Decentralized Intelligence - The Impact of User-Owned AI - Bangkok 2024

[Music] wonderful we got so excited we turned the mic off uh well hey folks I hate to be that guy standing between you and a break but I hope this can be worth your while the next 20 minutes or so uh my name is Jared Barnes I head up ecosystem efforts at the near foundation and have the privilege and opportunity to work with so many Founders that are building at the intersection of AI and web 3 really this talk is actually going to be maybe a little bit different flavor than some of the other talks today I actually want to take a few minutes and unpack like what is this intersection between Ai and web 3 why is it impactful and then really put it in context to our conversation and discussion today around public goods around impact and really what this means for the future of our world and so maybe just to help frame where we're at and why this matters um for those of you who do not know uh Dario emodi he is the co-founder and CEO of anthropic um he came out with an incredibly powerful article called Machines of Loving Grace highly recommend you check it out um his whole premise that he shared is that we're living in a compressed 21st century meaning that the pace of innovation in what typically takes 50 to 100 years with AI is happening in 5 to 10 and while that is absolutely incredible and exciting there's also some real challenges and questions around that right if we're living in a A Century of uh progress that is compressed into a decade who controls that future right um and what does it look like for actually all of us to control that future that's really what we want to impact our vision at near is uh to build that future where we all have the opportunity control to control that and we call that the user owned internet in other words a more secure safe and economically free world where everyone has access to their own data and how that data is used in AI model training and so to make that actually make sense and put some words to definitions right the world of Open Source and particularly in the context of AI incredibly powerful yet at the same time it's also hard to monetize if you're building something in open source and so the premise and focus that we have in really um the uh mechanics behind user owned AI or the user own internet is the incentive mechanism that would actually drive this distributed system right in other words uh we are Distributing the core functions of AI systems through decentralized AI we want to build the incentive mechanism around that right where it is both monetizable and controlled by many and so to make that make a bit more sense and kind of explain it to me like I'm five language if you do not have a PhD in machine learning um data is the most important thing in the context of AI the quality of the data set has a direct correlation to the quality of the model why is claw 3.5 Sonet and GPT 40 and 01 previews so good it's because it's trained on all the data that is accessible on on the internet right um and so this is the actual process of end to end to you actually interfacing and uh engaging with chat GPT this is everything that goes on behind the scenes and in order to uh live in a world that is user owned you have to distribute those systems and drive an incentive Network around it um so I'm not going to spend too much time here but why this is important and really where this intersection of AI and web 3 comes together is we're taking two very fundamentally different Technologies and we're putting them together right why Ai and particularly large language models are very different is because they're the first really non-deterministic software meaning uh whatever input you put in you're getting a variable output right you don't always know what the model is going to say that's very different than crypto very different than smart contracts that essentially are a deterministic if then statement and so we're bringing both of them together our Approach at near has been very pragmatic and Progressive decentralization not just decentralization for decentralization's sake but really focused on the use case and value that can be created um and so why does this matter is that we really can only understand models and how they generate content if we have a record of what data is actually underpinning that model right um so these are some of the core core uh challenges and really what is happening behind the scenes in this world of decentralized AI and we're at kind of this intersection of there are some really interesting interesting things happening right now in the space whether you've been following AI agents in that specific boom um but we're still early right and decentralized AI as a market is still very much at Day Zero so I think just saying that as kind of a primer and backdrop to what's happening that said I also think in the context of AI we're in 1995 but with a lot more upside and so these are some of the things we're really really excited about um and just to take two steps back and actually look outside of the crypto bubble and more broadly AI if you understand kind of What's called the scurve of innovation in other words you have to kind of start with like core infrastructure some foundational things that need to be true in order for in this sense AI to work to get to oh whoops getting to uh I think we lost our slides but we're good we're rolling oh we're back um getting to like where essentially web 2 AI is in some really really um interesting things we're still very much at phase one in this context of uh decentralized AI which is totally okay okay where we're at but I think it's important to put it into into context um and so why I bring that up is um Moors law if you're not familiar with Mo's law essentially it is um uh saying that every two years uh the compute will continue to double and compound and get better and better and better over time I think one of the things that I personally get really excited about in AI is everything that we're seeing now is still gen one like this is really the worst the AI will ever be which is kind of crazy to think about but if you put it in context it really makes you think and so I bring that up because there's some I think really powerful use cases that are happening um whether you've been following Runway figure or Hume that are kind of tackling very very different um uh ways and how we interact with a interact with software and interact with computers that I think are really really exciting really really powerful um as well so I say that as like backdrop but also what are some of the problems and what are some of the things that are stopping advancement um I'm going to skip this and actually oops go to uh this slide here there's kind of four fundamental problems that are stopping that is stopping AI advancement and really where I think there's opportunity to focus the intersection of AI and web 3 the first is power constraints this is a pretty complex graph but to make it make sense this bottom line here is kind of benchmarking what was used to make GPT 4 basically this graph is saying we need about 10,000 times greater power um to get to Frontier models uh over the next 5 to seven years right where're uh we're kind of trending towards in 2030 right assuming that we continue to grow at the rate that we have right to train llama 3 uh itself the model was released by meta it took over 25,000 Nvidia h100s that's like enough energy to power all of Bangkok for a couple days right so you need a lot of power and a lot of compute and energy you also need a lot of uh obviously Hardware but also a lot of data um and that is another huge huge gap in data scarcity um that again really addressing through decentralized Ai and then latency in other words how fast you're able to run computation over that data set right as well so these are really big audacious problems that I think the world is kind of grappling with and hyperscalers in other words like the big tech companies in the world are really doubling down to solve this if you aren't familiar there's about seven companies kind of called the mag 7 that drive majority of the value in the NASDAQ or the American Stock Market between Nvidia meta Tesla Amazon Google Microsoft and Apple that is kind of problematic now as we're actually thinking about the implications here where um if you ever played the game Monopoly and you got the railroad in the Monopoly board that's exactly what's happening right now with comput um and particularly um uh uh where all of the Power and essentially foundational models are are coming from right as well and so that matters because these companies are sparing no expense right this is actually taken from the Llama 3 paper and if you put it in context of what it took to create this model I said 24,000 h100s trained for 54 days 24 hours a day that's enough to power the entire C city of San Francisco for four days right just to get llama 3 15 trillion tokens that's 20x the size of the Library of Congress and then uh 240 p of storage that's about 15% of the active storage of file C just with the data set to train L 3 right and that's one of many models and so when you put it into context and actually understand here it's like wow we really got to help um frame this and understand what that means for us so for sake of time I'm actually going to skip ahead a little bit here um and get to one of some of the core things of why that matters for us I like slides and a lot of slides but I'm going jump to like what this what this means for us I'm an educator by trade and hard so I'm like big on backdrop and context so the audience understands but what does this mean for us we got to understand what our Tech is good for fundamentally three things that uh blockchains are great at privacy performance and verifiability how does that actually show up in the context of AI and particularly decentralized AI well you have these three kind of core triangles here and key use cases that fundamentally blockchains are suited for that Web Two uh alternatives are not suited for you think about performance and the ability to distribute both data labeling and data curation right like we said data is actually the most important thing to get Frontier models and to get uh the highest quality model you need the highest quality data set right you're able to distribute that um in the context of decentralized AI in addition to compute networks privacy which is something I'll touch a little bit further on is also a huge consideration right in how much data are you actually sharing with proprietary um uh language model providers and then verifiability how can you trust the output of a model right as well so these are some of the core things that are really important um and why this really matters is if you think about uh where the world has gone in terms of uh finance and I don't know if any here anyone here uses Apple pay um or using like an NFC chip to touch typically when you pay with Apple pay you don't think about like what's happening behind the scenes like oh like is this NFC chip is registering with my uh bank account and they're taking a transaction fee like you don't really think about that you just tap to pay I think in the next 18 to 24 months that is what AI is going to become it's going to become invisible it's going to be always on or persistent we're not going to be thinking about it and so why these core components of decentralized AI matter is because in a world where AI is invisible these are some of the things that we can really move the needle on and make a difference on as well um again going to skip this a bit just about some of the things happening in the near ecosystem um and some projects tackling some of those problems but for sake of time I really want to end with this and what does this matter actually for public goods right everything I said was a backdrop just to inform a little bit of context now it's like why does this matter well first just to again recap we do have a serious AI Monopoly challenge right with essentially 75% of the market cap concentrated in seven companies um how do we think about offsetting that um and how do we think about creating more accessible and Equitable access the second are there's some very serious research implications particularly um in the context of climate data particularly in the context of uh regenerative agriculture and particularly in the context of dii um and I also think now is an amazing time to start tackling some of these problems why is that well the cost of compute um is trending towards zero it's actually already uh very like very pretty much like Pennies on the dollar of what it was even 12 months ago um and the infrastructure around privacy is becoming more and more mature right as well and so let's make that a bit more real I'm going to again jump past this and actually talk about the use cases some of the things that again uh we have been focused on at the intersection of uh real world impact and decentralized AI um we actually this morning in our uh Workshop built a uh food security network in other words a privacy preserved Network that enables food banks um to uh share their data without um compromising those who uh would come in and actually use that uh for predictive Insight or Predictive Analytics um as well you think about Healthcare and collaboration there's an amazing uh group called crunch um they are solely focused on uh kind of Frontier medical research and curating a powerful community that would be able to create a general World model that's open source and particularly eliminating some of the barriers with climate action and breaking down some of those data silos that exist right and so there's a lot of layers I won't get into the tech but fundamentally what this means is we can leverage this architecture to fuel what the future of public goods could look like right in a really really powerful way um so I'll end with this this is actually an architecture uh piece that we put together for the hackathon that's been going on over the past month or so um I mentioned in our Workshop earlier today we actually built a MVP of this beta so for those of you who are there thank you um but what this is is an architecture of again a a uh uh open- Source uh privacy preserving Network specifically to help eliminate um food scarcity by enabling food banks uh to upload uh their data um in a privacy preserving way and get insights around who they serve right as well and so if you think about that in that use case is possible today what might we do in a world where the pace of innovation is happening so quickly and we're living in this compressed 21st century really truly believe it is on us maybe it's on us in this room or really on us that are focused um on these intersections of opportunity because centralized AI I think is also moving quickly as well so this really is a race and it does take a collective but I also think that's our greatest opportunity in that we can be a collective right that is not necessarily centralized or working just for one organization and really pull together the knowledge Capital um and the power of a collective to drive use cases just like this so if anyone is interested uh a in this architecture codebase open source happy to share it happy to chat more I know we're like pressing up against time here um but these are some of the things we really think about in terms of real world impact largely because most times you've been uh has been focused on speculation this is something not speculative that we're really really excited about um so thank you so much for your time and attention hopefully I didn't keep you from uh your snacks or coffee or anything um but uh looking forward to chatting with anyone as well so thank you [Applause] [Music]