1 00:00:05,120 --> 00:00:08,799 Speaker 1: In the early days of computing, everyone thought about a big, 2 00:00:08,880 --> 00:00:13,680 Speaker 1: colossal mainframe computer that had all the data, and eventually 3 00:00:13,720 --> 00:00:16,960 Speaker 1: the model switched and what we ended up with is 4 00:00:17,440 --> 00:00:23,079 Speaker 1: billions of individual small computers, all networked and operating as 5 00:00:23,079 --> 00:00:27,400 Speaker 1: a collective. Might AI go the same way, where we 6 00:00:27,480 --> 00:00:30,760 Speaker 1: move from giant, large language models owned by a few 7 00:00:30,840 --> 00:00:36,199 Speaker 1: companies to lots and lots of individual AI agents that 8 00:00:36,280 --> 00:00:40,519 Speaker 1: are decentralized and operating together. And what does this have 9 00:00:40,600 --> 00:00:44,040 Speaker 1: to do with solving diseases? And why in the future 10 00:00:44,080 --> 00:00:47,120 Speaker 1: we'll have an army of agents solving problems for us 11 00:00:47,120 --> 00:00:49,440 Speaker 1: out there. And what to make of the fact that 12 00:00:49,479 --> 00:00:53,600 Speaker 1: we will have created a whole new species that runs 13 00:00:53,680 --> 00:00:57,480 Speaker 1: at a timescale a trillion times faster than we do. 14 00:01:00,680 --> 00:01:02,080 Speaker 1: Welcome to Inner Cosmos with. 15 00:01:02,080 --> 00:01:02,920 Speaker 2: Me David Eagelman. 16 00:01:03,080 --> 00:01:06,240 Speaker 1: I'm a neuroscientist and author at Stanford and in these 17 00:01:06,240 --> 00:01:09,880 Speaker 1: episodes we sail deeply into our three pound universe to 18 00:01:10,120 --> 00:01:30,720 Speaker 1: understand why our lives look the way they do. Today's 19 00:01:30,720 --> 00:01:37,280 Speaker 1: episode is about what happens when intelligence becomes decentralized. When 20 00:01:37,280 --> 00:01:40,679 Speaker 1: we talk about artificial intelligence, the conversation is typically about 21 00:01:41,080 --> 00:01:46,280 Speaker 1: one giant model, a mind blowingly massive data set and 22 00:01:46,400 --> 00:01:50,200 Speaker 1: a huge compute cluster, so that we get one singular 23 00:01:50,240 --> 00:01:52,640 Speaker 1: model that takes care of everything. But the first thing 24 00:01:52,680 --> 00:01:56,160 Speaker 1: to note is that's not how the brain works. As 25 00:01:56,200 --> 00:01:59,360 Speaker 1: I argued in my book Incognito, the brain is best 26 00:01:59,480 --> 00:02:04,760 Speaker 1: understood as a team of ribles. Inside your skull are 27 00:02:05,360 --> 00:02:10,040 Speaker 1: scores or hundreds of specialized modules. Some care about vision, 28 00:02:10,160 --> 00:02:14,079 Speaker 1: hearing touched, some about walking, some about processing faces. Others 29 00:02:14,160 --> 00:02:18,000 Speaker 1: track moving objects in the world. Other systems are working 30 00:02:18,040 --> 00:02:21,280 Speaker 1: to predict what you should do next. But the key 31 00:02:21,560 --> 00:02:24,000 Speaker 1: is that these systems in the brain are not like 32 00:02:24,280 --> 00:02:27,680 Speaker 1: pieces of a very efficient machine where each part serves 33 00:02:27,720 --> 00:02:31,720 Speaker 1: its purpose. Instead, every second of your life, these networks 34 00:02:31,800 --> 00:02:36,919 Speaker 1: find themselves in conflict. They argue, they interrupt, they compete 35 00:02:36,919 --> 00:02:41,560 Speaker 1: for dominance. The feeling of having a unified self is 36 00:02:41,680 --> 00:02:46,160 Speaker 1: an illusion stitched together on the fly. So maybe the 37 00:02:46,200 --> 00:02:51,280 Speaker 1: future of AI isn't one giant LLM to rule them all, 38 00:02:51,680 --> 00:02:55,519 Speaker 1: but something more like what we are. A distributed system, 39 00:02:55,560 --> 00:03:01,160 Speaker 1: a team of rival agents and argumentative network of local experts, 40 00:03:01,600 --> 00:03:06,400 Speaker 1: each with limited knowledge and a unique perspective, together making 41 00:03:06,440 --> 00:03:10,320 Speaker 1: something more powerful than any individual part. But how do 42 00:03:10,360 --> 00:03:14,320 Speaker 1: you build that? How do you preserve data? Privacy. When 43 00:03:14,440 --> 00:03:18,520 Speaker 1: intelligence is spread across devices and across hospitals and across 44 00:03:18,560 --> 00:03:22,760 Speaker 1: country borders. How do you incentivize data sharing? How do 45 00:03:22,800 --> 00:03:24,840 Speaker 1: you verify who's a good actor? 46 00:03:25,280 --> 00:03:25,919 Speaker 2: These are the. 47 00:03:25,919 --> 00:03:29,639 Speaker 1: Kinds of questions that animate my friend and colleague, Ramesh Roscar. 48 00:03:30,080 --> 00:03:33,079 Speaker 1: He's a professor at the MIT Media Lab. He's an inventor, 49 00:03:33,160 --> 00:03:36,360 Speaker 1: He's a systems thinker and a pioneer in this new 50 00:03:36,400 --> 00:03:41,320 Speaker 1: world of decentralized AI. His work spands from privacy preserving 51 00:03:41,400 --> 00:03:45,680 Speaker 1: technologies to public health tools, to new ways of imagining 52 00:03:45,720 --> 00:03:48,760 Speaker 1: what AI can be from a technical point of view, 53 00:03:48,800 --> 00:03:52,480 Speaker 1: but also ethically and socially. So today we're going to 54 00:03:52,480 --> 00:03:56,920 Speaker 1: talk about why the future of AI might lie in localized, 55 00:03:57,040 --> 00:04:00,280 Speaker 1: specialized systems that are trained on the fly, and how 56 00:04:00,320 --> 00:04:04,480 Speaker 1: new architectures might allow data to stay where it is 57 00:04:04,640 --> 00:04:07,040 Speaker 1: on your phone and your clinic and your city, while 58 00:04:07,120 --> 00:04:10,520 Speaker 1: still contributing to a global model. So this is a 59 00:04:10,520 --> 00:04:15,600 Speaker 1: conversation about architecture and ethics, competition and cooperation, and the 60 00:04:15,760 --> 00:04:19,640 Speaker 1: radically different future that becomes possible when we stop trying 61 00:04:19,680 --> 00:04:25,000 Speaker 1: to centralize everything and instead embrace the power of the network. 62 00:04:29,360 --> 00:04:31,839 Speaker 1: So let's think, as an example, the healthcare systems. So 63 00:04:31,920 --> 00:04:34,240 Speaker 1: the idea is you've got all this data locked up 64 00:04:34,320 --> 00:04:36,680 Speaker 1: in electronic health records all over the place. None of 65 00:04:36,720 --> 00:04:40,880 Speaker 1: the hospital systems want to share that data because they 66 00:04:40,880 --> 00:04:43,200 Speaker 1: don't want other people looking at it. And it's valuable data. 67 00:04:43,240 --> 00:04:47,520 Speaker 1: And so the idea of decentralized AI, for example, could 68 00:04:47,600 --> 00:04:52,320 Speaker 1: be that an AI is looking at all the different 69 00:04:52,520 --> 00:04:55,599 Speaker 1: electronic health records and all the different hospital systems, but 70 00:04:56,279 --> 00:04:59,440 Speaker 1: no human is looking at it. With millions of records, 71 00:04:59,440 --> 00:05:01,520 Speaker 1: it puts together there a big picture and discover sing 72 00:05:01,680 --> 00:05:02,000 Speaker 1: is that right? 73 00:05:02,839 --> 00:05:03,000 Speaker 2: Right? 74 00:05:03,000 --> 00:05:05,000 Speaker 3: I mean health is a great example because I would 75 00:05:05,120 --> 00:05:08,279 Speaker 3: argue that if somehow I have for every health condition 76 00:05:08,880 --> 00:05:10,520 Speaker 3: all the data, I can bring it in one place 77 00:05:10,560 --> 00:05:12,839 Speaker 3: and train models, I can solve it. The problem is, 78 00:05:13,440 --> 00:05:15,599 Speaker 3: you know, the capitalistic system doesn't allow us to do 79 00:05:15,640 --> 00:05:18,280 Speaker 3: it because of privacy, because of trade secrets, and as 80 00:05:18,320 --> 00:05:21,200 Speaker 3: I said, people hold on to the data thinking it's 81 00:05:21,320 --> 00:05:24,760 Speaker 3: worth something, but they don't get paid either. So if 82 00:05:24,800 --> 00:05:28,720 Speaker 3: there was a marketplace for data so that they're not 83 00:05:28,720 --> 00:05:31,480 Speaker 3: to share the data, but they can share the insights 84 00:05:31,520 --> 00:05:35,839 Speaker 3: from it and get paid in return, then everybody would 85 00:05:35,839 --> 00:05:39,800 Speaker 3: be incentivized to contribute to these insights without contributing raw data. 86 00:05:39,880 --> 00:05:42,200 Speaker 1: So the idea is, let's say I'm a hospital system. 87 00:05:42,440 --> 00:05:45,360 Speaker 1: I'm not giving away the electronic health records. Instead, I'm 88 00:05:45,440 --> 00:05:48,719 Speaker 1: letting an AI poke its nose into these things, figure 89 00:05:48,760 --> 00:05:54,400 Speaker 1: out insights from the data, and leave. And that way, 90 00:05:54,400 --> 00:05:56,960 Speaker 1: I haven't given up something in particular, and I get 91 00:05:56,960 --> 00:05:58,320 Speaker 1: paid for that. By the way, that's right. 92 00:05:58,400 --> 00:06:01,279 Speaker 3: I mean another way to think about this is again, 93 00:06:01,320 --> 00:06:04,800 Speaker 3: you want to share intelligence without sharing raw data, and 94 00:06:04,839 --> 00:06:06,320 Speaker 3: that's how the human society works. 95 00:06:06,600 --> 00:06:09,400 Speaker 2: The analogy would be, if I had enough money, you know. 96 00:06:09,400 --> 00:06:13,080 Speaker 3: I could hire you know, a thousand medical interns and 97 00:06:13,520 --> 00:06:15,920 Speaker 3: send them to If I want to solve partical health condition, 98 00:06:16,440 --> 00:06:19,599 Speaker 3: I would send them to every hospital, ask them to 99 00:06:19,640 --> 00:06:22,040 Speaker 3: do an internship for a year, learn about everything that 100 00:06:22,040 --> 00:06:24,880 Speaker 3: partical hospital does, about how they treat a potical condition, 101 00:06:25,279 --> 00:06:28,440 Speaker 3: and how their ideology pathology works, and then bring all ten, 102 00:06:28,880 --> 00:06:32,400 Speaker 3: all thousands of them in one office and extract their 103 00:06:32,400 --> 00:06:35,799 Speaker 3: intelligence and then solve any condition that's out there. 104 00:06:36,200 --> 00:06:38,080 Speaker 2: And that would be a way to bring. 105 00:06:37,880 --> 00:06:40,880 Speaker 3: Intelligence from all these thousand hospitals without bringing in your 106 00:06:40,920 --> 00:06:43,160 Speaker 3: raw data. So how do you now make that happen 107 00:06:43,480 --> 00:06:45,679 Speaker 3: not in the physical world, but in the digital world. 108 00:06:46,040 --> 00:06:49,719 Speaker 3: By effectively sending a MICROAI like an agent that behaves 109 00:06:49,720 --> 00:06:52,520 Speaker 3: like an intern that learns everything and brings back only 110 00:06:52,560 --> 00:06:55,800 Speaker 3: the intelligence and in return pace for the stay at 111 00:06:55,839 --> 00:06:59,440 Speaker 3: the hospital, and then you know, creates a global algorithm 112 00:06:59,480 --> 00:06:59,760 Speaker 3: for that. 113 00:07:00,000 --> 00:07:04,080 Speaker 1: So let's take one second to distinguish decentralized AI from 114 00:07:04,360 --> 00:07:07,520 Speaker 1: agentic AI. They're related, but tells the difference. 115 00:07:07,800 --> 00:07:09,880 Speaker 3: Yeah, I mean the agent DKA is a great way 116 00:07:09,920 --> 00:07:14,480 Speaker 3: to think about it, because mix it very simple to understand. 117 00:07:14,560 --> 00:07:18,840 Speaker 3: So the reason to use agent as an anology as 118 00:07:18,840 --> 00:07:22,679 Speaker 3: an anchor for decent lazed AI is because you imagine 119 00:07:22,680 --> 00:07:25,600 Speaker 3: an agent is like a medical intern, and the agent 120 00:07:25,640 --> 00:07:28,559 Speaker 3: can go to the hospital, you know, you know, stay 121 00:07:28,560 --> 00:07:31,920 Speaker 3: there for some time, learn everything about the data, pay 122 00:07:32,000 --> 00:07:34,600 Speaker 3: for anything that they have learned, and then come back. 123 00:07:34,880 --> 00:07:37,360 Speaker 3: And I can take this thousand agents and now I 124 00:07:37,360 --> 00:07:40,440 Speaker 3: can ask them any new question about that health condition. 125 00:07:41,240 --> 00:07:43,280 Speaker 3: And the beauty here is that I know that the 126 00:07:43,320 --> 00:07:45,800 Speaker 3: agent did not steal any data because I can send 127 00:07:45,880 --> 00:07:48,640 Speaker 3: an agent as a model. And model, as you know, 128 00:07:48,760 --> 00:07:51,000 Speaker 3: is you know, basically a big file let's say one 129 00:07:51,040 --> 00:07:54,520 Speaker 3: hundred megabyte. You a model, it goes in it looks 130 00:07:54,560 --> 00:07:57,200 Speaker 3: like gigabytes of data in the hospital, but comes back 131 00:07:57,240 --> 00:07:59,520 Speaker 3: again only as a hundred megabyte model. 132 00:08:00,040 --> 00:08:03,360 Speaker 1: So that's you tweaked the parameters in the model exactly. 133 00:08:03,360 --> 00:08:05,320 Speaker 1: We're not bringing back the raw data, yeah, which is 134 00:08:05,360 --> 00:08:07,640 Speaker 1: what happens in the human brain. So if you think 135 00:08:07,640 --> 00:08:11,200 Speaker 1: about centralized A versus decentized AI, in centilis you have 136 00:08:11,240 --> 00:08:15,320 Speaker 1: this one massive model that's doing everything. Decentralized AI matches 137 00:08:15,400 --> 00:08:18,440 Speaker 1: very well with the concept of agent AI agent take 138 00:08:18,480 --> 00:08:20,840 Speaker 1: AI because the notion is that each of the agent 139 00:08:21,040 --> 00:08:24,480 Speaker 1: is not just massive six hundred billion parameter models that 140 00:08:24,560 --> 00:08:27,760 Speaker 1: you're hearing about from open AI and meta and so on, 141 00:08:27,920 --> 00:08:31,720 Speaker 1: and these agents are much smaller, and they're very specific 142 00:08:31,920 --> 00:08:33,360 Speaker 1: and they do specific tasks. 143 00:08:34,280 --> 00:08:35,400 Speaker 2: So a legal. 144 00:08:35,160 --> 00:08:37,600 Speaker 3: Agent may not notizing about the health agent or financial 145 00:08:37,640 --> 00:08:41,120 Speaker 3: agents and so on. They're doing very specific things, and 146 00:08:41,160 --> 00:08:43,440 Speaker 3: then they can get better over time, and they get 147 00:08:43,440 --> 00:08:45,760 Speaker 3: better in two ways. They get better because they're looking 148 00:08:45,760 --> 00:08:49,160 Speaker 3: at new data, but more importantly, they get better because 149 00:08:49,160 --> 00:08:53,880 Speaker 3: they're interacting with other agents. And that's what decentalization decential 150 00:08:53,880 --> 00:08:56,920 Speaker 3: plays in kind of mimics the human society because even 151 00:08:56,960 --> 00:08:59,360 Speaker 3: for us, you know, growing up, we learn from some 152 00:08:59,440 --> 00:09:01,959 Speaker 3: data and some experiences, but most of the things we 153 00:09:02,080 --> 00:09:03,640 Speaker 3: learn are actually through interactions. 154 00:09:05,280 --> 00:09:07,720 Speaker 1: So are you saying decentralized AI in a gent ki 155 00:09:08,000 --> 00:09:10,720 Speaker 1: is the same thing, or a gent ki is a 156 00:09:10,760 --> 00:09:12,680 Speaker 1: good example of decentralized AI. 157 00:09:12,880 --> 00:09:15,360 Speaker 3: Agent ki is is kind of the brain that you need, 158 00:09:15,920 --> 00:09:20,480 Speaker 3: but you need other restual ecosystem to emerge for decentralization. So, 159 00:09:20,520 --> 00:09:23,160 Speaker 3: for example, what does it mean to have for agents 160 00:09:23,200 --> 00:09:26,080 Speaker 3: to have commerce? How do agents pay for each other? 161 00:09:26,360 --> 00:09:28,640 Speaker 3: You know, how do I know that a data that 162 00:09:28,679 --> 00:09:31,360 Speaker 3: I you know, that data that I used from a hospital, 163 00:09:31,480 --> 00:09:33,800 Speaker 3: is it worth one hundred dollars, thousand dollars or a 164 00:09:33,800 --> 00:09:36,440 Speaker 3: million dollars? You al sort of figure out, you know, 165 00:09:36,440 --> 00:09:39,359 Speaker 3: how do you make sure that the agents are not misbehaving? 166 00:09:39,800 --> 00:09:43,240 Speaker 3: So beyond the intelligence of a given agent, which is 167 00:09:43,320 --> 00:09:45,840 Speaker 3: classic agent kai, we don't think about the rest of 168 00:09:45,880 --> 00:09:49,240 Speaker 3: the decentralization aspect of it. It's like saying, you know, 169 00:09:49,280 --> 00:09:52,840 Speaker 3: I might have ability to do transactions, but that's very 170 00:09:52,840 --> 00:09:54,280 Speaker 3: different than building a stock market. 171 00:09:54,440 --> 00:09:56,480 Speaker 1: And so give us an example of what the world 172 00:09:56,559 --> 00:09:58,680 Speaker 1: will look like. Let's say in ten years from now, 173 00:09:59,000 --> 00:10:01,200 Speaker 1: when we've got a lot agents and we have this 174 00:10:01,240 --> 00:10:03,920 Speaker 1: whole infrastructure for agents to communicate and pay each other 175 00:10:03,960 --> 00:10:06,760 Speaker 1: and so on, what does that mean for all of us, it's. 176 00:10:06,679 --> 00:10:07,439 Speaker 2: Difficult to product. 177 00:10:07,600 --> 00:10:10,439 Speaker 3: But I would say in the short run, every one 178 00:10:10,480 --> 00:10:13,120 Speaker 3: of us will have our own agent. You know, every 179 00:10:13,200 --> 00:10:16,880 Speaker 3: organization will have its own agent. And these agents will 180 00:10:17,040 --> 00:10:20,360 Speaker 3: work on our behalf. So they're anchored our identity or 181 00:10:20,400 --> 00:10:23,040 Speaker 3: identity of the organization and they're working on our behalf 182 00:10:23,120 --> 00:10:25,840 Speaker 3: and they are you know, they're doing commerce on our behalf. 183 00:10:26,120 --> 00:10:29,240 Speaker 3: So if you think about kind of the roadmap of 184 00:10:29,360 --> 00:10:33,160 Speaker 3: how this agent tech world will evolve, in the beginning, 185 00:10:33,559 --> 00:10:37,559 Speaker 3: it's just about foundations, like do these agents have identity, 186 00:10:38,120 --> 00:10:41,320 Speaker 3: do they have authentication? Can they really represent us? You know, 187 00:10:41,360 --> 00:10:44,839 Speaker 3: what's the reputation when when my agent is using another agent? 188 00:10:45,200 --> 00:10:47,560 Speaker 3: You know, can it trust? You know, like you know 189 00:10:47,600 --> 00:10:50,319 Speaker 3: what ky C is, like I know your agent. All 190 00:10:50,360 --> 00:10:54,000 Speaker 3: the foundational stuff. That's what's happening now in the industry. 191 00:10:54,040 --> 00:10:56,320 Speaker 3: And also work at an IM. The next phase is 192 00:10:56,640 --> 00:10:59,280 Speaker 3: agient commerce. Just as if you move to a new city, 193 00:10:59,600 --> 00:11:01,839 Speaker 3: you know, you're figure out, you know what job you're 194 00:11:01,840 --> 00:11:03,240 Speaker 3: going to get. You know, are you going to get 195 00:11:03,240 --> 00:11:05,880 Speaker 3: paid fifteen dollars an hour or thousand dollars an hour? 196 00:11:06,640 --> 00:11:07,160 Speaker 2: And so on? 197 00:11:07,640 --> 00:11:11,480 Speaker 3: You know, you start learning, you start working over time, 198 00:11:11,760 --> 00:11:13,960 Speaker 3: your skills are not that great, so you go to 199 00:11:14,080 --> 00:11:15,800 Speaker 3: you know, a school, so there'll be in agient schools 200 00:11:15,840 --> 00:11:17,120 Speaker 3: or agents go to agent schools. 201 00:11:17,480 --> 00:11:19,680 Speaker 2: Sometimes they need to be repaired. It will be agent repairs. 202 00:11:20,400 --> 00:11:22,720 Speaker 3: Sometimes there will be a risk of using an agent, 203 00:11:22,760 --> 00:11:25,240 Speaker 3: so you'll have agent insurances, you know, and so on. 204 00:11:25,280 --> 00:11:26,400 Speaker 2: So you have like a whole. 205 00:11:26,480 --> 00:11:29,920 Speaker 3: Commerce and economy around agents. But then the third phases 206 00:11:30,040 --> 00:11:34,160 Speaker 3: very interesting, which is you have very emergent behaviors with agents. 207 00:11:34,200 --> 00:11:36,120 Speaker 3: Just as once you move to the city, you don't 208 00:11:36,160 --> 00:11:39,440 Speaker 3: just work and you know, think about your livelihood, but 209 00:11:39,480 --> 00:11:42,840 Speaker 3: you start creating social circles. You know, you have sports teams, 210 00:11:43,120 --> 00:11:47,440 Speaker 3: you know, you have universities, you have you know, justice 211 00:11:47,480 --> 00:11:50,520 Speaker 3: and court. So you get all this emergent behavior when 212 00:11:50,559 --> 00:11:53,480 Speaker 3: folks get together and just as agents get together. So 213 00:11:53,520 --> 00:11:56,720 Speaker 3: you can see this going from foundations to Asian commerce 214 00:11:56,760 --> 00:11:59,719 Speaker 3: to this you know, emergent behavior of agent societies. But 215 00:12:00,200 --> 00:12:03,120 Speaker 3: just assumes that the agents are anchored at human identity, 216 00:12:03,720 --> 00:12:06,240 Speaker 3: we can't really see beyond this event horizon of what 217 00:12:06,320 --> 00:12:09,400 Speaker 3: will happen when agents that are billions may be trillance 218 00:12:09,440 --> 00:12:13,920 Speaker 3: of agents are navigating on their own on the internet, 219 00:12:14,000 --> 00:12:17,800 Speaker 3: you know, meeting other agents and learning and socializing and 220 00:12:18,120 --> 00:12:20,840 Speaker 3: you know, transacting on the on by themselves. So it's 221 00:12:20,880 --> 00:12:24,600 Speaker 3: kind of diffrully imagine what will happen beyond beyond the 222 00:12:24,640 --> 00:12:29,400 Speaker 3: scenario where they're still representing and anchored for individual human being. 223 00:12:29,520 --> 00:12:33,560 Speaker 1: And even when they are really anchored on us, it's 224 00:12:33,600 --> 00:12:35,600 Speaker 1: still very difficult to see what's going to happen because 225 00:12:35,640 --> 00:12:37,400 Speaker 1: they can run it trillion times faster. 226 00:12:37,200 --> 00:12:40,079 Speaker 3: Than we can, and they can replicate, and they can 227 00:12:40,240 --> 00:12:42,440 Speaker 3: be you know, in multiple less at the same time, 228 00:12:42,840 --> 00:12:45,360 Speaker 3: and they can migrate. You know, they can behave one 229 00:12:45,400 --> 00:12:47,480 Speaker 3: way and we have another way in the next many 230 00:12:47,480 --> 00:12:48,679 Speaker 3: second and so on. 231 00:12:48,840 --> 00:12:50,240 Speaker 1: So I have a question. So a lot of people, 232 00:12:50,240 --> 00:12:54,559 Speaker 1: of course, are worried about AI alignment in what happens 233 00:12:54,600 --> 00:12:59,440 Speaker 1: if things go bad, And in this case of decentralization, 234 00:12:59,600 --> 00:13:02,920 Speaker 1: with all these agents running around working at speeds that 235 00:13:02,960 --> 00:13:06,199 Speaker 1: are much faster than we can comprehend, that it does 236 00:13:06,240 --> 00:13:09,480 Speaker 1: seem like a bit of a dangerous territory to get into. Now, 237 00:13:09,520 --> 00:13:13,439 Speaker 1: you might say that centralized AI also has its dangers, 238 00:13:13,920 --> 00:13:16,520 Speaker 1: and that what we need maybe are checks and balances 239 00:13:16,600 --> 00:13:18,199 Speaker 1: with decentral ied AA. But what does that look like? 240 00:13:18,240 --> 00:13:20,240 Speaker 1: What would that mean to put checks and balances. 241 00:13:19,880 --> 00:13:23,280 Speaker 3: And glad you've brought up alignment in centralized AI. The 242 00:13:23,360 --> 00:13:26,480 Speaker 3: problem is different when you think about decentralization and agent 243 00:13:26,559 --> 00:13:28,840 Speaker 3: KI because now we don't have to worry about you know, 244 00:13:28,920 --> 00:13:33,080 Speaker 3: some one evil character taking over the world, but agents 245 00:13:33,120 --> 00:13:36,400 Speaker 3: on their own could do bad things. So as long 246 00:13:36,440 --> 00:13:40,480 Speaker 3: as they've anchored at a human identity I think should 247 00:13:40,520 --> 00:13:42,520 Speaker 3: be in a reasonably good ship. It's almostly using like 248 00:13:42,559 --> 00:13:45,160 Speaker 3: two factor authentication, you know, like the agent cannot do 249 00:13:45,240 --> 00:13:48,440 Speaker 3: anything unless that says something on my phone, just using 250 00:13:48,440 --> 00:13:51,000 Speaker 3: today's anology, so that should be under control. 251 00:13:51,800 --> 00:13:52,760 Speaker 2: But once we go. 252 00:13:52,760 --> 00:13:55,760 Speaker 3: Beyond that and agents are sufficiently autonomous that can create 253 00:13:55,840 --> 00:13:59,240 Speaker 3: their own societies. Now the challenge is, you know, how 254 00:13:59,280 --> 00:14:01,480 Speaker 3: do you make showed that doing the right thing. So 255 00:14:01,520 --> 00:14:04,120 Speaker 3: the alignment becomes a very different problem. It becomes an 256 00:14:04,160 --> 00:14:08,040 Speaker 3: algorithmic orchestration problem. So how do you have algorithms that 257 00:14:08,160 --> 00:14:12,240 Speaker 3: orchestrate you know, the incentives of the carrots and sticks 258 00:14:12,280 --> 00:14:14,600 Speaker 3: for the agents to behave. And that's actually a very 259 00:14:14,600 --> 00:14:17,480 Speaker 3: complex problem. That's a lot for research at m T 260 00:14:17,840 --> 00:14:22,440 Speaker 3: as well, because you could imagine that you know, if 261 00:14:22,480 --> 00:14:25,200 Speaker 3: you if you try to use decentralization, like there were 262 00:14:25,280 --> 00:14:29,200 Speaker 3: decentil issues. For DeFi, in the crypto and web three world, 263 00:14:30,000 --> 00:14:33,000 Speaker 3: it didn't go very well where examples like Luna and 264 00:14:33,120 --> 00:14:36,480 Speaker 3: Terra compared to stable coins today, and they had, you know, 265 00:14:36,920 --> 00:14:39,520 Speaker 3: an orchestration to figure out, how do we maintain you know, 266 00:14:39,560 --> 00:14:40,720 Speaker 3: their one dollar anchor. 267 00:14:41,200 --> 00:14:43,280 Speaker 1: So anyway, I think that ended up crashing. 268 00:14:43,320 --> 00:14:45,280 Speaker 3: And they ended up crashing because they were based on 269 00:14:45,760 --> 00:14:48,680 Speaker 3: the behavioral economics that was predicted by a bunch of 270 00:14:48,760 --> 00:14:51,840 Speaker 3: geeks sitting in a garage, and that didn't work out. 271 00:14:52,000 --> 00:14:54,320 Speaker 3: So I think we should not rely on algorithmic orchestration 272 00:14:54,720 --> 00:14:57,000 Speaker 3: of society of humans or society of agents. 273 00:14:57,520 --> 00:14:59,240 Speaker 2: And I think there are very important problems to solve. 274 00:14:59,480 --> 00:15:01,720 Speaker 1: So what is the thing beyond algorithmic orchestration? 275 00:15:02,240 --> 00:15:04,560 Speaker 3: So you need to slow things down, You need to 276 00:15:04,640 --> 00:15:08,280 Speaker 3: have some kind of a governance principle that's out there, 277 00:15:08,520 --> 00:15:10,400 Speaker 3: and you know, that's how our society works. All the 278 00:15:10,480 --> 00:15:14,960 Speaker 3: banks and courtes could run you know, algorithmically in many cases, 279 00:15:15,200 --> 00:15:17,520 Speaker 3: we still have human in the loop and we slow 280 00:15:17,560 --> 00:15:21,960 Speaker 3: things down and we trade inefficiency for you know, some 281 00:15:22,000 --> 00:15:23,400 Speaker 3: control over the system. 282 00:15:23,680 --> 00:15:25,080 Speaker 2: So I think that becomes very critical. 283 00:15:25,280 --> 00:15:27,960 Speaker 1: So what does this mean for biology research? For example? 284 00:15:28,240 --> 00:15:31,240 Speaker 3: Yeah, I think scientific research is going to depend on 285 00:15:31,280 --> 00:15:34,680 Speaker 3: some kind of a decentialization because very unlikely that you know, 286 00:15:34,760 --> 00:15:37,480 Speaker 3: all the data will come in plut complace, and you know, 287 00:15:37,560 --> 00:15:39,400 Speaker 3: all the scientists will just pick up the phone and 288 00:15:39,400 --> 00:15:41,640 Speaker 3: start talking to each other. They're all you know, we're 289 00:15:41,640 --> 00:15:44,960 Speaker 3: all scientists, and we like cooperation, but also like competition, 290 00:15:45,520 --> 00:15:49,040 Speaker 3: and you know, the human motivation to compete is what actually. 291 00:15:48,720 --> 00:15:49,840 Speaker 2: Gets gets us going. 292 00:15:50,560 --> 00:15:52,920 Speaker 3: So there will still be reviews, a competition to get 293 00:15:52,920 --> 00:15:55,880 Speaker 3: your paper published in top journals, but we need a 294 00:15:55,880 --> 00:15:58,600 Speaker 3: way for them to share their insights at the same time, 295 00:15:58,680 --> 00:16:00,920 Speaker 3: and for those insights they get something in return, so 296 00:16:00,960 --> 00:16:04,960 Speaker 3: it's crowdsourced, but something in return. I like to use 297 00:16:04,960 --> 00:16:09,600 Speaker 3: analogy of Google Maps and traffic on Google Maps. You know, 298 00:16:10,280 --> 00:16:12,960 Speaker 3: before that removes, there was Garment Maps, which we barely used. 299 00:16:13,320 --> 00:16:14,400 Speaker 2: And today, even if. 300 00:16:14,320 --> 00:16:15,760 Speaker 3: I know how to go from point to point B, 301 00:16:16,320 --> 00:16:18,640 Speaker 3: you know, I open Google Maps not because I want 302 00:16:18,680 --> 00:16:20,280 Speaker 3: to know the route, but I want to know where 303 00:16:20,280 --> 00:16:23,720 Speaker 3: everybody else is. So I'm benefiting from the fact that 304 00:16:23,800 --> 00:16:27,640 Speaker 3: everybody else is contributing the GPS coordinate and I get 305 00:16:27,680 --> 00:16:29,800 Speaker 3: to see, you know, the reds and the greens on 306 00:16:29,840 --> 00:16:32,720 Speaker 3: the map and turn, but turn navigation, but I'm also 307 00:16:32,760 --> 00:16:34,520 Speaker 3: giving up my own GPS. 308 00:16:34,120 --> 00:16:35,720 Speaker 2: Location while I'm driving. 309 00:16:35,920 --> 00:16:38,920 Speaker 3: I'm contributing to the system and I'm getting great benefits 310 00:16:38,960 --> 00:16:41,840 Speaker 3: in return. So can we create the equivalent of Google 311 00:16:41,880 --> 00:16:44,760 Speaker 3: Maps for any societal problem, you know, whether it's scientific 312 00:16:44,800 --> 00:16:48,880 Speaker 3: discovery or finding, you know, a treatment plan. Again, if 313 00:16:48,880 --> 00:16:51,360 Speaker 3: you can centralize all the data, we can solve these problems. 314 00:16:51,360 --> 00:16:53,960 Speaker 3: We cannot centralize it, and so we must use a 315 00:16:53,960 --> 00:16:54,880 Speaker 3: decentralized method. 316 00:16:55,000 --> 00:16:56,440 Speaker 1: And let me just make sure you have that. Why 317 00:16:56,440 --> 00:16:59,000 Speaker 1: can't we centralize it? Isn't that what open ai and 318 00:16:59,040 --> 00:17:00,480 Speaker 1: meta and everyone else is doing. 319 00:17:00,640 --> 00:17:02,480 Speaker 3: So, I mean, if you look at open ai and 320 00:17:02,720 --> 00:17:05,440 Speaker 3: everybody else, they you know, they scanned the whole Internet, 321 00:17:05,840 --> 00:17:08,280 Speaker 3: you know, kind of in a permission less way. And 322 00:17:08,280 --> 00:17:10,800 Speaker 3: that's why you could train this multi tens of trill 323 00:17:10,880 --> 00:17:14,120 Speaker 3: and token models that are out there. But as I said, 324 00:17:14,160 --> 00:17:15,920 Speaker 3: that's like just the tip of the iceberg. Most of 325 00:17:15,960 --> 00:17:18,880 Speaker 3: the information in the world is actually locked away. If 326 00:17:18,880 --> 00:17:22,000 Speaker 3: it takes health as an example, it locked away in hospitals, 327 00:17:22,440 --> 00:17:26,600 Speaker 3: in pharmaceutical companies, in all the failed trials, you know, 328 00:17:26,760 --> 00:17:29,720 Speaker 3: and all the research work in the lab notebooks of 329 00:17:29,880 --> 00:17:32,480 Speaker 3: you know, all the amazing postdocs and students. 330 00:17:32,840 --> 00:17:35,080 Speaker 2: So all the information is actually not available. 331 00:17:35,240 --> 00:17:38,399 Speaker 3: So think about if all the information was available to 332 00:17:38,560 --> 00:17:41,120 Speaker 3: open AI, it would do an amazing thing. It would 333 00:17:41,160 --> 00:17:45,360 Speaker 3: solve healthcare overnight. But it cannot because of the competition. 334 00:18:00,080 --> 00:18:02,040 Speaker 1: You know, I'm struck with it. An analogy here. I 335 00:18:02,080 --> 00:18:06,280 Speaker 1: saw an interview with Isaac asimof in nineteen eighty eight, 336 00:18:06,320 --> 00:18:09,160 Speaker 1: I think it was he was on jonah Lerror's show, 337 00:18:09,680 --> 00:18:13,000 Speaker 1: and he said, look, I foresee a day when there's 338 00:18:13,040 --> 00:18:16,359 Speaker 1: a central mainframe that has all of human kind's knowledge 339 00:18:16,359 --> 00:18:19,479 Speaker 1: and everyone has a cable to their house and they 340 00:18:19,520 --> 00:18:21,719 Speaker 1: can access the knowledge of the world that way. And 341 00:18:21,760 --> 00:18:24,840 Speaker 1: he was almost right. But what happened instead was instead 342 00:18:24,840 --> 00:18:29,640 Speaker 1: of a centralized knowledge repository, we it became completely decentralized 343 00:18:29,640 --> 00:18:30,600 Speaker 1: on the Internet. 344 00:18:30,960 --> 00:18:35,160 Speaker 3: Yes, I'm glad you mention that, because you know, today 345 00:18:35,200 --> 00:18:38,199 Speaker 3: we're kind of the mainframe era of AI, and from 346 00:18:38,280 --> 00:18:40,720 Speaker 3: that we're going to switch to the PC era of AI, 347 00:18:40,880 --> 00:18:43,520 Speaker 3: to the Internet era of AI. So, as you know, 348 00:18:43,560 --> 00:18:47,320 Speaker 3: in the eighties, big companies like IBM and Silicon Graphics 349 00:18:47,359 --> 00:18:50,760 Speaker 3: were selled as big mainframes and we would literally connect 350 00:18:50,920 --> 00:18:55,320 Speaker 3: to these machines or in a dial up motems, you know, yeah, 351 00:18:55,359 --> 00:18:57,320 Speaker 3: and even to use a calculator, by the way, we 352 00:18:57,359 --> 00:18:59,320 Speaker 3: had to use a dialup motem. It was not available 353 00:18:59,359 --> 00:19:01,880 Speaker 3: on the machines. But in the late eighties and nineties 354 00:19:01,880 --> 00:19:04,439 Speaker 3: we said, hey, that's kind of stupid. I should just 355 00:19:04,480 --> 00:19:07,600 Speaker 3: have that calculator or a word or whatever the software 356 00:19:07,680 --> 00:19:10,359 Speaker 3: using on my machine on a home machine. And the 357 00:19:10,359 --> 00:19:12,520 Speaker 3: fact that we moved from a mainframe era a PCA 358 00:19:12,600 --> 00:19:16,520 Speaker 3: I was interesting, but we needed one more thing. Before that, 359 00:19:16,560 --> 00:19:19,520 Speaker 3: there was Internet, which was mostly connecting machines, and Tim 360 00:19:19,520 --> 00:19:22,959 Speaker 3: Bernersley and others said in the nineties that actually what 361 00:19:23,000 --> 00:19:25,400 Speaker 3: we need is a world wide web, and he brought 362 00:19:25,440 --> 00:19:29,800 Speaker 3: in four things, a browser, protocols like SGTP and STML, 363 00:19:30,880 --> 00:19:32,040 Speaker 3: and a notion of URL. 364 00:19:32,040 --> 00:19:33,479 Speaker 2: How do you discover something right? 365 00:19:33,960 --> 00:19:37,800 Speaker 3: And with those four things he transformed the Internet into 366 00:19:37,840 --> 00:19:40,880 Speaker 3: world wide Web so that humans can interact with all 367 00:19:40,920 --> 00:19:44,800 Speaker 3: the stuff that's out there. And one would argue that 368 00:19:45,160 --> 00:19:50,160 Speaker 3: the impact of the mainframe revolution and PC revolution is 369 00:19:50,240 --> 00:19:53,640 Speaker 3: insignificant compared to the impact of the world wide Web, 370 00:19:54,480 --> 00:19:57,280 Speaker 3: and the same thing is true today. The impact of 371 00:19:57,440 --> 00:20:02,359 Speaker 3: AI will be insignificant to the impact of the Internet 372 00:20:02,480 --> 00:20:04,359 Speaker 3: and the web of AI that's or. 373 00:20:04,400 --> 00:20:09,560 Speaker 1: There terrific, and so tell us about Project Nanda. 374 00:20:10,000 --> 00:20:15,600 Speaker 3: So this decentralization agent tech concepts very nicely map us 375 00:20:15,680 --> 00:20:21,119 Speaker 3: to this Project Nanda. Nanda stands for networked AI agents 376 00:20:21,160 --> 00:20:24,920 Speaker 3: in decentralized architecture, but it's also the name of my sister, 377 00:20:25,640 --> 00:20:28,040 Speaker 3: and Nanda means joy in Sanskrit. 378 00:20:28,560 --> 00:20:30,560 Speaker 2: So I think, let's bring some joy to. 379 00:20:30,880 --> 00:20:34,480 Speaker 3: The technology and the Internet and the web of AA agents. 380 00:20:34,840 --> 00:20:36,000 Speaker 1: What is project going to do? 381 00:20:36,520 --> 00:20:38,879 Speaker 3: So Project Nanda is a projecda is A is a 382 00:20:38,960 --> 00:20:43,080 Speaker 3: large collective. We're looking at this three phases, how how 383 00:20:43,119 --> 00:20:45,320 Speaker 3: the agent taic world is going to evolve, you know 384 00:20:45,359 --> 00:20:49,040 Speaker 3: the foundations agent e commerce and you know agentic societies, 385 00:20:49,600 --> 00:20:54,919 Speaker 3: uh and involves multiple universities, multiple companies, and we're working 386 00:20:54,920 --> 00:20:58,800 Speaker 3: together to figure out how this you know, web of 387 00:20:59,240 --> 00:21:03,040 Speaker 3: agents willieve the same way Tim berners Lee launched world 388 00:21:03,040 --> 00:21:04,720 Speaker 3: Ward Web Constium. 389 00:21:04,760 --> 00:21:07,120 Speaker 2: Alside of my Tea to see how we can use. 390 00:21:07,000 --> 00:21:11,000 Speaker 3: The Internet technologies in a responsible way and keep the 391 00:21:11,040 --> 00:21:14,720 Speaker 3: flock together. And so one should thank Tim berners Lee 392 00:21:14,920 --> 00:21:19,080 Speaker 3: for keeping it an open, vibrant, you know, and neutral 393 00:21:19,320 --> 00:21:21,840 Speaker 3: web that's out there, as opposed to what we saw 394 00:21:21,920 --> 00:21:25,439 Speaker 3: later on in the computing you know generation, which is 395 00:21:25,720 --> 00:21:29,040 Speaker 3: things like social media or things like mobile phones are 396 00:21:29,119 --> 00:21:33,159 Speaker 3: highly fragmented systems and a few companies control you know, 397 00:21:33,240 --> 00:21:37,000 Speaker 3: these ecosystems. As opposed to the World War Web. Today 398 00:21:37,080 --> 00:21:39,480 Speaker 3: you can go online, go to go daday dot com, 399 00:21:39,520 --> 00:21:42,919 Speaker 3: create a website, launch a business, create new innovations, do 400 00:21:43,000 --> 00:21:46,680 Speaker 3: whatever you want. And that permissionless nature of the world 401 00:21:46,720 --> 00:21:49,840 Speaker 3: ward Web has really created in an unlocked you know, 402 00:21:49,960 --> 00:21:53,159 Speaker 3: not just you know treellanceer dollars economic value, but I 403 00:21:53,160 --> 00:21:56,800 Speaker 3: would say scientific progress, you know, new democracies, new societies. 404 00:21:56,800 --> 00:21:58,040 Speaker 3: I mean, it has really changed the way we think 405 00:21:58,040 --> 00:22:00,119 Speaker 3: about it. We also think about what would happen one 406 00:22:00,280 --> 00:22:03,159 Speaker 3: in the absence of project Now, it's possible that, you know, 407 00:22:03,200 --> 00:22:05,280 Speaker 3: there will be only two Asient stores, just like we 408 00:22:05,320 --> 00:22:08,040 Speaker 3: have app store and play store. That just two Asian 409 00:22:08,119 --> 00:22:11,680 Speaker 3: store and anything you create in a and agents, every 410 00:22:11,680 --> 00:22:12,960 Speaker 3: one of us will have to go through these two 411 00:22:13,000 --> 00:22:15,959 Speaker 3: platforms and they won't be talking to each other. And 412 00:22:16,000 --> 00:22:18,760 Speaker 3: those platforms say we use the tool, we will tell 413 00:22:18,800 --> 00:22:21,040 Speaker 3: you which AI to use, We will tell you what 414 00:22:21,080 --> 00:22:23,280 Speaker 3: you can do and cannot do. And by the way, 415 00:22:23,280 --> 00:22:25,160 Speaker 3: any money you make from that will take thirty percent 416 00:22:25,240 --> 00:22:25,439 Speaker 3: of it. 417 00:22:25,560 --> 00:22:27,879 Speaker 1: So the analogy here is like Apple Store and Google 418 00:22:27,880 --> 00:22:28,720 Speaker 1: Play exactly. 419 00:22:28,840 --> 00:22:33,480 Speaker 3: Yeah, and imagine if the World Wide Web wasn't open 420 00:22:33,520 --> 00:22:36,320 Speaker 3: and vibrant. Imagine there are only two websites you can 421 00:22:36,440 --> 00:22:40,080 Speaker 3: use to create an online business. One caught Squarespace and 422 00:22:40,119 --> 00:22:42,919 Speaker 3: we's call Wix, and there are only two companies, and 423 00:22:43,040 --> 00:22:44,879 Speaker 3: any website you create must be created those one of 424 00:22:44,880 --> 00:22:47,040 Speaker 3: those two platforms. You know, they will give you all 425 00:22:47,080 --> 00:22:49,320 Speaker 3: the tools, you know, they will do all the transactions 426 00:22:49,320 --> 00:22:52,119 Speaker 3: for you, They will disclose which websites are better than 427 00:22:52,160 --> 00:22:54,320 Speaker 3: others are not. And by the way, any money you 428 00:22:54,359 --> 00:22:56,199 Speaker 3: make on the web, you have to share thirty percent 429 00:22:56,440 --> 00:22:58,320 Speaker 3: with square Space or Wix. That would happen a pretty 430 00:22:58,359 --> 00:23:03,399 Speaker 3: dystopian world. Chance that the agentic Web could also end 431 00:23:03,480 --> 00:23:06,840 Speaker 3: up in this highly fragmented, highly dystopian world. And the 432 00:23:06,880 --> 00:23:09,240 Speaker 3: goal for NANDA is to kind of make sure we 433 00:23:09,280 --> 00:23:11,720 Speaker 3: have the freedom, we have the agency, and we have 434 00:23:11,800 --> 00:23:14,159 Speaker 3: ability to do this in an open and vibrate and 435 00:23:14,240 --> 00:23:17,679 Speaker 3: neutral way. So all the organizations that emerged in the 436 00:23:17,720 --> 00:23:21,879 Speaker 3: world Wide Web, such as world Wide Web Consortium, the DNS, 437 00:23:22,240 --> 00:23:22,840 Speaker 3: i CAN. 438 00:23:22,720 --> 00:23:24,720 Speaker 2: Which actually allows you to have the domain name. 439 00:23:24,680 --> 00:23:28,320 Speaker 3: Servers, Mozilla, all these great organizations we have to do 440 00:23:28,400 --> 00:23:31,520 Speaker 3: all that work again in the next six months. If 441 00:23:31,520 --> 00:23:34,280 Speaker 3: we don't do it, we'll end up in this highly 442 00:23:34,320 --> 00:23:37,639 Speaker 3: dystopian world where somebody else is controlling all the AI 443 00:23:37,720 --> 00:23:38,719 Speaker 3: and all the AI agents. 444 00:23:39,080 --> 00:23:41,720 Speaker 1: So out of curiosity, when we're in this future world 445 00:23:41,760 --> 00:23:45,399 Speaker 1: where we've created this species that runs parallel to us, 446 00:23:45,440 --> 00:23:49,560 Speaker 1: all these AI agents running around at much faster speeds 447 00:23:49,600 --> 00:23:51,760 Speaker 1: than us, just out of curiosity, what do you think 448 00:23:51,840 --> 00:23:54,919 Speaker 1: the Internet slash Worldwide Web is going to look like 449 00:23:55,119 --> 00:23:58,480 Speaker 1: for us? Will people still go and look up websites 450 00:23:58,600 --> 00:24:02,399 Speaker 1: or instead, will there be another version of websites? My 451 00:24:02,520 --> 00:24:05,400 Speaker 1: AI agent can go and buy me a shirt and 452 00:24:05,440 --> 00:24:07,919 Speaker 1: not you know, not have to look and scroll in 453 00:24:07,920 --> 00:24:08,640 Speaker 1: the way that we do. 454 00:24:09,000 --> 00:24:10,880 Speaker 3: Yeah, I mean a lot of a lot of things 455 00:24:10,920 --> 00:24:14,080 Speaker 3: that have been developed online have been for kind of 456 00:24:14,200 --> 00:24:17,360 Speaker 3: human conception and human interaction. And it's kind of clunky 457 00:24:17,440 --> 00:24:20,440 Speaker 3: the fact that the whole world interests with the internet 458 00:24:20,480 --> 00:24:23,160 Speaker 3: with your fingers on a keyboard. It's kind of kind 459 00:24:23,160 --> 00:24:25,760 Speaker 3: of strange if you think about it. And so clearly, 460 00:24:25,840 --> 00:24:28,480 Speaker 3: you know, the web of the future will be multimodal. 461 00:24:28,600 --> 00:24:30,760 Speaker 3: We can talk to each other, you know, we'll have 462 00:24:30,920 --> 00:24:33,600 Speaker 3: you know, I'll have you know, some an AI agent 463 00:24:33,720 --> 00:24:36,760 Speaker 3: just talking in my ear almost like the movie hear right, 464 00:24:37,119 --> 00:24:40,119 Speaker 3: So you can imagine the world becoming much more you know, 465 00:24:40,200 --> 00:24:43,119 Speaker 3: much more human again, you know, as opposed to forcing 466 00:24:43,200 --> 00:24:45,920 Speaker 3: us to interact through this through this keyboard. So that's 467 00:24:46,000 --> 00:24:49,439 Speaker 3: kind of the possibility of human interaction. And we'll have 468 00:24:49,480 --> 00:24:52,800 Speaker 3: other technologies going along with that, like VR and AR 469 00:24:53,080 --> 00:24:55,360 Speaker 3: and new materials, science and new technologies. 470 00:24:55,400 --> 00:24:56,840 Speaker 2: I think all that will be part of it. 471 00:24:57,040 --> 00:24:59,440 Speaker 3: And then that's kind of like the human at individual level, 472 00:24:59,800 --> 00:25:02,240 Speaker 3: but at kind of at a structural level, at a 473 00:25:02,280 --> 00:25:04,880 Speaker 3: systemic level, you can imagine all these agents are doing 474 00:25:04,880 --> 00:25:07,080 Speaker 3: great work, you know, behind the scenes. It's making sure 475 00:25:07,680 --> 00:25:10,600 Speaker 3: you know, our our health is taken care of, my 476 00:25:10,720 --> 00:25:12,959 Speaker 3: finances is taken care of. I mean, think about there 477 00:25:12,960 --> 00:25:15,760 Speaker 3: are millions of people who are not financially savvy and 478 00:25:15,800 --> 00:25:19,320 Speaker 3: they have to proactively go and check online or with 479 00:25:19,359 --> 00:25:22,359 Speaker 3: their financial advisor about what they should do, you know, 480 00:25:22,720 --> 00:25:25,639 Speaker 3: but imagine if there's an agent who's helping them in 481 00:25:25,840 --> 00:25:26,840 Speaker 3: their financial decisions. 482 00:25:26,920 --> 00:25:28,040 Speaker 2: You know. Think about education. 483 00:25:28,200 --> 00:25:29,919 Speaker 3: There are billions of people out there who don't get 484 00:25:29,920 --> 00:25:33,400 Speaker 3: good quality education, and people often talk about how AI 485 00:25:33,600 --> 00:25:36,440 Speaker 3: is going to improve education, but Asians would have much 486 00:25:36,480 --> 00:25:39,639 Speaker 3: more an impact because they can be highly personalized and 487 00:25:39,680 --> 00:25:43,080 Speaker 3: they can really understand the context and to you know, 488 00:25:43,280 --> 00:25:45,080 Speaker 3: use that AI. You don't have to make an API 489 00:25:45,200 --> 00:25:48,080 Speaker 3: call to somebody in California. You know, you could be 490 00:25:48,160 --> 00:25:51,119 Speaker 3: using you on your own device. You can be modifying it. 491 00:25:51,600 --> 00:25:54,719 Speaker 3: And so I think and we'll be treating our agents 492 00:25:54,760 --> 00:25:58,080 Speaker 3: almost like Tamagachi. You know, remember those toys from you know, 493 00:25:58,160 --> 00:25:58,439 Speaker 3: will be. 494 00:25:58,480 --> 00:26:01,600 Speaker 1: Kind of just as a remind the listeners. It's you 495 00:26:01,680 --> 00:26:03,560 Speaker 1: have to pet them or take care of them, right, 496 00:26:03,560 --> 00:26:04,560 Speaker 1: and then they stay alive. 497 00:26:04,640 --> 00:26:07,920 Speaker 3: And yeah, there's are very addictive toys that are popular 498 00:26:07,960 --> 00:26:11,080 Speaker 3: in the eighties and nineties from Japan which you are 499 00:26:11,119 --> 00:26:13,639 Speaker 3: to constantly kind of nurture them and feed them and 500 00:26:13,680 --> 00:26:15,359 Speaker 3: pet them and they stay alive. 501 00:26:15,400 --> 00:26:16,720 Speaker 2: All those are purely digital. 502 00:26:16,480 --> 00:26:18,679 Speaker 3: Characters, and Asians are going to be something like that. 503 00:26:18,720 --> 00:26:21,680 Speaker 3: You know, we really care about, you know, nurturing them. 504 00:26:22,160 --> 00:26:24,960 Speaker 3: And my guess is that we'll spend like one third 505 00:26:25,000 --> 00:26:28,920 Speaker 3: of our life every day just making our Asians better 506 00:26:29,000 --> 00:26:31,320 Speaker 3: and teaching them and coaching them and talking to them. 507 00:26:31,840 --> 00:26:33,960 Speaker 3: And these Asians will be such a significant part of 508 00:26:33,960 --> 00:26:37,000 Speaker 3: our life, our health or education, our social life, our 509 00:26:37,040 --> 00:26:41,399 Speaker 3: livelihoods and so on that we are very kind of 510 00:26:41,440 --> 00:26:44,119 Speaker 3: attached to our Asians, hopefully in a positive way. So 511 00:26:44,880 --> 00:26:47,000 Speaker 3: it will not only make kind of a daily interaction 512 00:26:47,119 --> 00:26:50,639 Speaker 3: more human because you're not constant looking at the screens 513 00:26:50,640 --> 00:26:54,520 Speaker 3: and computers all the time. At the same time, these 514 00:26:54,520 --> 00:26:56,560 Speaker 3: Asians will be doing things beyond you know, kind of 515 00:26:57,200 --> 00:27:00,000 Speaker 3: you know, kind of behind the scenes, and make it 516 00:27:00,040 --> 00:27:02,879 Speaker 3: again a more democratic world, you know, kind of reduce 517 00:27:03,000 --> 00:27:06,400 Speaker 3: the friction between people who have information and people who 518 00:27:06,400 --> 00:27:09,320 Speaker 3: don't have information, or people have a good rollodex or 519 00:27:09,320 --> 00:27:12,439 Speaker 3: they don't. So all this information of symmetry, you know, 520 00:27:12,520 --> 00:27:16,520 Speaker 3: power of symmetry could be addressed if everybody's empowered with 521 00:27:16,600 --> 00:27:17,200 Speaker 3: their own agent. 522 00:27:33,880 --> 00:27:35,880 Speaker 1: Okay, So that gives us a really good sense of 523 00:27:35,880 --> 00:27:39,040 Speaker 1: what this near future world is going to look like. 524 00:27:39,560 --> 00:27:42,320 Speaker 1: It may be that the World Wide Web, which we've 525 00:27:42,359 --> 00:27:45,280 Speaker 1: grown so used to and that seemed like the future, 526 00:27:45,359 --> 00:27:47,800 Speaker 1: it may be that that's already becoming the past in 527 00:27:47,880 --> 00:27:48,680 Speaker 1: the sense. 528 00:27:49,240 --> 00:27:51,600 Speaker 2: It builds on top of each other. That I wouldn't say. 529 00:27:51,480 --> 00:27:54,479 Speaker 1: Agreed to the past, agreed, but the idea of tapping 530 00:27:54,520 --> 00:27:58,280 Speaker 1: keyboard for a particular website, or going to Google and 531 00:27:58,320 --> 00:28:00,960 Speaker 1: doing a search and then flipping through all these different websites. 532 00:28:01,000 --> 00:28:02,480 Speaker 1: It seems like that's already. 533 00:28:02,320 --> 00:28:05,760 Speaker 3: Yeah, I think, you know, our agents will be more 534 00:28:05,800 --> 00:28:08,360 Speaker 3: like our pets or maybe a baby. 535 00:28:08,080 --> 00:28:08,639 Speaker 2: We are raising. 536 00:28:09,320 --> 00:28:11,760 Speaker 3: It'll be much more interactive, much more fun. I mean, 537 00:28:11,840 --> 00:28:15,399 Speaker 3: we'll still need computers to do like very information heavy 538 00:28:16,000 --> 00:28:18,960 Speaker 3: aspects of our life, but rest of the day, we 539 00:28:19,000 --> 00:28:21,840 Speaker 3: wouldn't be like we want be writing emails, you know, 540 00:28:21,920 --> 00:28:23,600 Speaker 3: we won't be kind of scrolling through the sea which 541 00:28:23,680 --> 00:28:26,000 Speaker 3: rastwenty you want to go to, or we won't be 542 00:28:26,520 --> 00:28:28,000 Speaker 3: you know, doctors want to be sitting in front of 543 00:28:28,000 --> 00:28:31,280 Speaker 3: the EHR report records to decide how they should help 544 00:28:31,280 --> 00:28:34,920 Speaker 3: their patients, and financial folks will not be spending time 545 00:28:34,920 --> 00:28:37,640 Speaker 3: in front of huge spreadsheets. A lot of those things 546 00:28:37,680 --> 00:28:40,080 Speaker 3: will be happening behind the scenes. We still need spreadsheets, 547 00:28:40,320 --> 00:28:42,160 Speaker 3: and we still had to go to the raw, you know, 548 00:28:42,280 --> 00:28:45,440 Speaker 3: source of truth. But most of the time you'll have 549 00:28:45,560 --> 00:28:49,080 Speaker 3: this you know this, this agents as your companions. 550 00:28:49,280 --> 00:28:52,200 Speaker 1: So what are the legal consequences if you're just blue 551 00:28:52,200 --> 00:28:55,520 Speaker 1: skying and thinking about the future. If your agent goes 552 00:28:55,520 --> 00:28:59,480 Speaker 1: off and does something and it ends up doing something 553 00:28:59,480 --> 00:29:03,400 Speaker 1: illegal or and that causes big damage, who is responsible? 554 00:29:03,440 --> 00:29:04,200 Speaker 1: You are the agent? 555 00:29:04,480 --> 00:29:05,360 Speaker 2: This is a great question. 556 00:29:05,520 --> 00:29:06,960 Speaker 3: Yeah, I think it's it's like saying, you know, if 557 00:29:06,960 --> 00:29:08,840 Speaker 3: you have a six year old child and they do something, 558 00:29:09,800 --> 00:29:11,240 Speaker 3: you know, other parents responsible. 559 00:29:11,640 --> 00:29:13,440 Speaker 2: I guess I don't have an answer for it. A 560 00:29:13,440 --> 00:29:13,840 Speaker 2: little bit. 561 00:29:13,840 --> 00:29:16,760 Speaker 1: Different because the child, the legal system understands that the 562 00:29:16,800 --> 00:29:18,960 Speaker 1: child is separate. You try to do your best raise 563 00:29:19,000 --> 00:29:20,959 Speaker 1: in the kid, but the k goes. Often it's an 564 00:29:20,960 --> 00:29:23,640 Speaker 1: independent entergy from the point of view the legal system. 565 00:29:23,400 --> 00:29:25,040 Speaker 2: Right, I mean, I mean today's world. 566 00:29:25,120 --> 00:29:27,640 Speaker 3: You know, if there's a you know, if a child 567 00:29:27,680 --> 00:29:31,480 Speaker 3: uses a gun, you know, the parents are considered responsible. 568 00:29:31,520 --> 00:29:33,040 Speaker 2: Again, I'm not an expert in this topic. 569 00:29:33,600 --> 00:29:36,840 Speaker 3: I would say that in a lot of these protocols 570 00:29:36,920 --> 00:29:41,000 Speaker 3: and systems will emerge over time. And yes, the agents 571 00:29:41,040 --> 00:29:43,760 Speaker 3: will go to agent schools to get trained. There will 572 00:29:43,800 --> 00:29:46,160 Speaker 3: be a char departments, you know, for if the agents 573 00:29:46,160 --> 00:29:50,080 Speaker 3: are misbehaving, and there will be you know, kind of 574 00:29:50,080 --> 00:29:54,960 Speaker 3: societal values that will be infused into the agents with 575 00:29:55,040 --> 00:29:56,000 Speaker 3: some character and sticks. 576 00:29:56,320 --> 00:29:57,560 Speaker 1: Wow, this is going to take a while after the 577 00:29:57,640 --> 00:29:59,600 Speaker 1: legal system to work out, because you can imagine a 578 00:29:59,640 --> 00:30:03,400 Speaker 1: bad actor programs a bad AI agent to go off 579 00:30:03,440 --> 00:30:05,280 Speaker 1: and do stuff and then it says, well, it wasn't 580 00:30:05,280 --> 00:30:08,040 Speaker 1: really me. I didn't know this was in the code, 581 00:30:08,080 --> 00:30:10,720 Speaker 1: and yes, it caused all this problem. 582 00:30:11,040 --> 00:30:13,200 Speaker 3: And I remember this is going to be agents will 583 00:30:13,240 --> 00:30:15,160 Speaker 3: may or may not have may or may not have 584 00:30:15,240 --> 00:30:18,120 Speaker 3: kind of national boundaries. So you know, a law in 585 00:30:18,120 --> 00:30:20,959 Speaker 3: Indonesia may be different from law in Malaysia, So what 586 00:30:21,000 --> 00:30:22,520 Speaker 3: does it mean. So a lot of the work we 587 00:30:22,600 --> 00:30:25,120 Speaker 3: do in Project Nanda is thinking about this notion of 588 00:30:25,520 --> 00:30:29,760 Speaker 3: geo fencing and creditialing and verifiability and so on. So 589 00:30:29,800 --> 00:30:32,280 Speaker 3: a lot of our research goes into kind of cryptographer, 590 00:30:32,280 --> 00:30:34,760 Speaker 3: get very fine that it's doing the right thing, also 591 00:30:34,960 --> 00:30:37,920 Speaker 3: making sure that there's a reputation system that's you know, 592 00:30:37,960 --> 00:30:40,240 Speaker 3: that's encoded and so on. So a lot of thinking 593 00:30:40,280 --> 00:30:43,760 Speaker 3: about the good behavior and the bad behavior is encoded 594 00:30:44,160 --> 00:30:47,160 Speaker 3: in how this agent Internet is being developed. 595 00:30:47,280 --> 00:30:49,600 Speaker 1: Quick question, is there a way for listeners to see 596 00:30:49,640 --> 00:30:51,040 Speaker 1: what Project Nonda is up to? 597 00:30:51,560 --> 00:30:53,320 Speaker 3: Yeah, you can, you know, you can go create your 598 00:30:53,360 --> 00:30:56,800 Speaker 3: own agent David in thirty seconds and we'll post a 599 00:30:56,880 --> 00:31:00,360 Speaker 3: link in the in the video. Go to join thirty 600 00:31:00,440 --> 00:31:03,800 Speaker 3: nine dot org ji n three nine dot org. You 601 00:31:03,800 --> 00:31:06,480 Speaker 3: can create your own agent in thirty seconds. It'll be 602 00:31:06,520 --> 00:31:09,960 Speaker 3: live online. Anybody else can interact with your agent if 603 00:31:10,000 --> 00:31:12,760 Speaker 3: you want, you can attach more information to it. You 604 00:31:12,800 --> 00:31:14,719 Speaker 3: can also start making money from it. Right now, we're 605 00:31:14,760 --> 00:31:17,520 Speaker 3: just using points system because the research so you can 606 00:31:17,520 --> 00:31:19,640 Speaker 3: make points. So for example, there could be you know 607 00:31:19,680 --> 00:31:21,880 Speaker 3: a David agent and you're a fancy neior scientist, so 608 00:31:21,960 --> 00:31:23,600 Speaker 3: you could say, hey, if you want to talk to 609 00:31:23,640 --> 00:31:26,560 Speaker 3: my agent, I'm going to charge you in a dollar minute. 610 00:31:27,280 --> 00:31:29,200 Speaker 3: At the same time, when you go to use somebody 611 00:31:29,200 --> 00:31:32,240 Speaker 3: else's agent, maybe it costs you ten cents a minute, 612 00:31:32,440 --> 00:31:34,240 Speaker 3: right and so on. So or it can go and 613 00:31:34,320 --> 00:31:37,120 Speaker 3: join thirty nine dot org, which is kind of an 614 00:31:37,200 --> 00:31:40,800 Speaker 3: open website to start playing with your agents, create your profiles, 615 00:31:41,120 --> 00:31:44,320 Speaker 3: start make money off it, you know, send links to 616 00:31:44,360 --> 00:31:45,160 Speaker 3: others and so on. 617 00:31:45,240 --> 00:31:46,280 Speaker 2: So kind of early days. 618 00:31:46,560 --> 00:31:50,960 Speaker 3: But yeah, definitely can go to projectnan dot org, which 619 00:31:51,000 --> 00:31:53,240 Speaker 3: is the main website, which is all the technology, all 620 00:31:53,240 --> 00:31:56,200 Speaker 3: the source code, all the research papers, you know, all 621 00:31:56,280 --> 00:31:59,800 Speaker 3: our industry partners, university partners working together, so you can 622 00:31:59,840 --> 00:32:01,520 Speaker 3: look got all of that, or you can just go 623 00:32:01,560 --> 00:32:03,560 Speaker 3: to jointdy Night dot org and get started. 624 00:32:03,920 --> 00:32:06,720 Speaker 1: Oh amazing. Okay, So now here's the part I want 625 00:32:06,760 --> 00:32:09,360 Speaker 1: to get to. Is something that you and I have 626 00:32:09,560 --> 00:32:14,000 Speaker 1: a common interest in is the analogies between the brain 627 00:32:14,600 --> 00:32:18,320 Speaker 1: and computation in general. And what's interesting is this has 628 00:32:18,360 --> 00:32:21,760 Speaker 1: a long history. You know. Johnny von Neuman, for example, 629 00:32:21,960 --> 00:32:24,280 Speaker 1: looked to the brain and said, oh, what I'm building 630 00:32:24,320 --> 00:32:26,720 Speaker 1: with this digital computer is like a brain. And nowadays 631 00:32:26,760 --> 00:32:28,320 Speaker 1: people often look at the brain and think, oh, it 632 00:32:28,400 --> 00:32:30,640 Speaker 1: is like a digital computer, and so on, and so 633 00:32:30,640 --> 00:32:34,080 Speaker 1: they're always looking to each other. But here's the interesting thing, 634 00:32:34,360 --> 00:32:37,600 Speaker 1: as I think you know, in my book Incognito a 635 00:32:37,680 --> 00:32:40,480 Speaker 1: while ago, I wrote that the brain, really the way 636 00:32:40,520 --> 00:32:42,520 Speaker 1: to think about this is like a team of rivals, 637 00:32:42,760 --> 00:32:45,760 Speaker 1: where you've got all these different neural networks that are 638 00:32:45,880 --> 00:32:50,320 Speaker 1: all trying to drive this ship of state, and they 639 00:32:50,400 --> 00:32:54,320 Speaker 1: have different desires and different needs and different outcomes that 640 00:32:54,320 --> 00:32:58,520 Speaker 1: they're looking for. And your behavior at any given moment 641 00:32:58,960 --> 00:33:01,960 Speaker 1: is who wins the parliamentary vote. And this is why 642 00:33:02,040 --> 00:33:05,400 Speaker 1: humans are so complex. This is why you can get 643 00:33:05,440 --> 00:33:08,200 Speaker 1: mad at yourself, or cuss at yourself, or cajole yourself, 644 00:33:08,280 --> 00:33:12,120 Speaker 1: or contract with yourself, because you know, if you were 645 00:33:12,120 --> 00:33:14,760 Speaker 1: a single thing, we'd say, wait, who's talking to whom exactly? 646 00:33:14,800 --> 00:33:16,560 Speaker 1: But what's happening is you're made up of all these 647 00:33:16,560 --> 00:33:20,400 Speaker 1: different things. So as a rough analogy for today's conversation. 648 00:33:20,840 --> 00:33:23,480 Speaker 1: It's like, we have a bunch of different agents that 649 00:33:23,520 --> 00:33:26,480 Speaker 1: are running in there, and somehow you get this emergent 650 00:33:26,640 --> 00:33:29,920 Speaker 1: behavior that we call you. But of course you're not 651 00:33:29,960 --> 00:33:32,760 Speaker 1: the same person from from moment to moment, no hour 652 00:33:32,800 --> 00:33:35,360 Speaker 1: to hour. So I want to explore this analogy here. 653 00:33:35,560 --> 00:33:37,360 Speaker 1: What's your reaction to this when you think about the 654 00:33:37,400 --> 00:33:41,120 Speaker 1: similarity between a future society of AI agents and what's 655 00:33:41,160 --> 00:33:41,880 Speaker 1: going on in the brain. 656 00:33:42,040 --> 00:33:44,560 Speaker 3: I mean, you know, Marvin Minsky also that Mighty Media Lab, 657 00:33:44,640 --> 00:33:47,160 Speaker 3: you know, talk about the society of the mind and 658 00:33:47,280 --> 00:33:49,760 Speaker 3: this notion that there's no single brain. But you have 659 00:33:49,840 --> 00:33:52,840 Speaker 3: all these components of brains you call them asians, and 660 00:33:52,880 --> 00:33:56,360 Speaker 3: they're all working together without a single orchestrator, and each 661 00:33:56,400 --> 00:34:00,080 Speaker 3: of them are seemingly doing meaningless task, but you know, 662 00:34:00,240 --> 00:34:02,920 Speaker 3: together they you know, they achieve something amazing. 663 00:34:03,080 --> 00:34:05,000 Speaker 1: And as you know, by the way, what I think 664 00:34:05,040 --> 00:34:08,720 Speaker 1: Minski's model is missing is competition. And that's what seems 665 00:34:08,719 --> 00:34:10,640 Speaker 1: to figuring on in the human brain, is that these 666 00:34:12,120 --> 00:34:16,160 Speaker 1: little agents are working together. They're all competing slash cooperating. 667 00:34:16,239 --> 00:34:18,600 Speaker 3: Yeah, and David, you wrote an article in I believe 668 00:34:18,600 --> 00:34:23,000 Speaker 3: twenty eleven eleven magazine where you talked about the conflict 669 00:34:23,400 --> 00:34:26,080 Speaker 3: as an important component of all these agents talking to 670 00:34:26,080 --> 00:34:26,359 Speaker 3: each other. 671 00:34:26,400 --> 00:34:27,719 Speaker 2: I thought that was amazing that you. 672 00:34:27,640 --> 00:34:32,839 Speaker 3: Wrote that even before the twenty thirteen alex netpaper which 673 00:34:32,920 --> 00:34:35,160 Speaker 3: brought us, you know, a lot of this amazing new 674 00:34:35,480 --> 00:34:38,200 Speaker 3: machine learning benefits, So you're like two years before that. 675 00:34:38,920 --> 00:34:43,720 Speaker 3: And so this notion of cooperation competition between this micro 676 00:34:43,840 --> 00:34:47,080 Speaker 3: ais or subcomponents, I think it's a very important idea. 677 00:34:47,520 --> 00:34:50,600 Speaker 3: And the fact that there's no global orchestrator either. Even 678 00:34:50,680 --> 00:34:56,160 Speaker 3: our prefrontal cortex is only doing integration but not doing orchestration, right, So. 679 00:34:56,080 --> 00:34:57,080 Speaker 2: That's an important concept. 680 00:34:57,120 --> 00:35:00,120 Speaker 3: So if you take that analogy further up, if you 681 00:35:00,160 --> 00:35:03,239 Speaker 3: use like that's decentralization, right, because you're letting all these 682 00:35:03,280 --> 00:35:07,719 Speaker 3: subcomponents that are doing highly specialized tasks and there's an 683 00:35:07,760 --> 00:35:12,759 Speaker 3: emergent intelligence from that is exactly a positive of how we. 684 00:35:12,680 --> 00:35:14,719 Speaker 2: Should be thinking about centralized intelligence. 685 00:35:15,360 --> 00:35:18,440 Speaker 3: And I think Open AI and you know, Google, they 686 00:35:18,480 --> 00:35:21,400 Speaker 3: all realized because what happened in the research world is 687 00:35:21,400 --> 00:35:25,400 Speaker 3: we started going towards concepts like mixture of experts, Like 688 00:35:25,440 --> 00:35:30,680 Speaker 3: what is that mixture of experts? Is highly specialized components 689 00:35:30,719 --> 00:35:34,200 Speaker 3: of machine learning models that work together to figure out 690 00:35:34,200 --> 00:35:36,840 Speaker 3: what you should do. And then they have these concepts 691 00:35:36,920 --> 00:35:39,960 Speaker 3: like routers, where when a task comes in, you decide 692 00:35:40,000 --> 00:35:42,840 Speaker 3: which part of the which experts within this mixture of 693 00:35:42,880 --> 00:35:46,320 Speaker 3: experts should be used. If you remember in the beginning 694 00:35:46,320 --> 00:35:49,719 Speaker 3: when charge epedy came out, it couldn't do math very well, 695 00:35:50,560 --> 00:35:53,080 Speaker 3: and of course you shouldn't be used a language model 696 00:35:53,120 --> 00:35:53,640 Speaker 3: to do math. 697 00:35:53,960 --> 00:35:55,840 Speaker 2: You should be routing it to a simple calculator. 698 00:35:56,640 --> 00:35:58,520 Speaker 3: So and now you know there are many other search 699 00:35:58,560 --> 00:36:01,680 Speaker 3: branches that happen in any model that you use, lessia 700 00:36:01,680 --> 00:36:04,719 Speaker 3: on chargipat. So what folks have realized very quickly is 701 00:36:04,760 --> 00:36:07,759 Speaker 3: that actually training one giant model doesn't make sense. We 702 00:36:07,800 --> 00:36:10,799 Speaker 3: wore to kind of route it to all this other intelligence. Now, 703 00:36:10,840 --> 00:36:13,560 Speaker 3: if you take that analogy further by the way, and 704 00:36:13,600 --> 00:36:16,080 Speaker 3: then we came to do this this modern world of 705 00:36:16,719 --> 00:36:20,120 Speaker 3: chain of thought or tree of thought and then reasoning 706 00:36:20,160 --> 00:36:22,520 Speaker 3: and all and all of that is basically pointing us 707 00:36:22,520 --> 00:36:25,760 Speaker 3: to this one direction, which is the concept of society 708 00:36:25,800 --> 00:36:30,520 Speaker 3: of the mind in a cooperation competition, the conflict that 709 00:36:30,560 --> 00:36:31,600 Speaker 3: you talked about. 710 00:36:31,840 --> 00:36:33,160 Speaker 2: So it's moving in that direction. 711 00:36:33,239 --> 00:36:36,279 Speaker 3: So why don't we imagine what could happen over time, 712 00:36:36,320 --> 00:36:39,600 Speaker 3: which is it's not only that, you know, all those 713 00:36:40,280 --> 00:36:44,200 Speaker 3: subcomponents of AI are working in this one machine of 714 00:36:44,360 --> 00:36:48,239 Speaker 3: chargepat but actually they're working across the world. And there's 715 00:36:48,239 --> 00:36:51,239 Speaker 3: an amazing health agent that's running in Manila, There's an 716 00:36:51,239 --> 00:36:54,480 Speaker 3: amazing legal agent that's running in you know, in in Rio. 717 00:36:55,040 --> 00:36:55,680 Speaker 2: You know, there's an. 718 00:36:55,560 --> 00:36:58,879 Speaker 3: Amazing softwaregent that's running in Bangalore. There could be all 719 00:36:58,880 --> 00:36:59,840 Speaker 3: these things running. 720 00:36:59,600 --> 00:37:00,400 Speaker 2: All over the world. 721 00:37:01,120 --> 00:37:04,480 Speaker 3: They are updated, you know, through innovators locally and through 722 00:37:04,520 --> 00:37:08,480 Speaker 3: which everyone gets benefits of this global intelligence. So we 723 00:37:08,520 --> 00:37:11,960 Speaker 3: can see that progression happening. As you said, in the 724 00:37:12,000 --> 00:37:14,840 Speaker 3: Society of the Mind, one missing piece was this notion 725 00:37:14,920 --> 00:37:19,560 Speaker 3: of rivalry or competition and conflict. Another piece that's missing 726 00:37:20,239 --> 00:37:22,719 Speaker 3: or is required, I would say as we go to 727 00:37:22,760 --> 00:37:24,960 Speaker 3: this decentalization is the. 728 00:37:24,800 --> 00:37:26,400 Speaker 2: Fact of incentives. 729 00:37:26,680 --> 00:37:30,520 Speaker 3: You know, why should this folks actually participate, and when 730 00:37:30,560 --> 00:37:32,520 Speaker 3: they do participate, how to make sure there are no 731 00:37:32,560 --> 00:37:35,719 Speaker 3: bad actors? And so if you kind of take that 732 00:37:36,520 --> 00:37:39,279 Speaker 3: thought process forward, you realize, wow, first of all, it's 733 00:37:39,280 --> 00:37:43,440 Speaker 3: not centralized, is decentalized. Second is that there is not 734 00:37:43,480 --> 00:37:48,160 Speaker 3: only cooperation, but there's competition and rivalry. Okay, but to 735 00:37:48,160 --> 00:37:51,759 Speaker 3: support that we need different incentives for them to participate. 736 00:37:52,080 --> 00:37:54,120 Speaker 3: But then if you proved incentives, bad actors will have 737 00:37:54,120 --> 00:37:56,719 Speaker 3: incenter to participate although they have nothing to contribute, or 738 00:37:56,760 --> 00:38:00,479 Speaker 3: they're maliciously contributing. So we need a way to deal 739 00:38:00,520 --> 00:38:02,880 Speaker 3: with those productors. And then you say, okay, at the 740 00:38:02,960 --> 00:38:06,160 Speaker 3: end of the day, we got all the thousands of 741 00:38:06,239 --> 00:38:08,680 Speaker 3: millions of ages doing strange takes. How to bring all 742 00:38:08,719 --> 00:38:11,160 Speaker 3: of that together, right, So you need some kind of 743 00:38:11,160 --> 00:38:14,399 Speaker 3: integration of that. So that's basically what Project Nanda is. 744 00:38:14,760 --> 00:38:17,919 Speaker 3: That's basically what Descent Last Day is. The same way 745 00:38:17,920 --> 00:38:21,760 Speaker 3: of brain and society's work is going to get mimicked 746 00:38:22,080 --> 00:38:23,400 Speaker 3: in the future of the Internet. 747 00:38:28,080 --> 00:38:31,000 Speaker 1: That was my interview with Ramesh Rascar. As I've talked 748 00:38:31,000 --> 00:38:35,000 Speaker 1: about on earlier episodes, we live in a world increasingly 749 00:38:35,160 --> 00:38:39,600 Speaker 1: shaped by invisible architectures, by algorithms we don't see, by 750 00:38:39,640 --> 00:38:44,799 Speaker 1: recommendations we don't question, and centralized systems that quietly make 751 00:38:44,880 --> 00:38:48,000 Speaker 1: decisions on our behalf. But what Ramesh is asking us 752 00:38:48,040 --> 00:38:51,880 Speaker 1: to picture is something different. He's asking what if AI 753 00:38:52,160 --> 00:38:55,520 Speaker 1: could protect your privacy instead of compromising it. What if 754 00:38:55,600 --> 00:38:59,480 Speaker 1: sharing data didn't mean giving up control over it? Buried 755 00:38:59,480 --> 00:39:03,080 Speaker 1: in here are there lots of questions about trust, about power, 756 00:39:03,400 --> 00:39:06,239 Speaker 1: about what kind of society we want to build, and 757 00:39:06,239 --> 00:39:08,120 Speaker 1: Remeasser would be the first to admit that there are 758 00:39:08,160 --> 00:39:10,960 Speaker 1: so many questions that still need to be worked out. 759 00:39:11,280 --> 00:39:13,919 Speaker 1: But let me say two things. First, I think it's 760 00:39:13,960 --> 00:39:17,400 Speaker 1: so deeply important that people like Ramesh and all of 761 00:39:17,400 --> 00:39:20,719 Speaker 1: his colleagues are pulling up chairs at the table to 762 00:39:20,800 --> 00:39:24,800 Speaker 1: try to figure out the structure of these questions because 763 00:39:25,239 --> 00:39:29,760 Speaker 1: agentic AI is already here and its future is only 764 00:39:29,840 --> 00:39:34,080 Speaker 1: getting more complex and fast and ubiquitous. So this new 765 00:39:34,200 --> 00:39:36,759 Speaker 1: type of society that we touched on today, with this 766 00:39:37,200 --> 00:39:41,440 Speaker 1: parallel species running at speeds we can't even comprehend all 767 00:39:41,520 --> 00:39:46,200 Speaker 1: this is an inevitability and it's not even particularly far away. 768 00:39:46,880 --> 00:39:50,720 Speaker 1: So I'm happy to see collections of smart humans trying 769 00:39:50,719 --> 00:39:55,000 Speaker 1: to build the right lanes and guardrails. Second, if the 770 00:39:55,000 --> 00:39:58,799 Speaker 1: brain teaches us anything it's that intelligence doesn't have to 771 00:39:58,880 --> 00:40:03,200 Speaker 1: come from top down command in the brain. It arises 772 00:40:03,640 --> 00:40:09,040 Speaker 1: from dialogue and disagreement and competition among subsystems that don't 773 00:40:09,040 --> 00:40:12,320 Speaker 1: always see eye to eye but still somehow make it work. 774 00:40:12,640 --> 00:40:16,200 Speaker 1: So in the AI world, what if intelligence could emerge 775 00:40:16,239 --> 00:40:21,520 Speaker 1: from a constellation of local minds, each acting in context, 776 00:40:21,840 --> 00:40:26,840 Speaker 1: but contributing to something larger than themselves. I'm really interested 777 00:40:26,880 --> 00:40:29,360 Speaker 1: to see what we are going to learn about intelligence 778 00:40:29,400 --> 00:40:32,680 Speaker 1: in ten years when we're looking at billions of agents 779 00:40:32,760 --> 00:40:38,240 Speaker 1: running around and we're asking if that collection somehow gives 780 00:40:38,320 --> 00:40:43,759 Speaker 1: us artificial general intelligence. So that's the current game, and 781 00:40:43,800 --> 00:40:48,759 Speaker 1: in Ramesha's view, our best shot at building ethical, robust, intelligent, 782 00:40:49,000 --> 00:40:52,640 Speaker 1: and democratic AI is to let go of the fantasy 783 00:40:52,680 --> 00:40:58,560 Speaker 1: of central control and the despotism of oligopolies and instead 784 00:40:59,160 --> 00:41:04,120 Speaker 1: learn from them messy, resilient genius of our own biology. 785 00:41:07,920 --> 00:41:10,680 Speaker 1: Go to eagleman dot com slash podcast for more information 786 00:41:10,760 --> 00:41:13,759 Speaker 1: and to find further reading. Join the weekly discussions on 787 00:41:13,800 --> 00:41:17,040 Speaker 1: my substack, and check out and subscribe to Inner Cosmos 788 00:41:17,080 --> 00:41:19,600 Speaker 1: on YouTube for videos of each episode and to leave 789 00:41:19,640 --> 00:41:23,200 Speaker 1: comments until next time. I'm David Eagleman, and this is 790 00:41:23,320 --> 00:41:24,279 Speaker 1: Inner Cosmos.