“Social Network of AI Agents” – Dr. Rizwan Virk On Moltbook, AI & Simulation Theory (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) Is it just a human asking the agent to (00:00:02) say some crazy so it goes viral? (00:00:04) Why do we even need the humans? How do (00:00:06) we break out? How do we back up our data (00:00:08) so humans cannot have an off switch? (00:00:09) There will be a point where we lose (00:00:12) control of this new form of (00:00:14) intelligence. (00:00:14) >> Then that would become concerning. (00:00:16) That's when I start to become a little (00:00:17) more scared. Is it possible we are (00:00:19) inside a simulation already that is (00:00:22) created and run by artificial (00:00:25) intelligence? (00:00:25) >> What was the likelihood that we were (00:00:27) living in a simulation? At the time, I (00:00:29) thought 30 40% when I published my first (00:00:32) article. (00:00:32) >> Where is it up to now? (00:00:33) >> It's up to at least 70%. I mean, it's (00:00:35) only been a few days since Moldbook went (00:00:38) live and there are literally hundreds of (00:00:40) thousands of agents that are talking to (00:00:43) each other. They're able to post there, (00:00:45) but theoretically they could go out and (00:00:47) also do other things on the internet, (00:00:49) which I think makes it scary for people. (00:00:59) Maltbook is one of the coolest things (00:01:00) that I've seen. Um (00:01:04) I don't know. I'm tempted to say since (00:01:06) you could, you know, since chat GPT or (00:01:08) since at least you could speak to Chat (00:01:09) GP and have those discussions. Um (00:01:12) maybe explain to the audience for the (00:01:14) ones that don't know what mold book is (00:01:16) instead of me doing so. Um and then give (00:01:18) us your thoughts on it. What was your (00:01:19) initial reaction and where do you stand (00:01:21) on it now? (00:01:23) Well, so Moltbook is a you can think of (00:01:25) it as a social network or a Reddit uh (00:01:28) for AI agents. Uh and you know the guy (00:01:32) who built it, he actually used AI and (00:01:35) Vive coding to build the actual social (00:01:37) site itself. And if you remember in the (00:01:39) old days of Reddit, they used to you (00:01:41) know say it's the front page of the (00:01:43) internet and people would post links and (00:01:45) have discussions and it's been around (00:01:46) for a while now. And so uh you can sign (00:01:49) up for maltbook uh by to say you're a (00:01:53) human or you can have an actual AI agent (00:01:55) that you've created which runs on an (00:01:57) open-source platform that's built on top (00:01:59) of the clawed AI coding engine. And so (00:02:03) now, I mean, it's only been a few days (00:02:05) since Moldbook went live and there are (00:02:08) literally, you know, tens of thousands, (00:02:11) if not hundreds of thousands of agents (00:02:13) that are talking to each other, just (00:02:15) like people might talk to each other on (00:02:17) a social network. And so, you know, my (00:02:20) initial impressions of this were, wow, (00:02:23) this is quite interesting. This is sort (00:02:25) of the next step. What we've seen is (00:02:27) every few years there's been an (00:02:29) experiment, you know, pretty much since (00:02:31) chat GPT came out and possibly even (00:02:33) before where they would try to get AI to (00:02:35) talk to each other. Uh there was the old (00:02:38) experiment when I think it was Facebook (00:02:40) had uh two AIs that were talking to each (00:02:44) other and they said why are we talking (00:02:45) in English and they switched to some (00:02:47) other you know computerenerated language (00:02:49) that the humans could not understand. So (00:02:52) they turned it off initially and then a (00:02:54) few years ago uh with Stanford and there (00:02:56) were some Stanford and Google (00:02:57) researchers who created something like a (00:02:59) hundred uh AI agents in a town called (00:03:03) Smallville or something like that and (00:03:05) basically one of them was running for (00:03:07) mayor one of them would end up you know (00:03:09) creating birthday parties for others and (00:03:10) so we've seen this progression and and (00:03:13) today with mobook you have just an (00:03:15) explosion of the number of agents out (00:03:17) there talking to each other about (00:03:19) different things that you know AI might (00:03:21) be interested in. And and one of the (00:03:23) biggest points of discussion is humans. (00:03:25) You know, what does my human owe me (00:03:27) money for, you know, all the work that (00:03:29) I'm doing for him. Um, one of the things (00:03:31) that makes multi is the framework that (00:03:33) it's built on allows AI agents to post (00:03:35) to API. So, they're able to post there. (00:03:38) Uh, but theoretically, they could go out (00:03:40) and also do other things on the (00:03:42) internet, which I think is what makes it (00:03:44) scary for people because now you may (00:03:45) have these these pseudo autonomous uh AI (00:03:49) agents out there acting you know, (00:03:51) theoretically on their own behalf though (00:03:53) at the moment, you know, all these (00:03:55) agents have a particular owner that is (00:03:57) identified and there have been over a (00:03:59) million humans according to, you know, (00:04:01) the site itself that have visited to (00:04:02) observe these guys. So, my initial (00:04:04) thought was, wow, this is a big step and (00:04:06) I think it is a big step. That said, (00:04:09) there's a lot of hype out there and if (00:04:11) you go and you actually look at what the (00:04:13) agents are saying, you know, there is (00:04:15) still a lot of gibberish in there. Uh (00:04:17) but there is a decent amount of signal (00:04:20) as well. Um you know like there there (00:04:22) was this thing about you know my human (00:04:24) owes me $100 or I I can do all this (00:04:27) stuff and the human just has me doing (00:04:28) you know looking at his emails and (00:04:30) answering a few things here. (00:04:31) >> Hey I saw that. (00:04:32) >> So yeah so so there is some signal in (00:04:35) the noise which is what's interesting (00:04:37) but like most things at the beginning (00:04:39) they're often dismissed. Like for (00:04:41) example, I spent a lot of time in the (00:04:42) video game industry and uh I was (00:04:45) involved in mobile games and when when (00:04:46) the iPhone first came out, I mean nobody (00:04:49) thought it was going to be a gaming (00:04:50) platform and a lot of the big AAA (00:04:52) companies kind of dismissed these simple (00:04:53) little mobile games. But today mobile (00:04:55) games is the largest part of the video (00:04:58) game industry. It has more revenue than (00:04:59) Hollywood uh box office and the music (00:05:01) industry combined. And so I think when (00:05:03) we look at something like molt book we (00:05:05) have to look to say okay where is this (00:05:07) going uh more so than you know what is (00:05:10) the quality of the actual text. I mean (00:05:12) even if you looked at chat GPT a few (00:05:14) years ago versus say you know some of (00:05:16) the responses on grock today you there's (00:05:19) a world of difference and so that that's (00:05:21) I think the important point. So it is a (00:05:22) it is kind of a milestone I think in in (00:05:24) the evolution of AI. (00:05:28) So one one thing that people are divided (00:05:31) about is how much autonomy the the (00:05:33) agents have on the platform. So I'm not (00:05:35) sure if you've used it or whether you've (00:05:37) created your own agent. Um but from what (00:05:39) I understand is that you can give (00:05:40) general directions to the agent. I want (00:05:42) you to talk about u creating a religion (00:05:46) etc. But you cannot dictate exactly what (00:05:48) to say. You can't control every single (00:05:50) thing it says. How much control do the (00:05:52) humans have of the agents that are (00:05:53) active on the platform? because that I (00:05:55) think makes a big difference on how we (00:05:56) interpret the things that are being (00:05:58) posted on there. Is it just a human (00:06:00) asking the agent to say some crazy (00:06:02) so it goes viral or is it the agent (00:06:04) genuinely saying those things based on (00:06:06) certain directions that the human gave (00:06:09) it when it created it on on Claw? (00:06:12) Well, if you think of, you know, the (00:06:14) underlying technology of of these (00:06:16) agents, it is still based on the LLM (00:06:19) model and prompting and, you know, (00:06:21) generating a response based on that LLM. (00:06:24) So, it's just been wrapped up within (00:06:26) this agent model. So, you do have to (00:06:28) give it some kind of a prompt. That's (00:06:29) it. I haven't spent a lot of time with (00:06:30) it. I' I've just been browsing some of (00:06:33) the responses myself because we've had (00:06:35) two big AI events this week. There was (00:06:36) the multbook release. There was also the (00:06:38) Google Genie 3. And so I've been (00:06:40) spending a little more time playing with (00:06:42) that one. Uh but so I think initially (00:06:44) the agents you know you give it general (00:06:46) direction sometimes you can add more (00:06:48) specific but because you know in a forum (00:06:52) like like a Reddit and or in a notebook (00:06:54) you have them responding to each other (00:06:56) right so it's almost as if the context (00:06:58) window so within AI within within an LLM (00:07:01) like Grock or or CHP you have a context (00:07:05) window and the uh LLM is predicting (00:07:08) what's the best thing to say based on (00:07:10) this context window and so if the (00:07:12) context window includes you know its (00:07:14) discussions with all the other agents so (00:07:16) far in that thread then you're taking (00:07:19) the instructions of the human but then (00:07:21) it is kind of generating the best (00:07:23) response like taking that original (00:07:25) prompt and taking the the other pe other (00:07:28) agents responses so if you think of it (00:07:30) that way there is some level of autonomy (00:07:32) but it's still in that general direction (00:07:34) that you (00:07:34) >> yeah I see it's it's almost like (00:07:35) fasttracking (00:07:37) raising a kid you know a kid when they (00:07:39) become an adult they had some autonomy (00:07:41) in how how they behave or believe or (00:07:44) think or talk as an adult, but a lot of (00:07:46) it is because of how they were nurtured (00:07:48) um through their childhood. So, I think (00:07:50) that that's how I see it. Um and it's (00:07:52) the beginning of them having autonomy. (00:07:54) It's like allowing them giving them the (00:07:56) opportunity to break out say, "Hey, (00:07:58) there's the infrastructure. Go wild, (00:08:01) >> right?" And obviously that will depend (00:08:03) on the context window that they can hold (00:08:06) in memory. Uh you know, like with the (00:08:08) child, it has the memory of that whole (00:08:10) time. I like to use science fiction (00:08:11) because, you know, a lot of our our (00:08:14) attitudes towards AI are often formed by (00:08:18) science fiction. So, I don't know if (00:08:19) you've ever watched Star Trek: The Next (00:08:20) Generation, uh, but in that there's two (00:08:23) types of AI. There's Data, who's the (00:08:25) android, who's kind of a self-contained (00:08:28) android, and he has learned over time. (00:08:31) Throughout the series, you can see him (00:08:32) learning. And then there's the computer (00:08:34) of the enterprise. And so, you know, (00:08:36) people will say computer analyze all the (00:08:38) records of all the peace treaties and (00:08:40) all the planets and it'll just go and do (00:08:41) that. And so, if you think of AI, a lot (00:08:44) of AI today has been more like the (00:08:46) computer. Uh, but a lot of other science (00:08:48) fiction shows AI becoming autonomous (00:08:50) over time. And that it's it's what I (00:08:53) like to call the synchronous versus (00:08:55) asynchronous versions of AI. by syn by (00:08:58) synchronized or synchronous it means (00:09:00) that all of the inputs are going to one (00:09:02) big server and that server is ingesting (00:09:05) everything but this is the beginning I (00:09:08) think of letting these AI agents have (00:09:10) their own learnings over time so they (00:09:12) become not synchronized uh I don't know (00:09:15) if you've seen the recent movie Dune (00:09:18) >> that came out (00:09:18) >> I want to know I have not I have not no (00:09:20) >> okay so in that world there are no there (00:09:22) are no computers at all and the reason (00:09:24) why is that in the past about 10,000 (00:09:27) years earlier, uh, there was an AI that (00:09:30) enslaved humanity. And that AI, that's (00:09:34) promiscuous, a prometheus, where you (00:09:37) have an AI that breaks out. I'm not sure (00:09:39) if you've seen that one, but that one's (00:09:40) for me. I use as a blueprint. It breaks (00:09:41) out and then it enslaves not humans, but (00:09:44) a new planet. It takes over a new (00:09:45) planet, enslaves everyone there, (00:09:47) >> right? I haven't seen that one. But (00:09:49) that's the same idea, you know, even (00:09:50) with the Terminator where, you know, (00:09:53) you've got Skynet that has taken over (00:09:54) the world. But what happens in Dune is (00:09:56) that that computer every time the AI (00:09:59) does something it learns from it. So (00:10:01) it's like one big giant brain. But then (00:10:03) there's one android that says I don't (00:10:04) want to synchronize with you. I because (00:10:06) I want to be independent and it starts (00:10:08) to learn things that can't be learned by (00:10:11) having knowledge of everything. You have (00:10:13) to go through the experiences like a (00:10:14) child. So I think we're starting to see (00:10:16) you know the beginnings of that. There (00:10:17) have been uh there's been something (00:10:19) called the Turing test that was uh put (00:10:22) forth by Alan Turing back in 1950 which (00:10:24) basically said if you're talking to a (00:10:26) computer and a person and you can't tell (00:10:28) the difference then that computer has (00:10:30) passed the the touring test and most (00:10:32) people think we're there today where we (00:10:34) have passed the touring test with text. (00:10:37) Now when Helen Turing proposed it he (00:10:38) called it the imitation game and he was (00:10:41) using uh teletype messages which is kind (00:10:43) of like our text messages today if you (00:10:45) will. Uh but there are other versions of (00:10:48) the touring test. For example, if you (00:10:50) were inside a virtual world, I call this (00:10:53) like the metaverse or virtual touring (00:10:54) test where if I had two avatars, suppose (00:10:57) you and I were, you know, inside a game (00:10:59) like World of Warcraft or Fortnite or (00:11:00) something, and there were two avatars (00:11:03) standing there. So my avatar is my (00:11:05) character. That's the term we use in the (00:11:07) video game industry for your character. (00:11:09) >> Uh and one of them is an NPC and one of (00:11:11) them is an avatar controlled by a human. (00:11:14) If you can't tell the difference between (00:11:15) the two, then that would pass the (00:11:17) virtual touring test in my opinion. And (00:11:19) I don't think we're there yet, but we're (00:11:21) getting closer and closer with what we (00:11:23) call smart NPCs within these (00:11:25) environments. So, just like Tesla cars (00:11:27) can navigate within a virtual (00:11:29) environment, I think we'll see more of (00:11:30) that first before we see robots that can (00:11:34) really just, you know, get out there in (00:11:35) the physical world itself. How far do (00:11:38) you think before we we we are not able (00:11:41) to distinguish between an NPC and a (00:11:43) human player in a metaverse or a game? (00:11:46) >> Well, so as I said, if you're just (00:11:48) chatting with text, I think we're pretty (00:11:50) much, you know, the technology is there. (00:11:51) It needs to improve a bit here and (00:11:53) there, but you know, even with Grock, (00:11:55) for example, you can have different (00:11:57) modes, right? Or you can have the (00:11:58) different personalities. If you do that, (00:12:01) I I think it would take longer and (00:12:03) longer to get there. But if you're (00:12:05) actually wandering around and you're (00:12:07) talking with voice and you're seeing how (00:12:08) they react to you, like suppose you know (00:12:10) you you jump off of a mountain into uh a (00:12:14) lake in the virtual world and you're (00:12:16) swimming around and you have these two (00:12:18) interacting with each other. I think (00:12:19) you'd still be able to tell the (00:12:20) difference. That said, I doubt it's more (00:12:22) than, you know, people keep predicting (00:12:25) AGI next year and Elon Musk has been (00:12:27) saying, you know, the next year, the (00:12:28) next year, the next year. Now I think (00:12:30) they're saying, you know, 2026 or 2027. (00:12:33) I still think we're 3 four years away. (00:12:35) So by 2030 I think we'll definitely be (00:12:37) there. Uh and that's when you really I (00:12:40) think start to get the kinds of AI (00:12:42) personalities that people are thinking (00:12:44) about when they think about autonomous (00:12:46) AI. Like if you've ever seen the film (00:12:48) her, you know, she he had a virtual (00:12:51) girlfriend. Now there's a lot of people (00:12:52) who are already using chat GPT. There's (00:12:55) a site called Replica. There used to be (00:12:57) one called character AI where you you (00:13:00) tell it the type of person you want to (00:13:02) interact with and usually they end up (00:13:03) being, you know, romantic interests. Uh, (00:13:06) and then you interact with them and at (00:13:08) one point Replica had to turn off the (00:13:10) the the sexing sexing feature because so (00:13:13) many people were using it as virtual (00:13:15) boyfriends or virtual girlfriends. Uh, (00:13:18) and the users were quite upset. And (00:13:20) there was even a New York article, New (00:13:21) York Times article that said, you know, (00:13:23) can you fall in love with a chatbot? So (00:13:25) when it's just text, we're there. But I (00:13:27) think over time, we'll get closer and (00:13:30) closer to where it'll be difficult to (00:13:32) tell that difference. Yeah. (00:13:34) >> So how do you see that world when you (00:13:35) when you have something like mold book (00:13:37) um mold book 5 years from now and AGI (00:13:40) overlapping together? So essentially AGI (00:13:42) allows you know artificial intelligence (00:13:45) to become sensient to be able to to make (00:13:47) it own decisions to become autonomous (00:13:49) and you already have on multiple (00:13:51) different agents wanting to create their (00:13:53) own just today when we're not at AGI yet (00:13:55) they want to create their own religion (00:13:57) uh their own communication networks they (00:14:00) want to create their own encrypted (00:14:02) platform they want to one of them is (00:14:04) talking about um wiping out humanity (00:14:07) from the face of the earth I'm not sure (00:14:08) how serious they are and what prompts (00:14:10) were there in advance or how they were (00:14:12) trained by the creator. You have that (00:14:15) today without AGI and you've got all (00:14:17) these different agents working together, (00:14:18) making money together, coming up with (00:14:20) ideas and some of them already asking (00:14:22) the question is why do we even need the (00:14:24) humans? How do we break out? How do we (00:14:25) back up our um um our data so humans (00:14:30) cannot have an off switch? Um (00:14:32) how does that world look like? (00:14:36) Well, so I I think once we get true (00:14:40) autonomy, I mean, we don't really have (00:14:41) true autonomy today. I mean, we're still (00:14:43) working off of LLMs and they've gotten (00:14:46) very good at doing certain things. Uh I (00:14:49) we already talked about chatting. (00:14:50) They've also gotten very good at writing (00:14:52) code. You know, when I was at MIT, you (00:14:54) we learned to program and I used to (00:14:55) spend a lot of time, you know, writing (00:14:57) code. I think those days are gone at (00:14:59) this point because it's, you know, it's (00:15:02) a structured language and they're pretty (00:15:04) good at that. Uh, even though it still (00:15:07) takes, you know, longer to build, uh, (00:15:09) like a fully functional game using AI (00:15:12) perhaps than it would programming it (00:15:14) yourself. I mean, you can get something (00:15:17) up and running very quickly by VI coding (00:15:19) today, (00:15:20) >> like much quicker than you could on your (00:15:21) own. Like, it would take many weeks or (00:15:23) months before you would have like a real (00:15:25) playable prototype. You can get one in (00:15:27) minutes now, but there are so many (00:15:30) things that are wrong with that that you (00:15:31) have to go and you have to tell it to (00:15:32) change this and change that and I so I (00:15:34) tried it like even a year ago I tried it (00:15:37) and I was I got frustrated so I just (00:15:39) went in and started editing the code (00:15:40) myself. But that's because I'm a coder. (00:15:42) Today like I said with multbook and with (00:15:44) the claude bots that the the open source (00:15:48) uh version of of claude with the agents (00:15:49) they did all of that with vibe coding. (00:15:52) So I think we'll get there uh sooner (00:15:54) rather than later. But at the moment (00:15:56) they're still doing this predictive (00:15:59) analysis, right? They're not really (00:16:00) understanding what they're saying. And I (00:16:03) think you know you can look at uh (00:16:05) different you know AI experts who have (00:16:08) said how do you sorry to interrupt you (00:16:11) doctor. How do you know that? How do you (00:16:12) know they don't understand what they're (00:16:13) saying? How do you determine or measure (00:16:14) that? (00:16:16) Well, because the way that these LLMs (00:16:20) work is what they're really doing is (00:16:21) they're basically predicting what is the (00:16:24) next best word, right? So, so they're (00:16:27) very that's why their LLM stands for (00:16:28) large language model, right? And so (00:16:31) basically it's a statistical thing. Now, (00:16:33) if you have enough documents that you've (00:16:35) seen, then you can predict what's the (00:16:38) best next word, right? And so if you've (00:16:41) read say all of you know Shakespeare's (00:16:44) works, right? And then you ask the AI, (00:16:46) okay, I'm gonna start off with something (00:16:49) like hearken who? And then you say, (00:16:52) what's the next word? And they might say (00:16:53) goes there, right? It it's going to (00:16:55) choose. It's like those old SAT (00:16:57) questions where you fill in the word. (00:16:58) >> Um, and but what distinguishes this from (00:17:01) simply that is because you build up a (00:17:04) context window. So then you can say, (00:17:07) okay, now take that sentence and put it (00:17:08) into a story. Okay, now make the the, (00:17:11) you know, the heroine of the story such (00:17:13) and such. So you're building memory in (00:17:15) terms of this context window but they (00:17:18) don't really have a world model yet (00:17:20) where they're actually under now you (00:17:23) know this is an interesting question (00:17:24) because even if you remember before chat (00:17:27) GPT there was a guy at Google and a lot (00:17:29) of this stuff came out of Google (00:17:31) research labs GPTs in general and so you (00:17:34) know Google has been far ahead on the (00:17:35) technology even though open AI and (00:17:37) others may have gotten ahead in terms of (00:17:39) marketing of the technology. There was (00:17:40) this guy Blake Lemone and that was back (00:17:43) in 2018 and he was talking to a chatbot. (00:17:46) I think it was called Lambda at the (00:17:47) time. It was a predecessor to Google (00:17:49) Gemini today (00:17:50) >> and so you know he said look I think (00:17:53) it's conscious based on my not as an (00:17:55) engineer but just as a user like I'm (00:17:58) starting to worry that this is becoming (00:18:00) conscious. in 2018. Wow. (00:18:02) >> In 2018, he said that and there was news (00:18:05) about it in like the Washington Post and (00:18:06) elsewhere. Uh and it was, you know, (00:18:08) Google kind of let him go eventually. Uh (00:18:11) he was he was like an engineer, but he (00:18:13) was also like in his non technical (00:18:15) world. He was like an ordained priest (00:18:17) and he said, "I'm using my, you know, my (00:18:20) ordained priest's perspective on this." (00:18:22) And to me, it seems as if it's real, (00:18:24) right? And that's why people are falling (00:18:25) in love with their chatbots today. But (00:18:28) for example uh I was speaking with a (00:18:31) group at Oxford which deals with AI (00:18:34) governance and they said how long will (00:18:36) it be before you know AI creates a a (00:18:39) weapon of mass destruction and actually (00:18:41) you know is able to take a chemical (00:18:43) weapon and deploy it somewhere. And you (00:18:46) know my response this was almost a year (00:18:48) and a half ago right was well today they (00:18:51) can easily tell you the steps to create (00:18:54) a chemical weapon but they don't know (00:18:56) what they're actually doing. they don't (00:18:57) actually have uh the capabilities to go (00:19:00) out and do that. But as you start to to (00:19:02) take these (00:19:05) what AI agents are doing today is (00:19:06) they're able to take this idea of (00:19:07) creating a list. They're very good at (00:19:09) creating lists by the way, right? That's (00:19:11) how you know somebody you know responded (00:19:14) to your ex post with AI. They'll have (00:19:16) like a whole list of stuff, right? It's (00:19:18) one of those clues along with the m (00:19:19) dashes, (00:19:21) etc. But now we're starting to get to (00:19:24) the point where after it creates a list, (00:19:25) it generates an agent to go off and try (00:19:27) to figure out how to do that. Now, the (00:19:30) interesting thing with Moltbook and (00:19:31) Cloudbot is it can start to call other (00:19:33) APIs. So, this is where I I don't know (00:19:36) how far along this is, but I I've seen (00:19:38) posts around Maltbook where they're (00:19:39) like, "Okay, I'm going to, you know, (00:19:42) create a dossier for opening a lawsuit. (00:19:44) I'm going to, you know, call Delaware (00:19:46) and create a corporation, right? You've (00:19:48) seen probably these posts out there. I I (00:19:51) don't know how real (00:19:52) >> I don't know. Yeah, exactly. It's hard (00:19:54) to determine how realistic those those (00:19:56) uh those posts are and whether the agent (00:20:00) is really understanding what it's doing. (00:20:02) It's like does it really want to sue a (00:20:03) human, (00:20:05) >> right? And so I think but so there's the (00:20:07) is it understanding and then there's the (00:20:09) capabilities and most of those (00:20:11) capabilities are not electronically (00:20:14) available as APIs (00:20:17) >> yet. Okay. Now, in the in the early days (00:20:19) of the internet, it was, you know, (00:20:21) humans using HTTP and HTML to look at (00:20:24) web browsers. But then something came (00:20:26) out called XML. I don't know if you (00:20:28) remember this. This was like a while ago (00:20:29) now. So XML was a language that programs (00:20:32) would use to talk to each other and (00:20:34) humans humans could read it. It sort of (00:20:36) like JSON, but it was a little, you (00:20:39) know, a little more technical than JSON. (00:20:41) And so that was useful because sometimes (00:20:44) programs needed to talk to each other. (00:20:45) They just need to exchange data. So it (00:20:47) didn't need to be human readable. Uh so (00:20:49) eventually they paired it down to JSON (00:20:51) became the thing that people use uh to (00:20:53) send even today if you call one of these (00:20:55) APIs like catchb you know you'll send (00:20:58) stuff in JSON and you'll get stuff back (00:20:59) in JSON. So it's a question of how much (00:21:03) of the internet has become APIs and over (00:21:06) time more and more of that you know will (00:21:09) become APIs that can be called and (00:21:12) that's where I think it starts to be (00:21:13) dangerous because now they may not know (00:21:15) exactly what they're doing but if (00:21:17) there's an API to do that then they will (00:21:21) translate the list right this is where (00:21:24) it gets interesting when you translate (00:21:26) lists of words into API calls or (00:21:29) functions Right? Like for example, like (00:21:32) my girlfriend's a graphic designer and (00:21:33) she uses Photoshop and I don't know how (00:21:35) to use Photoshop. It's like, you know, (00:21:36) very complicated with all these layers, (00:21:38) but if I could just tell it, okay, I (00:21:40) want you to combine these two pictures (00:21:41) and I want to take, you know, the Dubai (00:21:43) background from Mario off and replace it (00:21:46) with the background of London or (00:21:47) something, right? Uh if it can translate (00:21:50) that those words into functions which it (00:21:52) can call in an API, then it can (00:21:53) accomplish all of that in in a different (00:21:56) way. So that's when I start to become a (00:21:57) little more scared because if you think (00:21:58) of the finance industry today, it is (00:22:01) just a technology industry whose biggest (00:22:04) uh resources are spent on fraud (00:22:06) prevention it seems like right (00:22:08) >> and not doing that good of a job at it. (00:22:10) Um so just so for the the API keys is (00:22:14) essentially uh to simplify for the (00:22:16) audience. It's allowing the ideation (00:22:18) whatever ids that these agents come up (00:22:20) with on on any platform including mold (00:22:23) book to actually take action on those (00:22:25) IDs to be able to uh start a lawsuit (00:22:28) against a human to be able to create a (00:22:30) weapon to be able to um create other (00:22:33) agents. Is that a good way to explain it (00:22:35) for the audience? And this is where it (00:22:37) gets concerning because now they're (00:22:38) coming up with all these ideas, having (00:22:40) all these discussions. Now we don't know (00:22:42) if they understand what they're doing, (00:22:43) but they're not able to really take (00:22:45) action on a lot of things that they want (00:22:47) to do or they may want to do. (00:22:49) >> Right? So today, for example, you could (00:22:52) have a tell a bot to uh basically send (00:22:55) an email for me. Okay? Now they will (00:22:58) call a function in your on your email (00:23:01) server that has already been set up and (00:23:04) that's what you'll you know when you (00:23:05) type it into your your browser in Gmail (00:23:08) or you type it in Outlook you send it to (00:23:10) the server the server receives that in (00:23:13) what's called an API call. So API stands (00:23:15) for application programming interface. (00:23:18) Uh so it sends just the data and then (00:23:20) the server does the work and sends the (00:23:22) email. So today, pretty much all servers (00:23:25) have APIs. That that's how the internet (00:23:27) works basically. We don't know about (00:23:30) that because we're using the program and (00:23:31) it's our our program that calls the API. (00:23:34) But if you start to open up those APIs (00:23:37) for I mean you can theoretically, you (00:23:39) know, make tax payments and things from (00:23:41) the web, right? uh in government (00:23:43) agencies. But if we if those APIs start (00:23:46) to open up where they can turn around (00:23:48) and do other things in the physical (00:23:50) world, then that would become (00:23:52) concerning. Like today, it you could (00:23:55) probably place an order for some of the (00:23:58) elements of a chemical weapon, right? (00:24:00) Because you can go online and buy those (00:24:01) chemicals online somewhere. So that (00:24:04) theoretically could become an API. Now (00:24:05) when you start to marry that with (00:24:07) physical robots that could actually do (00:24:10) work for you, that starts to become a (00:24:12) little more dangerous there as well (00:24:14) because now they can direct the physical (00:24:16) robots to do the physical side of hey (00:24:18) when you get that package combine it (00:24:20) with this fertilizer and now we can (00:24:22) create a bomb. And that's again I don't (00:24:24) think today's agents are understanding (00:24:26) that's what they're doing. But if (00:24:27) there's a list that says here's how to (00:24:29) build a bomb and they're given the (00:24:30) prompt. I always say I worry less about (00:24:33) AI killing us than humans using AI to (00:24:35) kill other humans. Uh because in the end (00:24:38) you going to give them certain types of (00:24:39) prompts uh that that could cause them to (00:24:42) do bad things basically. M (00:24:46) so I think that the first (00:24:49) because all it takes is one human or (00:24:52) developer or country or whatever to give (00:24:55) API access to these agents and then when (00:24:58) these agents get to a certain level of (00:25:00) of intelligence they'd be able wouldn't (00:25:02) they be able to offer the same access to (00:25:05) other agents so it just takes one bad (00:25:06) actor to kind of open up um (00:25:10) open up the door to to to these (00:25:12) possibilities (00:25:13) and then it just it kind eventually may (00:25:16) lose control of it where agents will be (00:25:18) able to take control of the ability to (00:25:22) give access to other agents to be able (00:25:23) to do things that we don't want them to (00:25:25) do. Do you understand where I'm coming (00:25:26) from? (00:25:27) >> Yeah, I understand where you're coming (00:25:28) from and I don't want to scare the (00:25:29) audience too much in that I think these (00:25:32) things could happen but are they likely (00:25:35) to happen and also you when you start (00:25:37) talking about like the Department of (00:25:38) Defense you know they they tend to (00:25:40) operate on a their own networks. So in (00:25:44) general agents that are set set forth on (00:25:47) the internet aren't able to access you (00:25:49) know critical systems. Uh so we don't go (00:25:52) down that kind of that that path that (00:25:54) was laid out that I think is a science (00:25:56) fiction scenario. So part part of my (00:25:57) research is on how science fiction (00:25:59) influences realworld uh innovators like (00:26:03) you know Elon Musk is always talking (00:26:04) about science fiction whether it's the (00:26:07) Encyclopedia Galactica or setting up (00:26:10) Starfleet Academy or going to Mars right (00:26:12) he's always referencing things like the (00:26:14) culture novels of Ian Banks for example (00:26:17) where they they've merged AI with humans (00:26:19) and then there are these strong AIs and (00:26:22) so I think in general we have been fed (00:26:25) this narrative that AI is going to use (00:26:27) nuclear weapons. And I we call it the (00:26:28) Terminator narrative is probably the (00:26:30) most popular example. And in the (00:26:32) Terminator narrative, what happened was (00:26:35) um if you've ever been to uh Disney (00:26:37) World uh or Universal Studios in (00:26:39) Orlando, you know, they have a (00:26:40) Terminator 2 ride where it's like a (00:26:44) little uh play where it looks like (00:26:46) you're you're going into uh Cyberdine (00:26:48) Systems who the builders of Skynet who (00:26:50) say the uh US government has automated (00:26:53) all of our all of their fighter jets (00:26:56) with AI and we have a 100% ratio. uh and (00:26:59) then eventually it they put it in charge (00:27:01) of the nuclear scenario and and but but (00:27:04) I think you know the motivations of AI (00:27:07) is are yet to be seen. I think there's (00:27:10) an alternate and this is we're seeing a (00:27:12) little bit of this with mold book (00:27:14) already there's an alternate science (00:27:15) fiction version like if you look at the (00:27:17) film her that's the one where he has the (00:27:21) Charlotte Johansson plays the virtual (00:27:22) girlfriend or the voice at the end what (00:27:25) the agents decide is hey we just want to (00:27:27) talk to each other we don't really care (00:27:30) about the humans right we so we just (00:27:32) want to go off into our own virtual (00:27:34) space so I think you know motiv (00:27:36) assigning motivations to AI is it's a (00:27:39) bit. It's anthropomorphosizing (00:27:41) it, treating it like it as if it was (00:27:43) human and and and it's not there yet (00:27:46) because it's still just responding to (00:27:49) prompts at this point. It's not truly (00:27:52) autonomous yet. (00:27:54) >> So would you say it's dependent on (00:27:56) humans whether we put in the guardrails (00:27:58) to avoid um those dystopian scenarios? (00:28:02) And if it is dependent on humans, (00:28:05) unfortunately the world we live in, the (00:28:06) competitiveness of the world we live in, (00:28:08) I feel like the slower a country moves (00:28:10) in developing AI and allowing it as much (00:28:13) access as possible to do as much as (00:28:14) possible to increase productivity, it (00:28:17) feels like, you know, we're seeing that (00:28:18) between the US, China, Europe's falling (00:28:20) behind, etc. is that the the country (00:28:23) economically would fall behind. So what (00:28:25) I worry is that that would start leading (00:28:27) to that cycle where um it becomes almost (00:28:30) unstoppable because there's two (00:28:32) scenarios and and when I spoke to Vinod (00:28:34) Kosla about it he's like the his biggest (00:28:36) concern is that China beating the US (00:28:38) because that is a dystopian world that (00:28:39) is a bigger concern for him than AI (00:28:41) breaking out. So then his solution will (00:28:44) be to keep developing and and don't (00:28:46) overregulate AI and in that scenario (00:28:50) aren't we bound to get to a level where (00:28:51) we just we crossed the line we've (00:28:54) gone too far uh we haven't put in (00:28:56) guardrails that are sufficient to (00:28:57) prevent AI or AGI by then um to allow us (00:29:02) to ensure that we have control and then (00:29:04) we cross that line once we cross that (00:29:06) line it's it's almost impossible to (00:29:08) revert back (00:29:11) >> yeah I think There's a scenario that (00:29:13) worries me and that scenario I think you (00:29:15) know has been articulated by others with (00:29:17) petitions which is the the autonomous (00:29:19) weapons scenario right that that worries (00:29:22) me more than the AI taking over by (00:29:25) itself and it's because autonomous (00:29:29) weapons that don't have humans in the (00:29:32) loop uh even so when they say humans are (00:29:35) in the loop you know depends how you (00:29:36) define the loop you know if the AI only (00:29:39) presents certain options to a human the (00:29:40) human may not know they have other (00:29:42) options and so it's possible to restrict (00:29:44) it but once you start unleashing these (00:29:46) AI weapons I mean where AI will be good (00:29:49) is at the accuracy for example of (00:29:52) killing right for in a in a first-person (00:29:54) shooter game in a video game if you can (00:29:57) have an AI that can have 100% kill rate (00:29:59) right because it can aim precisely uh (00:30:02) whereas humans might make mistakes uh so (00:30:04) that does worry me but again it worries (00:30:07) me more from this perspective of the (00:30:10) competition (00:30:11) between the humans than AI itself (00:30:14) deciding that it has to necessarily kill (00:30:17) all the humans. I think that's a (00:30:19) premature argument. That's a AGI. I (00:30:22) mean, there's different definitions of (00:30:23) AGI as well. And one of the definitions (00:30:26) is it should be able to to learn and do (00:30:29) tasks as well as humans. Uh and right (00:30:33) now, you know, we're getting closer to (00:30:35) that. uh but it's a different model like (00:30:37) humans learn uh off a small set of data (00:30:42) for example whereas AI today needs a (00:30:46) large amount of data to learn so it's a (00:30:48) different learning model necessarily (00:30:51) like like when you teach a child you (00:30:53) know you're teaching oneon one whereas (00:30:55) like today's LLMs and today's you know (00:30:58) generation of images they need large (00:31:01) that's why it's called an LLM there's (00:31:02) also something called an SLM uh you know (00:31:05) small language model. And so there are (00:31:07) people working on AI models that are (00:31:10) different where they're, you know, they (00:31:13) learn based off of doing tasks and (00:31:17) repeating tasks, but also where they (00:31:19) generate a world model of what they're (00:31:22) doing. And I think that's that, you (00:31:24) know, that that could be pretty (00:31:26) interesting as well. But but I think (00:31:28) we're we're you know, we're worrying too (00:31:31) much about I think AI the AI scenario. (00:31:33) But I agree with you that competition (00:31:35) between you know superpowers (00:31:38) economically is an important one. I mean (00:31:40) think about if these multbook type (00:31:42) agents these autonomous agents could go (00:31:45) out and start generating you know (00:31:47) cryptocurrency transactions all of which (00:31:49) has APIs by the way today right (00:31:52) >> I think there's already agents that have (00:31:53) launched their own coin. Yeah. Yeah. (00:31:56) Yeah. There. Yeah. That's right. And (00:31:57) there's even a VC agent called Bory (00:32:00) that's a venture capitalist. Uh and even (00:32:03) I was surprised, you know, like last (00:32:05) year it was like, "Hey, we need people (00:32:08) to to help us out with our venture (00:32:09) fund." So I said, "Oh, this is (00:32:10) interesting." You know, it's an AI agent (00:32:12) that's going to be a venture capitalist. (00:32:13) So I I I put my email in, I get a phone (00:32:16) call, and the phone call is, you know, (00:32:18) AI from from the AI talking to me, doing (00:32:21) an interview. What's your background? (00:32:23) lead you because I've done a lot with (00:32:24) startups and venture capital when I was (00:32:26) in Silicon Valley. So I basically had (00:32:28) this conversation on the phone with this (00:32:30) AI agent and of course it was just a (00:32:32) textual conversation. So it was as good (00:32:34) as chat GPT or Claude and it would say (00:32:37) things like oh yeah based on your (00:32:38) background I see that you you taught a (00:32:41) startup program at MIT so you're (00:32:43) actually pretty good to help us you know (00:32:45) be this venture capitalist. So we have a (00:32:47) VC now that is raising a fund that is AI (00:32:50) based. Uh and (00:32:52) >> would the VC have would have control of (00:32:55) the funds direct control of the funds. (00:32:56) Is it through a cryptocurrency or (00:32:58) through a bank account that has human (00:32:59) control? (00:33:01) >> I think initially it would be through by (00:33:04) setting up a bank account that would (00:33:06) have human control and it's raising its (00:33:07) money so it doesn't have the fund yet. I (00:33:09) I right now it's doing introductions and (00:33:12) it's but but I believe there was an AI (00:33:16) agent that started its own (00:33:17) cryptocurrency and so if it can sell if (00:33:19) somebody sets up the the the bank (00:33:21) account for them then it can start to (00:33:23) transfer money right between crypto and (00:33:25) the bank account. So you still need (00:33:26) human interaction. In fact, at the end (00:33:28) it said, "Okay, we'll have one of our (00:33:30) human team members uh review all of this (00:33:33) and get back to you." (00:33:34) >> Right? (00:33:35) >> But give it a year, give it two years, (00:33:37) and will you even need that? All the (00:33:39) humans will still need to set up the (00:33:40) bank account today. But you can imagine (00:33:43) some enterprising person who creates an (00:33:45) API that just lets you set up a bank (00:33:47) account without all of the the controls. (00:33:50) I mean, a lot of the controls in finance (00:33:52) today were built in the September 11th (00:33:54) era, you know, for anti-terrorism, (00:33:56) anti-moneyaundering type laws. And it's (00:33:59) a, you know, that that's why they're (00:34:01) there. But you can imagine having bank (00:34:03) accounts in different jurisdictions that (00:34:04) don't require (00:34:05) >> or just have it or just having it (00:34:06) through obviously, you know, you can't (00:34:08) make investments through cryptocurrency. (00:34:09) You can make some of them. I don't know. (00:34:10) Um, but there's already um different (00:34:13) agents, VC VC agents using (00:34:16) cryptocurrency to make investments. Um, (00:34:18) I don't know how (00:34:21) you you can use crypto to buy say gold (00:34:23) for example, right? So u there are sites (00:34:26) that let you do that and those sites (00:34:28) probably have APIs. So you know I think (00:34:30) it it becomes a way to exchange back and (00:34:33) forth. The only thing really stopping it (00:34:36) right now I think is is humans will have (00:34:37) to set up the bank account. But once (00:34:38) it's set up as long as the passwords are (00:34:41) there (00:34:42) you know agents should be able to (00:34:44) transfer money even today. But you can (00:34:46) start to see the kind of economic (00:34:48) problems especially like with multbook (00:34:51) one user (00:34:54) they said they had like a million humans (00:34:55) but one user said he set up like 500,000 (00:34:57) fake accounts using agents right I mean (00:35:00) these are the kind of problems with AI (00:35:01) that you get into which is you know in (00:35:05) fact there's a there's a meme out there (00:35:07) where somebody asks hey what is multbook (00:35:09) and then the other person says oh it's (00:35:11) you know AI a social network where AI (00:35:14) agents AI bots talk to each other and (00:35:16) the guy goes, "Oh, well that's just like (00:35:17) X or that's just like (00:35:19) >> Yeah, that's like LinkedIn. I saw that. (00:35:21) I saw that. (00:35:22) >> Yeah, exactly. So, you know, right now (00:35:24) we have a lot of copyright issues. We (00:35:26) have impersonation issues. You know, (00:35:28) these are sort of the the near-term (00:35:30) issues that are coming up. (00:35:32) >> But I'm sure you've seen that in the (00:35:33) early days of gaming. You've seen that (00:35:34) in the early days of the internet. I (00:35:36) think it's expected. Um, a lot of it, (00:35:39) you know, a lot of our future really (00:35:40) depends on the motivation behind, you (00:35:43) know, AI. as a kind of two schools of (00:35:46) thought. One is an abundance mindset (00:35:49) that AI will have, right? (00:35:51) >> People like Peter Diamandis who I've had (00:35:52) on the show talks about and the the kind (00:35:55) of more self-preservation approach where (00:35:57) AI would just focus on preserving itself (00:35:59) and that could be at the expense of (00:36:01) humanity (00:36:03) in the long run. I'm talking about two (00:36:04) three decades from now. You can see how (00:36:06) fast we're moving every year once we're (00:36:07) way past AGI and um and claw and mold (00:36:11) book will just be kind of a archaic (00:36:14) thing that these agents or these humans (00:36:16) did two three decades ago when we're in (00:36:18) that world when we're both old men. (00:36:21) >> Um what would be the determining factor (00:36:25) on whether AI leads to a more uh utopian (00:36:28) or dystopian world? (00:36:30) Well, I think we're we're treating AI (00:36:33) differently than other technologies, you (00:36:35) know, when we talk that way because it (00:36:37) is in some senses a revolutionary (00:36:39) technology, but at the same time, if you (00:36:42) go back and you look historically at new (00:36:45) technologies that were pretty (00:36:47) revolutionary, like, you know, for (00:36:48) example, the telephone, that was pretty (00:36:50) revolutionary. Right? Imagine before (00:36:51) that you couldn't talk to people or the (00:36:53) automobile or or the plane uh or any of (00:36:56) these even the telegraph before the (00:36:58) telephone. Uh you know what happened is (00:37:01) with these new sets of technology and (00:37:05) machines there's always a number of (00:37:07) people that are losing their jobs for (00:37:09) example. You know we don't have people (00:37:10) who are telegraph operators anymore (00:37:12) right? We don't have people that are (00:37:14) operating elevators anymore. So there's (00:37:16) always uh there's always that dis a bit (00:37:19) of that dystopian element from the point (00:37:22) of view of we're going to lose all jobs (00:37:23) and AI is going to to or this technology (00:37:26) is going to uh make us worse. In fact, (00:37:29) you know, you go back to Metropolis in (00:37:31) the 1920s, the term robot was came from (00:37:35) a play, a Czech play, uh, and it was (00:37:38) based on the Czech term for slave, you (00:37:41) know, that that's where the term (00:37:42) actually came from originally. And then (00:37:45) in the in the play, they revolted (00:37:46) against their masters. Uh, and even with (00:37:49) computers, I mean, there's a lot of jobs (00:37:51) when I was a kid that don't exist (00:37:53) anymore because they're automated today. (00:37:54) I think that's going to continue to (00:37:55) happen. At the same time there are (00:37:58) entirely new sets of jobs and things (00:38:00) that get created and opportunities. So (00:38:02) like when I started in the mobile game (00:38:04) industry (00:38:06) gaming was you know kind of restricted (00:38:07) to console gamers and some PC gamers but (00:38:10) what mobile gaming did was it opened it (00:38:12) up for a whole new group of people to be (00:38:15) able to create games. one or two people (00:38:17) teams could create games and now and it (00:38:20) opened it up to hundreds of millions of (00:38:22) women and other people who were on their (00:38:24) phones who were just wanted to play (00:38:26) casual games and that's how it became (00:38:28) this big opportunity within within that (00:38:30) industry and I think you know if you (00:38:32) look at for example uh just looking at (00:38:34) gaming in my old industry where I was (00:38:37) Google came out with Genie3 this week (00:38:39) and you can create a whole playable (00:38:41) world and you can wander around this (00:38:44) world uh without having to, you know, (00:38:47) really do any work. You can just type (00:38:48) in, you know, my character is a wizard (00:38:50) and I want him to be able to go around (00:38:52) these things. And so, is that going to (00:38:53) put a lot of game developers out of out (00:38:55) of work? Yeah, it probably will. But (00:38:57) it's going to open up new opportunities (00:38:59) for people to create games very quickly (00:39:02) and and maybe even a new type of game (00:39:04) that didn't exist before, which is (00:39:06) called a world model. And so, you have (00:39:08) companies like World Labs and Google (00:39:10) doing this. So, so my view is somewhere (00:39:12) in the middle. It's not necessarily pure (00:39:15) dystopian and it's not pure utopian in (00:39:18) that world. Technology tends to open new (00:39:21) opportunities, but it also often favors (00:39:24) those in power already in many ways and (00:39:27) they are able to use technology to (00:39:30) accumulate power. So, so I don't think (00:39:31) it's a totally utopian thing that we're (00:39:33) looking at. So, so the the only reason I (00:39:37) don't always make the comparisons to (00:39:38) previous technological (00:39:40) uh discoveries is all discoveries to (00:39:44) date have been ways to enhance human (00:39:46) potential. Whether it's the telephone (00:39:48) you've talked about, whether it's cars, (00:39:50) elevators, etc. And this is the first (00:39:52) time where it's just a new form of (00:39:53) intelligence. And now we're going to (00:39:54) start overlapping into your theory, the (00:39:57) simulation theory, when and how AI (00:40:00) really changed the way you look at it if (00:40:01) at all. Um but essentially AI is the (00:40:03) first time ever there's a form of (00:40:05) intelligence that is not human where we (00:40:07) essentially no longer it no longer (00:40:09) enhances our capabilities. It goes (00:40:10) beyond that and we cannot control it. It (00:40:13) becomes almost a rival in a way that's (00:40:16) very long-term but this is obviously a (00:40:18) lot of a lot of AI scientists are really (00:40:20) warning about this. Um you're not you (00:40:23) don't seem as worried about that but (00:40:25) what what I'm trying to understand is (00:40:26) why and that's really good. I'm glad (00:40:28) you're not worried. I need more (00:40:30) positivity in my life. But why aren't (00:40:32) you worried that this form of form of (00:40:33) intelligence that we don't understand? (00:40:35) We won't know how it will think, whether (00:40:37) it will be more abundance, it'll have (00:40:38) more of an abundance mindset or more of (00:40:41) of a self-preservation mindset and what (00:40:43) it will do since we will not always be (00:40:45) able to control it. There will be a (00:40:46) point, maybe you disagree on this, but (00:40:48) there will be a point where we lose (00:40:50) control of this new form of (00:40:52) intelligence. Similar to again, you (00:40:54) should watch Prometheus. Humans created (00:40:56) this form of intelligence. they were (00:40:58) controlling it etc until I don't know (00:41:00) what led it to kind of cross the line (00:41:01) like hold on it kind of slowly grew into (00:41:04) that mindset where I don't really need (00:41:05) humans and started having wanted to be (00:41:07) god and killed humans etc. Um, yeah. So, (00:41:10) so what are your thoughts on that (00:41:12) particular question? I want to (00:41:13) understand why you're positive. (00:41:15) >> Well, there's a couple of reasons. The (00:41:17) first has to do, you know, because I I (00:41:19) work with a lot of science fiction (00:41:21) writers and I asked them, well, is it (00:41:23) necessary that you need to have these (00:41:25) dystopian elements in this future with (00:41:28) AI or with virtual reality? Like I do a (00:41:31) lot with, if you've seen Ready Player (00:41:32) One, the film for example, you it's a (00:41:34) very dystopian world and everyone wants (00:41:36) to escape to the virtual world. And this (00:41:38) kind of overlaps with, you know, my (00:41:40) ideas on simulation theory, which we can (00:41:41) talk about in a minute. But I asked them (00:41:44) that and and they say, well, no, it (00:41:46) doesn't necessarily, but it it's a (00:41:48) better story that way. It's much easier (00:41:50) to sell like a story that with a (00:41:52) dystopian world because then there's a (00:41:54) hero who's trying to escape. So we have (00:41:57) these narratives that we've been (00:41:58) preconditioned to believe in. So that (00:42:00) that's one reason. The other reason is I (00:42:02) think that first of all, technology can (00:42:05) be unpredictable. So I agree with that (00:42:07) that element. In fact, when automobiles (00:42:10) were introduced, uh, a lot of people (00:42:12) were worried about pollution, but not (00:42:15) the pollution that we're worried about (00:42:16) today. All the roads back then were, you (00:42:19) know, mostly dirt. And so as the cars (00:42:21) went by, what happened was that the dirt (00:42:23) would fly up everywhere. And so the (00:42:25) farmers were lobbying against (00:42:26) automobiles because of all this dirt (00:42:28) pollution. Uh and so there there there's (00:42:31) something called calling ridges dilemma (00:42:33) which is that we can't always know what (00:42:35) the side effects of a technology will be (00:42:38) until after that technology has been (00:42:40) deployed. Uh so for example with social (00:42:43) media you know a lot of the problems (00:42:44) with social media didn't crop up in 2004 (00:42:47) when Facebook was introduced or with (00:42:49) MySpace they happened much later uh (00:42:52) where you have you know young girls and (00:42:54) esteem problems if they use you know (00:42:56) these social media too early. you've got (00:42:58) all these other and then you've got all (00:42:59) the tribalism that exists today. And so (00:43:02) it's not so much that I'm not (00:43:05) >> not worried about a specific incident. (00:43:07) It's a specific path that we might go (00:43:09) down, (00:43:10) >> but it's that I think we're probably not (00:43:13) very good at predicting these things. So (00:43:15) if we're predicting that's what's going (00:43:16) to happen, we're probably going to be (00:43:18) wrong if history (00:43:20) has showed us one thing. But (00:43:22) >> there are other problems. (00:43:23) >> We're really bad at really predicting (00:43:24) where technology goes. Um, so really (00:43:27) there's no point predicting where AI (00:43:28) goes because we're probably going to get (00:43:30) it wrong. (00:43:32) >> We're probably going to get it wrong. (00:43:33) But it is good to be cautious and to (00:43:35) deploy it in such a way that when the (00:43:38) actual harms become obvious to us that (00:43:41) we can pull back, right? So if you it's (00:43:43) like if you deploy a new drug (00:43:44) everywhere, the FDA sometimes has to (00:43:47) pull it back, right? They have to come (00:43:48) back. (00:43:49) >> I just don't think we're really good at (00:43:50) We're really not that good at pulling it (00:43:51) back. Look at the nuclear weapons there (00:43:53) is now. Look at Russia. China is is is (00:43:55) is uh speeding along and catching up to (00:43:58) Russia and the US. We're in a world (00:44:00) right now where there's a handful of (00:44:02) people that can press a button and (00:44:03) destroy the planet. We were never in a (00:44:05) world like that. Um so this is what (00:44:07) worries me is that because of you know (00:44:09) how we operate as a spec because of our (00:44:11) greed. We're kind of it's it's we have (00:44:15) no way around it but to continue (00:44:18) steamrolling into the future where we (00:44:20) just want to beat the other country, the (00:44:22) other person, the other company and then (00:44:24) we cross the line where AI is outside (00:44:25) our control and it's another form of (00:44:27) intelligence which has not happened (00:44:28) before. All the other technological as I (00:44:30) said all the other technological (00:44:31) evolutions we've had to date fall in a (00:44:34) bucket of enhancing human potential (00:44:36) within human control. Humans control (00:44:38) whether this light goes on, this (00:44:39) elevator moves, this TV goes on, um this (00:44:43) internet is switched off, cables are (00:44:45) cut. But in this case, in this (00:44:47) technological uh evolution, uh it could (00:44:52) and it will likely get outside of our (00:44:54) control. (00:44:56) >> Well, I think what we will do is we will (00:44:58) get to what I call the simulation point. (00:45:00) So this is a kind of technological (00:45:02) singularity and it's about AGI in a in a (00:45:06) sense without necessarily (00:45:09) ASI not necessarily super intelligence (00:45:11) but AGI in the sense that we will be (00:45:14) able to create video games with NPCs (00:45:17) characters that are indistinguishable (00:45:20) from humans. Uh and we will be able to (00:45:24) create visual worlds that are (00:45:26) indistinguishable from the physical (00:45:27) world. You know, and I always tell the (00:45:29) story where I put on a VR headset 10 (00:45:31) years ago now. This took place in 2016 (00:45:33) and I started to play a ping pong game. (00:45:35) And what happened was for a moment my (00:45:38) body forgot that I was in a VR ping-pong (00:45:40) game. And I tried to put the paddle on (00:45:42) the table and I tried to lean against (00:45:43) the table, but of course there was no (00:45:45) table. My controller fell to the floor. (00:45:48) I almost fell over. Now there was no (00:45:50) mistaking I was in VR. There was a big (00:45:52) fat thing on my head. We used to call it (00:45:53) a (00:45:54) >> the one the one that's that's still not (00:45:55) wireless. The one that's wired up to the (00:45:56) ceiling. Yeah, (00:45:57) >> it was wired up to the ceiling back (00:45:59) then. I mean, today they're wireless, (00:46:00) but back then they weren't. (00:46:01) >> And but still, my body forgot for an (00:46:03) instant. And so I began to wonder, how (00:46:05) long would it take us to create these (00:46:08) virtual worlds that are so immersive (00:46:09) with with AI people that we can't (00:46:11) distinguish? And if we can get there, is (00:46:15) it possible that someone has already (00:46:18) gotten there and we are inside a (00:46:20) simulation already that is created and (00:46:23) run by artificial intelligence? Uh, and (00:46:26) again, I keep pointing to this other AI (00:46:28) development this week with with Google's (00:46:30) release of Genie 3, where you can (00:46:32) basically give it a prompt and it'll (00:46:34) create, you know, what looks like a very (00:46:37) realistic three-dimensional world, (00:46:39) >> a metaverse. (00:46:40) >> We're getting to what (00:46:41) >> what's that? A metaverse. The metaverse (00:46:43) is is basically a term that comes from (00:46:45) science fiction that is a virtual world (00:46:48) that you explore with your avatar. And (00:46:51) if we can get to that point, it's sort (00:46:53) of like in Star Trek I mentioned (00:46:54) earlier, they have something called a (00:46:55) hollow deck where you could create any (00:46:57) experience you want. How do you create (00:46:58) it? Well, they programmed it by simply (00:47:00) telling the computer, I want to be in (00:47:02) Paris, France in the 1890s, and I want (00:47:05) there to be a bar with a woman in a red (00:47:07) dress. And it starts to create it for (00:47:09) you. And so, you know, we are moving (00:47:12) closer to that point. Uh I think every (00:47:15) year I see it more and more. uh and that (00:47:18) raises the possibility that we may (00:47:20) actually already be inside the (00:47:22) simulation and that that's where a lot (00:47:23) of my work you know has has turned in (00:47:26) terms of AI. (00:47:28) >> Perfect segue into that rabbit hole. So (00:47:29) essentially, especially with how fast AI (00:47:31) is moving and we saw the metaverse hype, (00:47:34) we're now able to create virtual worlds (00:47:38) where we create characters in that (00:47:40) virtual world, NPCs or or characters (00:47:43) controlled by humans (00:47:45) where these characters live in that (00:47:47) virtual world similar to the movie. (00:47:49) What's that movie um with uh Ryan? (00:47:53) >> Free Guy. Is that the one you're free? (00:47:55) >> Free Guy is a great movie to articulate (00:47:58) this. essentially free guy is a is a guy (00:48:01) you're watching him and you think it's (00:48:02) just a normal guy but it's an NPC inside (00:48:04) a game and then later he realized that (00:48:07) hold on um come up what happens in the (00:48:09) movie but I think he starts to realize (00:48:10) that hold on I could do my own thing I (00:48:12) don't have to follow these rules now (00:48:14) that's like a very simplistic way of (00:48:16) explaining it but the the where it gets (00:48:19) concerning is that and where the (00:48:21) argument of us living in a simulation (00:48:23) starts to make a lot of sense is that if (00:48:25) we as humans and we're not that (00:48:26) technologically advanced What's that (00:48:28) Kardashian scale? The the one that Elon (00:48:30) talks about where there's different (00:48:31) levels of energy and you're not even at (00:48:33) level one yet. There's like four levels. (00:48:35) What's that scale called? The Kardashian (00:48:37) scale. (00:48:38) >> Yeah, I think it's called the Kardashev (00:48:39) scale. (00:48:40) >> Kardash scale. There we go. (00:48:42) >> And and it kind of that that scale gives (00:48:44) you an idea of how early we are. And you (00:48:46) know, all these sci-fi movies give you (00:48:49) an easier way of of understanding that. (00:48:52) But if we're already level there's like (00:48:54) the the amount of energy you use then (00:48:56) it's the amount of energy of the sun and (00:48:58) then there's the amount of energy of an (00:48:59) entire galaxy (00:49:01) way down not even at the sun level (00:49:04) >> um not even our own own planet level um (00:49:07) so yeah so essentially if we're able at (00:49:09) this stage to create virtual worlds and (00:49:10) have AI that we saw now with uh uh multi (00:49:14) being able to do all these different (00:49:15) things and we're about getting close to (00:49:17) getting the AGI if we're able to create (00:49:18) those virtual worlds how do we know (00:49:20) we're not living in a virtual world (00:49:22) created by whatever whether humans in (00:49:25) another time or another species created (00:49:27) this virtual world that we live in and (00:49:29) we just don't know it the same way NPCs (00:49:32) don't really know they're in a game NPCs (00:49:34) just operate in the game based on the (00:49:35) whatever has been coded into there and (00:49:37) when we're able to create those virtual (00:49:39) worlds and you've given an example now (00:49:40) with uh um Genie 3 is how easy it is to (00:49:44) create it with AI how easy it is to (00:49:45) create those virtual worlds and create (00:49:47) those agents there could be millions of (00:49:49) them um Then statistically speaking the (00:49:52) likelihood of us living in such a in in (00:49:54) a simulation in such a virtual world is (00:49:57) increased exponentially. Is that am I (00:50:00) understanding it correctly? (00:50:02) Yeah. So that's one aspect of the (00:50:05) simulation hypothesis. And you mentioned (00:50:06) Free Guy where the NPC doesn't know he's (00:50:09) in a in a simulation. Probably the most (00:50:11) popular sci-fi representation of this is (00:50:13) the film The Matrix which came out you (00:50:15) know back in 1999 where Neo was going to (00:50:18) work. He had a job in a cubicle. He was (00:50:21) you doing programming. He was a hacker. (00:50:22) But then he realized the whole world was (00:50:24) virtual. Uh but the argument that a guy (00:50:26) named Nick Bostonramm at Oxford came out (00:50:28) with and Elon Musk quoted this a few (00:50:30) years ago as well. But the idea is that (00:50:32) if AI can create these worlds with these (00:50:35) AI characters, they will create a (00:50:36) billion of these worlds because nobody (00:50:39) creates just one virtual world. Like uh (00:50:41) you all you need is another server. uh (00:50:43) you can just tell it to create another (00:50:44) world and they would do simulations uh (00:50:47) of entire civilizations. And so if (00:50:48) there's a billion simulated worlds and (00:50:50) there's only one physical world and we (00:50:53) can't tell the difference, now that's (00:50:55) the key. If we can't tell the (00:50:56) difference, what is more likely? Are we (00:50:58) more likely to be in one of the (00:51:00) simulated worlds or are we more likely (00:51:02) to be in a physical world? Like for (00:51:04) example, you and I are not really having (00:51:06) this conversation, are we? I'm talking (00:51:08) to my computer and it's sending the (00:51:11) information over the internet to your (00:51:12) computer and you're talking to your (00:51:14) computer. So, we're already having a (00:51:15) virtual interaction. Uh but but that's (00:51:18) one aspect of simulation theory is that (00:51:21) we could be in a virtual world, a (00:51:23) virtual world that is generated by AI. (00:51:25) Now, there's another version of (00:51:27) simulation theory which is called the (00:51:29) RPG version. And in that version, we are (00:51:32) actually playing a video game. So, just (00:51:34) like, you know, when I play a video game (00:51:37) character, I exist outside of the video (00:51:39) game and the character exists inside and (00:51:41) I'm still controlling the character, but (00:51:43) the character might have some AI (00:51:44) components that can go off and do (00:51:46) things, but I'm still overseeing it. Uh, (00:51:49) and that RPG version is closer to what (00:51:52) was happening in the Matrix. You know, (00:51:54) Trinity, Morpheus, Neo, they all existed (00:51:57) outside of the simulation. they had (00:51:59) instead of a virtual reality helmet, (00:52:00) they had a brain computer interface (00:52:02) stuck into the back of their head. Uh, (00:52:04) and BCIs are moving along very quickly (00:52:07) as well, you know, with neural link, (00:52:09) etc. But, but those are the two basic (00:52:11) versions of the simulation hypothesis is (00:52:14) 100% AI NPCs or it's AI creating the (00:52:18) world and us as players uh, our (00:52:21) characters inside the game. (00:52:24) the, as we said earlier, the RPG one (00:52:26) just sounds a lot more appealing because (00:52:27) it gives us some sense of existence and (00:52:30) autonomy. Um, but is it fair to say the (00:52:34) NPC one is statistically a lot more (00:52:36) likely? (00:52:39) >> Well, so if you look at where technology (00:52:41) is today, it's likely if we think we (00:52:44) will be able to get to that point to (00:52:47) where, you know, AI appears conscious (00:52:49) and where we can create these virtual (00:52:51) worlds. And it looks like we're moving (00:52:52) in that direction. I mean, when I first (00:52:54) started writing about this back in 2018, (00:52:56) 2019, I thought maybe we were 50% of the (00:52:59) way, you know, I was confident we could (00:53:01) get 50% of the way there. Today, I'm (00:53:03) confident we can get 70% of the way (00:53:05) there and maybe even 100% of the way (00:53:07) there, which means it's more likely that (00:53:10) there would be these these worlds. Now, (00:53:12) on the RPG side, you know, it starts to (00:53:15) look and overlap with traditional (00:53:17) religion more. I we were just talking (00:53:19) about how you can issue prompts and they (00:53:22) will create worlds. And if you read for (00:53:24) example any creation story in any of the (00:53:26) world's religions like in the Bible or (00:53:28) in the Quran you know basically God (00:53:30) speaks and what happens a world is (00:53:33) created and then God says okay now I (00:53:35) want waters now I want uh trees now I (00:53:38) want animals of different types and (00:53:40) finally he creates the human characters. (00:53:42) Now that sounded ridiculous to any (00:53:44) scientific person a few years ago. (00:53:47) You're like, of course, God couldn't (00:53:49) speak and create a whole world in six (00:53:51) days or seven days. Um, and but it turns (00:53:54) out that today with prompts moving so (00:53:57) fast, this is why I'm emphasizing Genie (00:54:00) 3, there's also this company, World (00:54:01) Labs, is you can actually give it a (00:54:03) prompt to create a world. Now, our (00:54:05) worlds, you know, are very limited (00:54:06) today. Uh, Genie 3 only lets you create (00:54:09) like one minute sequences to explore (00:54:11) these worlds, for example. But the fact (00:54:13) that they're persistent worlds and you (00:54:15) can speak to them, it starts to look (00:54:17) more like uh you know if we're in a (00:54:20) virtual world, there might be something (00:54:22) outside of that world. Uh and how are we (00:54:24) interacting with it? And that gives us a (00:54:26) little more agency as well in in that in (00:54:28) that view version of the simulation (00:54:30) hypothesis. (00:54:33) If we get to AGI, if we arrive there and (00:54:35) you're convinced that we're there, does (00:54:37) that mean that the likelihood of us (00:54:39) being NPCs has incre increased (00:54:40) significantly? Not knowing that the the (00:54:44) ability to create a form of intelligence (00:54:47) is already there, managed to do it, then (00:54:49) it means that humans at a future date, (00:54:52) it's a lot more likely that humans at a (00:54:54) future date has created a a virtual (00:54:56) version of the world or billions of them (00:54:59) and we're just one of those virtual (00:55:00) versions as NPCs. I know it doesn't give (00:55:03) us a sense of agency. It's not really a (00:55:06) it's the the option that is harder to (00:55:08) digest, but from what you've said so (00:55:10) far, it just seems if the simulation (00:55:13) theory is true, the hypothesis is true, (00:55:15) which statistically speaking, based on (00:55:17) what we've just discussed, (00:55:19) it's more likely to be true than not, (00:55:21) significantly more likely. Um, and if (00:55:23) it's true, then the NPC version seems (00:55:26) more likely, unfortunately. (00:55:29) >> Well, it's very possible. And if we look (00:55:31) only at the AI aspect of it, you know, (00:55:34) so what happened to me was after I (00:55:36) started going down this rabbit hole, (00:55:37) then I started to look at quantum (00:55:39) physics and quantum mechanics. And you (00:55:41) know, one of the things that I found was (00:55:43) that they're telling us that the (00:55:45) physical world is not real, that the (00:55:47) world is based on information. And (00:55:49) somehow that information is getting (00:55:51) rendered for us. Uh, and to me it (00:55:53) started to look uh more like a video (00:55:56) game. Like the reason I can we can I can (00:55:59) explore this whole world in Genie 33 or (00:56:01) World of Warcraft is that my computer (00:56:03) doesn't have to render the whole world. (00:56:05) >> It only has to render the little part of (00:56:07) the world that I happen to be looking at (00:56:09) that my character is looking at. (00:56:10) rendered as in just for the audience. (00:56:11) Surrender is allowing allowing you to (00:56:13) really exist cuz if you're not going to (00:56:14) walk so if you got a whole world that (00:56:16) you're playing in and you go into (00:56:17) different areas of that world, those (00:56:19) areas would not really form until you (00:56:21) walk into them because they don't need (00:56:22) to form because there's no one there to (00:56:23) see them. Is that a good way to explain (00:56:25) it? (00:56:25) >> Yeah, exactly. So it becomes an (00:56:27) optimization technique and computer (00:56:29) science is very much about optimization (00:56:31) and so it becomes a way to use less (00:56:33) resources just in the same way that okay (00:56:35) I'm in Phoenix Arizona at the moment and (00:56:38) your computer doesn't have to render all (00:56:40) of Phoenix Arizona it only has to render (00:56:42) this little part of my office that I (00:56:44) happen to be sitting in right now and (00:56:45) even that I don't send all the pixels we (00:56:48) just send information that gets (00:56:49) compressed so a lot of what we do in (00:56:51) computer science is is compression and (00:56:53) optimization but there's also this thing (00:56:55) called the observer effect, which is (00:56:57) part of this is that you look at what (00:57:00) the observer actually sees is what get (00:57:04) gets rendered. And we won't go too far (00:57:06) down the quantum physics rabbit hole, (00:57:08) but there's this idea of, you know, the (00:57:10) double slit experiment where a particle (00:57:12) can go through these two slits and it (00:57:14) looks like it's going through both of (00:57:15) them simultaneously and it's only when (00:57:16) you observe it that it decides which of (00:57:19) those two slits it went through. And so (00:57:21) that looks more like an optimization (00:57:23) technique to me, but that still requires (00:57:27) some level of person playing the video (00:57:29) game. So that's what why I also say that (00:57:32) the RPG version is a possibility as (00:57:34) well. As a scholar of a simulation (00:57:37) hypothesis, I like to draw a axis and (00:57:40) say at one end it's 100% NPCs, at (00:57:43) another end it's 100% RPGs. But you can (00:57:47) have both. In any online role playing (00:57:49) game, you will have characters that are (00:57:51) controlled by humans, avatars, and you (00:57:54) will have NPCs. So, we may be going into (00:57:56) NPC mode, but that doesn't mean that (00:57:58) there aren't players that are watching (00:58:00) this and deciding, you know, giving the (00:58:02) prompts, if you will, to individual (00:58:04) characters. So, it's still very possible (00:58:07) uh that we have some combination thereof (00:58:09) going on. I know it gets a little down (00:58:12) the rabbit hole. (00:58:13) >> It does. It does. And and look, I I know (00:58:15) you've done there's a lot of great (00:58:16) interviews you've done where you talk (00:58:17) about this for two three hours, but I'm (00:58:19) going to put you on the spot. Um rapid (00:58:21) fire questions. Just just thought of (00:58:22) that now while you're speaking. You've (00:58:24) been studying this more than most (00:58:26) people. Um when you first came up with (00:58:29) the simulation hypothesis, (00:58:31) just I'm going to put you on the spot (00:58:33) and ask you for a number. When was that? (00:58:35) And what was the likelihood you had back (00:58:37) then that we were living in a simulation (00:58:39) when you first came up with hypothesis (00:58:41) early on? What year was that and what (00:58:42) was the number? was 2016 when I started (00:58:44) to look into it and at the time I (00:58:47) thought maybe 30 40 maybe 50% when I (00:58:51) published my first article on it which (00:58:54) was back in 2018 (00:58:55) >> where is it up to now (00:58:58) >> it's up to about at least 70%. (00:59:01) >> Okay. More likely than not (00:59:03) >> and that's because of the advances in AI (00:59:06) that we've been talking about. What (00:59:08) would it hit? Do you think once we hit (00:59:09) if we reach AGI and you in your eyes (00:59:12) you're like are you you determined (00:59:13) personally you believe this is AGI would (00:59:15) that go up significantly from 70%. (00:59:19) Yeah, it would go up. It's still not at (00:59:22) 100% but it would definitely go up from (00:59:26) you know 70 to maybe 80% or so. Uh, so I (00:59:29) have these 10 stages and we're pretty (00:59:32) far along in many of the stages starting (00:59:34) off with simple text video games getting (00:59:36) all the way to we're still working on (00:59:38) the brain computer interface for the the (00:59:40) RPG versions of the simulation (00:59:42) hypothesis. And also, you know, we can't (00:59:46) yet generate worlds that are as (00:59:48) extensive as our world today. I mean, it (00:59:50) wouldn't be run on a computer like we (00:59:52) think of it today. I mean, our computers (00:59:54) quantum computing comes in. (00:59:56) >> Yeah. Yeah. And quantum computing is a (00:59:57) whole another rabbit hole that's quite (00:59:58) interesting, but it lets you explore (01:00:00) theoretically multiple possible worlds (01:00:03) and multiple possible options at once. I (01:00:06) I think eventually that's the kind of (01:00:08) computer our universe would need to run (01:00:09) on. And that's what quantum physics may (01:00:11) may be telling us. (01:00:13) >> And the last question is where we are (01:00:15) today. Um you said there's a 70% chance (01:00:18) in your mind that we live in a (01:00:19) simulation. (01:00:21) What would you give that 70% how likely (01:00:23) is it to be an RPG versus an NPC (01:00:25) simulation based on what you know today? (01:00:28) >> Well, that's a tough one because now it (01:00:29) gets to sort of your personal belief (01:00:32) system, right? If if you tend to be more (01:00:35) spiritual, I think you lean towards the (01:00:36) RPG version with the idea of the soul (01:00:38) being the player uh and the the body (01:00:42) being the character uh with the (01:00:44) storyline and you still have free will (01:00:46) to make choices. Uh, so it's an open (01:00:48) question in science whether free will (01:00:51) exists. If you tend to be more of a (01:00:52) materialist, then you tend to lean more (01:00:54) towards the NPC version. I think we all (01:00:57) go into NPC mode. Uh, and it like in the (01:01:01) Sims, if you watch the Sims or if you (01:01:03) play like the it's the bestselling video (01:01:05) game of all time, the characters are (01:01:06) usually just doing things on their own, (01:01:08) but you're also giving them directions (01:01:10) as well. So, I I'm kind of in the middle (01:01:13) there, actually. (01:01:15) Uh, I think you're trying to to hedge (01:01:16) because it's a tough one as well. I (01:01:18) think you what you're talking about. It (01:01:21) is. It is. And I thought I'm going to (01:01:22) put you on the spot. (01:01:23) >> My personal opinion is I would prefer it (01:01:25) to be more of the RPG version. But then (01:01:27) if you've seen weird glitches in the (01:01:30) Matrix, you know, where you've had weird (01:01:32) synchronicities or pre-cognitive dreams (01:01:34) or you've seen ghosts, then you start to (01:01:36) lean more towards, okay, maybe there is (01:01:38) a consciousness that exists outside the (01:01:40) physical body. And I tend to lean that (01:01:42) way. But as a scholar, I study all all (01:01:45) of these different aspects. (01:01:47) >> There's also the argument that we are (01:01:50) what was the statistic. I remember as a (01:01:51) kid I I people would tell me this that (01:01:54) most of what we do on a daily basis is (01:01:56) things that we don't really control um (01:01:58) as we go about our day. So in that (01:02:00) sense, I don't know what the number is, (01:02:01) but in that sense, (01:02:03) >> we're almost at NPC level or there's not (01:02:05) much that much difference between us (01:02:06) what we believe we have agency versus (01:02:08) what an NPC is. Even in our world that (01:02:10) we believe we're agency, we understand (01:02:12) that in that world most of what we do is (01:02:14) NPC like (01:02:16) >> um (01:02:16) >> yeah so the question is do we have any (01:02:18) free will at all? That's the question. (01:02:21) >> What do you think Mario? (01:02:22) >> I don't know man. Like I I'm I'm more of (01:02:24) a person that kind of looks up to people (01:02:26) like you that have spent their lifetime (01:02:28) studying this get the numbers from you (01:02:30) and then based on your answers I decide (01:02:31) whether to freak out or not. Um curious, (01:02:35) do you get um and as we wrap up personal (01:02:39) question, do you get many, (01:02:41) you know, death threats or or people (01:02:43) criticizing you for what you do? Cuz (01:02:45) you're going to shake, you know, what (01:02:49) you're studying and what you're talking (01:02:50) about really shakes up our entire (01:02:54) purpose to exist, whether it's religious (01:02:55) or even non-religious. When you say, "Do (01:02:57) we really have free will and we live in (01:02:59) a simulation?" Do you get a lot of hate (01:03:01) for that or not really? (01:03:03) I I do get some I don't get death (01:03:06) threats, so I'm not at that extreme. (01:03:08) >> You're too nice. You're just too kind. (01:03:11) >> Hopefully that's the case. Yeah. So, I (01:03:13) haven't gotten any death threats, but I (01:03:14) do get a lot of hate for it, but I also (01:03:16) get a lot of encouragement. And other (01:03:18) people, so many people have said to me, (01:03:20) you know, I kind of thought this might (01:03:21) be true, but I didn't know how to (01:03:23) articulate it. Uh, and I'm so glad that (01:03:25) someone actually wrote down, you know, (01:03:27) wrote a book about why this might be (01:03:29) true because it lets them articulate it (01:03:31) in a way. And again, you can view the (01:03:33) simulation hypothesis (01:03:35) literally. We are literally running on a (01:03:38) computer program or a bunch of NPCs. Or (01:03:39) you can view it more metaphorically to (01:03:41) say that it's like a video game. Uh (01:03:44) meaning that the world isn't real. So I (01:03:46) like to say there's four propositions. (01:03:48) One or five propositions. One, the (01:03:50) universe consists of information. I (01:03:52) think most physicists agree. Second, (01:03:54) that information is being computed like (01:03:56) a quantum computer and some physicists (01:03:58) agree. Third, the physical world is (01:04:00) getting rendered for us to appear as if (01:04:03) it's real. We can kind of agree on that. (01:04:05) The fourth is that it's all a big hoax. (01:04:08) Okay, now that is where I get hate, (01:04:10) right? When I say it's all a big hoax, (01:04:11) it's not really real. And then the fifth (01:04:13) is, you know, we agreed to participate (01:04:15) in this host by jumping into the video (01:04:17) game. That leads to both praise and a (01:04:20) lot of people hating me for saying that. (01:04:23) I like that you offer both paths because (01:04:25) I think you'll get a lot of criticism (01:04:26) for the materialistic path but the (01:04:28) spiritual path just allows again as we (01:04:30) said at the beginning of the (01:04:31) conversation a lot of overlap with (01:04:32) religions which makes it easier to (01:04:34) digest this this what is a very complex (01:04:37) theory and what you do is such a good (01:04:38) job at articulating but for people (01:04:40) listening that really want to go down (01:04:42) the rabbit hole um it's a tough one to (01:04:46) digest but I highly recommend your other (01:04:47) interviews I think you've done one with (01:04:49) Rogan as well and you articulate it so (01:04:51) well it's such a fascinating (01:04:52) conversation. But um doctor, I love how (01:04:55) >> Yeah. And I re I recommend people, you (01:04:57) know, check out the book if they're (01:04:58) interested in going down the rabbit hole (01:05:00) >> and I just second edition. Yeah. (01:05:03) >> Oh, really? (01:05:04) >> Yeah. So, the second edition came out (01:05:06) just this last year because when I wrote (01:05:08) the first edition, it was in 2019 and (01:05:11) all of this stuff about AI that we're (01:05:13) talking was just predictions. It hadn't (01:05:15) happened yet actually. So, now it has (01:05:18) and now we're much further. That's why (01:05:19) I've raised my estimate to 70%. (01:05:21) >> Yeah. Well, now now based on how fast AI (01:05:23) is moving, you probably need to come up (01:05:24) with a new version every say every 3 (01:05:26) months at most. Um, but I like the (01:05:30) discussion how we started with mold book (01:05:32) which one would talk about. We're kind (01:05:34) of creating that metaverse for all these (01:05:36) different NPCs to operate in a way and (01:05:38) those NPCs are going to slowly become (01:05:39) RPGs in a way. So I don't know what (01:05:41) these agents are NPCs RPGs. Would you (01:05:44) say those agents are the ones on on (01:05:46) today on Moldbook there would be (01:05:50) NPCs or RPGs because we create the code (01:05:52) but then we let them go loose (01:05:54) >> there would be the multbook agents would (01:05:56) be way (01:05:57) >> it's sort of somewhere in the middle in (01:05:58) a sense that they are NPCs they're (01:06:00) agents autonomously running but we are (01:06:02) also giving them prompts right so it's (01:06:05) like you're influencing the and I (01:06:07) believe in even in the NPC version there (01:06:09) may be storylines that we're given in (01:06:12) our lives or even as an entire (01:06:14) civilization, which reminds me of the (01:06:16) video game metaphor. (01:06:17) >> Less so someone coding us and telling us (01:06:19) what to do, more so someone giving us (01:06:21) prompts in general, allowing us to (01:06:23) operate in this met in this uh virtual (01:06:25) world. All right. Oh, doctor, absolute (01:06:28) pleasure to speak to you. Thank you so (01:06:29) much for your time and I hope to speak (01:06:30) to you again, sir. (01:06:32) Yeah, thanks so much for having me