WEBVTT - Ep121 "What’s the secret to intelligence (in brains and AI)?" with Ramesh Raskar 0:00:05.120 --> 0:00:08.799 In the early days of computing, everyone thought about a big, 0:00:08.880 --> 0:00:13.680 colossal mainframe computer that had all the data, and eventually 0:00:13.720 --> 0:00:16.960 the model switched and what we ended up with is 0:00:17.440 --> 0:00:23.079 billions of individual small computers, all networked and operating as 0:00:23.079 --> 0:00:27.400 a collective. Might AI go the same way, where we 0:00:27.480 --> 0:00:30.760 move from giant, large language models owned by a few 0:00:30.840 --> 0:00:36.199 companies to lots and lots of individual AI agents that 0:00:36.280 --> 0:00:40.519 are decentralized and operating together. And what does this have 0:00:40.600 --> 0:00:44.040 to do with solving diseases? And why in the future 0:00:44.080 --> 0:00:47.120 we'll have an army of agents solving problems for us 0:00:47.120 --> 0:00:49.440 out there. And what to make of the fact that 0:00:49.479 --> 0:00:53.600 we will have created a whole new species that runs 0:00:53.680 --> 0:00:57.480 at a timescale a trillion times faster than we do. 0:01:00.680 --> 0:01:02.080 Welcome to Inner Cosmos with. 0:01:02.080 --> 0:01:02.920 Me David Eagelman. 0:01:03.080 --> 0:01:06.240 I'm a neuroscientist and author at Stanford and in these 0:01:06.240 --> 0:01:09.880 episodes we sail deeply into our three pound universe to 0:01:10.120 --> 0:01:30.720 understand why our lives look the way they do. Today's 0:01:30.720 --> 0:01:37.280 episode is about what happens when intelligence becomes decentralized. When 0:01:37.280 --> 0:01:40.679 we talk about artificial intelligence, the conversation is typically about 0:01:41.080 --> 0:01:46.280 one giant model, a mind blowingly massive data set and 0:01:46.400 --> 0:01:50.200 a huge compute cluster, so that we get one singular 0:01:50.240 --> 0:01:52.640 model that takes care of everything. But the first thing 0:01:52.680 --> 0:01:56.160 to note is that's not how the brain works. As 0:01:56.200 --> 0:01:59.360 I argued in my book Incognito, the brain is best 0:01:59.480 --> 0:02:04.760 understood as a team of ribles. Inside your skull are 0:02:05.360 --> 0:02:10.040 scores or hundreds of specialized modules. Some care about vision, 0:02:10.160 --> 0:02:14.079 hearing touched, some about walking, some about processing faces. Others 0:02:14.160 --> 0:02:18.000 track moving objects in the world. Other systems are working 0:02:18.040 --> 0:02:21.280 to predict what you should do next. But the key 0:02:21.560 --> 0:02:24.000 is that these systems in the brain are not like 0:02:24.280 --> 0:02:27.680 pieces of a very efficient machine where each part serves 0:02:27.720 --> 0:02:31.720 its purpose. Instead, every second of your life, these networks 0:02:31.800 --> 0:02:36.919 find themselves in conflict. They argue, they interrupt, they compete 0:02:36.919 --> 0:02:41.560 for dominance. The feeling of having a unified self is 0:02:41.680 --> 0:02:46.160 an illusion stitched together on the fly. So maybe the 0:02:46.200 --> 0:02:51.280 future of AI isn't one giant LLM to rule them all, 0:02:51.680 --> 0:02:55.519 but something more like what we are. A distributed system, 0:02:55.560 --> 0:03:01.160 a team of rival agents and argumentative network of local experts, 0:03:01.600 --> 0:03:06.400 each with limited knowledge and a unique perspective, together making 0:03:06.440 --> 0:03:10.320 something more powerful than any individual part. But how do 0:03:10.360 --> 0:03:14.320 you build that? How do you preserve data? Privacy. When 0:03:14.440 --> 0:03:18.520 intelligence is spread across devices and across hospitals and across 0:03:18.560 --> 0:03:22.760 country borders. How do you incentivize data sharing? How do 0:03:22.800 --> 0:03:24.840 you verify who's a good actor? 0:03:25.280 --> 0:03:25.919 These are the. 0:03:25.919 --> 0:03:29.639 Kinds of questions that animate my friend and colleague, Ramesh Roscar. 0:03:30.080 --> 0:03:33.079 He's a professor at the MIT Media Lab. He's an inventor, 0:03:33.160 --> 0:03:36.360 He's a systems thinker and a pioneer in this new 0:03:36.400 --> 0:03:41.320 world of decentralized AI. His work spands from privacy preserving 0:03:41.400 --> 0:03:45.680 technologies to public health tools, to new ways of imagining 0:03:45.720 --> 0:03:48.760 what AI can be from a technical point of view, 0:03:48.800 --> 0:03:52.480 but also ethically and socially. So today we're going to 0:03:52.480 --> 0:03:56.920 talk about why the future of AI might lie in localized, 0:03:57.040 --> 0:04:00.280 specialized systems that are trained on the fly, and how 0:04:00.320 --> 0:04:04.480 new architectures might allow data to stay where it is 0:04:04.640 --> 0:04:07.040 on your phone and your clinic and your city, while 0:04:07.120 --> 0:04:10.520 still contributing to a global model. So this is a 0:04:10.520 --> 0:04:15.600 conversation about architecture and ethics, competition and cooperation, and the 0:04:15.760 --> 0:04:19.640 radically different future that becomes possible when we stop trying 0:04:19.680 --> 0:04:25.000 to centralize everything and instead embrace the power of the network. 0:04:29.360 --> 0:04:31.839 So let's think, as an example, the healthcare systems. So 0:04:31.920 --> 0:04:34.240 the idea is you've got all this data locked up 0:04:34.320 --> 0:04:36.680 in electronic health records all over the place. None of 0:04:36.720 --> 0:04:40.880 the hospital systems want to share that data because they 0:04:40.880 --> 0:04:43.200 don't want other people looking at it. And it's valuable data. 0:04:43.240 --> 0:04:47.520 And so the idea of decentralized AI, for example, could 0:04:47.600 --> 0:04:52.320 be that an AI is looking at all the different 0:04:52.520 --> 0:04:55.599 electronic health records and all the different hospital systems, but 0:04:56.279 --> 0:04:59.440 no human is looking at it. With millions of records, 0:04:59.440 --> 0:05:01.520 it puts together there a big picture and discover sing 0:05:01.680 --> 0:05:02.000 is that right? 0:05:02.839 --> 0:05:03.000 Right? 0:05:03.000 --> 0:05:05.000 I mean health is a great example because I would 0:05:05.120 --> 0:05:08.279 argue that if somehow I have for every health condition 0:05:08.880 --> 0:05:10.520 all the data, I can bring it in one place 0:05:10.560 --> 0:05:12.839 and train models, I can solve it. The problem is, 0:05:13.440 --> 0:05:15.599 you know, the capitalistic system doesn't allow us to do 0:05:15.640 --> 0:05:18.280 it because of privacy, because of trade secrets, and as 0:05:18.320 --> 0:05:21.200 I said, people hold on to the data thinking it's 0:05:21.320 --> 0:05:24.760 worth something, but they don't get paid either. So if 0:05:24.800 --> 0:05:28.720 there was a marketplace for data so that they're not 0:05:28.720 --> 0:05:31.480 to share the data, but they can share the insights 0:05:31.520 --> 0:05:35.839 from it and get paid in return, then everybody would 0:05:35.839 --> 0:05:39.800 be incentivized to contribute to these insights without contributing raw data. 0:05:39.880 --> 0:05:42.200 So the idea is, let's say I'm a hospital system. 0:05:42.440 --> 0:05:45.360 I'm not giving away the electronic health records. Instead, I'm 0:05:45.440 --> 0:05:48.719 letting an AI poke its nose into these things, figure 0:05:48.760 --> 0:05:54.400 out insights from the data, and leave. And that way, 0:05:54.400 --> 0:05:56.960 I haven't given up something in particular, and I get 0:05:56.960 --> 0:05:58.320 paid for that. By the way, that's right. 0:05:58.400 --> 0:06:01.279 I mean another way to think about this is again, 0:06:01.320 --> 0:06:04.800 you want to share intelligence without sharing raw data, and 0:06:04.839 --> 0:06:06.320 that's how the human society works. 0:06:06.600 --> 0:06:09.400 The analogy would be, if I had enough money, you know. 0:06:09.400 --> 0:06:13.080 I could hire you know, a thousand medical interns and 0:06:13.520 --> 0:06:15.920 send them to If I want to solve partical health condition, 0:06:16.440 --> 0:06:19.599 I would send them to every hospital, ask them to 0:06:19.640 --> 0:06:22.040 do an internship for a year, learn about everything that 0:06:22.040 --> 0:06:24.880 partical hospital does, about how they treat a potical condition, 0:06:25.279 --> 0:06:28.440 and how their ideology pathology works, and then bring all ten, 0:06:28.880 --> 0:06:32.400 all thousands of them in one office and extract their 0:06:32.400 --> 0:06:35.799 intelligence and then solve any condition that's out there. 0:06:36.200 --> 0:06:38.080 And that would be a way to bring. 0:06:37.880 --> 0:06:40.880 Intelligence from all these thousand hospitals without bringing in your 0:06:40.920 --> 0:06:43.160 raw data. So how do you now make that happen 0:06:43.480 --> 0:06:45.679 not in the physical world, but in the digital world. 0:06:46.040 --> 0:06:49.719 By effectively sending a MICROAI like an agent that behaves 0:06:49.720 --> 0:06:52.520 like an intern that learns everything and brings back only 0:06:52.560 --> 0:06:55.800 the intelligence and in return pace for the stay at 0:06:55.839 --> 0:06:59.440 the hospital, and then you know, creates a global algorithm 0:06:59.480 --> 0:06:59.760 for that. 0:07:00.000 --> 0:07:04.080 So let's take one second to distinguish decentralized AI from 0:07:04.360 --> 0:07:07.520 agentic AI. They're related, but tells the difference. 0:07:07.800 --> 0:07:09.880 Yeah, I mean the agent DKA is a great way 0:07:09.920 --> 0:07:14.480 to think about it, because mix it very simple to understand. 0:07:14.560 --> 0:07:18.840 So the reason to use agent as an anology as 0:07:18.840 --> 0:07:22.679 an anchor for decent lazed AI is because you imagine 0:07:22.680 --> 0:07:25.600 an agent is like a medical intern, and the agent 0:07:25.640 --> 0:07:28.559 can go to the hospital, you know, you know, stay 0:07:28.560 --> 0:07:31.920 there for some time, learn everything about the data, pay 0:07:32.000 --> 0:07:34.600 for anything that they have learned, and then come back. 0:07:34.880 --> 0:07:37.360 And I can take this thousand agents and now I 0:07:37.360 --> 0:07:40.440 can ask them any new question about that health condition. 0:07:41.240 --> 0:07:43.280 And the beauty here is that I know that the 0:07:43.320 --> 0:07:45.800 agent did not steal any data because I can send 0:07:45.880 --> 0:07:48.640 an agent as a model. And model, as you know, 0:07:48.760 --> 0:07:51.000 is you know, basically a big file let's say one 0:07:51.040 --> 0:07:54.520 hundred megabyte. You a model, it goes in it looks 0:07:54.560 --> 0:07:57.200 like gigabytes of data in the hospital, but comes back 0:07:57.240 --> 0:07:59.520 again only as a hundred megabyte model. 0:08:00.040 --> 0:08:03.360 So that's you tweaked the parameters in the model exactly. 0:08:03.360 --> 0:08:05.320 We're not bringing back the raw data, yeah, which is 0:08:05.360 --> 0:08:07.640 what happens in the human brain. So if you think 0:08:07.640 --> 0:08:11.200 about centralized A versus decentized AI, in centilis you have 0:08:11.240 --> 0:08:15.320 this one massive model that's doing everything. Decentralized AI matches 0:08:15.400 --> 0:08:18.440 very well with the concept of agent AI agent take 0:08:18.480 --> 0:08:20.840 AI because the notion is that each of the agent 0:08:21.040 --> 0:08:24.480 is not just massive six hundred billion parameter models that 0:08:24.560 --> 0:08:27.760 you're hearing about from open AI and meta and so on, 0:08:27.920 --> 0:08:31.720 and these agents are much smaller, and they're very specific 0:08:31.920 --> 0:08:33.360 and they do specific tasks. 0:08:34.280 --> 0:08:35.400 So a legal. 0:08:35.160 --> 0:08:37.600 Agent may not notizing about the health agent or financial 0:08:37.640 --> 0:08:41.120 agents and so on. They're doing very specific things, and 0:08:41.160 --> 0:08:43.440 then they can get better over time, and they get 0:08:43.440 --> 0:08:45.760 better in two ways. They get better because they're looking 0:08:45.760 --> 0:08:49.160 at new data, but more importantly, they get better because 0:08:49.160 --> 0:08:53.880 they're interacting with other agents. And that's what decentalization decential 0:08:53.880 --> 0:08:56.920 plays in kind of mimics the human society because even 0:08:56.960 --> 0:08:59.360 for us, you know, growing up, we learn from some 0:08:59.440 --> 0:09:01.959 data and some experiences, but most of the things we 0:09:02.080 --> 0:09:03.640 learn are actually through interactions. 0:09:05.280 --> 0:09:07.720 So are you saying decentralized AI in a gent ki 0:09:08.000 --> 0:09:10.720 is the same thing, or a gent ki is a 0:09:10.760 --> 0:09:12.680 good example of decentralized AI. 0:09:12.880 --> 0:09:15.360 Agent ki is is kind of the brain that you need, 0:09:15.920 --> 0:09:20.480 but you need other restual ecosystem to emerge for decentralization. So, 0:09:20.520 --> 0:09:23.160 for example, what does it mean to have for agents 0:09:23.200 --> 0:09:26.080 to have commerce? How do agents pay for each other? 0:09:26.360 --> 0:09:28.640 You know, how do I know that a data that 0:09:28.679 --> 0:09:31.360 I you know, that data that I used from a hospital, 0:09:31.480 --> 0:09:33.800 is it worth one hundred dollars, thousand dollars or a 0:09:33.800 --> 0:09:36.440 million dollars? You al sort of figure out, you know, 0:09:36.440 --> 0:09:39.359 how do you make sure that the agents are not misbehaving? 0:09:39.800 --> 0:09:43.240 So beyond the intelligence of a given agent, which is 0:09:43.320 --> 0:09:45.840 classic agent kai, we don't think about the rest of 0:09:45.880 --> 0:09:49.240 the decentralization aspect of it. It's like saying, you know, 0:09:49.280 --> 0:09:52.840 I might have ability to do transactions, but that's very 0:09:52.840 --> 0:09:54.280 different than building a stock market. 0:09:54.440 --> 0:09:56.480 And so give us an example of what the world 0:09:56.559 --> 0:09:58.680 will look like. Let's say in ten years from now, 0:09:59.000 --> 0:10:01.200 when we've got a lot agents and we have this 0:10:01.240 --> 0:10:03.920 whole infrastructure for agents to communicate and pay each other 0:10:03.960 --> 0:10:06.760 and so on, what does that mean for all of us, it's. 0:10:06.679 --> 0:10:07.439 Difficult to product. 0:10:07.600 --> 0:10:10.439 But I would say in the short run, every one 0:10:10.480 --> 0:10:13.120 of us will have our own agent. You know, every 0:10:13.200 --> 0:10:16.880 organization will have its own agent. And these agents will 0:10:17.040 --> 0:10:20.360 work on our behalf. So they're anchored our identity or 0:10:20.400 --> 0:10:23.040 identity of the organization and they're working on our behalf 0:10:23.120 --> 0:10:25.840 and they are you know, they're doing commerce on our behalf. 0:10:26.120 --> 0:10:29.240 So if you think about kind of the roadmap of 0:10:29.360 --> 0:10:33.160 how this agent tech world will evolve, in the beginning, 0:10:33.559 --> 0:10:37.559 it's just about foundations, like do these agents have identity, 0:10:38.120 --> 0:10:41.320 do they have authentication? Can they really represent us? You know, 0:10:41.360 --> 0:10:44.839 what's the reputation when when my agent is using another agent? 0:10:45.200 --> 0:10:47.560 You know, can it trust? You know, like you know 0:10:47.600 --> 0:10:50.319 what ky C is, like I know your agent. All 0:10:50.360 --> 0:10:54.000 the foundational stuff. That's what's happening now in the industry. 0:10:54.040 --> 0:10:56.320 And also work at an IM. The next phase is 0:10:56.640 --> 0:10:59.280 agient commerce. Just as if you move to a new city, 0:10:59.600 --> 0:11:01.839 you know, you're figure out, you know what job you're 0:11:01.840 --> 0:11:03.240 going to get. You know, are you going to get 0:11:03.240 --> 0:11:05.880 paid fifteen dollars an hour or thousand dollars an hour? 0:11:06.640 --> 0:11:07.160 And so on? 0:11:07.640 --> 0:11:11.480 You know, you start learning, you start working over time, 0:11:11.760 --> 0:11:13.960 your skills are not that great, so you go to 0:11:14.080 --> 0:11:15.800 you know, a school, so there'll be in agient schools 0:11:15.840 --> 0:11:17.120 or agents go to agent schools. 0:11:17.480 --> 0:11:19.680 Sometimes they need to be repaired. It will be agent repairs. 0:11:20.400 --> 0:11:22.720 Sometimes there will be a risk of using an agent, 0:11:22.760 --> 0:11:25.240 so you'll have agent insurances, you know, and so on. 0:11:25.280 --> 0:11:26.400 So you have like a whole. 0:11:26.480 --> 0:11:29.920 Commerce and economy around agents. But then the third phases 0:11:30.040 --> 0:11:34.160 very interesting, which is you have very emergent behaviors with agents. 0:11:34.200 --> 0:11:36.120 Just as once you move to the city, you don't 0:11:36.160 --> 0:11:39.440 just work and you know, think about your livelihood, but 0:11:39.480 --> 0:11:42.840 you start creating social circles. You know, you have sports teams, 0:11:43.120 --> 0:11:47.440 you know, you have universities, you have you know, justice 0:11:47.480 --> 0:11:50.520 and court. So you get all this emergent behavior when 0:11:50.559 --> 0:11:53.480 folks get together and just as agents get together. So 0:11:53.520 --> 0:11:56.720 you can see this going from foundations to Asian commerce 0:11:56.760 --> 0:11:59.719 to this you know, emergent behavior of agent societies. But 0:12:00.200 --> 0:12:03.120 just assumes that the agents are anchored at human identity, 0:12:03.720 --> 0:12:06.240 we can't really see beyond this event horizon of what 0:12:06.320 --> 0:12:09.400 will happen when agents that are billions may be trillance 0:12:09.440 --> 0:12:13.920 of agents are navigating on their own on the internet, 0:12:14.000 --> 0:12:17.800 you know, meeting other agents and learning and socializing and 0:12:18.120 --> 0:12:20.840 you know, transacting on the on by themselves. So it's 0:12:20.880 --> 0:12:24.600 kind of diffrully imagine what will happen beyond beyond the 0:12:24.640 --> 0:12:29.400 scenario where they're still representing and anchored for individual human being. 0:12:29.520 --> 0:12:33.560 And even when they are really anchored on us, it's 0:12:33.600 --> 0:12:35.600 still very difficult to see what's going to happen because 0:12:35.640 --> 0:12:37.400 they can run it trillion times faster. 0:12:37.200 --> 0:12:40.079 Than we can, and they can replicate, and they can 0:12:40.240 --> 0:12:42.440 be you know, in multiple less at the same time, 0:12:42.840 --> 0:12:45.360 and they can migrate. You know, they can behave one 0:12:45.400 --> 0:12:47.480 way and we have another way in the next many 0:12:47.480 --> 0:12:48.679 second and so on. 0:12:48.840 --> 0:12:50.240 So I have a question. So a lot of people, 0:12:50.240 --> 0:12:54.559 of course, are worried about AI alignment in what happens 0:12:54.600 --> 0:12:59.440 if things go bad, And in this case of decentralization, 0:12:59.600 --> 0:13:02.920 with all these agents running around working at speeds that 0:13:02.960 --> 0:13:06.199 are much faster than we can comprehend, that it does 0:13:06.240 --> 0:13:09.480 seem like a bit of a dangerous territory to get into. Now, 0:13:09.520 --> 0:13:13.439 you might say that centralized AI also has its dangers, 0:13:13.920 --> 0:13:16.520 and that what we need maybe are checks and balances 0:13:16.600 --> 0:13:18.199 with decentral ied AA. But what does that look like? 0:13:18.240 --> 0:13:20.240 What would that mean to put checks and balances. 0:13:19.880 --> 0:13:23.280 And glad you've brought up alignment in centralized AI. The 0:13:23.360 --> 0:13:26.480 problem is different when you think about decentralization and agent 0:13:26.559 --> 0:13:28.840 KI because now we don't have to worry about you know, 0:13:28.920 --> 0:13:33.080 some one evil character taking over the world, but agents 0:13:33.120 --> 0:13:36.400 on their own could do bad things. So as long 0:13:36.440 --> 0:13:40.480 as they've anchored at a human identity I think should 0:13:40.520 --> 0:13:42.520 be in a reasonably good ship. It's almostly using like 0:13:42.559 --> 0:13:45.160 two factor authentication, you know, like the agent cannot do 0:13:45.240 --> 0:13:48.440 anything unless that says something on my phone, just using 0:13:48.440 --> 0:13:51.000 today's anology, so that should be under control. 0:13:51.800 --> 0:13:52.760 But once we go. 0:13:52.760 --> 0:13:55.760 Beyond that and agents are sufficiently autonomous that can create 0:13:55.840 --> 0:13:59.240 their own societies. Now the challenge is, you know, how 0:13:59.280 --> 0:14:01.480 do you make showed that doing the right thing. So 0:14:01.520 --> 0:14:04.120 the alignment becomes a very different problem. It becomes an 0:14:04.160 --> 0:14:08.040 algorithmic orchestration problem. So how do you have algorithms that 0:14:08.160 --> 0:14:12.240 orchestrate you know, the incentives of the carrots and sticks 0:14:12.280 --> 0:14:14.600 for the agents to behave. And that's actually a very 0:14:14.600 --> 0:14:17.480 complex problem. That's a lot for research at m T 0:14:17.840 --> 0:14:22.440 as well, because you could imagine that you know, if 0:14:22.480 --> 0:14:25.200 you if you try to use decentralization, like there were 0:14:25.280 --> 0:14:29.200 decentil issues. For DeFi, in the crypto and web three world, 0:14:30.000 --> 0:14:33.000 it didn't go very well where examples like Luna and 0:14:33.120 --> 0:14:36.480 Terra compared to stable coins today, and they had, you know, 0:14:36.920 --> 0:14:39.520 an orchestration to figure out, how do we maintain you know, 0:14:39.560 --> 0:14:40.720 their one dollar anchor. 0:14:41.200 --> 0:14:43.280 So anyway, I think that ended up crashing. 0:14:43.320 --> 0:14:45.280 And they ended up crashing because they were based on 0:14:45.760 --> 0:14:48.680 the behavioral economics that was predicted by a bunch of 0:14:48.760 --> 0:14:51.840 geeks sitting in a garage, and that didn't work out. 0:14:52.000 --> 0:14:54.320 So I think we should not rely on algorithmic orchestration 0:14:54.720 --> 0:14:57.000 of society of humans or society of agents. 0:14:57.520 --> 0:14:59.240 And I think there are very important problems to solve. 0:14:59.480 --> 0:15:01.720 So what is the thing beyond algorithmic orchestration? 0:15:02.240 --> 0:15:04.560 So you need to slow things down, You need to 0:15:04.640 --> 0:15:08.280 have some kind of a governance principle that's out there, 0:15:08.520 --> 0:15:10.400 and you know, that's how our society works. All the 0:15:10.480 --> 0:15:14.960 banks and courtes could run you know, algorithmically in many cases, 0:15:15.200 --> 0:15:17.520 we still have human in the loop and we slow 0:15:17.560 --> 0:15:21.960 things down and we trade inefficiency for you know, some 0:15:22.000 --> 0:15:23.400 control over the system. 0:15:23.680 --> 0:15:25.080 So I think that becomes very critical. 0:15:25.280 --> 0:15:27.960 So what does this mean for biology research? For example? 0:15:28.240 --> 0:15:31.240 Yeah, I think scientific research is going to depend on 0:15:31.280 --> 0:15:34.680 some kind of a decentialization because very unlikely that you know, 0:15:34.760 --> 0:15:37.480 all the data will come in plut complace, and you know, 0:15:37.560 --> 0:15:39.400 all the scientists will just pick up the phone and 0:15:39.400 --> 0:15:41.640 start talking to each other. They're all you know, we're 0:15:41.640 --> 0:15:44.960 all scientists, and we like cooperation, but also like competition, 0:15:45.520 --> 0:15:49.040 and you know, the human motivation to compete is what actually. 0:15:48.720 --> 0:15:49.840 Gets gets us going. 0:15:50.560 --> 0:15:52.920 So there will still be reviews, a competition to get 0:15:52.920 --> 0:15:55.880 your paper published in top journals, but we need a 0:15:55.880 --> 0:15:58.600 way for them to share their insights at the same time, 0:15:58.680 --> 0:16:00.920 and for those insights they get something in return, so 0:16:00.960 --> 0:16:04.960 it's crowdsourced, but something in return. I like to use 0:16:04.960 --> 0:16:09.600 analogy of Google Maps and traffic on Google Maps. You know, 0:16:10.280 --> 0:16:12.960 before that removes, there was Garment Maps, which we barely used. 0:16:13.320 --> 0:16:14.400 And today, even if. 0:16:14.320 --> 0:16:15.760 I know how to go from point to point B, 0:16:16.320 --> 0:16:18.640 you know, I open Google Maps not because I want 0:16:18.680 --> 0:16:20.280 to know the route, but I want to know where 0:16:20.280 --> 0:16:23.720 everybody else is. So I'm benefiting from the fact that 0:16:23.800 --> 0:16:27.640 everybody else is contributing the GPS coordinate and I get 0:16:27.680 --> 0:16:29.800 to see, you know, the reds and the greens on 0:16:29.840 --> 0:16:32.720 the map and turn, but turn navigation, but I'm also 0:16:32.760 --> 0:16:34.520 giving up my own GPS. 0:16:34.120 --> 0:16:35.720 Location while I'm driving. 0:16:35.920 --> 0:16:38.920 I'm contributing to the system and I'm getting great benefits 0:16:38.960 --> 0:16:41.840 in return. So can we create the equivalent of Google 0:16:41.880 --> 0:16:44.760 Maps for any societal problem, you know, whether it's scientific 0:16:44.800 --> 0:16:48.880 discovery or finding, you know, a treatment plan. Again, if 0:16:48.880 --> 0:16:51.360 you can centralize all the data, we can solve these problems. 0:16:51.360 --> 0:16:53.960 We cannot centralize it, and so we must use a 0:16:53.960 --> 0:16:54.880 decentralized method. 0:16:55.000 --> 0:16:56.440 And let me just make sure you have that. Why 0:16:56.440 --> 0:16:59.000 can't we centralize it? Isn't that what open ai and 0:16:59.040 --> 0:17:00.480 meta and everyone else is doing. 0:17:00.640 --> 0:17:02.480 So, I mean, if you look at open ai and 0:17:02.720 --> 0:17:05.440 everybody else, they you know, they scanned the whole Internet, 0:17:05.840 --> 0:17:08.280 you know, kind of in a permission less way. And 0:17:08.280 --> 0:17:10.800 that's why you could train this multi tens of trill 0:17:10.880 --> 0:17:14.120 and token models that are out there. But as I said, 0:17:14.160 --> 0:17:15.920 that's like just the tip of the iceberg. Most of 0:17:15.960 --> 0:17:18.880 the information in the world is actually locked away. If 0:17:18.880 --> 0:17:22.000 it takes health as an example, it locked away in hospitals, 0:17:22.440 --> 0:17:26.600 in pharmaceutical companies, in all the failed trials, you know, 0:17:26.760 --> 0:17:29.720 and all the research work in the lab notebooks of 0:17:29.880 --> 0:17:32.480 you know, all the amazing postdocs and students. 0:17:32.840 --> 0:17:35.080 So all the information is actually not available. 0:17:35.240 --> 0:17:38.399 So think about if all the information was available to 0:17:38.560 --> 0:17:41.120 open AI, it would do an amazing thing. It would 0:17:41.160 --> 0:17:45.360 solve healthcare overnight. But it cannot because of the competition. 0:18:00.080 --> 0:18:02.040 You know, I'm struck with it. An analogy here. I 0:18:02.080 --> 0:18:06.280 saw an interview with Isaac asimof in nineteen eighty eight, 0:18:06.320 --> 0:18:09.160 I think it was he was on jonah Lerror's show, 0:18:09.680 --> 0:18:13.000 and he said, look, I foresee a day when there's 0:18:13.040 --> 0:18:16.359 a central mainframe that has all of human kind's knowledge 0:18:16.359 --> 0:18:19.479 and everyone has a cable to their house and they 0:18:19.520 --> 0:18:21.719 can access the knowledge of the world that way. And 0:18:21.760 --> 0:18:24.840 he was almost right. But what happened instead was instead 0:18:24.840 --> 0:18:29.640 of a centralized knowledge repository, we it became completely decentralized 0:18:29.640 --> 0:18:30.600 on the Internet. 0:18:30.960 --> 0:18:35.160 Yes, I'm glad you mention that, because you know, today 0:18:35.200 --> 0:18:38.199 we're kind of the mainframe era of AI, and from 0:18:38.280 --> 0:18:40.720 that we're going to switch to the PC era of AI, 0:18:40.880 --> 0:18:43.520 to the Internet era of AI. So, as you know, 0:18:43.560 --> 0:18:47.320 in the eighties, big companies like IBM and Silicon Graphics 0:18:47.359 --> 0:18:50.760 were selled as big mainframes and we would literally connect 0:18:50.920 --> 0:18:55.320 to these machines or in a dial up motems, you know, yeah, 0:18:55.359 --> 0:18:57.320 and even to use a calculator, by the way, we 0:18:57.359 --> 0:18:59.320 had to use a dialup motem. It was not available 0:18:59.359 --> 0:19:01.880 on the machines. But in the late eighties and nineties 0:19:01.880 --> 0:19:04.439 we said, hey, that's kind of stupid. I should just 0:19:04.480 --> 0:19:07.600 have that calculator or a word or whatever the software 0:19:07.680 --> 0:19:10.359 using on my machine on a home machine. And the 0:19:10.359 --> 0:19:12.520 fact that we moved from a mainframe era a PCA 0:19:12.600 --> 0:19:16.520 I was interesting, but we needed one more thing. Before that, 0:19:16.560 --> 0:19:19.520 there was Internet, which was mostly connecting machines, and Tim 0:19:19.520 --> 0:19:22.959 Bernersley and others said in the nineties that actually what 0:19:23.000 --> 0:19:25.400 we need is a world wide web, and he brought 0:19:25.440 --> 0:19:29.800 in four things, a browser, protocols like SGTP and STML, 0:19:30.880 --> 0:19:32.040 and a notion of URL. 0:19:32.040 --> 0:19:33.479 How do you discover something right? 0:19:33.960 --> 0:19:37.800 And with those four things he transformed the Internet into 0:19:37.840 --> 0:19:40.880 world wide Web so that humans can interact with all 0:19:40.920 --> 0:19:44.800 the stuff that's out there. And one would argue that 0:19:45.160 --> 0:19:50.160 the impact of the mainframe revolution and PC revolution is 0:19:50.240 --> 0:19:53.640 insignificant compared to the impact of the world wide Web, 0:19:54.480 --> 0:19:57.280 and the same thing is true today. The impact of 0:19:57.440 --> 0:20:02.359 AI will be insignificant to the impact of the Internet 0:20:02.480 --> 0:20:04.359 and the web of AI that's or. 0:20:04.400 --> 0:20:09.560 There terrific, and so tell us about Project Nanda. 0:20:10.000 --> 0:20:15.600 So this decentralization agent tech concepts very nicely map us 0:20:15.680 --> 0:20:21.119 to this Project Nanda. Nanda stands for networked AI agents 0:20:21.160 --> 0:20:24.920 in decentralized architecture, but it's also the name of my sister, 0:20:25.640 --> 0:20:28.040 and Nanda means joy in Sanskrit. 0:20:28.560 --> 0:20:30.560 So I think, let's bring some joy to. 0:20:30.880 --> 0:20:34.480 The technology and the Internet and the web of AA agents. 0:20:34.840 --> 0:20:36.000 What is project going to do? 0:20:36.520 --> 0:20:38.879 So Project Nanda is a projecda is A is a 0:20:38.960 --> 0:20:43.080 large collective. We're looking at this three phases, how how 0:20:43.119 --> 0:20:45.320 the agent taic world is going to evolve, you know 0:20:45.359 --> 0:20:49.040 the foundations agent e commerce and you know agentic societies, 0:20:49.600 --> 0:20:54.919 uh and involves multiple universities, multiple companies, and we're working 0:20:54.920 --> 0:20:58.800 together to figure out how this you know, web of 0:20:59.240 --> 0:21:03.040 agents willieve the same way Tim berners Lee launched world 0:21:03.040 --> 0:21:04.720 Ward Web Constium. 0:21:04.760 --> 0:21:07.120 Alside of my Tea to see how we can use. 0:21:07.000 --> 0:21:11.000 The Internet technologies in a responsible way and keep the 0:21:11.040 --> 0:21:14.720 flock together. And so one should thank Tim berners Lee 0:21:14.920 --> 0:21:19.080 for keeping it an open, vibrant, you know, and neutral 0:21:19.320 --> 0:21:21.840 web that's out there, as opposed to what we saw 0:21:21.920 --> 0:21:25.439 later on in the computing you know generation, which is 0:21:25.720 --> 0:21:29.040 things like social media or things like mobile phones are 0:21:29.119 --> 0:21:33.159 highly fragmented systems and a few companies control you know, 0:21:33.240 --> 0:21:37.000 these ecosystems. As opposed to the World War Web. Today 0:21:37.080 --> 0:21:39.480 you can go online, go to go daday dot com, 0:21:39.520 --> 0:21:42.919 create a website, launch a business, create new innovations, do 0:21:43.000 --> 0:21:46.680 whatever you want. And that permissionless nature of the world 0:21:46.720 --> 0:21:49.840