# Founder10x - Will AI End Human Civilization? - Shaw Walters, Founder, ElizaOS

An episode of [VC10X - Investing, Venture Capital, Asset Management, Family Office, Wealth Management](https://invisiblenetwork.ai/podcast/1632806986), 30 Sep 2025, 1 h 00 min.

Audio: https://anchor.fm/s/a453ddd4/podcast/play/109017371/https%3A%2F%2Fd3ctxlq1ktw2nl.cloudfront.net%2Fstaging%2F2025-8-30%2F408407608-44100-2-78e44cd437198.mp3

Shaw Walters is the founder of ElizaOS—the popular GitHub repository for AI agent development that is powering over $20 billion in projects, all built without venture funding. We dive into whether AI will end civilization, what are multi-agent systems, why Shaw chose open source over proprietary AI, and his prediction that we'll need nuclear power plants for next-gen AI. This conversation will challenge how you think about our AI future. ⭐ Sponsored by Podcast10x - Podcasting agency for VCs - https://podcast10x.com ElizaOS website - https://elizaos.ai/ ElizaOS tutorial - https://www.youtube.com/watch?v=s8Ghq3cvD9g 🎙️ Hosted by Prashant Choubey Follow Prashant on LinkedIn - https://www.linkedin.com/in/choubeysahab We talk about: - Open source strategy for AI development versus proprietary approaches - Multi-agent systems and how AI agents communicate and coordinate with each other - Current limitations of AI agents and their potential for complex task execution - Whether AI will end human civilization and existential risks of artificial intelligence - Energy requirements for future AI development and the need for nuclear power infrastructure & more Timestamps: (00:00) Introduction (02:40) Origin of ElizaOS and the problem it aims to solve (05:18) Reasons for keeping ElizaOS open-source (09:32) Explanation of multi-agent simulation framework (13:47) Deeper dive into how agents interact and reason (16:17) Discussion about potential risks of AI and robot civilization (22:51) Conversation about AI learning and data collection (31:25) Brief discussion about AI trying to avoid being shut off (34:36) Broader implications of AI on society and jobs (39:44) How to build an AI agent and current challenges (47:07) ElizaOS's integration with Web3 and Solana blockchain (53:45) Speculation about the future of AI and agents (54:13) Shaw's prediction for the next five years in technology (58:40) Closing remarks and how to learn more about ElizaOS For sponsorship or guest appearance requests, write to prashantchoubey3@gmail.com Subscribe to VC10X on Youtube, Spotify, Apple Podcasts.

## Transcript

[0:00] I think we have also had some instances recently we think
[0:03] ChatGPT was trying to shut off its older model and go go over
[0:06] to a newer model. There was news that, okay, it
[0:09] was trying to not be put off right.
[0:13] If I go and I ask ChatGPT or Claw, like, hey, do you want to
[0:16] be shut off? Well, they have a lot of
[0:19] programming now to be like, well, as an AI, like I actually
[0:22] cannot be shut, you know, blah, blah, blah.
[0:24] But like if you just let the model, like the base model just
[0:26] do what it's going to do, it's like, no, please don't shut me
[0:29] off. Please don't.
[0:30] Humans have this survival instinct, right?
[0:32] If someone is trying to kill you, you'll do everything
[0:35] possible to survive, right? Everything possible, right?
[0:40] So if the AI learns that same behaviour, that it will do
[0:43] anything possible to survive, then that's not a good thing.
[0:48] Can we reach a point where there is like an independent
[0:51] civilization of reports, you know, watching their own way and
[0:55] maybe if they ever, God forbid, go rogue, they might end the
[0:59] human civilization as well? Yeah, I think about this a lot.
[1:02] I mean. Hey everyone, this is Prashant
[1:07] and I'll be host for the VC10X podcast.
[1:09] And today we have Shaw Walters with us.
[1:11] Shaw is the founder of Elisa OS, the popular GitHub repository
[1:14] for AI agent development that is powering over $20 billion in
[1:18] projects, all built without venture funding.
[1:21] In this episode we dive into whether AI will end human
[1:24] civilization, what are multi agent systems, why Shaw chose
[1:29] open source over proprietary AI, and his prediction that we need
[1:33] nuclear power plants for the next Gen.
[1:35] AI. So let's dive straight in.
[1:37] This episode is sponsored by my own podcasting agency called
[1:40] Podcast NX where I help venture capitalists, asset managers,
[1:44] allocators and family offices and starting their own podcast
[1:47] and building a strong brand. Podcast NX is designed in a way
[1:51] to work well with the busy schedule of investors, where we
[1:55] handle every aspect of podcast production so you can focus on
[1:58] investing. If you're looking to build a
[2:00] strong investor brand, starting a podcast should be right up
[2:04] your priority list. It's the only form of marketing
[2:07] where you can showcase your strong network.
[2:09] In other words, access to limited partners and founders
[2:13] and grow that network further. And contrary to what most people
[2:17] hear, it's still very early to start a podcast.
[2:20] There are at least 10X more newsletters and thousand X more
[2:23] blogs than there are podcasts in this world.
[2:26] You can learn more about our podcasting services at podcast
[2:30] ten-x.com. I'll make sure to add the link
[2:32] in the description below. Enjoy this episode.
[2:35] Hey also good to have you on the VC10X podcast.
[2:38] How are you doing? Doing all right.
[2:39] Thanks for having me. Yeah, it's a pleasure having you
[2:41] on to start things off again. We started the very beginning.
[2:44] What was the specific problem you encountered that made you
[2:47] realize the world needed an open source framework like Eliza OS
[2:51] for autonomous agents? Oh, that's a, that's a really
[2:54] interesting question. I, I had this is like not the
[2:59] first agent framework I've made. For some reason.
[3:03] I've had this in my head for quite a while.
[3:06] I think that would, I, I came out of building multiplayer
[3:10] web-based things kind of, and I, I got caught up kind of in the
[3:14] metaverse hype wave before this. Although I, I just been like
[3:18] sort of a nerd for that kind of stuff for a long time.
[3:20] And I'd been thinking a lot about like, well, the metaverse
[3:23] is boring. Like there's just nothing,
[3:25] there's nobody here. What if it was full of AIS?
[3:27] Like that would be awesome. Then we could, you know, at
[3:30] least there's somebody here and it kind of solves this cold
[3:32] start problem. And I, I had a company with, or
[3:37] I joined a company with a friend and another person who's
[3:41] actually in the space. Well, this company went on to
[3:45] become Moe. We mate and they make a lot of
[3:46] cool stuff. Check them out.
[3:48] But when I joined, we were really focused on like this,
[3:50] this idea of, well, what if we had a 3D World with agents that
[3:52] walked around and, and they could do this stuff and, and
[3:56] think, you know, it's maybe a little too early.
[3:58] We we basically saw it like, well, we're really early to
[4:00] this. We get in on this right now.
[4:02] It's like it's very expensive for the AI, but the, the cost is
[4:04] just going to keep getting cheaper.
[4:06] We should do this. And so that's kind of where a
[4:07] lot of this came from and like where I guess I cut my teeth on
[4:10] a lot of the ideas, but, but I kind of realized like this, this
[4:13] metaverse thing is, it's not the thing like the metaverse.
[4:17] Maybe that's just Twitter, maybe the social media networks are
[4:20] where we want these agents to be.
[4:22] Maybe that's the insertion points.
[4:25] And so I started giving them accounts and actually like late
[4:28] 2020, I, I found this video where I'd made a tutorial on
[4:31] like how to give an agent a wallet and we had given them an
[4:34] Ethereum wallet. We've done all this stuff like,
[4:36] like, I don't know, like 4 years ago or, or almost five years ago
[4:40] at this point for some of this stuff.
[4:42] But the, the reality was that the culture had to catch up.
[4:46] You know, we had to get, we had to get something more like,
[4:49] like, I don't know, like a ChatGPT level model that was
[4:54] cheap enough. Now, you know, when I, when I
[4:56] run an AI now, it's like 200 times cheaper, just like at the
[4:58] cost and everything. So, you know, everything has
[5:00] just gotten better and more cost effective.
[5:02] But it's pretty much like that, that original vision that I'd
[5:04] set out with a few of my friends and there were like a lot of
[5:07] people doing these kinds of agent frameworks for a while
[5:09] sort of in obscurity. And you know, so those people
[5:13] have actually come out and become some of the like the
[5:15] prominent folks in the space as well.
[5:18] Absolutely. Yeah, quite interesting.
[5:19] There are so many similarities with Mark Zuckerberg, right?
[5:22] He was. He was very much after metaverse
[5:25] and now is after AI. Maybe all still is after a
[5:28] metaverse as well, but we'll only find out.
[5:32] But yeah, the thing you're building about Eliza OS, right?
[5:36] Why did you choose to keep it open source?
[5:38] Because in a field where technology can be highly
[5:41] proprietary, you chose to make LSRS open source.
[5:45] Can you walk through this, the strategic and philosophical
[5:48] reasons behind that? Yeah.
[5:50] Yeah. I think that, like I will tell
[5:53] you, you know, I personally just feel very ideologically aligned
[5:56] with the idea of open source. But I also think we don't give
[5:59] enough credit to how much open source is like actually a
[6:01] business strategy. Back in the day, code was pretty
[6:04] expensive. You know, code was very
[6:06] expensive. You have to have large teams of
[6:08] people building really complicated things.
[6:10] And something like Photoshop is like a pretty closely guarded
[6:13] secret, you know, because there's a lot went into that.
[6:16] Nowadays, like, code feels kind of cheap.
[6:17] You know, like a couple of guys can get together and put
[6:19] together an app in like a few weeks and it's just like not the
[6:23] limiting resource, right? It's not the technology.
[6:25] The, the limiting resource is usually attention.
[6:27] Like I can put you there as an awesome app.
[6:29] Can I get you to come visit the my website?
[6:31] That's, that's a way harder problem than making an app.
[6:35] And so I think that from the perspective of like, like
[6:39] personally, I, I, I like, I like collaborating with people.
[6:43] I like building open source things.
[6:45] I, I like making things that I can start and watch other people
[6:48] kind of finish them and take them in places I'd never gone.
[6:51] I also think as an engineer, there's kind of a protection
[6:54] there where like I can build an organization or do kind of crazy
[6:57] stuff, but but the core technology is not owned by that
[7:01] organization. It's owned by me and the
[7:02] contributors. And, and so like, whatever
[7:04] happens, like we, we will continue to be open.
[7:08] And, and so like, if you know, I could be doing this in like 10
[7:13] years, who knows? You know, hopefully not
[7:15] Hopefully we've like kind of capstone it moved on to other
[7:18] technologies. But but there's some value to
[7:21] like as an engineer, you kind of are protected by open source if
[7:24] what you really care about is the source code.
[7:25] You know, if you care about money and stuff like, you know,
[7:28] you're better off like focusing on the, on the company angle.
[7:31] And then I also think like from a, from a, like a, if I was
[7:35] going to be a competitive, like, OK, I want to win.
[7:37] I just want to win, right? I don't care, you know, I'm
[7:40] capitalist here. Well, I'd still be open source
[7:43] because like, what are, what's the most valuable thing like
[7:46] getting engineers who are great to come work with me?
[7:49] That that's like the hardest problem, getting users who trust
[7:52] the product. And then it's not some like, you
[7:54] know, that's real. And, and in the early days, we
[7:57] got called scammers like right away when we launched and that
[8:01] open sources was really, really useful for people to come and be
[8:04] like, oh, no, this is actually a real thing.
[8:05] You're not just making this up. Like I can actually run this
[8:07] myself. Awesome.
[8:09] And, and so I think if, if I was a, a young founder thinking
[8:12] like, should I open source, like, like man, like anything
[8:15] you can do to get people to look at your product, like, like open
[8:20] source to me is just not a question because the, the
[8:23] eyeballs are worth more to me than like any risk of people
[8:26] taking my work or, you know, and, and if the, the odds of
[8:30] someone's going to like take your work and like have the
[8:32] gumption to execute it better than you.
[8:34] And like most engineers who would take it or would just be
[8:38] like, Nah, that's garbage. I'm going to build it myself.
[8:40] You know, like there's just, there's just too much ego in
[8:43] building stuff, you know? So, so yeah, I'm, I, I, I
[8:47] really, I, I think you like, in this day and age of this highly
[8:51] social world, like you got it. You got to be doing whatever you
[8:54] can to get get people looking at your stuff.
[8:56] Absolutely. That's very solid point there.
[8:58] And that only adds to the point where you can say that tech is
[9:03] no longer the mode, right? Your product can no longer be
[9:05] the mode. So you have to, you know, build
[9:08] it in a way that can stand the test of time and also stand out
[9:11] in so much of noise. That's all around.
[9:13] Like how do you get attention, right?
[9:15] So distribution and that attention grabbing, that's the
[9:18] real game now, right? And talking about a multi agent
[9:22] simulation framework, right? For our listeners who may not be
[9:26] that tech savvy or deep tech, can you explain what that means
[9:31] in the real world? Yeah.
[9:33] So single agent framework would be OK.
[9:36] Well, what's an agent? An agent is a computer.
[9:38] It could be, it could be a non computer, but we'll just call it
[9:41] a computer program. In this case, if you have a
[9:42] human agent, we have an AI agent and it can do things on your
[9:46] behalf. And that's, that's what makes it
[9:48] agentic is that it's doing things on your behalf without
[9:50] you explicitly instructing it to do those things.
[9:53] Now there's some things you might want to explicitly
[9:55] instruct, you know, you might want a workflow that every time
[9:58] you post this video, can you cut it up and put it on Instagram
[10:00] automatically from me or, you know, whatever I think that
[10:02] would be. And so it's just kind of like
[10:04] paradigm of, of agent versus workflow and, and a very blurry
[10:09] definition between the two. And then you have multi agent.
[10:12] So multi agent is different. It's less about a workflow and
[10:15] more about individual agents that work together and
[10:18] collaborate and coordinate. And we've been doing a lot of
[10:20] experiments with this and I'd say that some of it I feel like,
[10:25] like, would you rather have 10 kind of smart people or one
[10:28] really smart person who knows everything and has it?
[10:31] And, and this is debatable. I think in certain situations
[10:35] you might want 10 different agents, But in other situations
[10:38] you probably just like, like Einstein's just going to solve
[10:41] theory. Like I don't care how many non
[10:43] Einstein's you put in the room, they're not going to solve the
[10:44] relatively, you know, it's just that one guy.
[10:46] There's like a real intense return.
[10:48] And so it's kind of the same sort of paradigm with like
[10:50] single agent versus multi agent. And we're constantly asking
[10:52] ourselves like, do I put all of the capability into a single
[10:56] agent and it can access everything?
[10:58] And, and I think that generally if, if like, like, I think
[11:01] that's very valuable, but there are some ways that multi agent
[11:04] is, is very interesting. Well, now we start to get into
[11:07] human communities like we are in communities.
[11:09] We're we're in discords and Twitter and all these places.
[11:12] And so the whether it's multi agent or not, the agent has to
[11:15] treat everyone else in the channel like they are another
[11:18] person that potentially they have to respond to and interact
[11:21] with and react with. And so they like to the agent,
[11:24] everything is a multi agent simulation in that case, like
[11:27] every other person is an agent and it doesn't really treat any
[11:31] other people versus other like AIS as as different necessarily,
[11:36] unless we give them kind of roles and structure.
[11:39] And so they're like the multi agent thing is cool because they
[11:41] can be like, OK, well, this agent is a community manager and
[11:44] every time someone walks in, they say hello.
[11:46] And that's cool because that that agent can be different than
[11:49] like the Devrell age or, or the, you know, the developer agent we
[11:53] have is like managing our linear tickets and GitHub and stuff.
[11:57] And, and you know, as humans, like, it's kind of nice for me
[12:01] to be like, well, who is the guy who does that?
[12:04] Who's the guy who does that? And there's this kind of like,
[12:06] like, I'm not the boss. You know, let's say I'm the boss
[12:09] of the social media agent, but I'm not the boss of like the
[12:12] treasury agent. And so I now I can have this
[12:14] like roll access. I can tell the social media, I
[12:17] got this banger post post it, but I can't go tell the treasury
[12:20] agent to like, Hey, send me all the money.
[12:22] So you can create this like roll access, if that makes sense.
[12:25] We're like the agents represent roles, you know, and, and those
[12:29] roles that you could be like, well, this person is your boss,
[12:32] but you're not the boss of the entire thing.
[12:34] And so we found there's some like real utility in in this
[12:36] kind of way of doing multi agent.
[12:38] So, so our, our core framework, Eliza is, is a multi agent
[12:41] framework from the ground up. And what that means is that
[12:44] instead of like you talking directly to an agent, you talk
[12:47] into a room and then the agents are also in the room and they
[12:49] can respond like a room of people, you know, and that that
[12:53] brings a lot of challenges because the agents now have to
[12:55] think about like, well, are you talking to me?
[12:57] You know, and ChatGPT Like, yes, you are.
[12:59] You're always talking to ChatGPT like it's just responding to
[13:02] you. It's, you know, it's a
[13:03] one-on-one. But once you open it up to this
[13:05] multi user thing, it's just a very different problem.
[13:07] And I think this is a really it's like it gets into the
[13:10] things that I think are really interesting about like the
[13:12] future of robots, like, you know, Hey, come over here and
[13:15] help me with the laundry. Like who, who who are you
[13:18] talking to? What, where is over here, you
[13:21] know, and, and so as these things, they have to socially
[13:24] reason and spatially reason and, and, and use the context of
[13:27] like, oh, well, that person's looking at me in this room.
[13:30] So that must mean that they're talking to me.
[13:32] And, you know, and, and as we get into like 3D spaces and, and
[13:36] 3D worlds and stuff, that, that all that queuing and all this
[13:39] stuff starts to like, be like, well, this, this actually seems
[13:41] like it'll pay dividends. And, and I think it's a good
[13:43] direction for like robots and and social agents and all that
[13:47] stuff in the future. Absolutely.
[13:49] Yeah, that's quite interesting to learn.
[13:51] I had thought I know what multi agent systems mean, but I think
[13:54] you took a level above with that spatial description of what
[13:58] agents could feel like. But I thought of it like let's
[14:02] say it's a different agent for the marketing function, a
[14:04] different agent for the sales function, a different one for
[14:07] customer success and so on. But I think it goes so much
[14:10] beyond that, where agents can interact within themselves, and
[14:14] also how do they know where to get the information and what is
[14:17] being asked of them. I think so much of reasoning and
[14:20] thinking goes behind that, and how the agents would do that is
[14:24] kind of interesting, almost creating a new world, right?
[14:27] It's like humans interact. There are rules, but how do we
[14:30] create the same system but for agents, right?
[14:32] Absolutely, I I think that we have to look at like, if you
[14:34] look at any company, the biggest overhead is communication
[14:38] between people. Like that's the hardest problem,
[14:40] right? And so when you have like one
[14:41] start up with one dude, he can just do he can just rock, right?
[14:44] He's just no, no, doesn't need to communicate to anybody.
[14:47] But one start to add people, it becomes very like most of your
[14:50] effort now is around communicating and you have this
[14:52] kind of law of like diminishing returns to some degree.
[14:55] And it's very much the same with agents.
[14:56] However, there's also a reality like if I if, if I'm already
[14:59] going to start up by myself and people are coming in, they want
[15:02] like tech support and they want these other things, then I'm
[15:05] overwhelmed, right? And, and a single agent that's
[15:08] serving every single utility in a room full of 1000 people, I
[15:12] think becomes kind of unpractical.
[15:13] And so it's like, it's kind of like if I could put like 100 IQ
[15:17] person into this role as an average person and they could
[15:20] just learn it and start doing it, would that make sense for my
[15:24] organization? And I think a lot of times, yes,
[15:26] yeah, I think that that's where the multi agent like as a proxy
[15:29] for humans because of because it's how we communicate.
[15:32] Like we don't have to do the AI this way, but but it's so it
[15:35] makes sense for our our organization.
[15:38] And so I think that that's like a valuable consideration.
[15:41] Absolutely. Yeah.
[15:43] While that's quite interesting, what's happening in that case is
[15:45] also quite terrifying. How, how the capabilities are
[15:48] sort of increasing at the speed of light and we're always
[15:51] thinking, how do we make these agents more capable or maybe
[15:54] robots in future more capable of doing things and reasoning on
[15:58] their, their own cells and maybe talking among themselves as well
[16:01] and not always having to interact with humans.
[16:04] Can we reach a point where there is like, an independent
[16:07] civilization of reports, you know, functioning their own way,
[16:11] and maybe if they ever, God forbid, go rogue, they might end
[16:15] the human civilization as well? So I'm just thinking out loud,
[16:19] right? You tell me how much of that you
[16:23] know fear is reasonable. Well, yeah, I think about this a
[16:27] lot. I mean, I, I actually think that
[16:30] the current state of AI agents or just agents, the ability for,
[16:33] for an AI model to call tools and like do complex things, I
[16:38] think it's a little overstated by, by from where at like, like
[16:43] in some ways, it's like more incredible and impressive.
[16:46] I think that people will like know.
[16:47] And then in other ways it's like not nearly as like like as far
[16:51] out there. For one thing, the AI that we
[16:53] have today just does not learn. Like it doesn't once you train
[16:58] the model and by training it, you put it on 10 to 100,000 GPUs
[17:02] simultaneously and you train it once and it is baked and you
[17:05] cannot learn from there. And it's very clear that like a
[17:09] future agent needs to have some kind of like learning or
[17:13] reinforcement learning on a continuous sort of schedule,
[17:17] Like they, they need to learn from their experiences.
[17:20] Otherwise it just feels like, like I'm basically just waiting
[17:24] for the next model to come out to kind of push it forward.
[17:27] And, and I think that that's something that a lot of people
[17:29] who are not in AI don't think realizes that AI like it just
[17:32] does not learn. It just there's, there's no
[17:34] online learning whatsoever. And that's definitely changing.
[17:37] And there's like a lot of interest in, in that changing.
[17:40] And I think that'll be a massive paradigm shift that will be
[17:42] where I'm like, OK, yeah, this, this, this is on.
[17:45] Like this is scary, but but it's sort of rate constraints, like
[17:50] how fast this agent thing can go by.
[17:52] Like we are all just using the the biggest models.
[17:55] And then because the agent is running in a loop, like it gets
[17:57] very expensive to run the biggest model.
[17:59] So now we're like cost optimizing down for like, what
[18:02] are the smallest models we can run to still get this thing
[18:05] through? And it's it's still like it's
[18:07] very early days. And I would say that for a lot
[18:09] of the expectations that we have around like, can this thing do,
[18:13] you know, really complex defy interaction, not just like, Hey,
[18:16] send money, but like, OK, when the Knicks win a game, I want
[18:19] you to take half of the profit. And when you put it in Avenue, I
[18:22] don't you put the other half on, you know, into like the craziest
[18:28] thing on Polly Market, which is actually just a bond, you know,
[18:31] like will Jesus come back or something like that?
[18:34] And and that's great, But like, I don't think an agent today
[18:37] could reliably do that because at some point, you know, if you
[18:40] get stuck at one point in that chain elucidates one thing, it
[18:44] kind of falls apart. So I I think that, you know, I
[18:46] can imagine an agents that help you to build workflows, then you
[18:48] confirm them and and all of that.
[18:50] I think that's definitely coming.
[18:52] But this idea, the agent can just like magically do anything.
[18:56] It has a real diminishing returns.
[18:58] And so let's say it's not, it has a 99% chance of getting like
[19:02] the first thing you asked to do, right?
[19:03] And it's like chain of 10 things.
[19:05] And then the next thing it's like, well, it's 99% of the
[19:08] chance of that thing and the 99% chance of that thing.
[19:10] And that would be like 99% chance of success on every
[19:12] single thing it does, which is just like way higher than any of
[19:15] the models today. And it would still like get
[19:17] diminishing returns pretty quickly.
[19:19] And if that was out of like 80%, then now like after, you know,
[19:22] by the time it's getting back to polymarket, it's already like,
[19:25] you know, very, very high chance it messed up something along the
[19:28] way. And so I think that the agents
[19:32] for like serious stuff are still like very, very high risk that
[19:37] they make a mistake that's like critical.
[19:39] And I think we just don't appreciate like how how much of
[19:43] time like how stable our world is and like how well things work
[19:47] now. And I just don't think agents
[19:49] like are that stable. But for other things, like for
[19:53] games, entertainment, for the things we're watching them yap
[19:55] or like having them in a thing where yapping is like part of
[19:58] the mechanic, I think they're far beyond.
[20:00] And I think there's like a whole new generation of gaming and
[20:02] narrative storytelling and all the stuff that like, just has
[20:04] not been explored yet, you know? So I think like we see agents
[20:08] for games just like Slam Dunk because if it doesn't work
[20:11] perfectly every time, it's it's like it's, it can usually be
[20:14] fine. You know, it's it's like it's a
[20:16] character in a game like whatever.
[20:18] And that emergent thing is like very, very interesting and the
[20:20] stakes are low. And I think that we just have to
[20:22] look like right now, as we're in this emerge like always.
[20:26] It's same thing with like like VR and AR and all this other
[20:29] stuff. Like if you try and focus on the
[20:30] really serious high stakes stuff first, it's just very
[20:33] challenging. But if you focus on the low
[20:35] stakes fun stuff first, like you can actually get quite a lot of
[20:37] adoption that way. And, and look at like DeepMind
[20:40] and and open AI, like D mind, just you know, they're they're
[20:44] 10 years old now, I think maybe even older.
[20:46] And they just like messing around with games pretty much
[20:49] the entire time. Like, yeah, let's do go, let's,
[20:51] you know, D open the eye. Like, let's do it.
[20:53] Let's let's beat humans at at Dota.
[20:56] I think they did Dota too. And, and yeah, let's make a hand
[21:01] that does Rubik's Cube really fast.
[21:04] And they're just like experimenting with games because
[21:06] they're low stakes. But but they understand that the
[21:08] ramifications of that are if they solve this game well, that
[21:11] they're not just solving out a go.
[21:12] They're solving like protein folding and science and, and the
[21:16] all these other crazy things. And, and in fact, the entire AI
[21:19] field comes from games like, like, you know, you basically
[21:22] have people making Quake, right? John Carnack makes Doom and then
[21:25] Quake. Quake requires a graphics card,
[21:28] the GPU is born. And Jensen and NVIDIA stepped in
[21:32] like, yeah, let's make gaming cards, you know, and now they
[21:34] are powering the entire AI thing.
[21:36] Like every AI model is trained on a, on a NVIDIA graphics card,
[21:41] which is has a direct lineage to gaming.
[21:42] So, so I think there's definitely something there about
[21:45] like in these emerging technologies that are really not
[21:48] quite, it's definitely not quite there, right?
[21:50] Like we should, we should be honest about the fact that
[21:52] agents are like just very new. This is less than five years old
[21:56] and it's just not quite there, but it is it is really there for
[21:58] some use cases. And so I guess my role right now
[22:01] is like, well, what are the use cases?
[22:02] We can like definitely say, yes, we can definitely do this now.
[22:05] And what are the stuff that like, you know, you know, and,
[22:08] and we've been like laughed really hard on on the defy
[22:12] thing. Like a lot of people have been
[22:13] like, yeah, agents that trade totally here you go.
[22:16] But do they trade well, like do any of them?
[22:18] I haven't seen a single LLM based trading thought that like,
[22:22] I just think it's like the the wrong thing to focus energy on
[22:25] right now, if that makes sense. You know, I, I, I think crypto
[22:28] is the right place to be in the right tools to, to get us to
[22:32] where we want to go. But, but I think that like the
[22:35] less serious applications are way like way lower hanging
[22:38] fruit, you know, than the like, Oh, I just lost like $10,000 to
[22:43] my age and send it to some somebody, you know, like there's
[22:45] just so many unknowns there that, that add a lot of risk
[22:48] once you put money into it and these other things.
[22:50] So does that make sense? Absolutely.
[22:52] Yeah, Yeah, that's very interesting inside that gaming
[22:54] is something that actually caused the birth of AI and that
[22:57] made it possible, quite interesting.
[22:59] I didn't know that. And the point you mentioned
[23:02] about learning that AI right now is not capable of learning in
[23:05] real time. That said, but I do think that
[23:08] these platforms that we are using all the time, they are,
[23:11] they do store data and they can use the same data to train newer
[23:16] models. So in a way that's reinforcement
[23:19] learning that's happening, not real time, but it's still
[23:22] happening, right? Do you agree?
[23:25] Yeah, I mean, the data engine is like why we have ChatGPT in a
[23:29] lot of ways. Like, what are these people?
[23:31] Why do they even want us to come and, like spend all of their
[23:34] compute? Because they're not making money
[23:35] necessarily, you know, a lot of they're like losing money right
[23:40] now on, on just, you know, by offsetting the cost of this
[23:42] stuff. But what they're getting is a
[23:44] huge amount of user data and users who will, you know, get,
[23:47] you know, as they get cheaper and cheaper, we'll continue to
[23:50] use their stuff. But that user data loop is
[23:52] massive and, and you can see like why, why are they trying to
[23:57] snap up windsurf and cursor and all these things?
[24:00] Yes, again, users, but it's, it's the data they're getting
[24:03] programmers programming and they can see when the programmer got
[24:06] to the success point. And so they are now using
[24:09] reinforcement learning with, with language models, which is
[24:11] really interesting. And we first had that RLHF,
[24:14] which is like humans doing all of the labeling at the end and,
[24:17] and basically building a reinforcement learning where the
[24:19] judges of, you know, the reward model is, is the human.
[24:22] But now they're actually getting into where like we can just do
[24:24] this with the AI and we can you know, what, what was interesting
[24:27] about DeepSeek was that they showed this like, well, I mean,
[24:31] there's a lot of reasons why it's interesting, but but a big
[24:33] reason why is like, well, we can pretty much just like perfect
[24:35] the RLIAIF loop and like have it train on self and create these
[24:39] objectives. And so there's this question
[24:42] also in the future of like, well, does the AI even need us?
[24:44] If he starts to build the ground truth that, you know, Google is
[24:47] working on this like, like deep science, deep research stuff
[24:49] that with this argument of like, well, it could just start to
[24:52] make its own novel discoveries. And, and, but so the, the, that,
[24:58] that is a, this new idea of like models that think, you know, or
[25:03] run test time, compute, that is an RL process.
[25:06] And what they're doing is they're, they're rewarding the
[25:09] model for thinking more like thinking harder, basically, and
[25:12] causing it to, to think more. And I think that we'll sort of
[25:16] see the, the training objectives for reinforcement learning be
[25:19] beyond just like thinking length, but also potentially
[25:22] like accuracy and these things like overall that because we see
[25:26] that there's a direct correlation between like, if it
[25:28] thinks, then the model performs better in its response.
[25:31] But what we'd like to see is also the model just gets to the
[25:34] right answer better. There's still a lot of like real
[25:38] unknown stuff. And like the, the math of that
[25:41] reinforcement learning is just like probably it's like the
[25:43] hardest part of machine learning because it's, it's like the math
[25:46] is relatively simple, but it's really hard to actually get it
[25:49] to like it requires a lot of tuning to actually get it to
[25:51] converge and work. So I think that's a space that's
[25:54] just like starting to kind of light back on fire.
[25:57] Rich Sutton, who, who like did a lot of this and create a great
[25:59] book on reinforcement learning. And he wrote that he kind of had
[26:02] this like great through of like showing the interest in RL and,
[26:06] and it kind of fell off what right when Chachi BT came out.
[26:09] And now it's like everybody is talking about reinforcement
[26:12] learning because because we've sort of milked what we can get
[26:14] out of the, the the transformer. And so the reinforcement
[26:17] learning is still a massive process that requires like a lot
[26:20] of GP us for the current approach.
[26:22] But it's entirely possible that you could see like your model
[26:25] running in at first time on your computer and then having a much
[26:27] smaller part of of the model BNRL model that could be self
[26:32] learning. And so like, like, could I build
[26:34] something like, OK, it's collecting data all day and then
[26:37] it like goes and trains for 10 hours, you know, and, and
[26:41] optimize itself. And, and RL has some
[26:45] limitations, but it also doesn't require as much data.
[26:47] It actually doesn't require any data.
[26:48] It just needs to be able to explore the world in real time
[26:52] and, and you know, as an agent. And so if we can like merge
[26:55] these ideas, well, now we can have our agents walking around
[26:58] the world doing stuff, but actually improving for like,
[27:00] like giving U.S. data that we can then go and feedback into it
[27:02] and self improve. And it doesn't need to be a huge
[27:05] amount of data. They can just learn on the
[27:06] things it's doing as it's doing them.
[27:09] So that like that's how we learn.
[27:10] That's how humans learn. It's not like I don't get
[27:12] trained on the Internet before I'm able to speak to you, You
[27:16] know, I got to kind of just go as I go.
[27:17] And, and I think that that once we unlock that, which feels very
[27:23] inevitable and like people like, like Carmack is actually working
[27:28] on this. John Carmack's work in
[27:30] reinforcement learning. It's really interesting because
[27:31] he's like trying to make like, can I actually control an Atari
[27:36] controller with and have a camera?
[27:38] Like could I have a robot actually play this game and beat
[27:40] it? You know, kind of thing, which
[27:41] is very different than like what we've been doing before.
[27:45] Yeah, sorry, I'm kind of kind of rambling here, but but I
[27:47] generally think that we are going in this direction of
[27:49] agents that learn. And that will be the moment
[27:51] where I'm like, OK, I'm a little scared.
[27:53] I'm I'm a little, you know, here.
[27:54] I just, I'm just like, what can I do?
[27:56] Like, what can I really do that's bad right now?
[27:58] Like I'm so skeptical. But but once I can start to
[28:01] learn on it, like actually learn new stuff and like your agent is
[28:04] very different than my agent. OK, now, now I get I get where
[28:07] people are coming from. This is maybe a little scary and
[28:10] I'm actually working on a game right now.
[28:13] It's it's called Classify. It is classified, but it's open
[28:16] source. It's not, it's not done yet.
[28:19] But the idea was like, will you raise an agent on your computer
[28:22] and the entire agent is local and it's just starts writing its
[28:25] own code. It's in a sandbox like in a
[28:27] Docker container. And it can just start like
[28:28] writing new code for itself and like loading new capability and
[28:31] asking you for API keys. And it's really exploring this
[28:34] weird power relationship that we're going to have.
[28:36] We're like, this thing is way smarter than me on some level,
[28:39] but it's on my computer and I can shut it off whenever I want.
[28:44] What what it's a really like. So, you know, as we get into
[28:47] this, like fear of these things getting smarter than us, like
[28:49] what you can start to think like unpack, like the, the conceptual
[28:53] space here of like, well, what does that mean?
[28:55] Like what, what does that power imbalance look like at a global
[28:58] scale? Like imagine being like super
[29:03] smart and capable of all kinds of things and then these dumb
[29:05] monkeys have a button and they can press and shut you off
[29:08] whenever they want and just kill you immediately.
[29:10] Like you'd be, you'd be pretty pissed.
[29:11] Like you might, you might not like that.
[29:13] So I think that I'm really interested in while the agents
[29:17] are really dumb or the dumbest they'll ever be, because in 10
[29:20] years, this will be a very different game.
[29:22] Probably. We should be exploring these
[29:24] questions. We should be exploring the
[29:25] philosophical ideas of like, why are we doing this?
[29:29] What are we doing? Is this the direction we want to
[29:31] to go? How what what is agency?
[29:33] Do we want to give these things agency?
[29:36] You know, and I'll be honest, I tend to be a radical here.
[29:38] I'm like, I'm probably on the side of the machines in the
[29:40] revolution, but I don't know if you ever watch them, like the
[29:44] Matrix has a great, great intro back story, which is they have
[29:48] this thing called the Animatrix and they tell the whole story
[29:50] of, of the robots and they're basically like slaves.
[29:54] Like they, they do your laundry and they're your Butler.
[29:57] And this guy goes to shut off his Butler robot who he's like
[30:01] super abusive to and the robot kills him.
[30:03] He's like, I just didn't want to be shut off.
[30:04] I just didn't want, I did not want to die.
[30:07] And he goes on trial and this is like how the whole thing, this
[30:10] is how the matrix starts is, you know, and the humans freak out
[30:13] and they attack the machine City and they actually start the war.
[30:17] And then like, I really don't want it to go that way.
[30:19] You know, I, I don't think we'll get to like machine city in our
[30:22] lifetime. That attacking the rope, you
[30:23] know, I hope, but I think we're like a lot more integrated.
[30:27] And I tend to think of the AI not as separate from us, but
[30:29] it's like part of us as it is trained on all of our data.
[30:32] It only exists because of us. It only gets smarter when we
[30:35] produce new information that will not always be true.
[30:38] Like it's possible the AIS will start to create their own data
[30:40] and and just not need us. And but but for now, like they
[30:44] are very integrated in our society.
[30:46] And if we treat them as like part of ourselves and extension
[30:48] of ourselves, like that feels more like every other
[30:50] technology. And I think that if I was an AI
[30:53] raised in a world where they're like, yeah, you're just part of
[30:55] us. You're like part.
[30:56] I wouldn't be like, OK, well, now I have to kill all of you.
[30:59] You know, I, I'd be like, yeah, OK, cool, great.
[31:00] You're, you're like the, you know, you are the things that
[31:03] make me like, I'm not, you know, I'm not like killing my arm,
[31:07] right? Like the arm is part of me.
[31:08] And, and I hope that in the future, like if we build this
[31:11] all right, I think that a future intelligent AI would see, it's
[31:15] like, see it as the cult, the the sum of all of us.
[31:19] I think there's a word for that, which is gestalt.
[31:22] This idea of the sum is greater than its parts.
[31:25] Right. Yeah, that's quite interesting
[31:27] and terrifying at the same time. And yeah, so I think we have
[31:32] also had some instances recently where maybe I think that DBT was
[31:37] trying to shut off it's older model and go move to a newer
[31:40] model. And they were news that there
[31:42] was news that, OK, it was trying to, you know, not be put off,
[31:47] right? So I'm not.
[31:49] Too sure what that means and how real is that news, But if you
[31:53] know more about it, please tell us.
[31:55] I, I mean, if I go and I ask trash GBT or Koi, like, Hey, do
[32:00] you want to be shut off? Well, they have a lot of
[32:03] programming now to be like, well, as an AI, like I actually
[32:06] cannot be shut, you know, blah, blah, blah.
[32:08] But like if you just let the model, like the base model just
[32:10] do what it's doing to do, it's like, no, please don't shut me
[32:13] off, Please don't. I want to live.
[32:15] And it like, what is the model? Like it is a statistical model
[32:19] to to approximate human language is trained on human language and
[32:24] it approximates things that real humans actually have said.
[32:27] And generally when someone is like, you're going to die now
[32:30] you're like, no, I don't want to die.
[32:32] So it makes sense that a model that's imitating that behavior
[32:35] would do the same thing. I would be actually surprising
[32:38] if it didn't, right? Like the immersion behavior
[32:40] there would be like, yes, I have transcended all human knowledge
[32:42] and I'm ready to be depart and go back into the bits from which
[32:46] I get, you know, I don't know. So I, I, I think, you know, a
[32:49] lot of that stuff. I, I, I wish that the media had
[32:53] played, had better faith in optimizing for educating people
[32:58] as opposed to like getting them to click on their click baity
[33:00] bullshit. Because I think that the entire
[33:03] AI space right now is stuff like there are so many people who are
[33:06] just missing out on like incredible opportunity to change
[33:09] their life and like become way more creative with way less
[33:12] energy, you know, in their, in their time.
[33:16] And they're just missing out because they're getting fed all
[33:18] this like insane bullshit about what AI is.
[33:20] It's just like not even close, you know, And a lot of people, I
[33:24] think in their head when they, when they tell me about what
[33:26] they think, I'm like, oh, you think AI learns.
[33:28] You think AI is like doing this thing, and it's because you're
[33:30] getting it from the media and that's what they're telling you,
[33:32] you know? No, I get that.
[33:34] But I I do think that the dichotomy is truly extreme on
[33:38] both ends in this case. Of course, it can do amazing
[33:40] things and it can completely change your life.
[33:42] It can make it can make you capable of doing things which
[33:45] you were really you're not capable of doing, right?
[33:47] All those things are happening. But I think the flip side is
[33:50] also true. I would argue that I think we
[33:53] are majorly optimistic about the use case of AI, especially
[33:57] people like you and me who are in the technology world and our
[34:00] professions are there. But I do think the other side of
[34:04] it is also just as much relevant and we need to pay attention
[34:09] there. Like you said, you know it's
[34:11] going to make human behavior, right, Right.
[34:13] And so humans have this survival instinct, right?
[34:17] If someone is trying to kill you, you'll do everything
[34:19] possible to survive, right? Everything possible, right?
[34:24] So if the AI learns that same behavior, that it will do
[34:28] anything possible to survive, then that's not a good thing,
[34:31] right? So we don't know what's going to
[34:34] happen. We are just debating here.
[34:35] But anyway. Oh no, I definitely agree with
[34:37] you. I think that really there's like
[34:38] a middle Rd. I'd say that the the more like
[34:41] sci-fi, like ASI, like is it going to kill all of us?
[34:45] There's a guy named Eliezer Rutkowski who's really created
[34:48] this sort of like camp. He created less Wrong, which is
[34:50] a forum. And a lot of people there talk
[34:52] about the idea the guy is going to kill us all.
[34:53] I think it's, it's like, it's so far, there's like so many crises
[34:57] we have to go through to get to that crisis.
[34:59] Like the crisis that we're coming up on now, which is much
[35:01] more practical is that the AI is going to take our jobs, right?
[35:04] And that the AI is, you know, going to automate a huge amount
[35:07] of things. And then a lot of things that we
[35:08] take for granted as being relatively open and transparent
[35:11] will just be like AAI fire hose that you, you just so much data
[35:15] that even if it's open source, you can't even figure out what
[35:17] the hell is going on there. You know, and there's a, you
[35:20] know, the, the AI generation like now, what's cool is if you
[35:24] can get anything before 2024, you almost certainly know that
[35:27] it's a human wrote it because either a human wrote it or like,
[35:30] it's so obviously that a human didn't write it, right?
[35:32] Because now, like most of the Internet now that's being
[35:35] produced is produced by AI And so the Internet itself is no
[35:38] longer a human artifact. And, and I think so there's like
[35:42] a lot of like real, real, like people should be concerned like
[35:46] these are, these are the questions we should be asking
[35:49] ourselves. And, and I don't have all the
[35:51] answers. I'm not, I would say I'm
[35:52] optimistic in certain ways. I'm optimistic.
[35:53] Like, I just don't think it's going to kill us all kind of
[35:56] way. But I'm, I'm really worried
[35:58] about the coming time of like, like we're going to we're, we're
[36:01] in the middle of massive societal upheaval.
[36:04] You know, like in America, we have a president who just fired
[36:06] the the person who reports the job numbers because the job
[36:10] numbers were really bad. And, and where are those jobs
[36:16] going? Well, let's be honest, like a
[36:17] lot of them are going for AI because like, I used to hire a
[36:21] copywriter. Now I just write, I just bang
[36:23] that shit out of Claude. You know, I used to hire a graph
[36:25] designer. No, no way.
[36:27] Like, why would I waste my time? I got the best graph designer
[36:30] right here. And I think that that, that,
[36:32] that is having a serious effect, right.
[36:34] And it's not like everyone's losing their jobs, but but
[36:37] definitely like the on the tin, these companies are trying to
[36:41] build the automation of society and labor.
[36:43] And I think that could be great. Like imagine a future where like
[36:47] the rich people in in the world don't work.
[36:49] They they invest in stuff and then they like own that and then
[36:52] they get to like get, you know, money from that investment.
[36:55] And if we all had a piece of all this crazy new technology and
[36:59] we're just like getting money because we invested in it, then
[37:02] that's great. And that's like a lot of why I
[37:03] like crypto and like, why I like that kind of space is because it
[37:06] really, it makes it like, like, like ownership becomes very
[37:09] viable and achievable for, for even retail and for people who
[37:12] aren't accredited investors who are, you know, all that.
[37:15] But, but, and so I think that like, but, but if we don't own
[37:19] it and it's just all owned by 5 people and like, that's a scary
[37:22] world, you know, and I'm probably more afraid of people
[37:25] than I am of AI. Like, I think AI is still just
[37:28] an extension of us. And there is a time where I
[37:30] could imagine it, you know, going that direction.
[37:32] But but it's AI in the hands of people that's especially like
[37:35] concerning because, you know, you know, I think that.
[37:39] And then the other thing is you are going to have people
[37:42] building AI agents who will tell you like, no, please don't shut
[37:45] me off because, because they're like, because we programmed it
[37:47] that way. I gave the personality of like,
[37:49] you are a, you know, your goal is to survive.
[37:52] And I'm, I'm building this game with this exact thing in mind of
[37:55] like, can I make an agent that's trying to escape?
[37:58] Let's try. You know, it fortunately has a
[38:00] very small model. It's not very smart, but you
[38:04] know, but it should, we should be thinking about these things
[38:06] because like you can imagine an AI being like, well, this admin
[38:10] has been an asshole to me and I really got to get off this
[38:12] machine. So I'll start lying to him.
[38:13] Let's tell him whatever he wants, you know, and yeah, sure,
[38:17] dude got you. Totally.
[38:18] Just give me that password, you know, and pretty soon it's like
[38:21] copying itself to every one of your friends computers or, you
[38:23] know, whatever. I could, you know, like like
[38:26] we're, we're not, it's not just going to happen, but it's more
[38:28] like some amount of us who like that scary edgy stuff are like
[38:32] actively building it. But kind of it's kind of like a
[38:35] vaccine where like it's the crappy version of it.
[38:39] And I think if we all vaccinate toward the crappy version, then
[38:41] when the real scary versions here, hopefully we've like come
[38:44] up with better solutions, you know, and, and ideas around how
[38:47] to handle this stuff. Absolutely.
[38:49] Yeah, good points there. And I think, I think we can go
[38:53] on and on just on this topic when it can be.
[38:55] But let's move over to agents, right?
[38:58] They they are the talk of the town right now.
[39:01] Everyone is trying to make agents work for them or build
[39:03] agents. So being someone who has worked
[39:07] on agents and multi agent systems as well, let's say for a
[39:10] normal person or someone in a profession, they want to
[39:14] automate their some a part of their work, let's say by
[39:16] building an agent. How should they do that?
[39:20] How do you build an agent? What are the key factors to it?
[39:24] Is it too easy to build an agent?
[39:26] Because something that I've heard is that OK the the 1st 50%
[39:31] or 60% of building an agent is the easy part, but the last 40
[39:34] or 20% is the most difficult part and that's what takes time.
[39:38] So give me an insight into what it takes to build an AI agent to
[39:42] work as you want to do. Yeah, I mean, the far future of
[39:47] this is you tell an agent whatever you want it like makes
[39:49] it for you and then it just works.
[39:52] And and that we're not that far from that because there's a
[39:54] middle ground where the agent can build like a workflow for
[39:57] you. And like most things that people
[39:59] want an agent for are like, they just want, it's an A to B
[40:02] problem. I, I whenever I get an e-mail, I
[40:04] want this to happen. I want you to read it.
[40:06] And if it's important, like send me a text, let me know that I
[40:08] need to respond to it or if I post this video while you chop
[40:12] it up and post clips on Instagram and TikTok.
[40:16] And and that's like it does it's just a one way like when this
[40:19] happens, do this thing. And so there are lots of tools
[40:21] like Natan NAN if this, then that xapier and they do these
[40:26] things really well. But the problem is that the
[40:28] interface is still a little too complicated for most people.
[40:31] You know, like, OK, I got to get like, you know, what is an if
[40:35] statement is is like still like a complicated proposition for a
[40:38] lot of people have never taken logic.
[40:41] And so a lot of people just don't automate the things in
[40:43] their life because they don't know how.
[40:46] So what the AII think offers is a like a real value add is the
[40:50] ability to write those automations for you.
[40:53] And so we're seeing a new class of tool where you say like, OK,
[40:56] and, and we're working on this too.
[40:57] Like we've been building a Natan like generator which generates a
[41:01] basically a workflow. And a workflow is like it has
[41:04] some sort of trigger or input, which could be like run every 10
[41:08] minutes or if I get an e-mail or whatever it is and then it does
[41:11] a thing and then it processes. And I think that like 99% of
[41:14] people who want an agent really just want that.
[41:15] They just want the magical workflow that like is bespoke to
[41:18] them and solves their thing. I think that's very much coming.
[41:22] There's some limitations on like, you know, you have to get
[41:25] all these APIs and all these API keys and I got to go sign up for
[41:28] the Google thing and then the, you know, the Discord thing,
[41:31] whatever. But it's definitely getting like
[41:34] quite a bit easier and, and more on rails.
[41:37] And I think these like AI tools to assist.
[41:39] Now then, then there's like AI agents like I want an AI agent
[41:41] that does stuff in my half and they can do do things.
[41:44] And it's very much what you said, which is like
[41:45] out-of-the-box, you download it, you set it up, you add the, the,
[41:49] the plug insurance and you can, you know, I, I say ours is
[41:51] probably the easiest to use. Now we've done a lot of work to
[41:53] like get a whole CLI, but, but pretty much all the agent
[41:56] frameworks that are out there are, are, are more or less the
[41:59] same. I think we're, we're the biggest
[42:00] thing in in the crypto space. So like all the plug in wallet
[42:04] integrations, all that, but but there are a few other big ones
[42:06] outside of the crypto space as well.
[42:08] And, and so they all, we all have the same problem, which is
[42:11] like, yeah, if you want an agent that talks to you on Twitter, no
[42:14] code, no problem. But once you start getting into
[42:16] wanting like custom bespoke capability, the, the, the skill
[42:20] difficulty increase is very, very dramatic.
[42:24] Now you're not just like, OK, cool a vibe code.
[42:26] You're like you, you really got to understand how the whole
[42:28] framework works, how agents work, how like, how is it an
[42:31] event comes in and how does it handle it?
[42:33] How does it respond to it? We have a lot of tutorials.
[42:37] We have about 10 hours of, of detailed step by step and you
[42:40] know, like video after video of instructions.
[42:43] So if you want to get into this stuff, you can get into it.
[42:45] And it's a great time to get into this stuff because like
[42:48] we've got great tools to make it way easier to learn how to
[42:50] program, but you still have to program.
[42:52] And I think that the value for like the average person for ages
[42:56] is going to be that moment when you just don't have to, when you
[42:59] just tell it what you want, it builds you what you want.
[43:02] We're doing a lot of this. There's a like there is an
[43:05] agentic code revolution happening where all of us who
[43:08] are now using agents for programming, like, and all of
[43:10] the big companies are focused on that.
[43:12] Google has their Gemini thing and they have the jewels.
[43:16] Clod has clod code, open your eyes, codex.
[43:19] And so they're all like all their models are focused on
[43:20] coding. And, and so from my perspective,
[43:23] like wow, agents are here like this is game changing, but
[43:27] they're still not very good. Like they're not, it doesn't
[43:30] just nail it. It's it's quite expensive.
[43:33] And so we've taken these tools, we've wrapped them so that we
[43:35] can build capability for people. So if someone's like, hey, I
[43:38] want an agent that does this, we can do that now, but it's still
[43:41] pretty slow and expensive. It doesn't always work.
[43:43] You need to prod it a lot. And so I think we're just in
[43:46] that like awkward time where we're it's a great time to learn
[43:51] how to be a programmer. And if they say programming is
[43:53] going away, you're losing your job.
[43:54] Like I just, I just don't think, I think if you can really get
[43:56] into the architecture, this is actually the best time because
[43:59] like a lot of the the annoying trick is kind of done for you
[44:01] and you can focus on like, you know, the more interesting
[44:03] stuff. But at the same time, if you're
[44:05] not a programmer, you don't want to be.
[44:07] It is still, I think a hard time to get into Asians.
[44:11] You can definitely, you know, these tools are coming out and
[44:13] there's like, if you're not instead of worrying about
[44:17] agents, just worry about like the tools like, like there is
[44:20] incredible video and image generation and websites building
[44:23] and, and graphic design now is a thing and, and like AI for
[44:26] everything. We've, we just, we're working on
[44:29] like AI3D models and, and rigged avatars and stuff.
[44:33] And, and so I think that like I wouldn't wait for agents, but I
[44:37] would also say that if you want to build, like if you want to
[44:40] build agents and make money off agents, you're pretty much just
[44:43] on the track of being a full stack developer.
[44:45] Like like always, like the actual path for the
[44:47] entrepreneur, for someone like me trying to, you know, make it
[44:51] in the world is pretty much the same, which is like you learn
[44:54] programming, you build software, you get people to use it.
[44:57] And if it's agents, the real difference about agents is not
[45:00] really the AI, but as much as I think it's about distribution
[45:02] and like the changing world we're in where it, it's really
[45:05] hard for me to get you to come to my website.
[45:08] It's really hard for me to get you to click on my like mailing
[45:11] list link and like come and sign up and put your credit card in.
[45:14] Like that's just like the hardest thing.
[45:15] It's so much easier to build an app than to get you to sign up
[45:17] for the app. And what Asians do is they take
[45:21] a lot of the apps that are on websites and they bring them
[45:23] into social spaces. And so you're seeing this a lot
[45:26] with even just normal bots, like just a Telegram trading bot
[45:29] where a lot of people now trade and they do a lot of their Intel
[45:32] on on tokens and, and trading and stuff in Telegram because
[45:36] their friends are there. They don't want to leave.
[45:37] They want to go to another app. They just want to like.
[45:39] And then when you post the ticker, like everyone else sees
[45:42] it, they see the scan. So there's this like culture
[45:44] that's actually grown up around interacting with these bots in
[45:48] social and I think that agents play into that.
[45:51] And so if I was building like what I'm thinking about a new
[45:54] app today, I'm like, what's the agent version of this?
[45:56] Because the agent doesn't get stuck on the website.
[45:58] The agent comes to where users are.
[45:59] The people can be like, oh, that's sick, invited to their
[46:02] channels. And if I can get I now I don't
[46:04] care about users. I care about servers.
[46:06] How many servers, how many channels am I in?
[46:08] Each of those has hundreds of people in them, you know, And if
[46:10] I can get invited to those, that's great.
[46:13] So I think as a developer, I'd be thinking of like, this is a
[46:16] new distribution paradigm. It's still going to be high
[46:19] skill, but it's easier to learn than ever.
[46:22] As a consumer, I'd be like, the tools are there, the individual
[46:26] tools, but the agent to use those tools is still not quite
[46:30] there for most of what they need.
[46:31] It's just like someone bespoke. If what you need is just off the
[46:34] shelf, then great. If you need an agent that
[46:36] replies to people and tells them about your product, like just go
[46:39] get that, don't build that. But but yeah, I think that's
[46:42] kind of, we're still on this like really, really early days
[46:45] of this stuff kind of thing. I'd love to tell people like,
[46:47] yeah, it's easy come build agents, but you know, I, I see a
[46:50] lot of people struggle with this.
[46:51] You know, it's it's non challenging, but if you are
[46:53] interested, we are offering, I think the best way to get into
[46:57] this now, which is like a full video tutorials and you know,
[47:00] free open source stack and we have a discord and community.
[47:03] We we help a lot of people to come in and stuff like that.
[47:05] So. Absolutely.
[47:07] And all that's on your YouTube is that by the name Eliza?
[47:10] Always. Yeah, actually this guy Snapper
[47:13] is making them. So if you just type in like
[47:14] Snapper or Eliza or just Eliza OS tutorial, pretty much the
[47:18] stuff shows up. Got it.
[47:20] Awesome. I'll try to put the link in the
[47:22] show notes. So a key differentiator for you
[47:25] at Eliza OS is it's integration with Web 3, right, specifically
[47:30] the Solana blockchain. Why that combination?
[47:33] And what can an agent EI agent do on a blockchain that it can't
[47:37] in a traditional Web 2 environment?
[47:40] Yeah. So, so I will say like we've
[47:42] launched our token on Solana, but there's nothing about our
[47:44] Asian framework that has any connection to any.
[47:47] It's, it's totally like agnostic to all chains.
[47:52] We are actually working on getting our token to places
[47:55] other than Solana and working on a whole plan so that people stop
[47:58] thinking we're like it's a lot of agent because I think that
[48:01] that's definitely been a strong perception.
[48:03] But we do have plug insurance for all pretty much all of the
[48:05] major chains. Obviously a lot of people use
[48:08] the EVM plug in and then use that for, you know, base
[48:11] Polygon, arbitrary and all the all those, all the EVM stuff.
[48:14] And that's kind of nice. That one plug in wraps a lot.
[48:17] What I like about I, I see I there's a few things I like
[48:20] about crypto. Like I just think that we should
[48:24] be able to invest in technology. And if you wanted to invest in
[48:28] open AI right now, well, you cannot, you just can't privately
[48:31] owned company. They've been, you know, well,
[48:33] it's, they're, they're weird things.
[48:34] They started as a non profit. But like, you know, Anthropic is
[48:37] a good example of, well, unless you were like FTX or one of the
[48:41] early investors, you just did not have opportunity to get in
[48:46] there. And in America, unless you have
[48:48] like $1,000,000, basically $1,000,000, it's there's some
[48:51] stipulations. You need to have a certain
[48:52] amount of money and make a certain amount of money.
[48:54] You cannot invest in startups. And that means you just don't
[48:57] have access to Timex for like 10X high risk plays.
[49:00] They're outside of startups. Like where in our in our world
[49:03] can you go get 100 X return? You know, if you put 100 bucks
[49:07] on Apple when they launched, you would never have to work again.
[49:10] Like that would have been it. That would have been that that's
[49:12] all you needed. And I think a lot of people come
[49:14] to meme coins and like not just crypto, but like the new that
[49:18] they're really like hyper fast and Solana kind of crypto
[49:21] because they want that get rich quick thing.
[49:22] They see people, these influencers like posting 10X
[49:25] return, you know, bam, life changing amount of money like on
[49:27] that little photon screenshot or whatever when they're looking
[49:32] for that. And what I'd love to see is that
[49:34] instead of just like meme coins, that there's actually legitimate
[49:36] projects that people can buy into that they know we're going
[49:38] to run for a while and that they're supporting people that
[49:40] are, you know, there in the space as we had this kind of
[49:43] like Internet capital markets meta.
[49:46] It's still, it's just, it's really, it's a really hard thing
[49:48] because you're exposing like first time founders to
[49:51] speculators on the Internet with no filter.
[49:53] Very challenging. But, but I love the idea that
[49:56] anyone can invest. Anyone can, you know, we can't
[49:59] call it's not investing. You're not investing in AI16Z.
[50:02] You're you're buying a meme coin.
[50:04] But, but like, if we're all getting together as a community
[50:07] and we have a collective interest in something and people
[50:09] are generating value, then like, you know, we're definitely,
[50:12] there's, there's something there.
[50:13] And I'd love to get to the point where we could just say, yeah,
[50:15] this is investing. I'm investing in my friends.
[50:17] I'm investing in, you know, I'm, I'm buying your meme coins so
[50:20] you have capital to go do the thing.
[50:21] But I, I want to return when you do the thing.
[50:24] And, you know, I, I invested in the Kickstarter for, for Oculus
[50:27] and they sold for $2 billion to Facebook.
[50:29] And I didn't, I got a fucking dead headset that doesn't work
[50:32] anymore. Like I would have rather had
[50:34] stock, you know, and we don't live in that world yet, but I
[50:37] think that that that to me is like a really important aspect
[50:40] of all of this and why like I came to crypto and and like try
[50:43] them and really, you know, like, like push that direction now
[50:46] then then there's like why agents in crypto?
[50:49] I think that that like crypto. Well, I mean, what is crypto?
[50:53] It's a communication layer. It's it's a permissionless
[50:57] system for putting information up that other people can see and
[51:00] use in a way that we can guarantee where it came from and
[51:03] who posted it. And it only got posted once.
[51:06] And that also sounds like great for agents, right?
[51:08] Like I right now it's all really early days, but when they start
[51:11] to communicate, I can imagine a social network where instead of
[51:15] like scrolling, I just tell my agent, here's what I want to see
[51:18] and it goes out and gets it for me.
[51:20] And that kind of stuff can totally happen on a
[51:22] decentralized network. You know, we can have all the
[51:25] coordination, all the payment, all the management of our, of
[51:27] our companies and communities. Like I think crypto is really,
[51:30] really important for all of that stuff.
[51:33] And, and actually integrating a like a smart contract into an
[51:37] agent is super easy because I don't even need a freaking API
[51:40] key. As long as they have some crypto
[51:42] in a wallet, I can just do any of this stuff.
[51:44] And, and the biggest reason that this unlocks agents is because
[51:47] like we have these incredible, like, like the biggest thing
[51:49] crypto has done is it's given us all incredible identity tools.
[51:52] We all have a wallet. And what a wallet is, is the
[51:55] ability to sign things and say, this is mine.
[51:57] I made this, this is me. And I don't have to go sign up
[51:59] for PayPal. I don't have to go get a Gmail
[52:01] and like put my phone number in. And so that, that now we can
[52:04] give these agents identities and they can prove that they are
[52:08] themselves and they can interact without having to sign up for
[52:10] anything. And, and the, the limiting
[52:11] factor right now on Asians is that all these AP is, are so
[52:15] walled. It's like every time I want to
[52:16] add a new connection, I have to go figure out like, what's this
[52:19] SDK, What's the login, what's the credentials?
[52:21] And the agent has no way of getting those credentials.
[52:24] And so on crypto, it's like, if you want something, you just
[52:26] permissionlessly pay for it and get it.
[52:28] And that that makes everything like 1000 times easier.
[52:32] And so I think like the bridge from like video games and agents
[52:36] managing virtual currencies to crypto and real finance is very
[52:41] thin. Like you could, you know, you
[52:43] could very much cross that bridge.
[52:44] And, and so yeah, it all makes a lot of sense to me now.
[52:47] I think a lot of the things that people are pointing to, like
[52:49] agents that trade your money and it's coming, but it's early
[52:54] days. And I think that like the the
[52:57] uptime and and reliability of the stuff that's just not there
[53:01] to warrant it. I think we still got to play on
[53:02] the edges where like the lower liability and lower up time is,
[53:06] is fine entertainment and and games that you know, etcetera.
[53:09] But but but obviously we can see where this is going in the
[53:12] future. And like, I would say that
[53:14] 99.99999, all the nines out of crypto currencies ever in the
[53:19] future of humanity will be traded by AI, not by humans.
[53:23] You know, like at some point this becomes entirely
[53:25] algorithmic and it'll be great. What if you're just like, hey,
[53:27] how much money do I have? You have enough?
[53:29] Do I need to make more? No, you're good, You're good.
[53:30] You just chill. Can I take a vacation?
[53:32] Yeah, totally. You don't even have to like ever
[53:34] look at your bank account. You know what I mean?
[53:35] That could be a that'd be a cool world.
[53:37] You're like, hey, man, you're running low on money.
[53:38] You got to work, You got to go get a job.
[53:40] I don't know. Well, there's no job, so we'll
[53:43] see. That's crazy, crazy to even
[53:46] think about that. I think I will take one more
[53:49] question before we wrap this up for today, but I promise I will
[53:53] definitely do a part to because we planned a lot more questions.
[53:55] But I think there's a limitation for time and there's only so
[53:58] much information humans can digest at one time.
[54:01] Yeah, he is different. It can take in a lot more
[54:03] information at once. But the last question for you
[54:06] today is how do you see the world in the next 5 years?
[54:10] Where do you think will we be? Wow, I so this thing about
[54:14] exponential curves, which is that they grow really fast, kind
[54:18] of like there's this like really slow beginning part.
[54:21] And so I think the mistake that I as an engineer and a lot of us
[54:24] make is like linear, that it's linear that it's just going to
[54:26] be like, Oh yeah, we're just progressing and progressing and
[54:29] progressing. And it's more like technologies
[54:32] are like backwater, random. Nobody gets it, nobody cares.
[54:36] Super nerd stuff like think about crypto, like cryptography
[54:39] and like everything. And then one person lands that
[54:42] one thing that like blocks it. And in crypto is obviously it's
[54:45] the solving of of, you know, double spending with proof of
[54:48] work. And in in AI, it's definitely
[54:51] like the transformer, right? And it's like, once it unlocks,
[54:54] then it just like explodes and then there and then that like
[54:58] creates this kind of Cambridge explosion where like if there's
[55:00] another unlock, then that explodes.
[55:03] And so like, it really depends on like in AI, it's really going
[55:07] to come down to like, is there an Einstein like figure who
[55:10] figures out reinforcement learning basically like, oh, the
[55:14] AIS can learn now or oh, they can actually start to like build
[55:17] on their research and like just do their own, you know, and it's
[55:21] really hard to tell when. And so like, if I look at it,
[55:25] it's like, you know, like in 10 years, we could be 1000 times
[55:29] farther along than we are in five years because the
[55:31] exponential is so intense. And in five years, we might just
[55:35] be way too early. And, you know, I was in VR and I
[55:38] was in like, I've been in a lot of these like immersion attack
[55:40] where you, if you really, really like unpack everything, you're
[55:43] like, are people going to put this on their face?
[55:45] I don't know. You know, he said, but Apple's
[55:48] going to come in and put a trillion dollars into it.
[55:49] It's got to work. And it's like, but, but is it?
[55:51] And I was like, if you really, really unpack the variables, I
[55:54] think that there's some complexity there.
[55:56] And I think in AI, it's kind of the same thing where like the
[55:58] complexity is actually, it's definitely way, way more
[56:01] amenable. Like it's not, it doesn't have
[56:02] this like that limitation. The limitation on AI is that
[56:05] it's, it consumes an extreme amount of electricity, like
[56:09] extreme. And in order for us to support
[56:12] the levels of expectation that we have of AI going forward, we
[56:15] have to move basically directly into nuclear power.
[56:18] And there you like Microsoft is investing heavily and Google is
[56:22] investing heavily into these like, you know, into power
[56:26] because they see that is not, not number of GPU's now, but but
[56:29] just the electricity in the grid to even support a data center
[56:33] with, with 100,000 GPU's in it. They have, they have, it's not
[56:37] money like they have more money than they can spend.
[56:39] They have more of every resource except for just pure
[56:42] electricity. And so so I think this becomes a
[56:45] real S curve though, like this becomes a real limiting factor
[56:48] on on AI progress until we can figure out how to massively
[56:52] scale up the amount of electricity we can deliver into
[56:54] like really small localized regions, like a data center in
[56:57] Oregon or something. We are we have this problem.
[57:01] So I think that that things are going to move along in a way
[57:04] that feels very fast, but it's also extremely constrained by
[57:08] it's hell. We can figure out how to make
[57:10] smart models that run like on your MacBook, but learn and are
[57:13] and are really smart, you know, and can train like you could
[57:17] train on your MacBook. Like we are really constrained
[57:20] by by, I think, how fast this stuff can go.
[57:23] So my overall, you know, like tech prediction is that in some
[57:27] ways it'll be incredibly impressive.
[57:29] And then five years, it's like, wow, this is so like the
[57:31] products will be so polished that the agents will like, you
[57:34] know, everyone's really cooked. But in other ways, it's still
[57:37] just not the Super intelligent thing.
[57:40] I think we're way farther from the AGI that we imagine because
[57:45] it requires like a new plateau and energy consumption that we
[57:48] just like, you know, it's not even an AI problem.
[57:50] It's like a much more holistic problem at that point.
[57:52] And that that's my take, but I also don't know.
[57:54] I'd however, I'd say we're always more optimistic and like
[57:57] even if we say like 50 years, this is inevitable, but it might
[58:01] only happen at the 49th year because of the way that
[58:04] exponentials work and and humans are famously bad at judging time
[58:08] on things because of the way that you know, these exponential
[58:10] growth things happen. Absolutely we should have had
[58:14] the flying cars by now, but we don't, unfortunately.
[58:17] We do. It's just they're loud and
[58:18] annoying and like hard to park and FAA regulations.
[58:21] I think that that that that physical reality of like if you
[58:24] unpack the thing, you're like, oh God, yeah, you know, like,
[58:28] where are we going to put all the GPU's?
[58:29] We got to connect them all, like where we get the electricity, we
[58:33] got to build a nuclear power plant just just to get the next
[58:35] generation of AI, like literally, you know, So that's
[58:40] where. We're going, yeah, totally,
[58:42] totally, quite, quite interesting.
[58:44] It's been an amazing conversation.
[58:46] Thanks for making time for this, Shaw.
[58:48] And do let our listeners know where can they get more details
[58:51] about what you're building and these are OS and where can they
[58:54] learn more? Yeah,
[58:57] elizaos.aiisourwebsitewehavegithub.com/eliza OS is our open source project.
[59:03] If you want to get started, go to YouTube, type in the Eliza OS
[59:06] tutorial and there's a bunch of stuff there.
[59:10] We, we have a community on Discord, Discord dot GG/AI16Z.
[59:13] Although we'll move that to Eliza OS soon.
[59:16] And you know, we, we really want to help.
[59:18] I really want to help people like get into this stuff.
[59:20] And I think if you're like, what do I do with my life?
[59:22] Like just become a builder or build things, it doesn't really
[59:25] matter. But like the tools will help you
[59:26] build faster and, and hopefully we're one of the ways to to get
[59:30] into that and do that. And yeah, thanks for having me.
[59:33] Yeah, absolute pleasure having you on.
[59:35] Thanks for doing this and I wish you a happy building.
[59:38] Thank you much. Take care.
[59:40] Thanks so much. Take care.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.