Marty sits down with Brian Murray and Paul Itoi to discuss the convergence of AI agents, graph databases as a solution to LLM memory limitations, and Bitcoin's Lightning Network as the native payment rail for the emerging agentic economy. Paul on X: https://x.com/paulitoi Brian on X: https://x.com/murr STACK SATS hat: https://tftcmerch.io/ Our newsletter: https://www.tftc.io/bitcoin-brief/ TFTC Elite (Ad-free & Discord): https://www.tftc.io/#/portal/signup/ Discord: https://discord.gg/VJ2dABShBz Opportunity Cost Extension: https://www.opportunitycost.app/ Shoutout to our sponsors: Bitkey https://bitkey.world/ OPNEXT https://tinyurl.com/tftc2026 Unchained https://unchained.com/tftc/ SLNT https://slnt.com/tftc Salt of the Earth: https://drinksote.com/tftc Join the TFTC Movement: Main YT Channel https://www.youtube.com/c/TFTC21/videos Clips YT Channel https://www.youtube.com/channel/UCUQcW3jxfQfEUS8kqR5pJtQ Website https://tftc.io/ Newsletter tftc.io/bitcoin-brief/ Twitter https://twitter.com/tftc21 Instagram https://www.instagram.com/tftc.io/ Nostr https://primal.net/tftc Follow Marty Bent: Twitter https://twitter.com/martybent Nostr https://primal.net/martybent Newsletter https://tftc.io/martys-bent/ Podcast https://www.tftc.io/tag/podcasts/
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
I don't know. We're like two months into it. I'm like, I'm going to have to set up a new one at some point because this is going to be outdated. Really? Okay. I think that's the conclusion I'm coming to. You can switch out the models and stuff like that, but I think the context memory is the big problem, I think, and users like me who aren't as technically competent need to figure out how do we nail that out of the gate. That's a good problem you're familiar with. Sounds familiar, yeah. Right? Yeah. That's what I mean. That feels like a good starting point. I think yeah what would be most helpful I think about
just catching up overall on what's happening in the space and talking about the pieces you've heard me talk about it for a long time so can we talk about the 1031 off site we're at sometimes yeah yeah I mean just I get a little embarrassed getting up there and showing graph stuff just in front of everyone I see everyone go oh god Paul another year the graph guy and you get a lot of grief for graphs on online because they've been tried so many times but um i've worked with neo4j since
2010 2011 sometime around then i think they were just starting out uh one of our technical guys brought it into the company and i hated it because it just crashed all the time so um but now 15 years later it's kind of seeing the light of day so i think you just have to have all these kind of of primitives in your toolbox and then when the time's right you pull them out and so we just think that the memory issue you just brought up graph databases just serve as a great scratch pad for that and it doesn't have to be
in a graph database it can be an obsidian files it's just the whole thing is relating one thing to another but anyway well I think it'd be worthwhile wild to go into differentiating like llm's how they work from these these graph this graph approach because i think again like we're just discussing before we hit record yeah a ton of capital time and uh effort has been put into llm specifically but i think some would argue your i
think yourself included that llms may not be the the best way to go about this problem yeah i think people anthropomorphize LLMs a lot you know because it's speaking language to you because you can talk to it you think that it's actually reasoning and especially when they call it a reasoning model and it does do an amazing job of mimicking logic but it does not know why it's saying what it's saying it's just the statistical output of the next word yeah yeah I mean how do you see it do
you when you think of an LLM do you think of it as do you find yourself thinking of it as a machine spitting out words or do you think of it as you know especially when you name a bar or something like that it really starts to feel like a human or something no I definitely don't think it's human like anytime I interact with our open call I'm like okay what context do I need to feed it to make sure that it gives me the right response like that's what I think most about it is like what do I need to preload this thing with today we should start with what's everyone running right now like that would be a good thing you're running a bunch of cool stuff right yeah I mean I'm using
I've of course yeah doing some cloud code but I think the the context issue is something that everybody's kind of running into they're able to just like stumble through it or hack their way through it but I think like we're all gonna feel this need of something better like something that's gonna be more accurate give us better answers help us do the next thing better so and you're starting to see it like you open X and there's more and more visualizations of
graphs I feel like the whole like ecosystem is drawn that direction but yeah those are some of the things I've been messing around with yeah it's you Marty I mean I for the context like interacting with my open claw bot like we're doing something i have to be very specific i'm like hey we're gonna write the bitcoin brief today i've dropped some stuff in a folder that i've named a specific name i'd say the name of the folder go to my tab the tab is named this like i have to give it like direct direction
and i've heard of people leveraging the obsidian bolts to to sort of solve that context issue i haven't dove down that but yeah we're using open claw we use claude 4.6 and then um combination of could we have codex sub agents that we can and we can ping if we need to build something and then 11 labs are the models that we're using now for voice yeah when you told me about your automation process it was pretty solid about what you do is that stayed the same since then yeah you want to
describe that a little bit or yeah i mean for um like we'll take this podcast this transcript and then we have prompts that we've been iterating on for literally two years at this point um in claude that will that very long prompts that will take the transcript and then pull certain sections quotes and be able to either um directly quote for a section of the newsletter or find something that we're talking about do deeper research on it to expand on the topic for another piece of content um and that is for the podcast like post-production
that's one of the things we'll do but other things we have for clips we'll give it the transcript with time stamps and claude will then use our third party twitter api that we have access to look at what people are sort of talking about what's trending and then it'll go back to our transcript see if we talked about something um during that episode that people are really on top of and then we'll be like all right here's the time stamp of a section of the conversation that
you just had that if you clipped it out and put it on twitter it would it would probably get good engagement and are using any then for that sort of prop that pipeline anymore or it sounds like you brought up we bought so we were okay we live coded our own like back-end dashboard that basically replaced n8n um shout out to ed on our team just put a lot of effort into that that's pretty cool but uh so when you're done like this episode when you hit finish or record the recordings over does it automatically kick that over to these workflows once we put the files and Dropbox it'll start and
then yeah you'll be or you can just if we put it in Dropbox you don't want to trigger automatically in the dashboard that we built and go like all right right around this process that's crazy yeah that's cool I was talking to Scott who runs you know Scott Foreman he runs a video lab they do a lot of work the Human Resource Foundation and he's trying to do a similar automation on the back end so everyone is trying to figure out this this process yeah but it is like it does feel we're joking you came here in
a way moan it was driving on the wrong wrong side of the road yeah I let you know I'm like I don't I stopped paying attention to the road when you're in the back of an uber or waymo and i look the car starts like jittering and i look up the waymo is on the wrong side of the road with traffic heading towards me because the austin has all these events going on so there's they were like coning off the streets um so you could tell that somebody took over via teleob yeah and then finally got me over and then as soon as we were getting over there was a traffic
cop who was it's like trying to direct the waymo and it was just this moment i was like gosh what the hell are we doing here yeah well that's every human directing a waymo and then some poor person sitting probably in the philippines operating the game on some playstation console from 1988 or something like that yeah it's crazy but one thing i was thinking on the way over was um and we had a team meeting yesterday we were talking about this i think we all find ourselves thinking about the current state and it's hard to think about the through line
like where things are going to be in a few months um and when i was in this waymo today i was thinking about when waymo first came out or or self-driving in general people have been talking about about this for a long time and dismissing it but it's like it's basically here um and of course there's going to be some kinks along the way but we've come a long way in a short period of time and it's pretty hard to imagine what a year from now is going to look like because things are changing so much um but yeah it's exciting it's like an amazing time to be alive i mean we
don't want to say the phrase but you have to say it gradually then suddenly it is and people just have such a hard time with i have a hard time with it so i mean we're doing coding automation stuff lots of people are but the platform that brian mentioned um the takeoff on it is so real it's so crazy i'll do a demo for you later on but i mean we pull up a voice assistant on our conference calls and you haven't even seen this yet um we nicknamed jamie after uh rogan's assistant which i don't think anyone's really got on that so i think it's a good nickname um but you
literally just talk and and jamie can read your code it's not just some note-taking assistant where you tell it to make a note of this or basically transcript manipulation. That's what most of these things are. It can read through all of our graph database, which is all of our code, all of our chats, all of our conversations, all the context for the conversation. So how many times have you been in a meeting where you go, hey, is that the way it really works? You don't remember? Hey, Jamie, can you look that up and see if the code actually does that? And then once you can get past that in a meeting,
you just have, it's such a flow state because you're not like, hey, let's figure this out and circle back together again. Yeah. You just know right away. Imagine if we were talking about your agent in this chat right now and you're like, yeah, I really wanted to do whatever. And then by the time we're done with the conversation, it's just done. Fixed it. I think that's what we all want, right? Yeah. I mean, you were demoing it to us in San Francisco in November and it was doing it. Come a long way. Yeah. Well, that's the thing as an operator, integrating these tools. To your point, Brian, like what is the through line
to where this is? Like you don't have to worry. like we i tell everybody at our team here like hey this stuff's changing rapidly like something's going to work one week and then something better is going to come out we're gonna have to like start from scratch to rebuild it and it's just like this constant iteration of like okay build something quick it works it works better than the thing we built before but oh here's something new like we got to go do that again so that's what like i'm curious like when are we going to get to the point where yeah you don't have to keep doing that well there's a book that really impacted
me early is called second machine age have you heard of this before yes i've read that book you read it it's been a while but yeah it is an old book yeah mit professors i think and they talked about this idea i mean don't read it but i think the main idea is combinatorial tooling so once you make tools that help you make other tools this is when you start to see the fast takeoff and so i think that book made the point of what we're seeing right now is that the software tools to make software tools is happening so quickly and you can automate workflows
you can automate all the stuff that you're talking about and I think that the people who figure this out and can get these tools are in a permit you know we talked about the k-shape economy have you heard of that mm-hmm talked about that so I was explaining that to someone yesterday it's so true if which line you're on could sort of determine what it's like forever or for a very long time if you're on the down slope of that k then you may not ever make it to the up slope which to me means that we're just at this really critical moment i was talking to the best
man at my wedding last night in georgia he works in construction management and i'm going are you using any of these tools have you heard of any of this nothing so i sent him a video here's how you get started so i think that um just the distribution of this tech is just so uneven right now what do we're in a bubble yeah that's what i was gonna say what do you think the penetration is like 0.1 it's like you go on x and you think everybody's like knows everything and then you go to like a family gathering you're like nobody knows anything um no i mean that's been my experience it's it's
uh we're sitting here driving around in waymos and experimenting with apple vision pros and playing around with claw bots and um it's it's pretty scary what i think can happen when you you diverge people this distinctly it's going to be tough super tough well that's again another question on the through line that i have like thinking through this social problem like will it get good enough where yes we're early adopters we're reaping the benefits of this massively right now but will it get good enough where it doesn't matter somebody like your best
man will be able in a year from now to download it up just do it like that with little learning curve it'll accelerate so quickly I think it'll be way faster than that I mean what's your take on the well okay so where are you on the doomer there's like a divergent point here so on the one hand you could say the k-shaped graph which is basically to say that people there's the people who know how to use these tools and the people who don't right on the other hand you could say all these the tooling is benefiting the creators because the tooling is getting more accessible and easier so then it lowers the bar to creation I
i hope that's how things come to pass and i think about my kids so i've got four daughters and the older two are old enough now to where i'm like working with them on some things in fact i'm working on this bitcoin project with them i want to tell you guys about but um like our local pool the community pool they have a website where you like go and see the pool schedule and times and order your swim gear and stuff like that it's a piece of crap website and so my
ten-year-old I was like hey why don't we try to fix this website and the tools are easy enough for her to sit down at the computer and we just like iterate on it so if she's able to do these sorts of things she has the intention and the desire and she has like a point of view on how things should be and she can do it um and that's today i just imagine a year from now it's going to be everybody with an idea can go and create their idea i hope that's what the future looks like did you see replets announcement
yesterday agent four yeah i mean i tried it yesterday it's really good i mean it's really good you can basically just start with a sentence and then it'll build the slides so he completely expanded what vibe coding or vibe building is is like you get the code you get the decks you get the images i mean a lot of it is stitching it together in a clever way but you stitch enough of that stuff together and eight-year-olds can create businesses ten-year-olds i mean what does
this do to well that's i had a 15 year old a 15 year old girl on the show a couple months ago she's alpha school she's uh oh yeah yeah i saw a clip for that yeah stella she's daughter of a of a good friend of mine here in austin and they're obviously alpha school is probably tip of the spear of integrating ai tools into education and it was blowing my mind like how how robust they're the aperture of like what they let the kids do at the school it has me thinking like so i have three
boys uh the oldest of which is six right now he's in kindergarten and he's just like a catholic school and i think it's a good school uh kindergarten it's montessori based so they're not really getting too much in the weeds of math and science and stuff like that but i really want to like go in there and be like hey we need to like sort of rejigger how you're teaching these kids with these and you need it in a great ai because this is what they're going to grow up But they're doing the opposite. The guy I had dinner with last night, his son's friend is going to one of the top boarding schools in the country
we've all heard of, and they're just so anti-AI. You cannot use it. They're not allowing it. You don't touch it. It's evil. Imagine teaching that to kids right now. I mean, you're just, it's unbelievable. I'm of two minds. So, Marty, you haven't met my wife, but Paul's met her many times. She's awesome. She on a crusade to reduce the amount of screen time for kids in school We go to public school which I all about I think just giving these kids Chromebooks and having them sit on a computer all day long
is not what young brains should be doing. On the other hand, I want them to be exposed to these tools and become capable with this technology because it's so cool. It's amazing. You can make anything. so I'm I'm like mixed on this you know how much to expose young people to like kids to this technology and encourage them to build but without having them sit in front of a computer screen all day yeah that's why I like the alpha model it's like two hours of intense like how they do integration then beyond
that's like social skills like group projects stuff like that yeah that's good especially if we if things happen the way we just described where it's like people can have a conversation and then you end up with a product that other people can use that's a lot different than the hacker spend it you know not sleeping spending all night coding something um i hope that's the the future we find ourselves in you know i i think the the focus needs to be on like first principles logical thinking like clearly you that's what i've come
to learn again going back to how i interact with my my agent is again thinking about the context and the direction i'm going to give it so like having to think from first principles okay what's the sort of logical order of instructions i need to give this thing to get the output that i need and i think reading comprehension uh is going to be way more important than stem moving forward unfortunately well you write every day so it forces you to clearly think through steps and i
think having that cape that muscle fully exercised when you write down it's you know the you've heard this term spectrum and development you're no longer encoding you just you have to clearly articulate what you want and that's really hard to do when i sit down and try to stack out and the system's asking you questions well you want to select all and then what if you select all then how do you unselect and you're just like oh man you're right but forcing you to think through the whole thing is actually where the mental friction takes place so how do you teach that to the next generation
yeah like how are you i mean you're is right so it's like how old are they and what we say to them i mean my oldest is he's just in the process of starting to read but i think never too late marty no but i think reading the classics again like uh getting like i don't know at what age you introduced of like literally socrates played on like how do you think like how you logically like aristotle yeah really old stuff yeah i mean i went to a jesuit high school and that uh and actually like learning latin like we had uh we were we had to take six years of language and two of which had
to be latin and like you get down to like the etymology of of language and uh the whole goal of our high school was like by the end when you graduate you're gonna be able to write competently like you're gonna be able to read something and then have an opinion on it and write articulately about it and i feel very fortunate it set me up like you said i write every day but i think that that sort of model is going to be more important um moving forward what survives after ai then i don't know that's the other question like is it going to make us dumber like you you see
people allocating their thinking to the machines and are we going to get find ourselves with the competency crisis a decade two decades from now where god forbid something happens to to the machines and then we don't know how to rebuild anything because you have a generation who never learned the first principles of software or math yeah but but just i i think most of society has gone that way like when the average person encounters a car engine they're probably i don't
know what the hell yeah what's going on in here and then you meet someone who knows what's happening and it's extremely impressive. And that person who knows what's going on can fix it. They can do a lot more things. So I hope that that kind of competence gets rewarded. In the future, I hope these types of people are better at creating things, at advancing things and encourages other people to do the same thing. And I think from like a parenting perspective, I think that's an important thing to bestow on your kids.
And honestly, my role model is Paul. yeah i remember the first time do you mind if i told the beaver story sure okay first time i met paul's younger son um came over to his house first time i went to your house there was a sous vide happening you know i was like what are you guys cooking like oh beaver tail like okay interesting and he goes on to tell me that his son just got he was like 10 at the time yeah 10 10 year old son just got back on a solo hunting trip to he trapped a beaver brought it
back was suviing the beaver tail so he's cooking for the family but then before that he like scanned the beaver and was making a video game out of the 3d scan of the beaver what and then over the years i've uh come to you homeschooled your kids yeah for a while up until uh high school really and i just i think i've learned that you took a very you and nicky your wife took a very intentional approach to how you brought your kids up and they understand
things deeply and now they're super resilient and capable um so i mean i feel like this is a tangent but also pretty important thing for people nowadays is to try and teach people these like yeah i mean i was i mean you guys are in a different phase in terms of upbringing but um I was right about one thing and wrong about another thing. I said, don't learn to code. I don't think that people will be writing these letters long term. So I wouldn't say that learning to write the letters is unnecessary
because I think that just like learning Latin, you're not writing Latin, but learning Latin, which I did too, teaches you the structure underneath. And so I really felt like the job of writing code won't be around long. That's why I told them when they were that age, just don't. don't bother worrying about that but i also thought they would never have to drive a car so um that took way longer than any of us thought you know i rode in a tesla you know i was driving down the freeway a decade ago and i thought oh we're so close so i think and
now that's basically here so the the question of timing is the one that's the tough part i will say back to our previous point just that it i know this from firsthand that you can automate basically two day with no advancements, no change of anything, 80% of every job done on a computer. So everything that we're all doing here, and 80 is pretty conservative. So go into the most, you know, esoteric job in some law firm, and you can automate it consistently to a degree that's
better than even doing it. And that's a fact. And whatever we do with that fact, I'm not sure. But, you know, again, my son's friend who's doing finance at a public university, and I'm looking at his homework, and it's manipulating spreadsheets. I'm like, oh, my God. You're paying to have someone test you on editing an Excel spreadsheet? And I dropped the files in the folder and said, here's this thing called Claude Cowork. Go to town, you know.
And then I had an Uber driver the other day paying $6,000 to $8,000, something like that. He saw the Waymos and was like, I have three. I've been driving for 15 years. I guess I have three years left before this whole job goes away. I'm paying $6,000 to $8,000 to learn to write QA code for the class at night. And I went, oh, buddy, don't do that. Don't do that. That's the job that's going away. Now, I mean, could you turn that into something else?
And he goes, he's sitting there, and he just says, well, what should I do? And I went, that is a very good question, man. And I had to go with, you know, what do you love to do in your free time? How do you turn your hobby into something? But literally that was when the ride was over, and I just said, it's just stuck with me. What do you tell people to do? What do you tell your kids to study? What do you major in college for? Do you go to college? I mean, these are things, you know, my kids don't seem to have plans to go to college, which I don't think is a bad thing.
I just, you know, I think they miss out on some of that critical thinking and socialization, that the college part, the collegiate part of meeting people is super critical. What they're teaching you, don't use AI to write this essay, is absolute crap, you know. So how do you separate the two? and so I think there are you know whatever these schools are popping up homeschools all that stuff I would say you know go I would advocate for that strongly I think that you don't want to have you don't want to be cookie cutter
in this neck yeah you don't want to be like everyone else because you're already in the training data well so are all your friends they're already in the training data like why do you exist well like talk about gradually then suddenly this has been around for like we don't have an attendance policy so I could a class and I just use Khan Academy to teach myself like oh really yeah I basically took all my Khan Sal Khan I graduated in 2013 so like Sal Khan covered everything I studied in college I was like all right I'll just teach teach myself and I were I wound up working during the day and but yeah I mean to
the point gradually and suddenly this has been possible this change has been emerging slowly and now it's hitting a critical tipping point where yeah it's it's it's here and to and I really want to dive into this like graph model versus LLM like sure 80% of the way there are you confident with that graph because again go back the LLMs that you can tell I was talking to Justin moon about this like the first thing he does whenever he tests out a new LLM model is tell me a joke and nine times out of ten the jokes very similar to the
previous model yeah and it's basically like it's using that distribution of okay this is what a good joke is and you end up in the middle of the bell curve and you get a similar joke every time um with the graph model it seems like it has way more yeah i mean there's graph pops up a lot so i mean i don't know if you guys want to dive into the details or not but it's i think i was practicing explaining this the other day but there's really um four maybe five pieces so should we go through yeah i think so um and you've seen this slide from i think 2022 is when i wrote this
slide so there's basically um people over in when chat gbt came out it was like six year olds playing soccer this model is going to do everything because you're so blown away by where you previously were versus what chat gbt can give you and so everyone was like these models just going to take over the world remember the thin wrapper thing don't be a thin wrapper because it's going to eat you know the model's going to eat everything no one talks about that anymore in fact i think open ai bought a workflow company late last year.
So if the model's going to eat everything, then why do you have all these extras? So in the corner, you do have the models, but I think that they are very good at interpreting human intent and then communicating back out the pieces of logic in words. So I look at it as like the input-output to these things. And you can extract a ton of logic from the language models because the sentences contain logic. So you have the models in one corner, and then you have the thing that everyone's excited about right now which is agents and loops
but any programmer will tell you is the more technical you are the more you look at this and you just go well what is an agent doing it's a loop with tools and language and you can do crazy stuff when this thing just has the right tools it can do crazy good it can do crazy bad but that's i think of agents as figure this problem out iterate and then get back to me that's what the agents are really good at. I look at them as like explorers or discoverers. I don't really know what the problem is. You go and do your thing and then tell me what happened. And those are agents are
incredible. So people call it agent harnesses and stuff like that too. So language models, agents, and then the piece that people overlook are the workflows. You actually don't want your agent figuring out how it did something yesterday every single day, right? I don't need it to go and figure out the way I look at it is if you didn't have Google Maps and you just had your cars driving around town with no map they will end up at the coffee shop one day but they're just driving and driving and driving once you know how to get to the coffee shop write it down create a list and
execute that list so the workflows help keep the agents on track so don't you know just go off the rails so when you know how to do something that's the workflow cron jobs and those are the kind of things schedulers you'll hear those words so that's that part of it and then run jobs is you would put that in the workflow bucket yeah that's basically triggering the workflow okay and just to just make sure I'm tracking so the agent and the driving around town analogy the agents loop is just go find a coffee shop and they're just sort of driving aimlessly whereas you give it the
workflow that's kind of like the instructions of how to get exactly to or the directions to get once you've figured something out got it so input output with the language models then you've got the agents doing discovery using tools and in language and then you've got your workflow that actually can do very complex long steps and you just when you know what you want you generally want it to go you put it into a workflow that's what you were doing with it yeah or that could be like a cloud skill or something could be anything yeah anything that's writing down steps skills can be tools though too i'm less familiar with how
cloud does the skills but then the last piece is memory right so um and memory right now is in the form of files so whenever you talk to your um claw bot or whatever you're calling these things is is just it's writing down these markdown files and it's reading them so the analogy people use is that movie um memento where the guy wakes up every morning and doesn't have any memories he has to read all of his tattoos. So every time the agent goes through the loop it's actually rereading the
files and figuring that out. The problem is is that you have the language model, the agent loop, and you have text memory. Those combinations results in the thing you were talking about before. The language model can only have the attention span of so many words. The agent is just spewing through words like crazy, writing down, reading, and writing down, reading. And then the more complicated it gets the longer it gets the model can no longer hear all the agent has to say so then what do you do you have to shrink or truncate or compact those are the words you'll
hear compassion um and it'll just arbitrarily take long sentences and turn them into short ones or crop them others many different strategies to do that that's when you start to see mistakes so now you have an agent with Alzheimer's because it's just missing chunks of his memory. So can you do files? What everyone's trying to figure out now is, are files really the best way to do this? Right now, they're incredible because the models and the agents
have really great file searching capabilities using rep and bash tools. So they're really finely tuned to do that. But it's almost like the tail wagging the dog. hey we're really good at text tools so let's use text for all of our memory it doesn't make sense so what we've explored for years is using graph databases so we started out with um micro task payments where we pay human beings to do small tasks using lightning right that was our company was founded on that so we were taking complex human processes breaking them down into
small pieces and then giving them to different human beings around the world and then reassembling the answer later which is exactly what you know these language models are good at and all these systems what you're describing is you're taking your process of of doing a podcast and breaking it up into small pieces giving them the different models and then reassembling it at the end we were doing that with humans from day one and so we were very well primed to then say oh we know how to break things up into workflows these skills humans are doing and some are automated and now just more and more is automated So we think that memory will end up in a graph structure
whether it's a graph database or not. And so that's what we've been working on forever. And again, you're demoing Hive to us in San Francisco, like comparing it to how people are using these harnesses, something like OpenClaw, like how would you describe the benefits of leveraging this graph database instead of all these text files? Yeah, the graph database can basically construct using a search the pieces of text you need instead of a file. The reason why we moved,
I mean, when you log into Airbnb, they're not reading a bunch of files to get the listings. They have a database, right? Because databases are way more efficient than text files. So somehow we've gone full circle back to pre-database land and going, we're now in the 70s going, oh, let's just use text files. So what do you miss with text files? Relationships, versioning, typing. It'd be like, I don't want an Excel file. I just want to put it all into a text file. You could, but you still lose some stuff.
So that's what we're trying to point out to people is the same things that drove people away from storing all of your information in text files 50 years ago apply to memory today. So you have to be able to query the pieces that are relevant. In the background, we do so much using workflows, janitor work to clean up the context so that it can be presented in a way. This is what Evan from our team has really excelled at, is pre-rendering the context in the way that the model can actually understand it.
So there's a lot of layers. So you have, say, every recipe you've ever cooked, all the ingredients in text files. But then you don't know how often you've cooked that recipe, right? This is a recipe I've cooked 800 times. Text files aren't great at that, right? Because they have the recipe. they have what you ate in a different text file let's say you're tracking your meals but those two things aren't linked together so how would you do that you you couldn't count and these things are trivial for databases to do but that's why your open claw um or your agent of any
kind might seem dumb because it can't keep those statistics yeah whereas any database and especially graph database can do all these things for you yeah does that make sense i mean it does and I um we have uh our agent has QMD yeah Toby uh Toby yeah from Toby running in his server to sort of solve some of that context not exactly it's a vector database or a rag yeah um I haven't loaded QMD but I was just gonna bring up rag um remember you guys invested in the company too
that does rag and so rag for people is just the the process everybody does rag all the time so So the term RAG is really retrieving stuff and feeding it to the model. The most popular way of doing RAG for years has been with a vector database. And no layperson touches vector databases, right? We've gone over this too. But it's basically turning text into points in a space, but you can't picture the space because it's not three-dimensional space, it's multi-dimensional space.
And the example I use is if you do back to the recipe analogy would be if you do a keyword search for the word avocado, guacamole will not come back. Those are two totally unrelated spellings. But a vector database, guacamole and avocado will be very close to each other in this space because they appear all the time together. Right? Yep. But when you take the word guacamole and take the word avocado,
one is from a farm bureau about avocado farming, and the other one is a recipe or menu in a Mexican restaurant, those two sources for those two words are completely unrelated. But the vector database cannot tell that those two things are unrelated. They'll just say they appear together very often. So if you use vector databases for RAG, you get these kind of random combinations. It's better than keyword search, but it actually orphans the word from the source of the word.
So I don't know if that helps or there's a bunch of different examples, but vector databases peaked like crazy about two years ago. And they took a precipitous fall because of this problem. It's orphaned information. um the best example i've heard and i'm borrowing this from someone else is if you do the berkshire hathaway annual report and you're searching for 2023 ebitda versus 2024 a vector database can easily mistake those two things because it's just one character difference and the embedding
might be very close to each other but those two numbers are very different meaning if you're trying to plug them into a stock picker yeah so that's the kind of thing that you get with vector and especially vector with code because code is so similar to each other if you search on um if you use vector databases purely for code search you cast almost like too wide of a net and you get a bunch of false positives so maybe that's too detailed no i think it's great that helped yeah but back to the four corners we've been living in this memory here so just to summarize you've got
text memory or you've got um database memory and then once you get into database memory how do you want to what kind of database do you want to use and vector databases are just we use all three so it's there's no right or wrong answer um vector is much faster but then when you want to know the ground truth the graph databases are really good for storing and keeping things that you've learned that you want to know for sure. I do care if it's 23 or 24 EBITDA.
Don't just randomly pick one. So that's where I feel like all roads are leading to graph for us. It's just the most convenient tool. So recap. Sure. LLM is like the input output. You've got the agents, which are the loops. They're out doing stuff. Discovery. Think of them like explorers. Explorers. That's why OpenClaw is so fun. You're like. just go crazy okay and then to direct them you got workflows yeah and then memory is the fourth
point yeah okay um i think there's like we can connect this back to bitcoin pretty soon uh if you you'll allow me so we're in the it feels to me like we're in this era right now where it's so fun everybody's trying their own thing like kind of isolated and independent tokens. Everybody's spending tokens to create their app or whatever it is. But we're redoing a lot of work. Like millions of people independently are spending tokens to do the same thing.
Now what if that thing that they spent tokens on, let's say they spent 100 tokens to do to create a transcript of the latest TFTC episode. You already do that. You spend your 100 tokens doing that you put it on an accessible graph with an l402 in front of it and charge 10 tokens to get that transcription that's where I think all of this independent work that we're all doing which is a lot of fun for everybody I think it starts to connect with each other that is an incredible
point because one of the things we'll do is we'll turn like if I listen to a podcast like hey this was a great podcast I think we should make Twitter aware aware of it and or acts aware of it and a lot of people don't want the clip they prefer to read it and so you just transmute the the content from audio video and ultimately a transcript into like a long yep a long tweet but we are going to YouTube sort of spending tokens to download the transcript
running it through our prompt but your point if you had a marketplace where it's like hey who already produced this transcripts I'll pay you 10 sats for it like a token arbitrage I think there's like this big thing that's this is this is this is very big yeah yeah it's big and I think I take a look at it yeah also yeah so this whole concept that and I think this is an exciting moment for just lightning micropayments all the stuff that we worked on for years mm-hmm
I can record for you, yeah. You want me to clap really loud? Just tell me when you can. Now. All right. So this is a graph database, and we have workflows. Well, first, we have a topic. So think of this as subreddit. And this also leads to the question of what will remain.
How do you make money in a post-AI world? It's not easy. I don't have a great answer for that question. So I think short answer is software goes to very cheap. Everyone's going to have software. Software engineers will perflate for a period, but five years, really? I mean, maybe not. So what do you end up with is marketplaces, and not just marketplaces, but marketplaces for agents. And so your agents, if you instruct it to save money, get my job done, but don't burn all my tokens.
So if you're downloading a transcript or you're downloading a podcast and transcribing and someone else is, you're just – everyone is wasting money and you're just burning electricity and tokens for no reason. So if there's a source, if you instruct your agent, go to this place, and they tend to have transcripts. Search and then get the transcript back, and it'll be cheaper than you doing it yourself, cheaper, faster, better. And so these content graphs can be built by agents or human beings.
And then what you can see here is you have – I'm using a shout out to Callie's, uh, claw interface. So I signed up for this. It's great. And you can see here, all I have to do is say search this graph or let's say. so clawy has an lsat that i gave it and it's using lightning to go out and retrieve data back
from the graph and paying 10 sats to get that data and then the person who uploaded the transcript you hopefully is getting paid some of that money back for providing that data so if you play this out and there it is coming back so this is everything the graph has on this thing so this is scraping all tweets podcasts um reddit you can put anything you want into the graph but in order to make money back from the graph by putting it up there you have to stake a little
money otherwise you get garbage so everyone is talking about how do you have an internet when content goes to zero the cost of content goes to zero you're just going to get spam right so um if creating profiles is free and creating content is free then how the heck is your agent going to find real information or not and so you'll have to have these wall gardens think of them as sub a protected subreddit it could be on k-pop on fly fishing on skiing on machine learning whatever
the topic is and then you have a community of people and agents building this shared memory together and then if you instruct your retrieval agent to just get it in the most efficient way then they'll say yeah it'll cost me 10 cents of tokens to get this or i can spend two cents here but then you if you upload tftc now you'll be making sats for years to come every time someone retrieves it you get it back this is what um i talked about with adam curry five years ago this
was and the way i told him i said hey there's going to be podcasting 2.0 but really that's a stop in the journey this is the final destination of the journey it's all content it's um gated so that the creators make their money without having to have advertising and this just wasn't going to happen I didn't believe with human beings so we built the first prototype of that podcasting 2.0 and streaming sass with you and I loved it this is you can feel how this could work but also human
beings are going to change their habits but now you don't have to convince human beings change their habits the agents figure it out on their own so people like us who aren't great at marketing myself i'm speaking about myself this could be the golden era you just have to build something better and the agents find it and then once you have this rolling and your graph becomes the most popular graph on that topic honestly it's like open it's like owning beach farm property if you keep the best graph going it's the most thorough high quality reputation
all that stuff the agents will keep coming back if you're down for a week agents are going to look for someone else this is the full information meritocracy but there's pay to put it up there but you earn it back it's like an investment and then you pay to retrieve you have to pay your way so this really is the value for value for the agent world anyway and but this is where like the human in the loop is like the curator at the tastemaker exactly right and that's going to be again going back to what i was saying earlier like understanding logic and like what people need and
thinking about humanity and the age of machines like what is our edge it's funny to think about i was one of the many listeners of that episode with you guys talking about podcasting 2.0 and being involved with the Bitcoin and Lightning community for a long time. We've been talking about machine-to-machine payments forever. But I think everybody, if we're being honest, when we were talking about it, we were thinking about a refrigerator is going to pay a Roomba or something like that. It was like an IoT use case. But really, what it's come to be is agent.
Machine-to-machine is agentic payments. And so I think this graph concept, what we're talking about here, for me it was really eye-opening to it just makes sense like these agents are going to be going around on the internet trying to get information or skills or things to complete a job they're going to be spending tokens and they're economically rational actors the agents so if it costs them 100 tokens to do the thing or they can get it for 10 tokens they're just going to go get it for 10 tokens but the as paul was saying the cool opportunity is for people to cultivate these
graphs and make sure their their graph is really good and has the highest quality content so the agents keep coming back to them um so yeah beachfront property is pretty if i were to do anything i've ever to say to my kids what to do i would say build these graphs on things that you know about my son's super into restoring toyota land cruisers and he was on uh claude yesterday or two days ago building a toyota maintenance app a web app i mean he's just going you can just build this and i said yeah but anyone can build that so what is actually useful collecting every
single tip and trick about how to do this building a community and then making it available to humans and making it available to agents if you have those two things and you have the lead this is the k-shaped thing if i try to start the second subreddit on open claw no one's going to come to in my subreddit because there already one and it already huge So there a real advantage to be first mover It kind of like TFTC is this for Bitcoin for humans
Right, like the humans who want to learn about Bitcoin come to TFTC to learn about it. What is the TFTC for agents? It's probably like this graph thing that's like easy to traverse, easy to do an economic trade to get the information. Like agents are probably not listening to ad reads, right? they're just like going to get the information. Yeah. So yeah, the information curation as like a job, I think it's gonna be very rewarding because people will focus on the things they're naturally interested in,
whether it's Land Cruisers or Bitcoin. Yeah. Hopefully nobody from Zero Hedge is listening to this, but I have like a Zero Hedge Pro subscription and with that you get access to all the banking reports that they have. You just get access to- Zero Hedge's agent is listening to this. it's like i've been like so i've been running cron jobs and all the banking reports that get uploaded to the zero hedge pro subscribers so like i've got insight like with gold and city like all the european banks are saying like every day and like i just wake up to a report from martin the sophisticated marty agent right that uh it's like what are the banks thinking today
he's like here's what they're all saying well people will do this with play it forward and i'm I'm saying that I would, but people will upload articles from behind a paywall, right? Same way a friend will print it to a PDF and send it to you. Agents will figure all this stuff out whether you want them to or not. Yeah. You know? So the next step, I think, so build graphs for agents with L402s in front. One shot, open claw, once I pasted the LSAT, knew what to do with a 402 response from the web server,
which is, do you have a LSAT? Do you have this string of characters? I paste it in, it goes. The next step will be, can it buy one on its own and all that stuff. I think there's actually a real business for selling L4O2s where you take Bitcoin and turn it into an L4O2 in case your wallet doesn't support it. But I think the big next step is your personal graph. So this is your shared machine learning graph or your fly fishing graph.
and think of it like a magazine. In the old days where you would buy a fly fishing magazine and read about what's going on, you know, the physical magazine. So this is like a collection point for all that information for enthusiasts on any topic, a business topic, personal topic, whatever it is. And then I get so sick of going to YouTube and to Twitter to try to keep up with anything I'm interested in. Surfing, we both surf. I want my personal graph to know what I'm interested in, give it a budget and just go get me the best surf clips from Instagram.
I don't want to go to Instagram and YouTube. I just, I want five minutes of great surf clips every day. And I want, you know, surf clips from where my friends live. So I see what they're, you know, that kind of thing. So you'll store that on your personal graph. And so your graph will be talking to your agent. We'll have your graph and talk to the shared graph and exchanging value back and forth. And then it'll know what you know. So applying this to say the example I use is like the Joe Rogan podcast because he's a ton of interest that I like.
I'm a bow hunter. I love bow hunting stories from Joe Rogan. There are certain ones he tells too many times. So my graph knows that I've heard this bow hunting story. so when i listen to joe rogan's podcast i can have it shrink that section of the story down to hey he's telling that same story again that you've already heard 50 times so i can just set to skip it just like i can skip ad reads right well so as you're describing this
i'm thinking because like you brought up surfing and uh i just want to get the pinch my salt clips i think uh yes i think those guys are hilarious funny it's like but i don't want to depend on the instagram algo to like surface it to me or me to have to go exactly find it via search uh pinch myself how's it funny like a top tier comedian is running a surf podcast yeah this is like so funny and it's uh but it's like to your point like you can get it and this is like another thing that's been a big topic is like how do you get away from the algos exactly and you just create your own by
curating what you want that's your that's some sort of personal agent and it's some sort of graph or memory i think of it as graph um is what people try to do with obsidian remember rome this is what you know really intense people want how to run their life but what if this were more passive and what if this were just as you're listening to things it's collecting what you know what you don't know um there's i think it's called math academy or something i don't know if
you saw yeah yeah yeah third grader doing calculus or speed running six years of math all by doing a skills graph where it knows the skill depends on this skill. So if you want to learn, there's a Google Maps for your brain. It's like a spiral vector, right? Exactly. That's a graph database. And so if you have my math, my math graph would be a tiny fraction of their math graph. But I have this node colored, and it knows by referencing the shared math graph
that the next step for me would be to learn this concept. So I have five minutes, and I've told my personal agent while I'm waiting for this plane, I want to learn a little math. Give me a math thing. Well, it knows that I'm sitting at a terminal, and so I can actually watch a short video. So it could construct a lesson using the video and content and just play me a video in the style that I like, the playback speed, the talking speed, the accent I want, male, female, whatever you want. just generate the content and feed it to me and the only way to get that is you have to have know
what you know your agent has to know and store that memory and then you have to have shared you know someone has figured out all these math skills i don't want to do that right it wouldn't make any sense and so those two interactions i think for the first time we're seeing this open claw phenomena really open up the idea that people want this control they want their version of this which i think is super encouraging yeah well bringing it back to machine payable web like i bought a 21 co computer back in 2015 uh when pelagi first launched it and we're talking about machine
payable web via on-chain bitcoin an idea that was great too early um brian you're on the board of lightning labs like obviously l402 has been around for years now and it's been funny with the emergence of OpenClone, everybody talking about the agentic economy. Like, oh, now we have X402 and all these other payment protocols for agents. And as we were talking about, I've had Cali, Justin Moon, Matt Corallo on in the last month talking about agentic payments specifically. And I think massive opportunity for Bitcoin,
but I think it's still unclear to me how this two-sided market emerges. Like, and how the, I think demand from Bitcoiners is there in terms of we have bitcoin we're willing to pay for this stuff with sats over the lightning network or whatever it may be ultimately the the path to make the payment um but i think that the chicken and egg is like how do we make sure like enough people have like payment gates that are leveraging this tech yeah a lot to say on this topic so i remember um when elizabeth and lalu first were
telling me about l402s i was like what are you talking about this was back in the refrigerator paying the Roomba machine to machine concept um but turns out they were very prescient on this point so you talked with matt in your last podcast about the concept of 402s we don't want to replay that but um having a system where you can trade value for resources where agents can trade value for resources is going to be a very important part to realize this future that we just described.
I think that there's still a lot of work that needs to happen to make it easier to get up and running. Like you said, the three of us were Bitcoiners. We've made transactions in the Lightning Network. We run nodes, etc. That's still pretty hard to go from zero to one right now. It's pretty complicated. I would encourage people in this ecosystem, and I know Lightning Labs is working on this right now, to make it as easy as possible for the next Vibe Coder to integrate L402
payments. So that's not just about putting an L402 gateway in front of some API. They need to be able to spin up a node. They need to be able to manage their channel balances and then liquidity. I still think there's quite a bit of friction there. I think that's getting solved very quickly, I'm excited for some announcements coming up pretty soon that are gonna lower those barriers. But I think if we can make it easier, there's a team called Money Dev Kit, Nick, who's those guys are doing a lot of good work on this front. But if we can just like lower the
barrier for developer to get going on this, I think well that's just gonna be a massive accelerant for all these things. And the vision that we're talking about with the these interconnected knowledge graphs that that's where we're going but to get there faster we need to make it easier to just get started in the first place yeah i mean my our clanker he's running a phoenix d server and it was i mean it it was fascinating also that process because to your point like the lightning node management stuff for some reason i had the ldk docks up on my on my screen and i was
all right take these docs and like see if you can spin up an ldk node and it completely like we banged our head for an hour and it completely um completely failed and i remember i was like all right let's think through this logically like instead of me giving it something while i have it do research like what would work best for your server setup and like what do you think you could actually do and then like went did some research came back it's like hey this phoenix d server looks looks good like i think i can give enough space on our disk like it makes sense it seems like it's easy enough to set up i was like all right go set it up and then set up it's like
okay it's it's here it's running but we don't have any bitcoin in it it's awesome you need to open up a channel um i was like okay like what's the best way to do that he's like oh i found this bolts tool where you can send on chain bitcoin then it will submarine swap and automatically open up a lightning channel with async this is the stuff we need yeah it's got to be like less yeah it was like all right all right get an on-chain address from bolts and tell me where to send it and i was like boom and then it was up and running yep this is where i i just pictured the cat opening cash app sending
bitcoin getting a channel yeah you know you asked us yesterday how did you get the inbound liquidity first question yeah so no one's gonna want to deal with any of that you're just gonna have to have it work your agent will have to say do you have this app do you have any bitcoin in it do you have any money go buy some now turn it into a channel yeah if you want to run one yourself i want to just like so my dream is to be able to let me back up a second so first some a few shout outs like shout out people like uh graham at voltage jesse at amboss the team at alby like a lot of people have
been working on the key bits of infrastructure to make this easier we just got to keep going on that front I can't wait for the moment where someone they're making like viral apps so they see something that happens on X and then they make an app to respond to that thing they saw so I have an example of this that I worked on which is do you guys remember seeing that that tweet of there's a speed reading video it's like you can read something really fast if your eye focuses on the center letter. So I was like, okay, how fast can I make an app to do this? And a bunch of people
did this too, but I wanted to like integrate L402 payments. So the website is called speedread.fit. So you can see it in the, in the wild, but I want to be able to take that idea and have like a production app ready with L402 payment gateways with the right liquidity and everything that's necessary in under an hour. I think that we're pretty close to that. It took me like a day to get it live because of some of the friction but I think if like if we can head in that direction where people have ideas and they can quickly convert
those ideas to cool things that people can use we're just gonna see an explosion of innovation yeah I wish we had two more hours I know you got to go soon yeah I think yeah that was great I think I guess just final thoughts on all this the where we are uh in terms of approaching the suddenly moment of ai like where this will go by the end of the year and how bitcoin payments actually become an integral part of this agentic economy i mean everything takes longer than i think so you know my timelines
are uh are always off but um it's everything that we planned on where speech and money are combined into one protocol and that's what we started with Sphinx and we continue to work on we pulled everything back knowing we needed to re-architect everything but speech and money become more important than ever because when content is infinite then you have to gate it somehow so people start to look for that I think that we'll have this concept of staking is going to come up
a ton micro payments even nick zabo i think recently retweeted something that was counter to his original criticism of micro payments where people will never dedicate their brain space to micro payments and now the era of you don't have to dedicate your brain space to micropayments allocate it to the age finally here right so lightning l402s staking marketplaces graph databases that act as shared agent memory you can see how this just speed runs
this is not stuff that we have to cook up every piece exists right now we're in the assembly phase and it's built on i think you know we all agree this is the best money out there it's global um and it doesn't put money into someone's pocket who made up that money right and so I think that the agents will be doing their owners a favor by operating on a Bitcoin standard. Yeah. I think especially for the listeners of this podcast who are Bitcoiners by default,
it's like, why put your energy into this? Well, you're advancing the core principles of Bitcoin if you help build, because now we're all kind of contributing to the ecosystem of Bitcoin. The more sats are flowing around the internet, the more Bitcoin is successful and the more we're all achieving the bigger picture goal that we care about as people who believe in the power of Bitcoin and the reason for its existence. So I think there's like this really neat underlying motivation that everybody in this community has, which is not tied to corporate profits. It's tied to the
mission of Satoshi, which makes it, I think when people are mission driven like that, like even if it's sort of deep down in the psyche, it makes it more interesting, more fun. People are more motivated to to see things through so you know like we've been saying it's a really fun time to be out there and I think everybody should be trying things trying to make you have an idea try to make it make it happen it's never been easier I need to set up a graph yeah we'll help you it'd be fun yeah and I'll
just reiterate what I said with with Matt and I think the one core advantage that all this Bitcoin tech has compared to other payments mechanisms is the interoperability and i think that it's not it doesn't mean it's a foregone conclusion it's going to win but i think it is a massive edge that people need to lean into more it's you don't have to the agents will find the pathways that work the best and bitcoin and lightning have the best pathways it's just purely factually correct you know and so um now's the time and bitcoiners like
myself aren't good at marketing so you don't have to agents will just find you so yeah we can't wait five years and let's not brian thank you for coming on let's harvest yeah thanks for having thank you thanks marty peace and love freaks
Théo Mogenet of Bull Bitcoin joins Marty to unpack The Cost of Lies, a manifesto comparing the fiat system to a Chernobyl reactor that was never supposed to explode. They walk through the debt trap math behind why...
Garrett Ziegler, the private investigator behind the Biden laptop report, returns to expose the least scrutinized branch of government. His new directory, thirdbranch.icu, profiles all 1,400 federal judges with...
James Check returns to explain Bitcoin's air pocket from $87K and why on-chain data points to seller exhaustion rather than a new leg down. He breaks down the base at the 200-week moving average, why $72K matters, and...
Pollster Mark Mitchell called the midterm blue wave months early, and he breaks down why the Trump administration still refuses to pivot. Mark explains how independents broke 20 points for Democrats, which pollsters got...
Everything's bigger in Texas. Marty and John Arnold break down the blue wave midterms, what a Democratic Texas governor could mean for the AI build out, and why Bitcoin may have several ways to win no matter how it...