Nodestar: Turning Networks into Knowledge w/ Andrew Trask

(upbeat music) Hey there, welcome back. This is our last episode in our "Node Star" series, although as with other series, there's different people we wanna talk to and also people come out of the woodwork when we do these episodes with really good ideas for conversations they wanna have. So maybe not our last for good, but our last in this series. So go back and listen to the other episodes. If you haven't already with Mike Maznick and Rudy Frazier, this conversation is gonna be focused on decentralization, but we're gonna kind of take a left turn away from social media and towards AI. I think most people me included when we think about AI and especially large language models and our current AI hype cycle, I think about all of the work that's been done to show that concentration of power and consolidation and commercialization and corporate control is essential for these technologies to do what they're doing. So I just wanna caveat that skepticism I have about overall how centralization plays a role in large language models. But on the show today, we have Andrew Trask, who's the CEO of OpenMind and he's been thinking on a pretty long arc of technical development and trying to find ways to disrupt that concentration by building protocols to allow for a more decentralized training of models so that it becomes essentially impossible for these companies to maintain the level of control that they've had over the processes necessary to make models. So we get into how important it is to try and make sense of lots of information at once, which is one of the main challenges of our information environment and it's actually one, if that's an interesting problem to you, I think Eric Salvaggio and I sat down in December, I think I'm talking about the age of noise and Andrew is working on the problem of what do we do in an age of noise? How do we not just broadcast more information into these spaces? How do we build the infrastructure necessary to do broad listening? Two of sides, one, during our childhoods, it turns out Andrew and I grew up within a five-minute drive of each other in Memphis, Tennessee, which is a random piece of information. The second thing is that in our conversations over the last few months, as I've gotten to know more about open minds mission and work, I joined their board very recently so just to flag that disclosure. So with that, let's dig into it with Andrew Trask. (upbeat music) - Hi, my name is Andrew Trask. I'm the executive director of the Open Mind Foundation where a not-for-profit software shop that builds open source software, looking to decentralize power over AI at the recombination of kind of deep-parning cryptography and distributed systems. - So do you wanna talk a little bit about, I mean, we could start with just your background, like how you got into open mind and then I'd love to hear a little bit about the problem you're trying to solve. But like, how did you end up founding a nonprofit that's trying to build technology, which is like the two most difficult things to do? - I wonder. (laughing) - That's a good reminder. I'm a bit of a one-hit wonder. My first job at undergrad, well, I was in undergrad and really wanted to do AI stuff, this is in like 2011. I took an AI course. Got very lucky because everything was still Bayesian then and I was like, "Nural networks seem cool. I should do that." And I was like the one kid in class who was interested in all that and had to do all this weird side stuff to try to do it. This is before GPUs and all this kind of stuff that didn't really work. Got into it, tried to find the one company in Nashville that did AI, found it, and they did a particular brand of AI, which was AI for data that was too sensitive to go to the cloud. Thus my journey on like access to private data was born. Work there for three, four years, went back to grad school to do a PhD. My second stroke of luck, which was basically, I got into language modeling. Nural networks and language modeling started in 2013, 2014. So I published a paper at ICML, which was my third stroke of luck 'cause right after WordDevac there was a period where it was particularly easy to publish. And so I did a very good WordDevac and got in. And yay, a first author paper at a leading conference kind of helps you jump into grad school, which was great. And then my fourth stroke of luck was my PhD advisor led the language modeling team at DeepMind, and they funded my PhD. So a year after that, I started working at DeepMinds on the language team. A little bit before that, I discovered home morphic encryption and differential privacy and some of these algorithms. Definitely a person with a hammer looking for a nail, except for the fact that I spent year or four years doing access to private data stuff already. Started open minds as a non-proper just to basically find people who knew about these different fields 'cause they really didn't talk to each other. They were two different conferences, did different things. This is this kind of stuff. And then for the last eight years, it's been a tech transfer journey from then to now. So basically, a bunch of raw ingredients produced in a raw form with problems statements that are basically pretty ill-framed. Fast forward eight years, I think the problem that we're working on is attribution-based control, which is this idea that someone who has data can know and control whose predictions in the world, they want to make more accurate. So it's actually not about sharing data. It's about knowing and controlling the, someone else's aggregating data and using it to make better predictions about the world of those statistics or AI or whatever. Or just even by hand, like just learning stuff and making decisions based on it. It's about this idea of attribution-based control, meaning, or AEC, people who have data can know and control who they give more power to, more decision-making power to, along a certain dimension. And vice versa, the people who are receiving information can know and control who they're relying upon for each particular insight. That's inherently a decentralizing principle because it's have a direct-to-market principle. It's sort of people with data directly connected to the people who are seeking insights without necessarily intermediaries obfuscating the middle, which is really the main breaker of attribution-based controls. Usually, I'll grab this thing right away in the intermediaries. So yeah, that's sort of the journey in a nutshell. Happened to start working on private data 15 years ago and AI and LMS, because it was a natural language processing company for access on really sensitive data and I basically never left. - I appreciate the note of humility in it, but one hit wonder is like the opposite of what you just described. - Oh, okay. - 'Cause you at every stage basically have found the thing within the thing, like you've had some lucky breaks, but like every time you've iterated, you've been doing something impressive at that stage that has accumulated into something that it makes the next thing possible. So in terms of the problem you're trying to solve, if you want to paint as a picture, within like 10 years, you all have like a confluence of luck where a circumstance is all line. You build technology, you find the right partnerships, the timing is good. What does the world look like if open mind is like wildly successful? - Broad listening gets solved and it becomes possible for you to have an interactive conversation with a million people all at the same time. The most mature version of this tech is communication technology. So it's like a deep learning becomes a communication technology. - Do you want to describe the difference in broadcast and broad listening? 'Cause I feel like we've had conversations about this in the past that I'm like, it's taken me like two days to process it and then I'm like, oh shit, that was actually really, really inspirational. - Wow, well thank you. Yeah, broadcast and technology is, in the history of information technology, it's 250,000 years since we started using words to like now, right? It's important to understand the problem that information technology is trying to solve, which is we were people living in the jungle and one of us with either poison berry and it would make our stomach feel bad. And we needed the ability to tell others, hey, this berry is bad, don't make this decision because it leads to bad things. But the problem is we'd come around the corner, see our local tribes people and have no ability to tell them and so they would live the same patterns that we did largely, right? We didn't have ways of better living that would traverse across time and space. Our kids would live the same life that we did largely. And so when we got language, this was a huge upgrade relative to gestures and tradition that gave us the ability to broadcast and broad listen information, broadcasting, being a half piece of information. I mean, I make copies and I will give it to other people. And broad listening being, I'm gonna take in information from different people around me. I'm gonna synthesize that into a better model of the world and use that to make better decisions. Since that time, quarter of a million years ago, the main change is increasing scale of our ability to do that. So if the first broadcasting or broad listening technology was language, it was at the scale of yelling. You know, it's like, I could broadcast to 100 people if I went to the densest part of the community and just yell at the top of my lungs and we still call that protesting to this day. It's an important thing. And broad listening is still a little bit less scale 'cause you can only really listen to one person at a time, the exception are things like choirs or chanting or whatever. It's like kind of social things that we do to make broad listening more palatable so that you can know that 100 people all agree at the same thing at the same time. So yeah, broadcasting broad listening is like, this journey of increasing scale and the project of broadcasting is almost done. It's almost possible for one person to send a message to every other person on earth for free and instantly. We're like a couple orders of magnitude off from that thing is thing. And when it's done, it'll be actually done, done. There won't be just like infinite rocket shit up of like more broadcasting 'cause it's like broadcasting to who. You know, it's like, what's the point? Like you can send it to more rocks and outer space, I guess. And maybe if we discover aliens out there that that'll be interesting. Broad listening, however, is so far behind. Oh my gosh, it's still so far behind. We still largely listened to like one person at a time. The contrast between broadcasting and broad listening is the source of enormous numbers of problems and it is a huge source of the centralization of power in society because it's this really, really upstream problem that requires us to centralize control of information to get anything done that then makes all the doing really centralized in terms of its power. There's kind of free main technologies that we use to broad listen right now. It's a literary canon, statistics and sampling. So by literary canon, what I mean is 10 people witness something. They eat right articles. One person reads those 10 articles and writes a new article and then someone reads 10 of those articles and writes a new article. And there's like some entrance criteria into some literary tradition that allows you to eventually read a paper that was sourced by 10,000 people but it might take 10 years for a hundred years for it to ultimately come together, right? And that's the cat's meow of broad listening because you can fully synthesize nuance perspectives about the world. And the other two main ways of broad listening are there really crappy versions of that. But the exchange for the crappiness, you get to go a lot faster and a lot higher scale. And so one of them is sampling where instead of summarizing anything, you just grab a sample, grab one example, one interview, one case study, one person and say, this is representative of the group. Doctors say that you should try mylanta because it's gonna make your life better. And you know, if you want to know where advertising in public relations came from, this is guy named Edward Bernays from the 1920s, he realized, oh crap, I can get one person to say something and totally mold public opinion. And like this is where like PR advertising also came from. It's a wonderful documentary called Century of the Self that I highly recommend watching it cost about that. The third and the final one is statistics. Love statistics, statistics is great, but it requires you to throw away all the nuance when you're converting the fuzzy world into specific facts and figures, specific numbers, like you know, say age is just a number and like they're kind of right. If people are actually older or younger, physically or mentally or whatever, but we throw away all that nuance, we create an age and then we can go, what's the average population of Nebraska? We know that. Does that tell us their average maturity level or their average life expectancy or their average, whatever, like not necessarily, it's this myopic view, like through line on life. It's not really a full description of what's going on. And so those three technologies are like powerful and each one of them has been introduced. The world has radically changed. I mean, the introduction of the literary canon was like, you know, it's a long time ago. So we wouldn't, you know, it's the beginning of kind of recorded history. So we don't really have a ton of the before or after because before that we didn't have literary canon. So we had no idea what was going on. But radical, radical, radical change, right? But they're still very flawed. And today there are still kind of free big problems that are preventing broad listening scale. It's the problem of information overload, information privacy and information veracity. And when those three problems are solved, it will become possible for you to listen at scale to the world's public and non-public information with the ability to verify the values of who you're receiving information from, who you're trusting to rely upon for facts about the world. And when that becomes possible, you can listen to the whole world. You could have an interactive conversation with everyone in the state of Nebraska. And by that, I mean, people load their data into devices they control. That creates an interface for automatic communication. - Okay, so I get the capability that you're trying to design for an individual to participate in the system. Do you want to talk a little bit about the technology? Like what would that require? And what are you guys building? - So there's a collision of three blobs of technology or three kind of fuzzy groups? - Love a good blob. - Love a good blob. You know, I wish I could say these were like really crisp just technologies, technology. You can buy what this thing, but like it's tech transfers. So everything's a little fuzzy. One is kind of the deep learning statistics blob. One is the cryptography, privacy and technology's blob. And the other one is distributed systems blob. And so for deep learning, unfortunately, when you actually dig it, it's like eight or nine different technologies that are all colliding as different ingredients. I'm not inventing any of these technologies. The best analogy is the people who aspire to build the Apache web server back in the day were combining a whole bunch of like recently developed individual ingredients that weren't mature enough to cause a general purpose web server to happen before then, but when they could combine all of them into this one general purpose thing, magical things happened. Same thing with the iPhone. Like the iPhone wasn't held back by like any particular genius thinking, oh, the iPhone, everyone knew tablets were coming. It was in cartoons and animations and all sorts of the battery technology and screen technology. I'm like, all these like individual ingredients that got small enough and low power enough, all this kind of stuff to find and make the phone possible. And Apple did a great job of like pushing a few of those way further forward in the synthesis of them together. And so we're in the same type of moment. That's the nature of the challenge of trying to describe what's going on. So what is a partition form of deep learning that doesn't have a name? My personal favorite is deep voting because I think it lends credits to the type of power that we're trying to create, which is this kind of weighted combination of people's models. And so this is already happening around four or five different dimensions because of efficiency reasons. But the sum total is instead of having one blobby model with all your black box weights within one black box file that you run on a big GPU cluster, that you have one deep learning model. It's actually separated into tens or hundreds or even thousands of submodels that are each controlled by different people and likely trained on different subsets of data. So if you've heard of mixtures of experts or rag or model ensembling or model merging or get rebased in or like these different paradigms, they're all doing the same thing. They're all splitting up data into partition sections and combining them at runtime, as opposed to combining everything during training. That's the kind of the first blob of technology is basically this deep learning tech that is partitioned. The second is a blob of cryptography technologies and try to make or maintain encrypted computation work. So you've probably heard of end to end encrypted message transfer stuff like WhatsApp and signal where it's like, I have a message and I can encrypt it and it'll be encrypted until it gets to you. End to end encrypted computation is like that, except you're also controlling some sort of information processing between the sender and receiver of information. There's four different categories of cryptography that need to come together to make this possible. These categories are called input privacy, output privacy, input verification and output verification. We probably don't have time to get into all of them, but if you've heard of things like home warfare encryption or secure on clay as a differential privacy or serialized proof, so that these are all the little sub components that have ingredients that make this possible. This is the part where the holders of specific private keys can control how their information is used after it's technically left to our computer and by use, you mean which encrypted computations you're going to choose to allow. And you have to have a whole little bundle of cryptography technologies to actually do that well. That bundle is sort of still coming together. And the last one is distributed systems. And distributed systems is really about, one is all the stuff you need to have for the data to actually stay federated in its raw form. So if you actually want people to maintain control a bit how the data is used, then actually has to stay on computers they control. You only leverage it in an encrypted form for specific encrypted computations, they're going to produce specific insights, specific context that you want to have happen. And doing distributed systems is just like really hard, it's just like a bunch of really important subproblems to doing that well and doing it in a modular way that doesn't pin you into one particular network effect or one particular opinionated protocol. And then the other big problem that's being solved there is what people call the web of trust. So being able to sort of overlay your actual social graph of like who you trust to send or receive information with in reality. And the grand prize here is basically word of mouth at scale. The ability for you to ask your friends of friends of friends of friends of friends something. Or even not that, that could be an organization who knows a person, who knows an organization as a person, that could be any type of social hop across these different edges. That's where the solution to problems like disinformation and intermediation information is kind of stuff can be found because in theory, you can find many redundant paths back to the same event or the same source. So you could talk to 1,000 people who all use the same product or you could talk to 100 eyewitnesses of the same event or whatever, via your different networks. And if you have the ability to do that, it significantly reduces the chances that there's a coordinated disinformation program that can be successful because they would have to get so many people on board to convince, to have the same corroborated story, right? That it just becomes logistically very difficult people to kind of the see view. As opposed to when all of your information funnels through two or three bottlenecks, it becomes a lot easier for people to the see view, especially if those two or three bottlenecks have the same information it sent us. Anyway, those are three big kind of blobs of technology. One is partition deployment technologies and to end encrypted computation technologies and distributed system technologies. And that's roughly deep learning, slash statistics, cryptography, and then network infrastructure. - It's so cool. It kind of reminds me of, do you know entangled particles in physics? - Oh, I know about them from like a popular mechanics level. I'm not like a specialist. - Me too, me too, me too. - Yeah, yeah, yeah. - But like the way you get the, I think it was Chinese scientists that took one, so basically two particles get entangled. And if you change the state of one, it changes the state of the other and distance matter. Actually, these like Chinese scientists flew one and love the two entangled particles into space and then change the state of the one on Earth and it changed the state of the one. - That's a while. - I know. - That stuff boggles my mind. - It makes you wonder what the world is. I don't know, but it's so interesting that once a piece of data leaves your control or your immediate control, that you could continue to sort of change the state. - Yeah, that's a great analogy. That is the dream, right? And even when it's been combined with many other particles, right, in this case, three data gets combined with many other data points into a model or whatever that you still have your degree of agency or like the degree to which your contribution contributes to the whole, I think this is very much to the whole. - Yeah, super cool. So what's the, 'cause I feel like you also use the phrase network effect. But like in our first conversation, I remember like my big takeaway was you thinking about the history of decentralized technology and public protocols for the internet. And like where we are now in terms of the network effects that are created and which entities are in control of the protocol layers of this or our challenge. Do you wanna talk a little bit about maybe what you've learned from the last 20 years of concentration of power in terms of like protocol control and like what you see maybe in the next like 10 years is like the fights that are gonna happen around the protocol for these kinds of things? - I think the right backdrop for that conversation is to acknowledge that everything that I just described is going to happen regardless of whether I'm here or not. Like these technologies are colliding because there are very big important problems that they solve, thousands of people have been working on them for decades and they're reaching a point of maturity where something exciting is about to happen. And this is the story of technology in general, but it's a wonderful documentary from the BBC from like the late 90s, late 90s called Connections. The documents basically the whole history of technology from the beginning of the birth of civilization to now that basically makes this one point which is nothing can be invented before or after it was because it was waiting for all the ingredients and the political moment to make it happen. The other thing if you look at history of technology that happens over and over and over again is that right when a new information technology is birthed whoever the most powerful people are of the day tend to have an outsized amount of control over it for some non-trivial period of time. You could say literacy in the dark ages or you could say you know the main thing of computer and IBM or you could see the radio, television, whatever. That's like almost an absolute truth because it takes a bunch of R&D and power and whatever until like build and implement new technology. It's the library, the Library of Alexandria, this is this concept. The one big exception was the internet and this is like wacky and wild 'cause it shouldn't have happened that way. In a sense that like ARP and Net was prototype on top of Bell Telephones existing monopoly which was like so strong and so stringent that they were literally broken up a few years later as a result of that monopoly. Yet because the US government through a bunch of money combining a bunch of ingredients before they were kind of naturally going to be combined they got ahead of commercial industry and traded public goods early. Normally this is ARC where like it starts off is like a little experiment the thing. It becomes centralized for a while then it becomes very centralized for a while then people start like this is a terrible thing. It should be decentralized and then it like arcs down and becomes more federated and then it becomes like a public good. Like literacy is like pretty much fully there although it's never there as much as people seem to think it is even just like knowing the language of our most important texts. You know used to be bladdened and all of a sudden it kind of sounded like culturally controlling things. You know knowledge of computers, stuff like radio and broadcasting is decentralized with the advent of social media and all this kind of stuff. The question is kind of like almost like from a COVID analogy perspective like how did the internet cut the peak, right? How did they skip all the way to the end? The big thing was they got a network effect crazy early with free open source tech. I mean just like so so so early. This network effect was so powerful even though it was kind of kicked off in large part by funding from the US government in 1991 and then too when the US government tried to issue you know another official standard for the internet called the OBS it's reported that like basically the world kind of ignored them because everyone had ordered deployed some version of the protocol and they weren't all gonna reinstall all at the same time. I mean think about like TCP/IP having 32-bit IP addresses like talk to Vince Sir if he'll tell you that was the prototype that got away. It was not meant to have 32-bit IPs and even though it benefits every single person in the world to upgrade a 64-bit IPs it is still taken three to four decades to like do it. The lesson here is that basically every time there's a new set of technologies there's an arc of centralization and an arc of decentralization. The main task is not to be the lone genius that like invents a bunch of new technologies or whatever it's to cut the peak like how can you skip to the end and have safe but federated tech as early as possible and that's primarily a function of who builds it first and the people who build it first are usually the ones who are just crazy wealthy and powerful at the time. Every once in a while and by every once in a while I mean like one time we did better. - Oh, which is helpful, yeah? - Yeah, yeah. The odds are very much stacked against us except for the fact that it happened recently and so it's still in people's, you know, it recent every this kind of stuff. So I don't want to like trivialize all this is just like a fundraising problem but it definitely starts there. Ultimately the magic happened when World War II ended and the US government said, that seemed to go well, why don't that go well? Superior science and technology. Let's always do that and then they started ARPA and just through crazy amounts of money during the Cold War. And my dad was in the military. He told me in all the military simulations like the people who invested in R&D and technology were always the ones who won in the long run. He just had these like systemic advantages. And so, you know, for me the thing I look out now is like, are we going to do that? Are we going to have enough of a cohesive vision to build the ingredients that haven't happened yet as opposed to throwing money kind of chasing LLMs or something like that to get ahead of it far enough where you can actually get these sticky network effects before the next obvious thing will happen, which is basically a startup that scales these protocols out of Silicon Valley. In the grand scheme of things, there's this old phrase, I think it's like a military phrase, which is like for novices focused on strategy and experts focused on logistics. And what they were talking about was like, in war, the people who win are the ones that keep their troops fed and watered better than the other guy. And strategy is like about what you don't do. Like, oh, we're going to go left and instead of going right and all this kind of stuff. And logistics is like how much resources you're pulling in. And so I really do believe that this is the stuff that's actually going to control how AI gets used in the long run 'cause it's control over the supply chain. If whoever builds the protocol that accesses the world's data and compute and talent and networks it together, that protocol is going to have a massively outsized impact on like who gets power, who doesn't, whether it's good or whether it's bad, all this kind of stuff. And the particular algorithm you happen to be using, whether it's a transformer or LSDM or whatever, is going to kind of matter, but it's going to matter only in so much as it changes who owns the supply chain. And I'm deeply more passionate about that problem than the problem of just like numbers go up and they make the model smart. It gets kind of boring staring at the line after a while. People think about chatbots as like talking to the chatbot. And I think that's something I really buy right up until I read this paper that was kind of the founding paper of the internet that gave a richer description of what communication actually is, it just changed my life. I would say the most important paper I've ever read in my entire life is called The Computer as a Communication Device. And it was paper written in 1958, I think. There was right before a JCR liquid where it went to RFID, great RFID. The intro to the paper basically says, you think you know what communication is, but you don't. Communication is not descending and receiving of bits, which is what everyone would say communication is. Like, you know, dialing someone on the telephone or you know, sending them a text message to whatever that's communication. He said, no, no, no, no, no, no, no, no. Communication, and he was a psychologist also, the person who like funded and coined the idea of the internet, psychologist, not technologist and just interesting. Communication is the alignment of mental models between two people. You have a mental model of the world. I have mental model of the world and we throw bits at each other until our mental model is aligned. And that's what communication actually is. And as soon as you make that connection, all of a sudden, deep learning tech is an amazing communication technology. Federated learning is an amazing communication technology. Chatfots are an amazing communication technology because it's actually about you being able to put your mental models into a software program that then someone else can download those mental models when you aren't around. - This is also reminding me of when I was a kid, I used to think all the time, if weather, the blue I see is the blue you see, or if you need French color and call the same thing, because we-- - Totally, I wanted that too. Yeah, yeah, yeah. (laughing) - You think about what's happening with Groc, for example, like this idea of language models and chatbots. The purpose isn't, I put my mental model, or maybe for Musk it is, but like there is something happening that is I want my mental model to be the hegemonic mental model of the world. Ah, yes. - Which is a very different to hear a proposition. - This is also an extremely common pattern in the invention of the information technologies. The Encyclopedia invented in the 1700s in the Encyclopedia's movement in France, and it was to create the official dogma of like what the world is in one book. That goes along with another pattern that I saw in the information of the information technologies. Like every time there's a new synthesis technology, every time there's a new way to bring information to pick together, some genius has the bright idea. Like let's put all human information, all human knowledge into one version of these things. Like the first synthesis technology, I would argue is the library, because it allowed you to put lots of ideas in geographically the same location, and you could go in and synthesize. And of course, Alexander the Great is like, let's put all human knowledge into one building, right? One doing to one library. And then like when someone invented the book, the Encyclopedia's were like, let's put all human knowledge into one book, then the computer scientist of like, you know, the Google Books project, like let's put all human knowledge into one computer, all this kind of stuff. We're still living in the, in the era of that. Everyone has that bright idea and thinks that they're geniuses, and then ultimately they realize that's a bad idea, but you get to have an amazing instrument to propaganda for a little while. Even like even literacy was that way for a while. Oh, literacy. Oh, well, only the elites should have, you know, the one centralized all, you know, whatever in one place. And I think it is potentially dangerous, but this is what trimming the peak is all about, right? So if you're just trying to like, cut off the peak of centralization, it is kind of the inevitable ebb and flow that is kind of the path-dependent thing, which is not that we're going to be centralized forever, but there's going to be a period that really is crappy, right? And we have, we have the chance to kind of skip that and allow our children to benefit from it instead of our grandchildren, instead of an addition to our grandchildren. I feel like broad-listening as a concept. Maybe it doesn't miss this, but I think about a lot about feminist epistemology and the idea that there is no distilled truth, everything is a combination of multiple perspectives and experiences and sensessionality. And then like the pursuit of a single point that we can all agree on is kind of missing the point of society. (laughing) - Yes. - So I wonder a little bit like when you say one of the pillars, I think you said was veracity. Like how do you think about preference and pluralism and complexity of social experiences within that? - So I'm definitely a pluralist. There is also definitely an interesting open question of what happens when we get to the end of information technology of like what that does to that. Because since the beginning of time, no one has had the overview. No one has had the ability to take in all the information in the world and like make a perfectly well-informed decision. HG Wells had a series of essays called The World Brain. He had this great hope that if information technology reached its climax and everyone could see the same information, they would come with the same decisions it would have world peace. So it's not a new idea. I guess all I'm trying to say is like in my opinion, and this is where that kind of web of trust, stuff that I mentioned earlier ago kicks in, is when information technology is done and kind of complete. There's still an ongoing problem that people need to work on, which is what sources of information do you trust or not trust? What is your personal weighting of information you're willing to take in and rely on for decision-making power? And there's this really tricky problem that we all die. And as a result of that. - And not where I thought you were going to happen, yeah. - No, no. Well, we all live short lives. None of us live long enough to actually learn fully what sources of information are true or not true. - That's why academia is so cool to me because the idea that you make your contribution for 30 or 40 years and then you die. - Oh, literally Ken. It's amazing, right? That's fantastic. But I think there's a, even really dumb bars number, you can only remember like 150 brands or like, 500 faces or something like this, right? This kind of thing. And I think this gets back at the same core idea, which is that like at some point, all the world's trust systems are grounded in one person, just taking a risk and believing another person and then learning after the fact whether that person was actually a reliable source of information. There is no way to predict it in the future unless you have that ground source data to extrapolate from. - And so should trust as distinct from being correct? - Well, and even this question is like, what is correct? And I would say, what is truth? Let's just go all the way there. What is truth? And there's like two competing definitions in my opinion. And what are the leading ones? And one of them is truth is repeatable verifiable information. I can put in the same information, I get out the same information. This is like the rounding of like the scientific method and all this kind of stuff. And the other one is truth is whatever information that I take in that allows me to like flourish in life by my own definition. I take in this information, I use it to make decisions so the decisions work out. And so that's like a totally, it's like it's an end to end kind of truth. And like that's the difference between like, secratic thought and like traditional thought. And it's a difference between science and religion and politics. These two definitions of truth are in like a 2000 year struggle to figure it out, they're not going to go away. But we are more trapped in the latter than we are in the former. And since up like, how do you know that one eye is deceiving you, right? If your brain is just a vat with sensory information coming in, the way that you know is, it gives you inconsistent information with what your other senses are telling you. If you only have one sense, if you only had one eye and no touch or taste or whatever, you have no ability to filter out that conviction. You just have to trust your eye because that's the one you got. This is like, you can extrapolate this all the way up to like, how do I know to trust that journalist? Well, I don't trust that journalist because they're attached to this institution, which is attached to this idea of this institution, which I've trusted in the past. It like led to good things for me. Or I trust them because people who I trust, trust them. These are like the two big things. It's like, your whole world is constructed based on like 50 to 150 relationships that you have personally tested yourself. And then you extrapolate that to all the relationships that they trust and the source of information that they pull in. And that's like a, uh, Iki thing to like grapple with. But it also means that cancel culture is broad listening. Yeah. Any type of information that sits in the scale is broad listening. It can absolutely have challenging themes. I think there's like the whisper network thing of like, someone shitty things to different people. And then you frankly know about that person. And maybe there was no formal process whereby that person was not accountable. But I know that like that person is someone that I will avoid. Because that's so interesting. Yeah. So it's a product reviews are broad listening, voting is broad listening, statistics is broad listening. Any any type of like many people say something. And I'm waiting it via the relationships being filtered through and ultimately synthesizing it so that I can make better decisions. All of that is broad listening. The big, big, big problem is that it happens at crazy low scale. And so because it's so low scale, we delegate the free big problems to centralized institutions instead of doing it ourselves. And the free big problems are information overload privacy and baracity information overload is there's too much information for me to read. What should I look at news feed? What should I look at? New station, broadcaster, whatever, researcher. The second problem is privacy, which is most of the world's information that could be useful to me to make a better decision. No one's going to send to me, right? Because it's like in the personal lives of people's personal affairs. And I don't know where have trust with those people. So think like every decision you make in your life, there's probably a billion people in the world who have faced a very similar thing within the last 10 years. But you can't go and talk to them about that because it's like there's too much information and they don't want to just send it to you in the raw because it's it's it's it's it's scary and could cause all sorts of problems. It could give you power that they don't want to give you power. The final one is the inverse of that, which is you receive information from people but you don't know which pieces of information are true. You have two tools with which to do that. One is test some information channels yourself manually, which is painful and scary. And like that is the trustability exercise. And the second is redundant pass to sources of information that you can't see. So that if you get 10 eyewitnesses to the same thing, 10 people all try the same product or a hundred or friends or friends of friends, like larger lists, then it become more redundant. So what word of mouth continues to be so popular is by the fact that it's so low scale. It's harder to coalesce these types of centralization issues. And so yeah, broad listening is like each of those three things we delegate to a centralized authority to like process more information that we could or if it's private information, we have the CIA or healthcare institutions or banks or whatever the whole kind of the world's private data and can centralize it and give us insights about it. And the third one being veracity, we'll hire people to go out and see what the stuff is really true. And we have journalistic institutions and research institutions and all this kind of stuff. And you have to ask yourself, do I trust that a research institutional like that stuff? But unequivocally, we have to solve these collective problems. We delegate it to centralized institutions. It's hard to think about the world in a world where like broad listening technology is fully accessible and you no longer need to centralize that power. It makes the whole political economic constraints radically different. Why are markets inefficient? You can't talk to every other potential buyer and seller. The people who are building products can't talk to every prospective customer about what they actually wanted the world. There's huge, huge information bottlenecks that make it difficult. Why is the political processor jam? Well, when you go to vote in this multiple choice question that you get to send to the government, are you actually fully expressing everything that you need from your government and their people? Like, no, you're doing statistics. It's an information bottleneck. It's that statistics can synthesize information once every four years in the 1700s. But like, if you just wrote an essay with like, here's what I need from my government. The government would get, in the case of America, 360 million essays. And like, it's a totally, it's a cumbersome system, right? And that support conversation happens anyways, the big, whatever people call the political dialogue or whatever. It's pretty decoupled from systems of power. It's not the way in which they are appointed or not appointed. They're appointed or not appointed by votes. And what's that old phrase you show with the incentives and I'll show you the results place quite true there. So anyway, the whole world changes if we get broadlisting tech. And so to get back to your original question of like, what does it look like if I open my as rapidly successful? We get to the end of information technology early. And we get there based on public goods before they were able to be co-opted for, you know, a few decades or a couple of centuries. The experience of it feels a little more like synthetic media because that's the most accessible form. This won't be the exclusive way, but my guess is that it looks and feels like a signal app, but where you get to choose massive groups of people filtered by your own friends and friends of friends of friends to have a conversation with. And it might be audio, visual, text, all this kind of stuff. And it's not you talking to the AI. The AI is just an interface to all the concepts that they've written down. And at the end of the day, it just becomes communication at scale. When you want to go solve a problem in your life, you get to have a conversation immediately with everyone who had that problem very recently, even if it's a problem in your personal life because you're not actually learning personal data about any particular person. You're having a conversation about your own life. And they're just informing it, basically the latent patterns that they'd experienced. Who, when hearing this conversation, are you trying to attract into a conversation? Just so people know, I should reach out to Andrew. If I am a health institution trying to figure out bloody blah or I'm a politician, who are you looking to talk to about this? There are not that many programs in the world that can trigger this network effect fast enough and early enough. The main bottlenecks are the small teams of people who are specialized in one of these core ingredients that needs to come together and be combined with the other ingredients. So the 150 leading experts in differential privacy, the 100 leading experts in home orphan encryption, the like, these like little tight communities in these out of the origin window tech stuff, they need to like realize that they're a part of a much bigger picture and start building that way and researching that way as opposed to continuing to work on the pure version of their own thing in isolation. The next community are people who can fund public infrastructure at the scale that is necessary for this stuff to be finished and adopted. In my opinion, there's probably only like three to five programs that I know of that are plausibly able to do that. So this is like stuff in and around the National AI Research Resource in the US. So this like kind of public AI pool of funding and program being run by the NSF that like has the potential to set aside a few billion dollars to drive public infrastructure around, you know, federated control over AI or decentralizing AI. And it's the democratizing idea. It's the phrase that use anything around that. DARPA NSF and IH, these places they can write in theory can, I know the procurement of climate is challenging, but in theory, you can write big enough checks and have in the past to cause this to happen. In the UK, there's a thing called the National Data Library Project that I think is probably the leading contender in the UK to manifest this type of vision. Also, a lot of people outside of the UK maybe don't know that ARPA Net had a sister project in the UK. It was incredibly important called NPL Net. It was basically the same type of packet switching prototype. At packet switching actually was developed in the UK. And I would say the global network effect got solidified and the public good got solidified. Not when the US built ARPA Net, but when the US linked internationally to other nations that had similar analogous protocols and the whole word internet is actually from a paper called the interconnected network of networks. And that was when all these little test prototypes linked together and that's when the world became open and the trajectory of the 20th century really changed. So I would say, yeah, there's a kind of a complicated set of dental funders in Europe that could build public infrastructure. You probably know more about them than I do, Alex. And then I have a vested interest in also slowing down this information transferring to those who would be competitive to it. So I don't want to see any VCs standing up ever-odd listening program anytime soon. That would be disastrous, I would say. Now, once the highways are laid down, go nuts. Go crazy, right? This is great, but the software playbook is so centered around owning marketplaces and owning the comments, this kind of stuff. And it just, the whole ball game is about can public funding front-run VC funding with enough time that the highways can get laid down so that commerce can happen on the highways and not in the highways, I guess. Policy makers, unfortunately, there's not a lot from regulatory perspective that I think can really change things. People will talk about standards bodies, but in the history of the internet standards bodies, they were not relevant in the beginning, they were relevant in the middle to late and in perpetuity after that. But we're using TCP because they were the ones who built it first, not because the ones who sat around and had a conversation. And I love standards bodies, they're really important. Everything that is being constructed right out needs to find its way to being stewarded by a standards body, absolutely needs to happen. But we should not get our cart in a course backwards. The power is in the hands of the builders and the funders. And it's actually mostly the funders, so I would say, that's where the bottleneck is. - I feel like I always take away a lot from our conversation, so thank you. This is a type of decentralization that I don't think that many people know is possible or think about in this way, because I feel like most decentralization conversations happen at the level of broadcast and centralization, not at the level of meaning making and hybrid stores of data. So I feel like it's just a really helpful, different kind of approach to a problem that other people I think think a lot about, but in very different contexts. So thank you, this was lovely. I will be talking to you soon. (upbeat music) - Thanks to Andrew for wrapping us up in this three-part series on Node Star. Next, we are actually doing another series, which thanks to Georgia, Yakovot and Sarah Miles, as ever for producing these individual episodes, but also for architecting these bigger series where we can go deeper on particular topics with lots of different conversations with different folks. So our next many series is on scams. The first in that series is gonna be an interview with Mark Hayes, who is an anti-crypto lobbyist in DC. And I'm hoping that a conversation on crypto is kind of situated really nicely in between decentralization and scams. So we'll get into that next week. And if you happen to be in New York for climate week, reach out, we're gonna be around. And if not, we're gonna have a live show on data centers. Obviously, the primary environmental conversation we think we should be having during climate week about AI. Even if you're not there in person, we will have a live stream, which we'll link to in the show notes and we will be streaming it on YouTube. We have a new is YouTube channel. It's always helpful to subscribe and also share this episode and all the episodes that you love that we do, just to reach the number of people that engage with these topics and the people that we platform. So do subscribe to our YouTube channel and also sign up for the live stream and also let us know if you're gonna be in New York for climate week. And we will hopefully see you there. [BLANK_AUDIO]