Why Does AI Need Crypto? | Ben Fielding - Gensyn

hey hello um going to kick things off um just a very quick note for everyone on this session uh this session was originally supposed to be a fireside chat uh but as you can see I'm here alone so I'm not doing a fireside chat with anybody else um I figured given the short notice probably uh best to try and involve you find folks and make this a Fireside chart between me and you um so I will I will give a bit of an intro into my thoughts on why AI needs crypto I'll give you a bit of an intro to myself and kind of why I care about this and uh I'll probably do that for five 10 minutes and then hopefully hand it out to you guys and I'm happy to answer any questions about AI crypto decentralization why AI needs those things um and also what we're doing at Jensen um and obviously I'll explain a little bit more about that so um yeah I'll kick things off I'll give you an intro to me and then uh and why I think this is an important problem and then hopefully hand it to you guys if you do have questions please I think come down to the front and they'll hand you a microphone or just sort of put your hand up and wave and I think someone will bring you a microphone and then uh we can get you involved we can make this a kind of interactive sort of fir side chat ideally um so to kick things off uh I'm Ben I'm the co-founder of Jensen um my background originally is in the machine learning space so Ai and machine learning not the crypto space uh I'm a bit of a recovering academic so um I started out doing a PhD in machine learning uh back in 2015 when deep learning was exploding for uh for machine learning in general and for computer vision um I think as we all know now deep learning and deep neural networks are kind of very exciting most people will have interacted with them most people will have used chat GPT and image generation models and things like that but back in those days um it was kind of within the academic Community um and one of the the biggest things I realized whil doing my PhD was the compute crisis that we were starting to be in then and we're definitely seeing now um my research focused on on evolving and growing the structure of deep neural networks which is an incredibly computationally intensive task um essentially you're training and retraining candidate neural networks over and over and over again I was doing this as a PhD student with four gpus under my desk in a workstation I'd built myself and Google was starting to do it with thousands of gpus in a data center that they would run for weeks at a time um and what I realized uh doing that was that I was being significantly held back by reasons that to me weren't necessarily valid they didn't need to be the case um the main things holding me back were access to Capital and access to resources and assets rather than knowledge or ability uh and I realized if that was the case for me it was the case for a lot of other people in the world and it was generally holding back the machine learning space and human progress um so Jensen now as a company solves that problem uh we are a decentralized machine learning compute protocol so we connect up all of the machine learning capable compute Hardware in the world and provide that as a resource to anyone training machine learning models when I say training I mean essentially distilling value and data into model parameters or weights um and the current situation the current way that you do that is to rent time on devices owned by a very small number of people so AWS Google Cloud people like that they buy up all of these chips and then they rent them out to people for exorbitant prices Jensen is creating a new infrastructure that sits in between the device owners so the supply side and the machine learning Engineers the demand side and the idea is to be a protocol almost in the web one sense it's a communication protocol that allows those two parties to interact and it eliminates all of the centralized bottlenecks and the ways that the companies right now manage to maintain profit um and this kind of leads us on to uh why I guess the the topic of this which is why does AI need crypto when I say crypto I mean crypto decentralization web 3 kind of whatever we want to call it blockchain or the space that we all exist in and I think there's there's two main reasons in my mind the first reason is kind of backwards looking it's looking at the market for machine learning and why people have uh a monopoly why they can continue to create profit over the machine learning space we're generally as a society okay with that if that company is bringing something new and novel to the market what we don't like is rent seeking and what we see in the machine learning space is lots of rent seeking on top of core resources uh compute is one of the most vital core resources in our minds and a way to eliminate that rent seeking behavior is to make a much clearer connection between the owners of the compute and the people who want to use the compute and stop the cloud Giants from being able to essentially sit on top of that compute and make money on it without really doing any additional work because that's not competition working in the in the kind of sense it should so that's the kind of backward looking view it it also includes the the kind of chip shortages that we hear a lot about there's uh we're not making enough chips it's very hard for us to make those chips Nvidia have a very strong hold on the AI Market because they're able to do that um the supply chain of chips themselves is just very difficult um so if you were to just graph the demand of machine learning Hardware against the ability for Humanity to create more machine learning Hardware those things are diverging very rapidly two ways you can address that one is make more chips the other is make the usage of the current chips more efficient and that's what Jensen focuses on so that's the backward looking view that's looking at the market right now and saying hey with decentralization can we change that and and we can the forward-looking view is how is machine learning space going to change what's going to be different in 5 10 plus years one of the things that we think is very important is that we need to make it so that models will be able to interact with the core resources themselves right now everyone training a machine learning model is a person and they're perfectly happy to go through various kind of like human layer aspects in order to do that so they'll go through people they'll use legal contracts to secure trust uh they'll interact with businesses they'll use infrastructure built by people they'll interact with customer service agents Etc in the future we think models will be the ones building other models for that to happen all of those human aspects need to be automated away we need to replace those things with technology uh and that's what Jensen's doing one of the absolutely crucial things there is verification we need to know that if I have sent a part of a machine learning model to a machine somewhere else in the world that machine has actually done the training work that I expected it to do um right now we do that through lots of kind of human mechanisms we do kyc we have contracts Etc and it's very very slow and it's it's a bottleneck so if we repl that with technology we can improve that situation and we can make models much easier to develop and we can let models build models it leads to a very rich ecosystem of autonomous agents all interacting all building models themselves uh and this essentially embedding space that covers the entire world and all of the compu in the world um I've been speaking for a while now uh I'm happy to continue monologuing to you to you folks but if you have any questions I'm also very very happy to to answer those questions so put your hand up come down to the front if you want to ask a question otherwise I will just continue telling you my views on machine learning and decentralization um but yeah please would love to uh open this up a little bit uh so one question that I wanted to ask is that there are a few solutions that enable people to rent out their compute that they have for example like 490s or if you have like v100s or A1 100s lying around from the use Market uh for individuals like vasdi or you know vasdi would be a good example of this solution how's your protocol going to improve on that and specifically another item that I was uh looking into is that the network requirement for individuals is kind of strict because you know uh AI is very bandwidth constraint and high networking speed is something that AWS has on the pinch and a lot of consumers and a lot of you know small owners of compute don't really have because you know your internet provider is not necessarily going to provide you with gigabit speeds on demand and that's going to be difficult for all so how would you go around these two challenges basically improve on existing Solutions and solve the networking issue yeah two great questions um so I guess to paraphrase the first one is how do or how would what we're describing be different from the current centralized aggregators essentially where vast AI will allow anyone to list a GPU on their Network and then somebody else can use that in comparison to a centralized provider where AWS has all the gpus so it's kind of an open network uh so first question is that second question how do you overcome the latency and bandwidth constraints of a truly open network where maybe you've got 309s in people's homes running that's obviously very different from aws's Big Infinity band cluster of h100s um so on the first question I think the vast AI approach is a a solution to the the backwards looking half of what I described as being the problem so you look at the market right now and you say I have to go to a centralized provider they have a strangle hold on the market and they're able to make enormous profit off me if we can unlock some latent Supply we can reduce those prices because we can bring Supply into the market I think the centralized approach like vast AI is a good approach it's an improvement on what we currently have but it's still bottlenecked the way that vast a ensures trust is via centralized kyc essentially so they have to onboard every person who brings a GPU to their platform uh and they have to essentially ensure like ensure the um the user of the platform against malicious behavior from those supply side participants that sort of works right now but it doesn't scale because at some point that kyc burden overcomes the benefit of growing that network uh and you hit a ceiling and we saw this back in 2014 there was a company called Vector Dash that came out of Y combinator they had this exact same approach um and they ended up pivoting into the cloud gaming space because they realized that overhead of centralized kyc and ensuring trust is just too high that's why decentralization and blockchain is so interesting because it allows us to solve that with technology rather than with people so we can ensure that trust via a chain itself it's it's it's easy to say it's hard to do in reality you have to terminate disputes over very massive machine learning kind of non-deterministic tasks onto the a chain consensus mechanism which is incredibly difficult but if you can do that you've replaced those uh people portions with technology and you've made it instantaneous um so that's why you need decentralization on top of something like vast to scale um vast as well if you did just a completely decentralized um but kind of token based alternative to vast you would still only solve that portion of the problem you need the verification and you need the um the kind of programmatic transactions to scale to the second portion of what I said earlier which is machines interacting with machines it's machines triggering that training and able to use this infrastructure without a human ever involved in that case they need wallets they need tokens they need uh access to the other core resources like data and expert knowledge and things like that and we need all of those portions before we can unlock the true new machine learning future so vast kind of improve what we currently have have we view them as like a 10 to 20% Improvement on what we currently have with Jensen we view as building the next kind of Step change the idea is change the way machine learning is done completely that leads into your second question which was the latency bandwidth constraints um and if you look at the way that machine learning development has um has developed and progressed over the past 5 to 10 years we've gone down this path of Highly centralized models getting bigger and bigger and proving to the world that we can do very interesting things um so we view that as the T as the tip of the spear for research so the first thing that we do is say hey world look we can build models that are incredible like this the way that we've done it is super fragile and it requires us to put loads of gpus in a data center together and make them like really fast interconnected but it works that's the first step after that we go back to kind of general computing principles we look at distributed systems and we say how do we actually scale this properly uh and what Jensen thinks is that scalability will be highly distributed highly decentralized it won't be adding another thousand gpus into a data center it will be removing that kind of Step that small step change and saying what if you can access a million MacBooks now could you build a model differently over those devices and I think one of the things that's interesting is the machine learning space has a really short memory uh people will look at the progression of these models and they'll think oh this is the only way to do this but actually if you look back we've only been doing these hyper centralized models for a few years it was a really interesting way of doing it but there were plenty of other interesting ways to do it so there's a there's a kind of chicken and egg problem here and there's a naivity that can come if you're building something like us where you say hey we'll just build this massive network of all the devices in the world and then people will just turn up and build like crazy decentralized models over it that's true to some extent but we also have to prove it so we're doing research to say rather than building a hyperconnected model that's fragile and if any piece kind of Falls the whole thing fails let's build a massive model where it's lots of different pieces all into operating and replication between those pieces allows us to drop out pieces and bring in other pieces combine different modalities and generally make this huge kind of Rich sharded embedding space over the entire planet's compute resources we think that's the true kind of future of AI rather than these fragile centralized models that we're sort of focusing on as Society right now that's that view of AGI as being like one model in open ai's data center we think is just flawed it's a very shortterm view the future is Rich bits of models everywhere on everyone here phone and then if you're running something through those models you're going to connect up all the different pieces as and when you need to to get the answer that you need um hey I was wondering how you would ensure the Privacy like I send someone my model and then perhaps someone steals it and now there are like disputes and stuff like that great question um another another tricky one um so that's something that we we care about a lot um as a brief aside my previous startup was actually in the Privacy space um so I was building individual consumer privacy tooling I think in the future we should we will have a lot more control and agency over our own data I think we've gone down a path with particularly with social media where we've we've become accustom to giving up a lot of our data in exchange for services but not in a kind of granular consent based way and I think we do need to move back to that I think it's slowly happening um what I learned from my previous startup is that a startup's probably not going to do it um it's going to have to come from one of the big kind of giants who already have a lot of us as users Apple are doing pretty good work on this they've got it as a brand value and they're doing really good research so it's starting to happen um but when it comes to our infrastructure and the the decentralization of AI what we found when we talk to potential users of this system is that actually a lot of machine learning work is happening in the open we're seeing a lot of Open Source models being developed uh they're being developed on open source data compute as a barrier is actually much larger than the access to data and access to model architectures so a lot of people are using open versions of those two things but still can't access compute so what we found is 50% of those people are happy to use completely open compute where everyone can see what they're doing just because they're unable to do that right now and they want to be able to build these models after that 50% the remaining 50% are split into two so all of them care about data privacy like the individual training data going into models and half of them care about model privacy which is the weights or architecture of the models themselves the kind of proprietary uh design of the model that gives us a road map to say we can develop in the open we can build The Primitives the verification the hard things the the bandwidth and latency kind of uh research to overcome those those constraints on completely open models at first and then we can move to data after that we move to models and the road map that we have on the technical side side reflects that so right now uh building in the open obviously you don't need extra technology to uh build on top of private data there are various techniques that are coming to fruition so you can do things like differential privacy it's very tricky but it allows you to uh essentially anonymize data that goes into models so you can take out the kind of individual aspects and use the value and the data without identifying people it's subject to attacks and it's it's there's lots of reasons why it's it's hard to implement but it's possible another one is functional encryption where um there's a real interesting technique you can do with a neural network where you replace the first layer of the network with a functional encryption key and that key then decrypts the data but only for the operations of the network so it's homomorphic over the rest of the execution of the network very long story short that lets you train on Cipher text so you can encrypt your data with an unencrypted model you can send that encrypted data to a party who has the model and the compute power they can train the model on that data and they can send the model back and they've never seen the data they have no way of decrypting it so that's solves the data privacy problem it also comes with downsides you then have a model that every time you want to run data through it you have to encrypt it first which is a bit annoying and you kind of lose a little bit of information in the data but it works um the final piece is the model privacy piece and that's where fully homomorphic encryption comes in fully homomorphic encryption is very early as a technology it's possible for model inference right now it's very very slow and it's incredibly slow for training so from our perspective it's maybe 5 years out before we'll get that really realistically applied to training but it gives us a solution and it gives us a road map to work towards um so there's some pretty exciting techniques in summary um and there's also we have a road map for what the users actually want and surprisingly half of the users in the early cases just want to build open source models on open source data but at the moment they can't get compute I guess it's going to be me again uh one of the problems that I saw in the space is that the you know the fact that the uh compute comes in many different forms and it's not uniform and that specifically plays very strongly into nvidia's hands basically right now if you look at this space that there are kind of two points uh you know in terms of the training data and that kind of gets all all the other vendors of compute out of the game so basically right now if you want to get state-of-the-art with say good uh FP or in a capabilities you look for h100 and if you're looking for a more cost uh sensitive especially on the supplier side that's probably going to be like two generations back the 100 would be the best choice for you uh however you know there are like a lot of solutions for by AMD by Intel you know they are coming up with their God line and the question here is that uh how do you ensure that in this case you get you know if most people are optimizing for NVIDIA how do you make sure that people having compute from other vendors say apple is completely upt right now and they have pretty good machine learning capabilities uh how do you make sure that those people get you know a fair return on their uh computer because in that case if you have like a lot of a models optimized for NVIDIA they disproportionately reward nvidia's hard uh Hardware in terms of performance as a result if you have am and D struggling to get you know good Returns on your because the model just does not provide just do not rotten well enough with AMD hardware for example yeah another great question um I think one of the the issues here is nvidia's ability to capture the uh capture the market with Cuda so the software side of the market and I'm going to have to be very quick here because I only have a few seconds left but um in essence that's down to the protocol design so we're building the software layer that allows that to happen our long-term goal and I think this is probably why we differ from the networks that you see that are only addressing half of the the two things that I mentioned at the beginning they're only addressing the current situation so the way that models are currently built and the uh the way the market currently works and they're not looking forward to the way the market is going to work in the future if you're looking forward I think we need to focus on commoditization of compute if you look at most machine learning uh Engineers who are after Hardware they think they need the h100s the h20s the most powerful devices that they can get from Nvidia but in reality the tasks that they're doing don't need those uh those pieces of Hardware uh they could work work over AMD Hardware they could work over Intel Hardware they could work over some of the custom chips but the software layer doesn't exist and so the proper protocol design involves solving the the trust problems but also solving those software problems and I think that's one of the things that's really difficult and one of the big problems is it takes a long time and it's very expensive to do because the Nvidia has been very careful to capture that stack with Cuda you essentially have to disrupt Cuda and you have to build around it so that you can bring in the other devices it's something we're focusing on very heavily we're building an entire stack to do reproducible machine learning over heterogeneous devices and the idea of that is it's it allows machine learning Engineers to stop thinking about which device they're using and just think about the fact that they need to distill value in data into models and that's it they should only have to think about that it's a little bit similar to the serverless shift within Tech in general where Heroku allowed people to not think about that I'm being told to wrap up so I will wrap up there but please ask me any questions I'll hang around for a while and and happy to talk more thank you