then it will respond. But one thing is specifically when it comes to like the strawberry question is it for some reason, maybe it was in the training data, really, really believes that there are only two Rs in strawberry. And when it spells it out, it can see they're free, but it doesn't know that it needs to believe itself, right? Like it doesn't have that kind of capability. So it will actually sometimes say, for me, it's random. It will sometimes get the answer right, but sometimes say it wrong. And you might be going like bars and strawberry, why is that? There are reasons for that because the LM's don't actually see letters, right? So they have to from their training data kind of know the spelling of it because it showed up somewhere. And if there's enough misspellings of it in its training data, then it might become unsure. And like I mentioned before, there's a sampling process where the LM, you know, it has to deal with whatever comes out of this random sampling process. So if the sampler picked two or five, then it will continue its chain of thought after that mistake and just be like, yeah, that's what I totally thought, right? It has no way of really doing that. And yeah, and there's various things, right? Like these are very incredibly capable, but they're not that smart, you know? So four or five billion parameter models are giving you something that's like a one, but it's not yet at I think a level that a lot of people, like for some people when, you know, I don't want to belittle the work that some people do, right? For some people, they use a one and like, this is perfect, it's so smart, it has every all the knowledge I need. But that's like knowledge knowledge, right? Like text book, remember stuff and spit it out knowledge. They're, they are, you know, wiping the floor of humans at a four or five billion parameter level, right? So like a PhD student needs to know a lot about the human body to be a doctor or, you know, chemistry and quantum mechanics to be able to do things is able to utilize that knowledge very well because it has just so much space to store it essentially as is, right? Like we can interpret it, but it's storing all that stuff in it and it can access it very readily. But when it comes to like figuring something out, so one test we use is Sokovan. So it's it's a Japanese kind of game where you can push a box, but you can't pull a box. So if you push a box into a corner, you're stuck. There's nothing you can do with that box anymore. And the goal is to push all the boxes into these specific places on the map. Oh, three mini, which is the latest model from from OpenAI is able to pass level three in our test. And if to give you an idea of what the test is, there is a wall and it has to push the box left and then up and then right, right? That's it. Level four is there's a wall, but there's a hole over here. And there's one goal here and one goal here. So if it pushes the box through that hole because it's the closest hole, it's going to get stuck into the corner. It has to go to this side and push it past the original goal to get both boxes on the two things, right? So I guess people who are listening to audio might not get the the hand signs that I can see Eduardo making in video. But basically it's a it's a it's a visual like, you know, a game with blocks that require some visual reasoning. Let's let's call it that. And apparently LLMs are not very good at that, but you know, to come back to the the conversation of like the size, okay, like, you know, the 400 gigs will give you deep seek for the most part or or a cutting edge LLM as of 2025. And that size, I guess, is only about to inflate over time for the most part. Possibly. Like I come up with this new way of training them, you know, yeah. Are you architecture? I totally believe that we will be seeing smaller models very soon, right? Like, because so from from our perspective, right, we're looking at this. We, we, you know, my background is actually in deep learning, right? Like making a deep learning frameworks. So I can understand how these things work. And, you know, the the fact that our LLMs are not able to do backspace, right? Like they can't delete what they just wrote or something like that. They have no ability to observe that is a big limitation. And one of the reasons why we need such big parameter models, so it doesn't make mistakes. But if we had a model that could fix its mistakes, right? Like, it sees the mistake and it fixes the mistake. That gives it a lot less. Like, you don't need so much exact knowledge, right? Like, you can you can go by this kind of high level understanding of whether not what you're writing makes sense. And then the other thing is there's recently this idea of test time training, which is, it's a complicated word. But what it means is that you're learning while you're doing it, right? Like LLMs today, they are trained first. And then when they are being used, they are not learning anything, right? They are able to manipulate the weights a little bit based off of the context to get the the right answer more probable. But they don't really learn new things as as, you know, you're teaching them. And that's where the next kind of level will come, I think. Once we have the ability for an LLM to see mistakes, fix them themselves, and then incorporate that mistake into itself, right? Like, reweight its own weights so that it doesn't make that mistake again. And do that while it's being used, then you can start having way smaller models that are, yeah, that are going to be just as capable as 400 billion premium models. So I don't think we're, you know, we're not seeing that, yeah, if you don't have a terabyte of hard drive space or terabyte of remote memory, you're never going to run an agent. I don't think that's the case. We will totally be running agents in like 16 gigabytes of ram or eight gigabytes of ram within three, four years at most. And they're going to be just as good as R1 Mini is. Yeah, you know, there might be a there might be a Moore's law or its equivalent, you know, at work in the AI space as well, right? You know, and Moore's law for those not familiar is like there was a guy who from Intel or AT&T Bell Labs, maybe, I think the latter who predicted that, you know, the size or the number of the number of transistors on a chip will double every year. Like on a certain area of, you know, of wafer will continue to double every year. So, you know, chips will keep getting smaller and smaller and smaller. And it's finally, I think, slowing down now, but it was true for a long time. We people kept finding new ways of making chips smaller and more compact for like decades, which is why we have smartphones and smartwatches and AirPods today, but we might have similar, you know, things play out in the AI space, okay? You know, for a fact that the human brain uses 20 watts of power or so. Yeah. So, you know, that's the limit. A human should be possible, at least physically, because there's evidence of it everywhere to get a human level amount of intelligence in 20 watts. So, yeah, like there's definitely room to improve if we do. Exactly. So, now let's let's step away from, you know, the discussion we've had so far, and Lail, let's talk about you and your company, right? So, let's start with your background. Like, what's what's your story been to get to, like, how long, like, let's give it a quick history of Express AI. How's the journey been so far? Yeah. So, we found Express AI in 2020, and yeah, like started with this idea of kind of like 2020. Yeah. This is before LLMs, by the way, guys. Before LLMs, yeah, exactly. So, like, the mission has always been to enable any company to leverage AI as the core business, right? Like, in their core business. And at the time, you know, after doing many AI projects for various companies, right, the amount of expertise needed to actually deploy some AI thing was a little crazy. And so, our focus has been, how do we make this easier? And vector databases kind of came up as one of the ways we could do this. So, that's why we created our vector database. What's a vector database for anybody who's non-technical, you know? Yeah. So, there's a couple ways to do it. I think a lot of people. Let's compare it with, let's compare it with, let's let's for context, you know, to make it an easier, easier to explain. Let's compare it with an SQL database or no SQL database. Let's just have that as the comparison. Yeah. So, an SQL database, it allows you to search through matches, exact matches, right? So, if you want to look up the, you know, fifth most popular product, it will line up all the products that you have ordered by their popularity. How many downloads they have? And it can give you that exact fifth most popular product. And it's doing that. And SQL database is basically a table, right? It's basically a table, or a set of tables, like a spreadsheet. Yeah. For a really terrible analogy. And you can find stuff in a spreadsheet. That's an SQL database. And a no-SQL database, I guess, is a dark, is a table with documents, you know, or documents with tables that you can kind of look up similar to an SQL database. Like, let's put it that way. Yeah. I put the, so an SQL database is like a spreadsheet where, you know, you're, you have one table in it and it's fixed, right? Like, if you want to put more data into that spreadsheet, well, you need to add a whole column to everything that you already have in that. And then you put more data into it. And that allows you to look things up specifically, right? Because everything has the same data. Yeah. With a document database, it allows you to have different rows, right? Like different number of columns for every row of data that you have, if you were to do it that way. Yeah. And so you can't be guaranteed that everything will have a download count thing, data, but you can do it. It's kind of like looking for, looking through the documents in your computer. And, you know, not every document is going to have the same format, right? Like most, exactly. The documents can have different formats. But if you're looking to a spreadsheet, which has fixed columns, you know that every row will have the data in more or less the same, you know, the same format as the next row. Yeah. Yeah. Exactly. And so the main advantage of that for SQL is that, well, if you know that everything is exactly the same size, right? And you search through the middle, you find the middle and it's like, well, that's less than what we need. Well, then you've eliminated half of the things you need to look up, right? And so you can order it with a date with a document database, unless you have some other index, you're going to have to go through one by one. Now, a vector database allows you to search not by exact values, but by meanings. So it's a much more fuzzy kind of concept. So let me, so as a simple example, let's say I'm, we have a movie database, let's say we're Netflix, right? And I want to find a cyberpunk movie. So I put the word cyberpunk and I'm going to get a bunch of cyberpunk-y type of movies, right? The there's no keyword cyberpunk in like the description of the matrix, for example. But yeah, the idea of the matrix and the description that it has, you know, fuzzily matches the cyberpunk aesthetic or the ideas of it, right? Even you could even go as so far as saying like the image of the movie poster of the matrix has a cyberpunk vibe to it. So a vector database allows you to search by these vibes. You can find things that are similar or you could either even do math like in our database of saying, you know, like you can encode the idea of an accessory to something, right? So like a mouse is an accessory of a computer, a pair of sandals accessory of a dress, right? These kind of things. Yeah, it makes sense to make sense. Yeah. So it's like, it's basically like the a vector database is similar to like a like a vector based, you know, what we, like what I understand is vectors is mostly in the machine learning world, like when you're doing like NLP, right? So essentially, you know, we have this concept, you know, for like, let's say you're looking for the word dog, for example, is very close to the word rotwiler or dashdund or German shepherd I guess, you know, and it's also similar to wolf and cat. Maybe it's not very similar to, I don't know, shrimp, you know, or something like that or tractor or, you know, so basically words have relationships to each other kind of fuzzy. They're closest closer to each other, right? And I guess a vector database captures that in memory so that if you're looking for one thing, you can kind of find things around it. I guess like that's a, that's one way to put it, put a vector with a database. It will encode the meaning of a thing in numbers in a way that similar things have similar numbers and that's what really gives it this power. Okay, so that's what you started bit, okay? Yeah, so the, you know, the interesting thing about vector databases is, you know, as someone who did deep learning in NLP, we deployed vector databases as like the result of our recommendation system all the time, right? But the interesting thing about the, the language models at the time was that they could be pre-trained on just general data and give you very general vectors. You didn't have to very specifically train one. And for us, I was like, oh, that makes it much easier for people to take advantage of this phenomenon in a single database. And that was a big step up. And then as the chat to BT eventually came online, right? We were using the OpenAI APIs for a long time. You know, this idea, we can actually use this to, to think, right? Like, there's a paper that, and it's the only paper I'm going to mention that, that language models are in context learners, is, is one of the conclusions of that paper. What that means is that if you give it an example A, B, and then example, you know, C, D, you could then ask E, and then the language model will predict an F, right? And it should follow. So if you say, like, yeah, English, and then you put English word, and then put J, Japanese, and Japanese word, and then put another English sentence without the Japanese, then the language model will be like, okay, I need to, you know, the next most likely word is Japanese, and then it will start translating the thing, right? So, taking advantage of that, and using that in an agentic system meant that you can start to automate things, and I, we did lots of things with RPA in my previous job, and I thought, this is, like, this can revolutionize the RPA space, because now it can work around issues instead of just hitting right into the issue, and then failing. So, yeah, that's where we kind of landed into our current incarnation, I would say, is kind of the, so you started out by, so you started out with RPA, in a sense, and you were using, you know, oh, what kind of applications were people using Express AI for in 2020, and what's the evolution being like in brief as you could mention that? Yeah, so the original things was like recommendation systems, right, so like matching and these kind of things. So we had this video system, so you could take in just a bunch of video feeds into the thing, and then be able to search for things puzzle. What's an example of a use case? Yeah, so, so for the security use case, let's say they know that somebody in a green jacket was, is doing something suspicious, so where is these images, right, so with the system, you could just say man in a green jacket, and then find all the instances of a man in a green jacket in this video feed, for example, or an e-commerce, right, like we had the ability to say, if you look up chairs, you're going to get office chairs and furniture chairs, but then you could just click on the image of an office chair and say, hide things similar to this, and then you would get just the furniture chairs. That kind of stuff. I see. And how does that evolve over time? Yeah, so you know, from the kind of vector search, like recommendation space, when we started incorporating LLMs into it, we started mixing in this automation platform, so we originally had our platform for training models, and then we just kind of evolved it to not just train the models that do the inference, but in, you know, workflows, so you could just kind of create a workflow of, this is step one, give that to an LLM, see what the LLM says, and then do step two based off of that. Oh, I mean, what I mean, I guess, because, you know, most of the audience might be non-technical, I mean, I can, I can make sense of what you're saying, but like, I guess what people might be, what you want to know is like, what kind of things are people building, you know, with XPCI? I guess is the question. Okay, so today, people are building things like, you know, assistance for like finance and, you know, data extraction workflows. So for example, extracting data from receipts or PDFs or legal documents, for example, and building that into their applications and workflows. So, you know, you can have a chatbot that gets like, that's listening on the slack, and when there's an incident happening, it can actually look into your knowledge base, search for things, and say, this is the person that you should talk to for this kind of incident type of thing. So being able to hook these agents into your systems is what, you know, is what our XPCI is doing today. I see. And there was also somebody who was doing some neuroscience research who you talked about. What was, I mean, not on this podcast, but like, I saw it on LinkedIn. What was that about? Yeah, so we have, you know, so we're working with a team that's doing expansion, microscopy, and things like that. And the interesting thing about that is that it's, the thing is moving and growing while you are trying to image it. And you got to put it together in a 3D kind of space, right? The main goal of that is to do with essentially $20 worth of diaper materials, what needs a million dollar electron scanning microscope to do, right? And the way to bridge that gap is with computation. And so through machine learning and imaging and things like that, we're able to align and resize and fix these things together for that particular system. And, you know, like the interesting thing is how, you know, there's a lot of qualitative work that scientists are currently doing to get those images to match up when, especially when they're kind of weird and, you know, like, is this the same direction kind of thing? But now that we have like multi-modal agents that, you know, we can give very generic instructions to, and they can produce scripts to, you know, get it into the right orientation. We can now do this automatically, right? So instead of buying a million dollar like electron scanning microscope to view proteins of neurons, you can, you know, do a quick little wet lab test, scan things with, with a microscope, and then use a thousand dollar computer to compute this almost the same kind of level of fidelity in the results. So very exciting stuff. Okay, okay. And what's the, let's talk about the entrepreneurial journey. Let's step away from the world of AI and, you know, agents and everything altogether. And let's talk about your personal journey as an entrepreneur so far. 2020 was pandemic time, right? And how has, how was the, and I remember that AI, I think, was hot when self-driving cars were really taking off, but 2020, I remember, was the year when the certainly industry suddenly was like, okay, we're not putting any more money into self-driving cars that that trains going to, that trains missed, you know, missed the station and all like, it's going to be a while before we see all the magical benefits of machine learning and convolutional neural networks that we thought we would see. So the machine learning industry, AI itself took a nose drive when, you know, like when autonomous robots lost their, you know, sheen, and I think crypto was really taking off, you know, was really like getting hot with, and that's when we had experienced AI being born. So, have you raised any VC money at all so far, for example? So, so we're about to announce our first seed round, essentially. So, okay. I started Express AI in the bootstrap kind of manner, for various reasons. Like, essentially, I started Express AI because I lost my job at a Y-combinator startup in 2020, essentially. The pandemic really closed off China. And, you know, in December of 2019, I was in Shanghai. We were going to make the SkyMine Neurobiotics Lab with the China Academy of Science. We signed a bunch of papers. We made a bunch of deals. And then when the world started closing off, that essentially killed all of those things, right? Like just to zero. And, you know, the company was not able to handle that, right? Yeah. It was doing like the VC-funded thing of grow, grow, grow, spend all a lot of money. And that's a very fundamental, that's something, for something that fundamental to happen to a VC back startup, it's very fragile in that way, right? Yeah, exactly. I don't think many founders can pivot at a big level at all. We're going to build something completely different with all this money that we raised, right? So, I get that they just decide to fold, although there's probably a parallel universe where they could have said, oh, fine, we'll just build something completely different. Yeah, exactly. But, you know, from that experience, I was like, I'm going to build a, you know, the technology AI is very much valuable, right? Like I've done amazing things with many, many different companies. So, you know, I decided at that time, I was just going to do things, bootstrapped, we're going to keep things small, we're going to, you know, have the team that we need and make the products that make our customers happy type of thing, and keep things focused that way. And, yeah, so that's where I started. But, like, so before then, you know, like the picture of what I was leading at the time, I was the CIO, the Chief Innovation Officer of SkyMind at the time. We had like 200 something people. When I joined SkyMind, we were like 10 or something like that. So, it was really growing. And, you know, we were flying all over the place to China, you know, talking Apple and these kind of things. So, for the people who don't know what SkyMind is, we were one of the first kind of deep learning frameworks. We created a deep learning framework called DO4J, which is deep learning for Java. And, at the time, there wasn't TensorFlow, there wasn't PyTorch. Your options were use something that's written in Lua, right, or write MATLAB code, you know, like that's kind of the options at the time. So, this was to have been like 2015 then, 2015, 2016. Yeah, exactly. So, that that's when, yeah, I think I joined in 2017. So, just before I think TensorFlow did their release, we were, you know, talking to Sonson and things like that about deep learning was pretty big, right? And it wasn't LLM's, it was like, you know, just now, yeah, I think we were, yeah, seeing it, seeing it. Yeah, exactly. So, really, really early. But, you know, like the fundamentals of taking large amount of data, training models to that, you know, taking vectors, making things that are incredibly valuable to companies, like I think Boeing was on our comp orange, you know, like really incredible systems, including software and things like that. So, we did a lot of things in there. And before I joined SkyMind, I was actually working in a Japanese system integrator called the Japan Business Systems. And at the time, I was working on an AI assistant that would look at your emails and kind of extract your to-do's and find dates and look into your calendar in these things. So, I was like doing agent things way back, you know, when our tools were sticks and stones. And we had to train LSTM models just to find these things. And, you know, at the time, I was collaborating with Stanford University with Professor Jeroski, you know, a guy who wrote the book on LLP. And yeah, they were like, oh, bi-directional LSTM's are a thing. Here's some MATLAB code. So, my job was to take the MATLAB code and make it into something that could be run in a bank. And that's where I was like, well, they used Java. So, how can I take this MATLAB code and make it into Java? And that's where I reached for DL4J. And because I was such a big DL4J user at the time, they hired me, they poached me from that company. And yeah, and before JBS, I was at Panasonic where I did research on user interfaces and things like that. And that got me into this concept of domain-specific modeling where you drew your code, right? Like you draw the interface and then that would get compiled into code. And I find Express AI is putting all of these threads. Yeah, kind of like that. Because you can drag and drop your code. Yeah, exactly. It's full circle, or taking diagrams of things, compiling it into code, but making it built into LLMs and deep learning along with the research and products that are making back in JBS stuff. And I did this all in Japan, right? So, yeah. So, let's talk about the bootstrapping journey a little bit. And especially, I mean, I have been a bootstrapped entrepreneur for my previous company, which did not fly off the wheels eventually. And I pretty much decided that it's one of the, I mean, I don't think I'm built for bootstrapping. I think it's only for the hardcore. That's my self-awareness. Basically, it probably shaped years of my life, my lifespan, you know, to bootstrapping company. So you've got to start with zero revenue, right? Yeah. Same here. And so, what was that? Like, especially hiring, you know, especially hiring for a hardcore tech company like Express AI. And I guess I would call it mostly hardcore tech, because you need good engineers to build the kind of stuff that you've been, right? Just face it, right? You can't build this off of people in Fiverr, right? So, how was that experience? And what were the biggest lessons that you learned that somebody else who has the, who is, in my view, stupid enough to build a bootstrapped company might use them. They don't journey the future. Yeah. I totally understand the depth of the meaning of being stupid enough to make a bootstrapped company. Yeah. Like, there's definitely, yeah, it's, it's complicated. So, so when you have a bootstrapped company, so I would say like the, the, the flaws, right? The, the difficulty of doing a startup where you get the big VC money upfront. And like getting into the Y combinator in particular, because I see, you know, yeah, having SkyMine B and Y combinator was a huge boon to it for getting started. But it doesn't guarantee anything, right? Because it means that you can, if you don't, if you're not careful, use those advantages that you got. And you get like this atrophy. So it's like skipping leg day every day for like three, three years, you're going to be totally buffing your arm. So maybe, you know, like because you have that, you can raise money, your marketing engine is amazing. But you, you can't walk because your legs are just so weak. So, yeah. In, in SkyMine, that kind of manifested with the, the, the sales and the marketing, right? Like because it had this status, you know, we could just go and say, why come back by one combinator? And then that would open doors that weren't really open, right? Like they're like, oh, we should probably talk to these guys. They're, they're really good. But we didn't, couldn't follow through on that. And I think that, that's something that when you're bootstrapped, you got to do that from day one. And it's really hard, right? Like there's, you got to learn it very quickly because you don't have a lot of runway to work with if you have any of that, right? And, you know, so from day one, you have to be able to talk to people who don't want to talk to you and sell your thing. When you don't even have it, right? And, and, you know, pulling it up from your bootstraps, especially if you don't start with a ton of money. For me, I started with $10,000, right? So I had two months of my own salary to do something with. And if I couldn't get it up, if I couldn't, you know, get the engine turning by the end of two months, I had to find something else to do, which at the time was perfectly reasonable. Yeah. But, you know, was able to use my network, was able to convince someone to, to let me work on some stuff for them. And, and we did an interesting IP deal. They were spending way less than they would have spent if they were doing something similar. And we would just share the IP. And that's how things really got started. So I'm not sure there's any, especially because like I'm focused on this like enterprise thing, right? Like the goal has been to make companies and enterprises be able to use AI. And I have no idea how, I don't have a playbook to how to make an enterprise B to B bootstrapped, right? Like I was able to do it, but that's not normal. It's really weird. Yeah, yeah. It's also a Japan thing, right? Because in Japan, I see a lot of startups, you know, near me and a lot of young kids starting startups, which is good. And their definition of a startup is, oh, we're going to start by being a service provider or like a basically a dev shop of like a tech, tech dev shop, a cutting edge tech dev shop for enterprises. And then we will find a problem to build into a product eventually later down the line. But the Japanese entrepreneurs from what I've seen so far has been very much service first. And then eventually product as opposed to I guess the more, I don't want to call it Silicon Valley model, but I would rather call it the innovators, the dilemma model, which is like, I have an idea for something I want to see in this world. And I will give it a shot and keep, you know, like ring and repeat, like as long as it takes. And that's like, so I would say that is very much true, right? Like Express, I started that way as well, essentially, right? We didn't really start with Vecto. We started with service. And then that morphed into Vecto based off of what the service happened or what we did through the services. And I would say it has this advantage, right? Because like when you're young, you haven't worked in a big company. You don't know what their real problems are, right? So the people who are able to stick that landing, right? First try is ridiculously small. And it's a really good way to lose a lot of money, especially if you're trying to bootstrap it with your own funds. So going the services route means that, you know, you have these customers and experiences that you can leverage to get this knowledge of a problem space that you don't originally have access to because you're not inside of that company. But through giving them these services, you're able to interact with them. And if you do this right, you can find the problems that they're most willing to pay money for. And then try to solve that on the side. And solving that on the side is, you know, probably one of the hardest things, right? It's like, it's definitely one of the hardest things because, you know, your company dies without the service revenue. And it's a real struggle to, you know, you have to, like, it's like a first stage. It's like a multi-stage rocket, essentially, right? Like the services will get you to a certain level height and not in orbit, but above the ground. And then your product has to be there ready at that stage separation. And you need to be able to launch it. And, yeah, you need to kind of get the extra momentum to take yourself into orbit. And if not, you just have to re-enter. And, you know, I would say our vector database is a case where we started that separation and launched, but then, you know, missed the opportunities couldn't get the funding that we needed. You know, vector databases are now getting hot in Japan, kind of thing. When in Silicon Valley, nobody is investing in vector databases anymore. So, yeah, so we had to go back, re-evaluate, start again, and do something. So, yeah, it's definitely, it's suddenly tough, right? Like, if you start with a big enough rocket, you can then get yourself an orbit. And that's what I think VCs give you. And, yeah, when your bootstrapped, you have to kind of think in multi-stages. And the problem is, you know, when you're making a business, you don't, you don't really want to be working in your business. You want to be working on the business. But when you do it this way, you have to work on two businesses, and you have to keep that in mind. The people that you hire in the first stage, they might be perfect for that, but not necessarily for that second stage. And that transition is not easy nor cheap, right? And how to structure that is something that I didn't appreciate when I started, but now, you know, I do, right? And I try to, you know, make myself learn the skills that I need to get into that second stage, but also make sure that everybody else is building like product teams build, not like service teams build. And those very different ways of, you know, building code as well. And sound. And if we, if we have time for a couple more questions, is that, is it fine to continue? Yeah, yeah, yeah. So I do want to get further into the hiring part and the team leading part, right? Because I know, I mean, it just brings back, like, you know, traumatic memories of, like, how difficult it is to find, like, just peace together, the funds to hire and sell good people on joining your company and stuff like that. But what was, what were the biggest revelations, let's say, you had when you were looking for good people to join you, especially in Japan. Let's also put this in a Japanese context, right? Even though you speak, you probably speak Japanese, like, you know, at a pretty bilingual, level of fluency. I learned what was your, like, what was that, what was that like, that process like? And what did you, how did you grow, how did you develop, or what did you learn, you know, from how you were like and do, how you operate, what do you believe now? Yeah. So, yeah, Japan is an interesting challenge in various ways, right? So, I would say that express AI would not be possible in its form if it wasn't for remote work. Like, remote work is what really allowed express AI to get the talent that it needed to make the products that we have, right? And if, you know, if from the start, we were limiting ourselves to the talent pool in Japan, that would have been a no starter. The reasons for that are, you know, varied, but, you know, essentially, the average Japanese talent is a salary man, and the stability of the company and the, and, and also Japan has a lot more importance on status than money a lot of the times, right? So, you know, status versus money. Yeah. So, for example, like, I was talking to a CEO of a startup here in Japan. He has access to the American club. And for him, that's like a very important perk to be able to get into the American club, you know, you have to be invited by a member. And more important than any bonus that he could get, right? So, like, even though his salary as a CEO is probably a lot less, you know, at his particular stage in their revenue than you would expect for in a global state, he's very, very happy with the status that he gets because of that. Like, he can access people, he can call certain people. So, as a startup and a bootstrap one of that, you have no statuses to give, right? And you have no benefits to give, and, and no job security to offer either. The only thing you can offer is a lot of hard work, you know? So, that's, that's very difficult. It's a tough sell. It's a tough sell. Yeah. Yeah. If the salary was high, right? That could be a thing. You'd be like, you know, she's like, yeah, you either, you know, sell sugar water for the rest of your life at a thousand dollars a month or something, or, or join us and make some extra money. But your bootstrap, you can't really offer anything crazy. Yeah. But with remote, you can offer something that everybody really wants. Like, so one of our employees used to work at a fang company, right? And he just hated having to go to the office, right? It's the one thing he hated more than everything. And during COVID, you know, he, when they started the remote stuff, you know, he was kind of, that's when he realized how happy and how important it was to him to have this freedom, right? And now that they stopped doing that, now he's, you know, really, you know, interested and then joined because of, because of that. So, yeah, like remote is the interesting, like the remote and work life balances really interesting. The other thing that we use is contractors, right? Like hiring is, is tough. And, you know, if you have like real laws for, for contractors after a certain number of years and things like that, that makes your your life much harder. But having remote contractors, you know, that, that just gives you that extra knob. It gives you that extra time to do things with. And, you know, it gives the people that are working with you the, the non-exclusivity option. That makes it a much smaller pill to swallow, right? Because you can be like, yeah, just do it 40 hours a week for now or 20 hours a week. And then that can get people on board in the beginning. And then you just ramp it up if they're, they're really good. And then when it comes, for me, sorry, go ahead, sorry, I could, I could go ahead. And, and I would say the, the final thing is, well, then how do you find these people? And it's tough because, you know, freelancers aren't on job boards, right? They're not scouring these things. And my playbook is always conferences. Because with conferences, you get the, the people who go to the conferences are usually the really good people like it, like big companies. And a lot of the times they're not happy with their job. And if they see your company and you're doing cool stuff and they, they, you know, they can see themselves and, and be doing this work. And then you can tell them that all this is also remote. Then you're selling this, this freedom for them, this ability to grow and get this freedom at the same time and, and start their own entrepreneurial journey. That's something that I think, you know, you can leverage, or I think most companies that are bootstrap business need to leverage to really get this off the ground. Because if you just go and try to compete with other companies on salaries and bonuses and perks and, and job securities, you're going to fail in all those fronts and then only get bottom of the barrel people that failed to get into those places, right? And then you're going to be stuck with bottom barrel. And you definitely don't want that because this is harder than all of those jobs, right? Like, so yeah, like that. That is, I think, one of the keys of why I've been able to keep the company running for as long as it's been running. Yeah, so it's interesting because I have had a similar experience with contractors like I have been, I have hired contractors in the past. And maybe it's just part of the evolution that the longer you get into the journey, the more you have to phase out the contractors, even if you have to be alone. I mean, that's at least been my part because I, I just, I'm kind of over contractors at this point because I realized that you can't build a lasting culture and a lasting like foundation for a company on like mercenary type. I mean, we try to keep it real like let's, on this podcast, like that's that was my, kind of realization, or maybe that's my management style or that's my, that's how I, that's kind of meaning I attach to work and company culture and like, you know, and maybe there's a way to build a really successful company, really profitable and everything. We're just a lot of flexible, you know, a part-time or non-committed workforce, what they call it, right? Like very human resources type of approach. What if, what are your thoughts overall in general on building a company culture and foundation with like, you know, contractors, especially with people who are remote, right? So there's two parts, right? There's a contractor in the remote part, right? Now, I would say this, this works specifically with the bootstrap business in the multi-stage part, right? So because you have to start with services and because services work is different than product work. You, you, you are advantaged by having contractors because you can, you know, put them for a specific project and then if you have another project, then you continue and if, you know, if they're real, real contractors, you're doing that. Not everybody's actually a contractor contractor. Sometimes you're able to pull them from, you know, another job or they're a full-time employee and then they start working for you as a contractor. And those are the ones you have to think about as they're going to be the ones who stick with you from stage one to stage two. And so you got to build them up to stage two. So, you know, you have to identify like who are the ones who are going to stick to help you get to stage one and then you drop at stage one. As long as you're doing it professionally and they know what they're getting into and you're doing it the right way, there should be no real problems with that, right? They'll just move on to the next, the next project because they're like, you know, full-time contractor contractor type people. And then the, the people who are interested in staying for longer, you identify through the work that they do, right? And they have the kind of grit that's necessary. And you can start offering them, you know, like, you can offer them a contract that has shares like future share grants on a condition on raising a round, for example. Like you can set up contracts like that to kind of get them to get this idea that they're in this for the long term. And then also take advantage of the fact that once you do have some of the resources to make your products of stage two, you can hire different contractors for that, right? So, and they can be temporary as you just need maybe some, someone temporary to do some initial designing work, but you don't have to pay their salary for the, you know, the entire months where you're still working out things or talking to customers, you know. So you got to, you got to do the 40 chest there and figure that out. And so, yeah, contractors are great for getting that, but, you know, like you mentioned, they're mercenaries if they're professional contractors. So you can't put too much expectations on them. And then the other thing is, yeah. Sorry, go ahead. I was going to move on to the remote part, so if you had a question about this, go ahead. Yeah, I mean, I guess like my reaction to that is like, it's, it sounds like you from the beginning were focused on, you know, I mean, the setup of the, like the contract between you and the person, the legal side and the how is actually the deal is set up is one thing, right, which is, I guess, in your case was, you know, like a typical contractor freelance relationship, but the, the new ones there was that the, the interpersonal relationship with that person was not really like a contractor. You were actually the expectation and the, the undertone promise was that this is going to be a long term thing. I want you to come on board eventually if you can, right? And that's what that's the premise of the relationship. So you were, it wasn't like you, like a mercenary hires mercenary and let's get this job done, make some money and get out. It was more like, you know, the way you frame the relationship, besides the legal contact and how you get paid and that's like, that's the, the, you know, the other stuff. I guess that's what helped you set a foundation. Yeah, am I reading this right? Yeah, yeah, exactly. Like, so, you know, like the, the person who is X-Fang, he had a stable job, right? And so I offered him a stable job like contract. I will pay you a fixed amount of money. You give me exactly these many hours, right? You can start part time, it's remote. And at this, you know, six months and one year, we will expand that and then you become full time. And you can wean yourself off of those things and get ready for, for this thing. And from their perspective, they're, you know, they're like, you know, going remote. And yeah, you do have to be careful because there are like, you know, rules even in Japan about hiring contractors long term, right? Like that, they're, they eventually become essentially full-time employees. It's the same for America. My advantage is that the remote contractors are remote in other countries. So I get to take advantage of that and, you know, it's not about taking advantage of the people is like just staying alive, right? Like, you got to make it work sometimes. Yeah. Yeah. And yeah, because it's like that for some of the hires, not all of them, others are like, you know, they, they work hours. So what is the project? What do you need to be done? And then they, they do the thing and then they leave, right? So, you know, I'm, even though I'm hiring people as contractors for some of them, they are really like, I treat them as much as full-time as employees as possible. I try to instill that culture because I want to keep them on board. And that can be a transition that you go to. But other times, you know, if you, if you do pull them from, from, as full, from full-time to contract work, you know, you're there with them on that journey and you treat them that way. But yeah, that's, that's a, that's a tough thing. I didn't, I didn't appreciate that, that, that, that when I started at least. And now looking back, I can, I can see that that's how things played out. But yeah. Yeah. I mean, it seems very straightforward in, in foresight. Like, yeah, I'm going to, you know, find some people and hire like contractors and get some stuff done. But then looking back, it's always like much harder than we expected. I think that's been kind of like, my, I wouldn't call it mistake, but just try to like, for me, it was always a struggle to establish that vision, I guess. And that's also one more thing that I have personally learned where it's really important to paint a vision. Not just for the employees, but also for myself. Right. Yeah. For me, it was like, I've pretty much wasted, let's not cut, I've pretty much underutilized several years of my life, building a boot shop business, when I did not really have a vision beyond survival. Yeah. And I guess even when you are in survival mode, you need still a vision that you're working towards, like what's beyond the survival, right? Because that's, I guess, you need that to inspire or motivate your team and to also inspire and motivate yourself to make coherent decisions. We're not changing the strategy every five minutes for the whole company. Any, any, does that resonate with you? Yeah. Yeah. For sure. Yeah. So for sure, like when I started to express, I didn't, I wasn't like, I'm starting a services business. And then I will find a product and whatever. I mean, yeah, like in a sense, yes, that is kind of what it, what it was, but the goal has always been to create a product, right? And I had ideas of products in my mind. And I was going to be paying people to make those products into a reality from day one, not just the services that we got. And yeah, having that vision, that expectation, right? We're going to, that there is a stage two, and we're working towards that. You have to build stage two while stage one is running essentially, it's like building a rocket while it's going to space type of thing. So yeah, if you, if you start building stage two, one stage one is out of, out of, out of fuel. You're, you're going to really, really bad place, right? Like there's not a lot of time. You have to build it little by little. And what that means is you, you can't beat the competition in go to market timing, right? Like you can't be the first, even if you were the first with the idea. So, you know, one thing about Express AI is I think we were, you know, the first to be like, hey, you can drag and drop agents together and LLMS, right? I don't remember seeing any, any company doing that before Express AI back in 2023, 2020, late 2022. But there are, so there's certainly a lot of those now, right? Like now even AWS comes with like a low drag and drop thing that is completely useless from my perspective. I don't know why anybody would use it because it lacks a whole lot of features that we've been building little by little. Yeah. But yeah, the go to market speed that you can achieve is going to be slower. And so you have to keep that in mind, right? Like the, the first to market advantage is one that you need to know from the beginning that you're not going to have. But their advantage is to not being the first to the market that you can use. And because you're kind of using your services revenue as part of that, as funding your, your growth, yeah, you just like take advantage of that in your, in your own marketing materials and things like that. Yeah, it's tough, but it's important. You got to have that idea in mind because you don't want to just start building stage two to late, right? It has to be something from the beginning. Yeah, I guess, I guess that's, you know, that's the part where it shows the, the real caliber of a founder to know what, like to know the secret why the others will not work and why you have the advantage, even if it has nothing to do with first mover advantage, right? Like you have to know some kind of thing, some kind of secret about how the product works or how people actually use a product or how people, how people's brains work. Like in terms of what they will adopt and what they will not adopt, that gives you the confidence that, oh, it's okay. If the news headlines are, you know, talking about this new company that's going to build, you know, like my friends are sending me messages about, oh, did you look at this company? They're doing something really similar to yours and they're doing it first and they have so much funding in this net and like you have to have, you have to have some secret knowledge in order to sit back and say, yeah, fine, it's okay, you know, they're going to fail or they're going to be, they're going to have to pivot because they, unless they fix this thing that I know that they cannot fix or that I have fixed, I'm good. Like I can just, I can chill, I can go at my own taught eyes, taught eyes, and I guess that's what you're talking about, like when you talk about all those little things that you have to build into the product, those little tweaks that really create usability or delight or whatever you call it, right? Or some, in some times it can mean the difference between a product that sounds really good and a product that actually is good, right? Yeah, if you can take advantage of that symbiotic relationship that you can sometimes do with services and your product, which at first they start separate, but as you learn things in the service and you can, and if you do things right, you find what is that real pain point that your customers have had and start building towards that, eventually you can start doing things symbiotically. And a lot of the times you have the customer that those startups wish they had, right? Like they are going to have to gather data through tons of small people and very small signals to get the same insight that you get because you're so deeply embedded into the actual solution space. So when you do have that ability to kind of symbiotically work on those two, you have this secret sauce that becomes very valuable, I think, and you don't have to worry about it because you found your niche, right? And the world is full of niches, you could just like, everybody can be making a lot of money on their own niche. And so, like for me, I get the question a lot, like, how are you going to compete with AWS or Google? And it's a really good question, especially in AI because with SaaS companies, they don't really care about your email automation thing for marketers, right? It's too niche, they don't care about it a lot. But when it comes to AI, there's something different. Google, Microsoft, Apple, they care about that and they feel that that's their wheelhouse. But there's still so many niches inside of that solution. And all you have to do is find the really good one. And if you can get to that really good niche through the services that you're providing, I think that makes them the most sense. And I'll help you in the bootstrap journey, I think. Makes sense. Last question, who's the funniest member of your team? The funniest member of the team. That's a good question. I like to think it is myself because I make a lot of dad jokes on the team. So I try to to line the mood. And I think sometimes it's important, right? Because like, you know, sometimes customers don't pay or there's a competitor or something like that. And you know, as a founder, bad news happens all the time, right? And it's stuff that you don't want to hear. And you need to wake up that day and do it. Because what else do you give up? Right? It's stupid. And as a founder, you might have that conviction, but not everybody in your company is going to have that. And it's important that everybody kind of has that confidence in the company itself. Because one, you know, you definitely don't want to lose your team, especially when they're remote, right? So keeping that, keeping that going through comedy is one of the ways I do it. I don't know if I'm actually funny to these people, but I try to be the meme lord in the channel and stuff like that. Yeah, you know, it's also an interesting balance. And by the way, I apologize that there was not the last question. Could I just recall a question that I was going to ask you earlier? And I kind of like, forgot when I was listening to you, but just to a reaction to that, like, I found it a really kind of weird balance that, of course, as a founder, I can't I can't be seen sweating too much, right? Yeah. I have to hold the fort in some way. I have to hold even though I try to be as honest I mean, every founder like should be as honest as transparent as they can be with every single member of the team. There should be no secrets as such. But even then, you have to understand that people are people and they look, they need leadership, right? They need to see, they need to have a certain perception that, you know, you have to sweat in private for, you know, as long as you can, although some people take it to the extreme, like, you show up to work on a Monday and then, oh, guys, we ran out of funding. Goodbye. You know, like, that happens as well. So and I see that that happened, right? Like, oh yeah, it was quite like, you know, fast, the, the, the, the funding running out and things like that. And that was a big, a big problem. You have definitely owe your employees and even your contractors, you know, not just the contract, the things in the contract about like lead times and things like that. But you need to be responsible for these things. And you have to, you know, answer that question for yourself, like, okay, this is pretty bad. But, you know, is it 100% bad? Is there a way we can get around this? Is there anything in the pipeline? And if there is, you make it work, right? And you, you want to be doing that as soon as you can, as always. And there's always going to be choke points. And people can be flexible sometimes to take advantage of that flexibility when necessary and get it, get through. But when, yeah, when the writing is on the wall, you have to, you have to really own, own up to that. But you don't have to be, yeah, if you're just projecting the worry to the employees, you're not going to have a good time, you know? Yeah. There's, like one framework I found, there's two frameworks that I found that really inspired me in that context. One was if you ever get a chance to get to read a copy of Made in Japan by Akiya Morita, who's the founder, one of the co-founders of Sony, he talks about how during bad times, the cuts start at the top. So the, the, the person who's lowest on the payroll is the last to be laid off, and the last to get a paycheck, a paycheck. Completely flipped from the American model. That's one model. And then the model that I kind of, kind of devise myself was to, in my last company, was to have a rule that if we ever get to a certain maximum runway, like, we have let's say, for me, it was, I think, two months, right? Or I would actually stretch that in hindsight to four months, let's say four months of runway. At that point, I would call a meeting and tell everybody, hey, this is the situation. I need suggestions on what to do. Yeah. If you have a solution that I am not thinking about, right? Tell me, let me know. And then at the two month mark, I remember having this written down rule in the handbook that if it gets to two months of runway, I will give everybody a paycheck, you know, and shut down the company. Like, better to give your employees two months buffer to look for the new job, look for a new job, then to tell them on a Monday or Tuesday that, hey, guys, bad news. Yeah, exactly. And I've had that conversation before as well, like, express say, I used to be 15 people. And like early on, right? And that was, you know, we had extra contractors because we had all this contract work. And then there was the vector database. But yeah, I like told people, like, okay, here's the game plan. We need to sell X amount of these things, right? And we have this deadline to do it. And if not, then I'll let you guys know, we do things. And then you, you know, we keep you on for a couple of months. And if you need any help with recommendation letters, you know, who to call, right? I will write you the floating is the floating, like, they won't even be able to put it down with the recommendation letter. It keeps floating because of how good it is, the type of thing. And had to do that once. And, you know, that was necessary at the time. And now the company, you know, we were still going, right? It's not like it was the end of the world there. But yeah, you have to make those tough decisions. And being like that, that's where the true color show, right? Like for the other employees. Yeah. Yeah. So if you, if you can't show that the management takes this responsibility seriously and does this right, it doesn't matter what kind of culture you otherwise do, right? Because like, you're going to just kind of create the, you know, like an apathetic, under culture to whatever surface culture, your memes on Slack are going to do, right? So yeah, yeah, it's, it's important to keep those, those fundamentals in scope. And I do think because I've been in Japan, essentially, my entire professional career, the, the book that you were talking about and how that cuts that start from the top, that's something that I've done, right? Like I said, okay, my salary is now the lowest of everybody, right? Let's, let's get through this. And, and, and if, you know, things aren't bad, I'll, you know, we'll work it out, right? People have priorities, the ones have families, you know, like you do things in a professional way, and you can negotiate. So with some, with one of our employees, he was, he's single, right? Lives with his parents actually said one, two months, not, not a problem for me. And I'm like, great. Well, I'll take advantage of that and give it back in space later. And yeah, you just do those kind of things, just do it responsibly and transparently. Yeah. I have way more thoughts on that and I could keep talking to you, but I know, like, you know, we've kind of run our time on your part. But thanks Eduardo. Let's talk, let's talk off line. That was the podcast, everybody say thanks to Eduardo on LinkedIn, Eduardo Gonzales, founder of Express AI, Express without an E in the beginning at X, start with X, the letter, and then press AI, right?