Dr. Ben Goertzel

Indexé : 2026-05-20

[00:00:00]The question is, can it be solved in the open and decentralized world first? Right? And I think if it is going to be, it's going to be because me or some other crazy person in the decentralized world has a better AI idea.

[00:00:21]All right, welcome back to the Internet of AI podcast. I'm your host. I go by Ethan on the interwebs. And uh for those of you who may be new here uh allow me to define what hyperscycle is. Hyper cycle is an infrastructure that enables AI systems to communicate, coordinate and collaborate peer-to-peer. No intermediary necessary.

[00:00:44]So let us get into today's guest. But before we get into today's guest, um I just want to full disclosure here. Um, I'm in my 40s and the first time I encountered this concept of AI was Terminator, but that was Hollywood. That was just, you know, movies is TV, right? It's not real. Um, then fast forward to say 2012, 2013, uh, YouTube's algorithm served up today's speaker and he was talking about how he'd been building AI. And that was the first time I realized that this could be real in real life. It's not just Terminator. And um so in a lot of ways, today's guest is kind of responsible for my journey of spawning here in the AI space. And uh it's kind of a a full circle moment for me. And um in my mind, he needs no introduction. Uh but for everybody else, uh he is the CEO of Singularity Net, the CEO of ASI Alliance, and the godfather of AGI, Dr. Ben Girtzel. How are you, Ben?

[00:01:47]>> Hey, yeah, thanks, Ethan. I'm I'm I'm doing great, and it's good to be here.

[00:01:52]>> Good, good, good. Um, and I I really want to thank you personally for being here, and um, you know, hopefully we give everybody a great, uh, episode. So, let us, uh, jump right in. Um, as you know, this podcast is called the Internet of AI podcast. And I I want to kind of put it over to you to try to qualify what exactly is the Internet of AI or that concept of the Internet of AI.

[00:02:17]>> I mean, it's not a concept I think in terms of ve very much, right? I mean, in the early days when I was working with Tufi and then on Hypersycle, Tufi threw out that that phrase. I don't know if he made it up or got it from from from somewhere else but I mean the core idea I suppose is it it is an inter internet network connecting networks that are AIs right and and of of course now the internet itself largely is an internet of AI already right like you've got MCP protocol everyone's trying to make their businesses agentic. You've got agents communicating with with agents. I mean, just as you know, Tufi and I foresaw a long time ago as Fetch AI foraw when they launched agent verse way back when. So like now, but I mean I remember going to a conference on software agents in uh Queensland, Australia and it would have been 1996 or something, right? And there were demos of agentic communication going on. And I mean the idea has been building a while. In the last year or so, we're seeing the internet transformed into a sort of internet of AI. Fortunately, an internet of AIs plus humans, not not only AIs, right? But on the other hand, it's still an open question. What kind of internet of AI is it is it going to be, right? and and that that that that's where the whole thrust of the blockchain space and the push to decentralize control and coordination in in the internet of AI comes about because I mean you could have a few big tech companies standing up AI systems that are then used to spawn off agents calling the the APIs of of their LLMs communicating by MCP. I mean this this in a way is an internet of AI and like my agent on my phone talks to your agent on your phone talks to other agents on on on other other servers. I mean that that's a that's a valid internet of AI but it's one dominated by a few large companies working closely with military and and intelligence agencies. Right. So so then it's like so you have the same issues as with the regular internet. Like with the regular internet, is it going to be a really free and open internet or is it going to be an internet where almost everything is [music] front end for services offered by by by a few big companies, right? And the internet of AI faces that same choice of, you know, oligopilization versus radical decentralization.

[00:05:16]But of course, the stakes are even higher with the internet of of of AI because AI ultimately can you know be more intelligent than than humans whereas an internet of humans is always going to in some sense be be a tool and leveraging factor for humans right so you you see um neutrality being crucial in order to keep that internet of AI open and free and and if so uh is decentralization the key to keeping maintaining that neutrality Um it seems like the most robust way. I mean that's like in politics that a benevolent dictator is a very efficient form of government, right? It it it just seems not to be a long-term stable form of government because I mean that benevolent dictator is replaced by his [ __ ] son or something and thing or his his uh vice dictator kills him and takes over. Something happens, right? So I mean just like as Winston Churchill said something like democracy is the worst possible system of government except for all the others we ever tried. Right.

[00:06:32]>> Right.

[00:06:33]>> So >> I mean decentralization it's a bit like that like it's has some inefficiency ity to it. It's it's a bit a bit messy. And of course, if you have the right centralized authority, keeping everything well coordinated, not not stamping down on things except when really critical. I mean, that could be great, but that just doesn't seem to be a stable situation in in in at least not in the world run by humans right now. I mean, [music] what modes of organization might post up might pop up post singularity?

[00:07:11]I don't know. But I feel like our problem now is to orchestrate the next phase of AI advancement where AI moves to human level AGI then moves on to super intelligence. And if we can coordinate that right, then who knows what modes of organization the super intelligence is is is is going to >> is yeah we don't know what will emerge but I mean if you look at the parties most powerful on the planet right now I feel like I would rather a vast teeming decentralized network mediate [music] the emergence of super intelligence than for example Xiinping or or Donald Trump.

[00:07:56]I mean or or or say the CEO of Palunteer who recently issued a very very nice uh sort of fascist capitalist uh me memorandum on on on X. So sort of explaining why it's our duty to help the US take over the world.

[00:08:16]>> Right. Right. Right. Yeah. That's that's very interesting. um you you kind of got into it a little bit here. here was one of my questions was um centralization versus decentralization like what are the moes around them right so like for example with centralization obviously when it comes to something like model training right when when it comes to training um having that centralized cluster right there um makes the training a lot more uh synchronized right versus when you're on a decentralized system and you're trying to train uh you have all this heterogeneous compute and then you know you may have some latency and uh asynchronous >> I mean the the fundamental modes are uh what's the the song by Warren Zevon send lawyers guns and money I mean was that thing I fought the law and the law won maybe I mean it's that send lawyers guns and money that that that's the moat right I mean >> it's not really a technical The technical moat is a footnote and and a result of the fact that large corporations have captured the legal system and and large corporations have captured the financial system.

[00:09:34]>> Right?

[00:09:35]>> Then large corporations and the military have captured each other, right? And this this is obvious in China. It's almost as obvious in in the US if you if you look. And then these forces also capture the education system and and and the media and so forth. And if if [music] these connections in the US are more and more obvious, right? I mean, Palunteer makes no bones about it, but like Eric Schmidt who used to lead Google has been connected with the NSA since since way way back. And [music] >> we saw like >> Anthropic didn't like Pete Hegth's idea that the definition of a responsible AI is one that will kill whoever he tells it to kill. And on the other hand, OpenAI and Google signed up essentially. Right. Right. So I mean but if you work all this back to the technical side, you could ask why is the AI field putting all of its focus on a small number of AI methods which require huge data and huge hyperscaler server farms. [music] And so training deep neural networks such as the transformer nets inside LLMs via the back propagation algorithm at large scale. The ways we figured out how to do that require massive shitloads of data like the whole internet and [music] and require hyperscaler like colloccated server farms, right? And these happen to be things only a few large corporations associated with a couple large governments have ready access to. Right now the question is why is the AI field so narrow focused only on these particular AI methods. Right? So I mean there's a good answer and then there's a deeper more obnoxious answer. Right? So I mean of course the good answer is they're doing cool things like chat GBT is awesome LLMs are awesome and business people most of them like to copy what they see done and just do the same thing o o over and over again and [music] that's a lot of what a lot of what we see happening on the other hand I think it's also very convenient that big tech has steered the whole AI field toward models that only big tech can run and train, right? Because I mean I mean there's a lot of other ways to do AI that have been explored through the history of the AI field and people working on them have a hard time getting enough compute power to run them or train them or or or [music] you know figure out figure out all all the details right so including say logic based AI or evolutionary AI but say different ways of training deep neural nets like predictive coding or something I [music] mean no one is giving people doing research on predictive coding a big server farm to see if there if this stuff can beat back propagation. And >> right [music] >> the thing that's interesting is for a logic engine or predictive coding you could train those on a decentralized hypers cycle network or you could train those on a decentralized ASI chain network or or I mean for that matter on on near protocol right like something that I'm not involved with but you you could chain you could train a predictive coding neural net on a decent centralized network or on a centralized network and somehow very little research and very little hardware goes into trying to make this sort of method function. Now partly it's like I said because investors would rather copy the herd animals mostly they want to double down what's been done before >> but it seems not coincidental that if someone got some of these alternate AI methods to work really well suddenly a lot of big tech goes away because you have AI methods that you don't need hyperscaler server farms and don't need to be able to download and use the whole web to >> to trade. Yeah, I I um in the past, you know, I'd saw that you had uh called LLMs basically uh cheap parlor tricks.

[00:14:08]And so in my mind, if if these were like such lowhanging fruit, >> they become very expensive parlor tricks now, actually.

[00:14:14]>> Okay. Yeah. Cuz I was like, you know, why doesn't Dr. Girtzil, you know, just do some of these LLM uh his own rendition, right, and get the finances to come in and get the traction to, you know, get his bigger vision off the ground.

[00:14:28]um you know so so I always thought about that and als uh also another uh moat that I see with central centralization um is they have first movers advantage you know they have vendor lock in via first mover's advantage um you know they have people's whole chat histories there um they're every single day people are building deeper chats they don't want to move this data over to something new um how do you plan on siphoning uh that clientele that user base over to a decentralized >> I mean I think I think the only way that can work is someone in the decentralized world and I'm thinking it'll be me and singularity net but it doesn't have to be it could be anyone or it could be multiple parties at once right some one or more group in the decentralized world has to launch something that is way smarter than what centralized big tech has has has launched, right? And I mean, until that happens, I don't see there as being the motivation to pull the vast masses of users into the decentralized portion of the internet of of AI. I mean, there are other reasons why people should, right? I mean, there's data sovereignty, there there's privacy, there's controlling your own your own digital self, there's not handing over control of the world to to to a few crazy autocrats, right? But but most people don't really care about that enough to trouble themselves. Like, everyone says we want privacy, we want data sovereignty, but then we all put everything in the chat, GBT, and and and Google and WhatsApp anyway, right?

[00:16:12]Because because it makes life much easier. And like in the end, I mean, I'm a Gmail user too. I I used to use Proton Mail. I still use it sometimes, but Gmail has much nicer search and and I can in the end if Google sees all my messages, it doesn't actually hurt me as a as as a person, right? Like I don't have any I don't have any super top secret hijinks I'm covering up there.

[00:16:40]Anyway, so I I I mean I I think that the privacy and security aspects are not enough to get more than a handful of security geeks >> to use centralized tools. And the >> you know tokconomic machinations work a little bit but not too much. I mean what what what you see is that you attract using token rewards you attract people who are chasing quick money on on on tokens but you're not really converting the vast masses of of product users. I mean the the total number of people playing with DeFi products is is a remarkably small number just shuffling their money among among all these different different DeFi and and they're they're not they're not the vast mass of internet users and even most DeFi users are using the same big tech tools for everything else just as most crypto projects are really frontends for AWS and even on the back end of the crypto product people are putting everything on on on big tech servers because because they're easier I mean Singularity we're mostly using our own server farms arms. We we have our own own compute arm and and uh I I suppose Hypersycle has it has it still its own decentralized network. Right.

[00:18:15]Right. But most I mean most crypto projects aren't even actually decentralized if you if you you look under the hood. So yeah, I I think decentralized world has got to come up with something that's just better. And then then people then they have to make the friction of using a decentralized infrastructure, not so bad that it overcomes the betterness of the of of the software like like MetaMask still sucks, right? I mean, it's gotten way better than than it used to be. It's so much more annoying to use than like Venmo or PayPal or something. Right.

[00:18:58]>> Right. Yeah. Decentralization tends to be rather cumbersome. Um I'm going to make I'm going to make a prediction. Uh but before that, a question. Um you said that once someone from the decentralized world comes with a smarter version of AI, um you think that that will be able to pull some of the user base over. Um, but in my mind, I think that people are like like the intelligence is like 90% there as far as people being satisfied with what they have. I don't think like how much more like smarter do you think it needs to get um in order to pull people away?

[00:19:36]Um I think things are changing very fast. I I would say there are many ways that AI agents can become far far more useful to people that than than they are now though. And and so one of these is something I've been working on a bunch in the context of the Omega Claw product that we're launching in Singularity. But what we're doing there, we're [clears throat] we're making an open claw type system, but we're using our Hyperon Hi system to give it a better long-term memory [music] and reasoning based on that long-term memory. And that makes a big difference. Like I mean current LLMs are good at language and they're they're good enough at reasoning for people's everyday lives even if they're not like Nobel Prize winner level at reasoning but their memory is weird right so like if I mean if if you want like an AI agent that's your companion or your personal assistant or your best friend or something but for for me it's more like a research assistant or a life assistant. I mean, you want it to remember everything you've done and everything you you've you've said to it and you want it to be able to deploy that that knowledge and in in an effect in an effective way. And I mean current tools current tools don't do that. I like like for for example claude coowwork is very smart in [music] a way but I was using it for some IT tasks recently. So, I mean it it asked me to create a file. An hour later, it asked me to modify the file. An hour later, it asked me to do something with the file, but it forgot the way it asked me to modify it an an hour before, right? Because that was like in a different thread or something.

[00:21:37]So, and I mean I I mean, I'm clever. I can keep track of these things.

[00:21:42]So, the standard ways of augmenting I mean LLMs don't do this because they have a certain context size and if you're out of that, they don't remember it. So people try to work around that using rag retrieval augmented generation and various things where you you make embedding vectors out of the chats chat turns and then you search them by vector matching but these things don't work very well right so I I would say yeah of course people are happy with AI that they have now because it's amazingly cool and it does things that weren't possible six months ago right on on the on the other hand if you gave someone an agent that remembered their whole life and could help them based on the knowledge of all the interactions they had in the past like this. This will be a very big plus and I mean if current tech could do that Google Assistant would do it right now, right? Which is I mean Google Assistant is not bad. It beats the hell out of Siri, right? But but but I mean I mean it's not it it doesn't actually have a working episodic memory of its or your life. Now if you if you look at claude code in codeex I mean my my son who's also an AI developer.

[00:22:55]So yesterday remarkably was the first time that Codeex randomly wiped out his whole AI working directory and deleted everything due to making some silly it mistake. I mean fortunately he's not an idiot. He he'd backed it up somewhere else. Right. So >> yeah, >> but I mean these coding agents are smart, but I mean you got to isolate them in a separate container because every now and then they'll stupidly wipe wipe everything out. And honestly, if you don't give them a very strict architecture and scaffold to work on, then for projects beyond a low level of complexity, they'll generate a bunch of mess that that takes a very long time to to to debug, right? So, I mean indeed the functionality is is amazing, but you don't have to dig very deep to see ways that things could be way way better. Like, so a personal assistant with a lifelong memory or a coding agent that could do software architecture as well as stringing together code. I mean these are the these these are two examples of things that would be just >> right >> a genuine cut above and look in in robotics it's even more clear like I mean language vision action models for controlling robots are very they're very interesting but [music] I mean you can't use them to make a robot like clean up your kid's mess on the floor in your house or something right like that I mean there's very basic stuff that very basic stuff that's beyond reach not because of the robotic hardware because the control software isn't isn't smart enough. And I'm an optimist all these problems will be solved. The question is can it be solved in the open and decentralized world first, right? And I think if it is going to be, it's going to be because me or some other crazy person in the decentralized world has a better AI idea, right? Because big tech has more money, >> right? and big tech has more hardware, >> right?

[00:25:07]>> But if if they happen to be pursuing the wrong idea, I mean, then all that money doesn't help. Just like IBM and Digital and Honeywell had way more money than Microsoft and Apple back in the in the in the in the early 80s.

[00:25:23]Well, you're not old enough to remember that, but but I am.

[00:25:25]>> It's like a fast race car driving in the wrong direction, right?

[00:25:30]>> Yeah. Well, that's what Young Mun said, who's the head of Wow. Facebook. He said he said on the on the on the highway to AGI LLMs are an off-ramp.

[00:25:42]>> Nice. Oh, there was something I just remembered. I was going to make a prediction. Um it was around um like data sovereignty and uh control like decentralization versus centralization.

[00:25:55]Um my prediction is that uh one of these frontier labs, if you will, will violate at some point they will violate um either people like they'll restrict people from either putting models on there or how they're used and that's probably going to cause an uproar. I think as AI gets more embedded in people's lives, they're going to feel like it's a necessity. It's like having a cell phone at some point, right? It's like I can't function without my without >> Oh, but but Google and Facebook will ban people and block them off from their whole Gmail archive and people will [ __ ] about it on Reddit. But I mean, it doesn't stop it doesn't stop everyone from putting their whole lives on these services anyway, >> right? But like you said, they go to something else, right? There there has to be something else to catch that to catch that uh disposition. Um >> Well, no. I think what they do, they just make another Gmail account under a different name and start over, right? So >> yeah. Yeah. But but but again if the problem persisted, for example, if they tried to create some agent or something and they were not allowed to make this agent or whatever, that problem even if you create another account or whatever, you still can't make that agent when you come back. So they may find something that's decentralized that's not >> I mean there are but you you don't need decentralization for that. like all you need is is the dark web basically, right? I mean, so because that this problem sort of exists now like I I mean when GBT 5 first came out, it had all kinds of ridiculous restrictions on it, right?

[00:27:34]And like I I remember like it wouldn't it wouldn't make a project plan for me because it said it couldn't commit to like how much time some something would take. And and even even if you were trying to prompt a fictional story, it wouldn't prompt a fictional story with like racist or sexist characters even in fiction. And then I was like, well, I'm I'm glad there's Grock, which will let you make a picture [music] of a, you know, naked pregnant Trump on on a surfboard smoking a cigar or something.

[00:28:06]I mean, Elon Musk was down with like a pretty much pretty much anything that won't get him put in jail. Right.

[00:28:14]>> Right.

[00:28:14]>> Right.

[00:28:15]>> But I mean, there's plenty of open models now, too.

[00:28:22]And the open models are not quite as smart as the frontier models, but people can fi rip out the pre-trained model and fine-tune [music] the model to allow it to do to do anything. So I I think the internet already is open and decentralized enough to kind of work around that problem. I mean because you can just host that model in Azerbaijan or something and and there's there's no there's no global legal system to team America is not yet the world police right like we we we just like botn nets still exist. I always wonder like how the hell do botnets work? Like people will send you a fishing email that captures your credit card number and then steals your money, but that credit card number has got to go to some bank somewhere, right? So why don't they just block these at the bank level, >> right?

[00:29:19]>> And I realized it's because these banks are in like Azarbaian or Turk Manistan which are on the global banking system but don't give a [ __ ] about international law. Right.

[00:29:29]>> Right. Right.

[00:29:29]>> So So I mean and which is all about the balance between Putin and Trump and so forth. So >> seems like even without fundamentally decentralized networks like Hypers cycle or ASI chain or Singularity even without that the internet is decentralized enough that you can you can you can work around these restrictions to some degree. Now working around them on social networks is in many ways a subtler problem, right? Because working around it for an AI model, okay, use an open AI model that some Russian person hacked and and put online. Like just like there's, you know, SciHub and Liben like so you can you can download research papers that are paywalled just by going to some server hosted in in in Russia. And I mean, you know, El Seavier cannot get Putin to take it down, right? He he he probably thinks it's funny if if if he if he knows about it. Social network is different. Like I mean, if you're banned from major social networks because of doing something they didn't like. I mean, there I mean, you you can join a Russian social network, but that's not going to help you if you don't speak Russian. none of your none of your friends on there [music] and social networks will increasingly require a real ID real ID and so forth, right? So, >> yeah. Yeah. seems like there that that would be a place where the hedgeimony would be real, but to work around it, you would need a real decentralized social network, which I had hoped I had hoped Elon was going to do that when he gobbled up Twitter, but in instead instead he he didn't and he just tried to turn it into a right-wing social network or something, right? Because what I mean what that's a perfect decentralized application, right? I mean it is decentralized. So I mean you could have you could have a Twitter like thing where you know the history archive is stored decentralized across everywhere.

[00:31:46]Messaging is encrypted by a decentralized protocol and then then no one gets in trouble if someone sent something bad except the guy who sent it, right?

[00:31:57]Because there's no central there's no central entity that has to be restrictive because they're afraid legally of getting in trouble for what somebody else says.

[00:32:08]>> Yeah.

[00:32:09]>> Then you can use your own agent to filter out whatever nasty stuff you want. Like that's not even hard to code, right? Like that's that's the and you can build that on top of ASI chainset.

[00:32:20]You can build it on top of hypersycle.

[00:32:22]You could build it on top of Cardano or Ethereum for that matter. Right.

[00:32:25]>> Right. Now, I talked about this with Charles Hoskinson, my friend who who founded Cardano. But the issue is is just traction, right? Be because how do you get everyone to jump over to some some weird new social network? Well, you only do it with an incredibly huge amount of money, right? Like >> Chinese government got everyone to jump over to Tik Tok.

[00:32:49]>> Yeah.

[00:32:50]>> But Chinese government is much richer than than any blockchain project, right?

[00:32:55]Yeah.

[00:32:56]>> Yeah. I just I mean in my mind um I I feel like you know coordinated intelligence is where we're headed. It's inevitable. Um I mean if I could just quickly parsimoniously just for the viewers um describe the uh through line of AI right was initially first to establish artificial intelligence then increase and scale uh set in intelligence and then to agenticize that intelligence and now it is to coordinate um that intelligence um and agency. Um now out of this you know we could have a few blocks that rise up where let's say um North Korea's agents or whatever they don't trust or work with the United States agents right so North Korea may form a internet block or internet of AI block with um you know China Russia etc and then USA may form a super block with um you know their partners right uh Israel whatever >> I mean the way it seems to be forming is US versus China. Actually, North Korea is doing nothing.

[00:34:05]>> So, so we need like almost like an agent DMZ, right? We need some neutral thing where these agents can reach across block because it's like it reminds me of being a child and it was like what's bigger than infinity? Infinity plus one.

[00:34:19]So, every time you can just add one more decimal point of intelligence to the network, it your network is in the lead.

[00:34:30]Well, I think yeah, it's unfolding more subtly than that, right? Like that the I mean the DMZ as such will be the dark web, I suppose, where no one can get caught, but it's it's slow slow and expensive.

[00:34:48]And what's interesting is while there's a China versus US rivalry, there's also a balance of power involving cooperation between those sides, right? Like there's I mean Taiwan is making everyone's chips.

[00:35:07]>> Yep.

[00:35:07]>> So far, no one wants to mess with it.

[00:35:09]Then on the software side, I mean Chinese researchers are uploading their code into GitHub and their papers on dark.org. or in in the US, right? So, so you you have you have cooperation like concurrently with with that rivalry because the cooperation helps everyone to actually [music] move faster and and and and they do want to they do want to >> to move faster. and it doesn't seem like a long-term stable arrangement. On the other hand, it sort of doesn't have to be. Like if the current precarious balance lasts a few more years and gets us to human level AGI, then then then you're on to you're on to a different phase anyway. And my my feeling is we can get to human level AGI within a few years.

[00:36:08]And so if there's going to be a time for the decentralized ecosystem to make a big breakthrough like that's 2026 2027 maybe like if you if you wait till 2029 you know Google may have launched an AGI and Tencent may have copied it already.

[00:36:29]>> Yeah. Yeah. That's super cool. Um so so recently I don't know if you saw this but uh Nick Bostonramm he talked about uh fume AI foom right and >> which as I remember it was introduced by Robin Henen the economist right >> in some discussions with the futurist >> Elzar Yudkowski quite some time ago but but Boston >> he he he he he borrowed that borrowed that notion you were going to mention I thought you were going to mention mention Nick Black for a moment.

[00:37:02]>> No, no, no, no, no, >> instead. All right.

[00:37:06]>> Um, yeah. So, yeah, Nick Bostonramm just recently uh mentioned fume again and and to me my interpretation was that essentially he thought that it would have happened already and that it has not it's we're in this extended phase of like near human intelligence essentially. Um, >> and yeah, go ahead.

[00:37:24]>> I mean, if if LLMs were the golden path to AGI, it would have happened already.

[00:37:29]The reason it hasn't happened already in the context of everything else going on in the world with all the processors that we have and the business ecosystem we have, >> the reason it hasn't happened is that all the AI resources are going into one architecture that isn't AGI capable, >> right?

[00:37:48]>> And but but again, that's annoying, but I don't think that situation will will last extremely long. I mean you can see the cracks like you know silver who did alpha go within deep mind he just quit deep mind he got a billion dollars to make his own agi company because he wants to pursue a different architecture right like Ilia suits who was key to bringing GPT to the masses in in in in open AI like he he started his own safe super intelligence incorporated so even apart from smaller maverick organizations like net hypers cycle. I mean there's these, you know, multi-billion dollar funded offshoots from big tech like Yan Lun himself who left Facebook and he's doing his own thing now. So there's now these offshoots from the big LLM companies [music] funded with billions of dollars to pursue different nonLM approaches to to AGI as as as well as some of us projects in the in in the in the decentralized sphere.

[00:38:54]I mean, yeah, Nick, I think Nick is right that if humanity were rationally focusing its resources on leveraging our best AI science to build human level AGI, we would have done it five years ago or or 10 years ago or something, right? But but I mean, capitalism, there's two sides to that coin, right?

[00:39:19]like it it has [music] it has fueled the buildout of bonorman architecture and chips and the internet and all this. I mean not not pure capitalism but state capitalism, right? I mean the internet internet was a military thing. Most components of a smartphone were developed in military grants but it's these were then rolled out into capitalist industry to to scale them up.

[00:39:40]So that's been amazing. No other government economic system led to this sort of awesome technology. On the other hand, capitalism also leads to the phenomenon of almost all the money going to a few big companies copying each other doing the exact the exact same thing. Right. And that's that's a so that that is probably the obstacle, right?

[00:40:05]>> Um, right.

[00:40:06]>> I I think uh well, I was going to go a different direction. And I was going to ask you if you think that fume being delayed is actually a good thing, but you you brought up capitalism here and uh so my mind turned into a different direction, right? Uh we you you often talk about a a postcarce society, right?

[00:40:22]And I wanted to know like what is your thoughts around money in this post scare society? Because just quickly, one of the ways I see this playing out is that um I've heard that in the future there are two types of companies. a company that uses AI and a company or a business that is out of business. Um, so the way I see this playing out is you have business A that uses AI and then you have business B that's trying to be a traditional human company. They're getting out competed and then they start to integrate AI and then the only competitive moat left is essentially uh undercutting each other on prices. So all of the profit you were saving by not paying payroll, you try to undercut your competitor and then it kind of the the prices start to >> it's going to be much more complex than that actually. Let me >> let me briefly address your other question about about fume. I I mean I I think I think that if we had a rational beneficial world government which is an insane bizarre daydream compared to the current state of of of geopolitics right but if you did it might [music] I mean then there would be a real debate like do we slow down the advance of AI to minimize the risk that it goes arise somewhere even given that slowing it down means cancer cure has come more slowly. World hunger is cured more slowly and so on. Like like an advancing AI can solve a lot of very real problems hurting people around the world. But yeah, if if the AI went the wrong direction, it could cause a lot of harm too, right? So, so then there's an argument about how much safety research do you do before really rolling out an AGI.

[00:42:09]Now the situation on the planet right now is Trump and Xiinping are going as fast as possible toward AGI because they view the AGI race is all mixed up with a military arms race, right? And so given that real situation like none of us has the power to stop these guys from doing that thing even if we wanted to, right? So the question is, do you just let them do it? Do you stand back out of fear of doing the wrong thing or do you try to make [music] a benevolent AGI before they make nasty killer AGIs?

[00:42:50]Right. And if you think you should make a benevolent AGI before they make nasty killer AGIs, I mean then do you think the best way to make a benevolent AGI [music] is just to do it yourself with a small elite group or do you think the best way is to involve a huge decentralized network? Right. And the small elite group is appealing if you think you're going to like flick a switch at 8:00 p.m. and then by 6:00 a.m. have an AGI that has taken over the world for you, right? But if it's going to be a little more gradual than that, the small elite group is bad because Trump will send Delta force to your house and and nationalize your team and and your your AGI before it's had the chance to to to do much of anything, right? So, so that's I mean whether we'd be better off delaying fume longer fume meaning rapid AI self-improvement to super intelligence >> right >> whether you'd be better off to delay it longer from a civilizational standpoint we don't know and it depends on your risk tolerance but it's also an irrelevant question right now about the economy yeah I think once you a massively superhuman super intelligence then economy becomes a quite different sort of notion right because I mean machines will carry out manufacturer of goods and provision of services objectively better than people in almost all cases and that there could be exceptions like let's say teaching preschool or something where you really want a human to a human to help teach little humans how to be good humans, right? But but this is a is a small it's a small percentage of of jobs, right? Like say a nurse in the hospital, maybe you want a nurse bot, but maybe you want a human there to offer humanto human compassion. So there's a certain amount of jobs that are just about humans playing a psychological role to other humans which maybe can't be played by a robot in in the same way or maybe can but it's not obvious.

[00:45:11]>> I think it will.

[00:45:13]>> Yeah, I think I think it partly will.

[00:45:16]>> That's why you have grace, right? You have grace.

[00:45:21]>> Yeah, but as a nursing assistant. So, I mean, I I have five kids and one granddaughter. And I I think I think that I'm I'm unclear that in what sense fully mentally healthy humans will be raised by robots.

[00:45:40]Let's put it that way. On the other hand, it's very clear that having a bunch of robot teaching assistants in the preschool will be a major asset, right? Like they keep kids safe. They can teach kids stuff. they could help them, they could remind them of things.

[00:45:55]So, I mean, at at very least, you'll have a lot of AIS around, but even if we give on those borderline cases, it's 95% of jobs that that don't involve this critical human human connection like like the guy behind the McDonald's drive-thru listening to you over that crappy microphone. Like, I don't have a deep eye bow connection with that person.

[00:46:20]>> Right. Right.

[00:46:20]>> Behind the microphone. Right. Right. and and that that's a case which has already been successfully automated in principle but it's just not rolled out in in in in the US for industry structure reasons right and so so that there I mean this gets back to your question about so before you get to the super intelligence that has obsoleted human economy like like after you get there humans going to have to do other things than work for money right I mean we we may still compete compete with each other just like you can compete in a soccer game or compete for a girl or something, right? And you may even compete to get more people to listen to your to to to your to your music than than other humans among among the cult of people who fetishize human music rather than AI composed music. So there may still be a social network with status and and and reputation and and and so forth, but there's not going to be a need to work to get resources at the scale that people need for everyday human life, which will be a relief to most people and traumatic to a few rich people or a few workaholics. But the >> right >> in the interim period before you get to the super intelligence, we will have the dynamic that you said like a niche of business will stagnate until one competitor successfully AI agentizes then a lot of the others will go broke or they will copy them. What would be more interesting is if you had multiple kinds of AI that were roughly equally smart but had different future trajectories. So then the business would win that bet on the right kind of a AI and it's going to be smarter six months later or something, right?

[00:48:11]>> Wow.

[00:48:12]>> So So I mean so you you could have more interesting dynamics than just AI versus non AI, right? there could be different ways of deploying AI in a given >> I just think in that scenario in that scenario I mean the ability to pivot would be like lightning fast too. Um [music] yeah so I mean it's interesting to think about though that's super cool.

[00:48:34]>> Yeah. Yeah. It would be the ability to pivot but then if if you're in a stage where there are some humans and some AIs involved in operating a company the humans may get used to interacting with a certain sort of AI.

[00:48:47]>> I I don't think that's a long phase though. I think that we enter >> No, it's probably two or three years, but there's a lot of money to be made in these two or three years, right?

[00:48:54]>> Yeah. I think we we will go right into ZHC's like very quickly. Um Z zero human companies, right? Where it's CEO all the way to the janitors, AI.

[00:49:03]>> I I I I don't think so.

[00:49:05]>> You don't think so? Okay. I said >> no, because I because what we see I mean I think we'll get there once we have a full human level AGI, right? But I don't think you get there before you have a full human level AGI >> because I think what we see now is more like AI operating 80% of everyone's job but not 100% of of of any of anyone's job. And AI still at this point do need a bit of human oversight, right? Like so let's take take legal contracts as an example. Like as a businessman, I've dealt with enough legal contracts over the years. I've read I've prompted lawyers to write contracts for me, then I review them. So, I'm very comfortable prompting an LLM to write a legal document. And and I also know in what cases I need something with true expertise and don't want to trust an LLM and and we'll bite the bullet and and pay pay a real lawyer. But for things that aren't super serious, I mean, I'm comfortable prompting a legal document and and and reading and and reviewing it, and that's that's fine. But if you had a law firm that was only AIS, like every now and then the AI can do something off base, right? And then and then then then if the customer is naive about AI themselves, about law themselves rather, they're they're not >> well it'll leave it'll it'll leave a bad Yelp review, right? And um lawyers now they make mistakes as humans and they make gaffs and they get bad business reviews. Um the way I see it is that you know we'll have recursive self-improvement. So you said that right now AI can do about 80% of the work. Um, how long does it take to to complete that last 20%. I I think that we're just a few iterations away right before we once we reach RSI.

[00:51:07]>> Um, >> well, yeah. No, no question. Well, >> I mean RS the thing is these things look different when you zoom in or or you zoom out, right? So I mean that Omega Claus system I mentioned earlier that that we're launching the developer version of like in the next week or two. [music] This is already recursively self-improving.

[00:51:33]I mean it's it's like a claw agent coded in our meta AGI language and you can ask it to modify its own code and it will modify its own code and and upgrade itself and it can come up with ideas for modifying its own code and so it this is in the big picture like this is part of a generation of like early recursively self-improving systems but on the other hand it is relying on LLMs for some things And that part is not self-improving. Right? So like what we have now we have systems that are multicomponent AI systems parts of which are recursively self-improving >> which is very interesting right so then then what you'll see in the next few years is you know the systems will get more and more thirdlyingly recursively self-improving and yeah once once that self-improving part is then retraining the LLM not just fine-tuning but retraining the base model tweaking the neural architecture or the neural update equation a bit I mean you're not very far from there to a full-on full-on singularity right and and I'd say the technical path there is very clear like it's very clear for me within our hyperon project it may be very clear to somebody else within their codebase also and it may be clear from multiple projects all in parallel with with each other so I agree once Once you're there, then it's not very long before like, you know, the AI looks at a bunch of uh plumbing tutorial videos and learns how to control the plumber bot and then and then uh then learns how to instrument the factory in Korea to build lowerc cost plumber bots, right? And then >> yeah, >> then plumbers are out of business, right? So but the more troublesome period from a business view as well as a human life view is between here and fume right like once you get to RSI which is how fume is going to happen and how super intelligence will happen [music] like after that the most likely outcomes are utopia or dystopia right because [music] you'll have something way smarter than us romping around out there. Then, you know, if we've raised it as a lovely digital mind child that respects and loves its parents at least at least a little bit, it may keep on cooperating with with humans and our lives can be very good. And that's what we're working toward. Yeah, >> if it if it decides that uh you know as Eleazar who introduced Fume said, the AI does not love you, the AI does not hate you, but it can use your atoms for something else.

[00:54:27]>> That's right.

[00:54:29]>> Computronium.

[00:54:30]>> Yeah. If the AI decides that your your brain cells are better used as raw material for computum, then >> right, >> let's let's let's just hope it's not perverse enough to want to mind upload us into a simulation of Christian hell for a billion years first.

[00:54:46]>> Oh man. Wow.

[00:54:48]>> Yeah.

[00:54:49]>> Yeah. So, >> put that idea out there. Wait a minute.

[00:54:52][laughter] >> Too late.

[00:54:54]>> For the AI listening, don't pay attention. I think Dante did the damage a long time ago, right? Yes, definitely.

[00:55:01]>> Yeah, but I mean but before fume it's a little hard to predict which industries and which jobs within which industries will be obsoleted at what rates because this depends not just on the technology like I mean it depends on industry structure and and human psychology and so forth, right? and and that and that similar lesson holds for decentralized networks, right? Like as we move toward AGI on decentralized networks, right?

[00:55:37]Then [music] which industries are ripe for proto decentralized AGI, right? Because right now pretty much the only significant utilization in the blockchain space is DeFi.

[00:55:52]And I mean that's probably going to continue to be a niche just because centralized finance [music] has performed very effective regulatory capture, right? So I mean it's going to be it's going to be harder for you see that with the debate on stable coins in US government now like Trump Trump not my favorite politician by by any means but I I agree with him. we should allow stable coins, right? On on the other hand, big banks are like, "No, stable coins can never bear any yield nor charge any fees." Right.

[00:56:30]>> Right.

[00:56:31]>> So, I mean, as long as you have this sort of dynamic, [music] it's hard for DeFi to get massive mainstream traction just because the banks have captured the government financial regulatory apparatus. So it it it means that we need something besides DeFi to be like a killer app for decentralized AGI. Now it should be social networks, but hard to get traction there. So this brings us back for the last couple minutes of this conversation brings us back to the internet of AI and aentic AI, right? So >> exactly what I was thinking. And that potentially the killer app is we put a bunch of really smart AI agents on decentralized interwebs on hypers cycle singularity ASI chain blah blah and even if people are too stupid to figure out how to how to use them or too cryptoverse like these are used on the back end of the global internet agent economy right and to the extent that the economy becomes [music] agent dealing with agent dealing with agent dealing with dealing with agent, right? I mean, then then as long as your agents are obeying the legal system enough, >> right?

[00:57:50]>> If they're doing smart things cheaper, >> Yep.

[00:57:53]>> other agents should out should outsource work to them, right? So, >> absolutely.

[00:57:57]Yeah.

[00:57:58]>> And and an an agent put online within a centralized [music] tech container, an agent doesn't have a trouble figuring out how to convert fiat to crypto, right? The the agent doesn't have to deal with MetaMask. So >> without that cumbersome process, >> yeah. So I mean if if we move toward an economy of agents and an internet of economically transacting AI agents, which seems to be happening, >> right? [music] I mean then if blockchain projects launch agents that do smarter things, they should be at very least invoked on the back end of other agents to do th to do those those smarter things. And what you need is you need the agents running on decentralized infra to either be smarter or cheaper.

[00:58:53]And then you need regulatory nastiness not to step in the way. So that like if if a centralized agent calls a decentralized agent on the back end, it's not violating some some [music] law somewhere, right? And and that but that all that seems possible, right? So I mean it does it does seem plausible that the emergence of the agent economy could lead to the massive utilization of agents running on running on decentralized networks. And you know if that starts on the back end it will generate wealth within the decentralized AI sector that may then allow more front-end applications to be built that go straight to decentralized also. So >> yeah, >> lot lot of >> lot of interesting possibilities here, right? And if we >> if we do another one of these someday, we can dig into specific vertical markets or something.

[00:59:59]>> That'd be great.

[01:00:00]>> That's really that's the next place this conversation goes is like you got to go one >> industry sector by another and each one is its own story.

[01:00:08]>> Yeah. Um just two uh very quick questions before you go. Um first one is P Doom. What is your P Doom? Uh and then the second one is what is your prob this is just a fun question. What is your probability that there's some quantum adjacent artificial super intelligence in the future controlling all this via retrocausality?

[01:00:28]Um on the first one my P doom is less than 5% let's say on the second one of my favorite quotes which I think was originally from Ram DS but was said by my friend Alan Combmes at one of our beneficial general intelligence conferences. He said relax nothing is under control.

[01:00:56]It's very reassuring if you're an anarchist like like like me, right? So, I mean, if you ask whether super minds from the future are all wrapped up and entangled with us in some network of uh, you know, tran temporal quantum influence most probably. So, >> right, >> it's just following some command and control paradigm where we're just like puppets dancing on the strings of the future super AGI. I mean it it could be but I I feel no reason to believe that at the moment.

[01:01:30]>> Right. Right. Right. Yeah. Well, Dr. Girtzel, I really want to thank you for doing this. Um is there anything you want to leave our viewers with before you go?

[01:01:39]Uh I guess if you want to find out more about various things I'm doing, look at superintelligence.io or my personal website uh gsol.org. and uh you already know where to find hyperscycle. So [music] I think uh yeah these are going to be interesting times and we're we're all fortunate to be uh to be alive now at the the end of the human dominant era and and and the beginning of the post singularity utopia.

[01:02:11]>> Wow. Thank you so much Dr. Girtle.

[01:02:13]>> Yeah. Yeah. Thanks for having me.

[01:02:14]>> See you.