The DePIN Advantage: Supercharging Crypto's Role in AI | Doug Petkanics - Livepeer

all right hello everyone my name is um Doug panx I'm the founder of a project called live Pier which is a deepin infrastructure Network that's been building the world's open video infrastructure for the last um seven years one of the longest running deepin projects and um if anyone saw the previous talk I'm going to talk about a similar topic something I care deeply about which is the sort of importance of opening up and decentralizing access to artificial intelligence and I'm going to talk about why that's really critically important right now we're in an important moment and I'm going to talk about how crypto networks or dpin networks actually are a critical um part of the solution to doing that a lot of people this week have been wondering you know what do crypto and AI have to do with one another is it just a a meme is it crypto search for a narrative is it a pivot um I'm going to make the case that it's not I'm going to make the case that it's not the total solution but that decentralized Technologies and uh blockchain networks can be a you know really pivotable pivotal um you know impact into decentralizing artificial intelligence I'm going to talk about sort of two reasons why it's really important one is the risk of regulatory capture that prevents access to decentralized projects International projects startups from having access to artificial intelligence one of the most transformative Technologies of Our Generation and I'm going to talk about the um lack of access to the compute that's necessary in order to actually train AI models do run AI inference to anyone outside of the big Tech Monopoly players first I think we can all acknowledge if you've been paying attention artificial intelligence is is accelerating at a rapid Pace if you look back just 18 months ago you would see a lot of this stuff didn't exist and then late in 2022 consumers first experienced interacting with large language models or llms through open ai's release of chat GPT where you could type a prompt ask a question and have a chatbot type experience that access the world's information and can engage with you and give you give you answers and how interactive that would be and then right around that same time and sort of following very quickly was the introduction of um kind of image generation diffusion models which introduced heavier and heavier forms of media um Beyond just text and you can type prompts generate images and you can see how that's changing the world of design and creativity these things require a pretty significant amount of compute but can still be performed pretty quickly gpus can give you an answer to a text query in you know hundreds of milliseconds that can generate images for you in seconds with the turbo models or maybe 30 seconds of GPU compute time and some of the more advanced models but recently you're seeing this accelerate even faster to heavier forms of media and you're seeing really mindblowing things like what open AI released with Sora just a week and a half ago where you can type a prompt and they can generate 10 20 60c videos um just from The Prompt and this was really inspiring to the world and it was really eye openening in terms of how much it's going to change entertainment creativity content production um Tyler Perry one of the most successful you know media Moguls of Our Generation saw this and literally canceled a $700 million investment into his Hollywood Production Studio because he said we need to pause and we need to rethink how AI is going to change and reshape video creation before we go and throw our money uh toward the old school way of of developing things um this power doesn't come for free though in terms of compute I mention it takes you know maybe 30 seconds to generate a a beautiful image using AI to generate a you know 30- second movie might take an hour of GPU compute time and anyone who's actually tried to pay for and access GPU compute right now knows how expensive it is and how much of a shortage there is this is reflected really heavily in the price of one company stock Nvidia and by the revenue that Nvidia the company that makes all of these high performance gpus for AI training and inference um the the revenue they're generating each quarter what you see here is this quarter alone they generated $18 billion doar of Revenue selling data center gpus for AI training half of which was bought by four big tech companies Amazon Google Facebook and uh micros moft and this is only accelerating because these video generation models which are the heaviest of anything we've seen aren't even available to the public to play with yet and as they are imagine the explosion in this this curve isn't slowing down anytime soon so this shows how much demand there is for compute what it doesn't highlight is how limited access is to anyone outside of big Tech Cloud companies that have to buy hundred million do purchase to access this compute from Nvidia um that's just simply not accessible to to the rest of us and that's where I'm going to connect the dots to how decentralized Technologies and you know blockchain uh blockchain networks actually matter a lot in making sure that we have democratized access to this transformative technology the first reason I'm going to hit on and this is kind of ties to why now we're in a really important moment in terms of making sure that we ensure access is that um open and decentralized access is actually at risk due to regulatory um actions being taken here in the United States the president has signed an executive order already it's already in effect that regulates how startups can interact with AI or how companies can interact with AI it contains a bunch of sort of draconian uh Primitives including that infrastructure as a service providers that want to provide AI inference have to do kyc on their users and they have to actually report foreign Nationals that are trying to um use AI inference in their data centers it says that there's actually regulations that you can't train models of certain sizes without complying with six different government agencies and uh that the size of the models is defined by how many floating Point operations you need in order to train it which is just an arbitrary number chosen by the government at this ear early stage of AI iteration um anyone who's tried to build a startup a decentralized tech project a hack do anything Innovative knows that you certainly don't have the time or the money to comply with six different government agencies and wait for permission from the secretary of agriculture before you can innovate in the AI space and this is going to be challenged it's going to be fought in the courts um but we need to make our voices heard and we need to create Unstoppable platforms that sort of render this regulation uh you know less relevant to early stage innovators right now this whole thing is not just the government alone they're being influenced by the big tech companies who have already spent billions of dollars training their models creating proprietary products and sort of owning access they want to own access to the space they have lobbyists and they have you know law firms and uh in-house lawyers at hand that they can pay to do the compliance with these agencies and that actually creates a mode because no startup Challengers no innovators no individual hackers have the ability to uh you know pay and wait and and comply with sort of arbitrary Draconian restrictions um and it's really important that the open source AI movement and decentralized Tech communities represent Solutions not just for reg regulatory reasons but also for the second reason which is to ensure that there's open access to this important technology um Elon Musk highlighted you know gpus are harder to get right now than drugs I don't know if it's a super high bar but it's still um it's a true statement uh why is this you have companies like meta who just are are committing to spending 10 billion to Nvidia to order the highest density compute units the H h100s um they say they're going to buy 350,000 of them I think you know a couple weeks ago they just announced they just bought 60,000 of them these are the same machines that you can't get access to through the clouds because they're they're being sold out and reserved by these big tech companies and while that may be the case it's an Open Marketplace inidia can sell who they want to there's tons of idle compute out there in data centers um for use cases like gaming rendering video encoding that all can be put to use on open and decentralized marketplaces to train AI models to perform inference to do things like generate images generate videos run llms for chatbot Comm communication um and all you need is sort of the right protocols to coordinate them and the righten incentives to bring that capacity online and that's actually the big opportunity that decentralized physical infrastructure networks that are blockchain coordinated have they can create deep in infrastructures that allow AI training and inference at scale and they allow for a place to run open models that are not censored that are not biased that are not dictated by the interest of big Tech advertis iing models or big Tech sort of legal regulatory restrictions for those of you who don't know a lot about deepen networks um it kind of describes a category of blockchainbased projects that coordinate real world infrastructure it can be things like mobile access points in the helium Network Automotive car sensor data in networks like demo or in this case large networks of gpus or graphic proc graphical processing units which are used for AI compute um these networks have a couple superpowers that actually you know the game is changed via crypto across you know a number of vectors that matter scale reliability cost and access I think the biggest superpower actually comes in scale and it comes from the ability for these networks to bootstrap the entire supply side using cryptoeconomic incentives prior to their being demand that's really powerful if you didn't have crypto you didn't have blockchain and you didn't have incentives if you were to try to compete with Amazon web services you would have to invest billions of dollars to buy data centers buy tons of Hardware from Nvidia buy the bandwidth build software orchestration layers on top um a huge ambitious and risky undertaking with crypto incentives where you're actually making infrastructure providers owners in your network through Distributing token to them in the early days and you're not doing it for your own centralized control and profit but you're saying those who bring the infrastructure are actually the owners of the network as long as you show up and you're available make your capacity available you actually can see super scaled networks of gpus um show up available for your use case as the demand is ramping up because they're they sort of compensation is in this sort of speculative bet that they believe this network will be useful in the future and they want to be a big owner in it because they'll receive the fees that are that are coming in in the future and um this is not made up we've been doing this at live Pier for six years we have a huge network of thousands of gpus that do video transcoding I'll get to that um and it it's existed across many other dpin networks as well and we're going to apply it to AI um the fact that these networks are open anyone can put compute on them and they compete on price to do jobs means that you actually get the lowest cost possible you have not operators competing with their idle Hardware that's not otherwise being used to say I'll charge a lower price because I want to get the job and earn this additional revenue and so for video encoding you see that you know the live pier network is 10 to 50 times more cost effective than um what you would see in public clouds and that's going to be really powerful for this super super heavy form of compute where you need hours of GPU time uh to generate videos images Etc so excited about that networks can be reliable when they're that cost effective because you can afford redundancies you reduce single points of failure and of course you have open access which allows you to train and run inference on open models allows people to pay as they go instead of having to reserve capacity months in advance like you do with the clouds and it enables people to build personal agents personal llms free from the biases that are uh you're seeing in products like Google Gemini that you know are trying to be um sort of subject to the interest of of Google or the mainstream or the advertisers that don't want to disrupt Google's traditional business model so as I mentioned you know live Pier is an example of this in practice when it comes to video en coding excited to bring these same capabilities to AI uh Global Network of thousands of gpus we transcode over 3 million unique minutes of video per week that means at any given moment hundreds of video streams are flowing through the platform and delivered to to viewers at a significant scale 3 million minutes a video watched per day across hundreds of applications built on the network um excited to bring generative video to the network in the next couple of weeks and and months actually um this is not just you know made up non-real crypto stuff it's powering real use cases that millions of people connect with every day there was a concert last week on the lot radio one of our users with Fred again and Lil yotti um Skrillex has been on there through that platform so it's actually you've onboarded millions of users to crypto without them realizing that there there's anything crypto related it's being used under the hood to power and coordinate these networks um on the left here you see fish tank. live a hugely popular Internet TV show had millions of viewers generated millions of dollars in subscriptions in order to watch a real world Style Show um huge with the Reddit crowd um you know all powered through a cost-effective infrastructure like live Pier um in addition to powering many of the leading web 3 social and video and media applications as well um quick slide about how we're going to bring AI compute onto the network we're going to use a subnet type model where anyone running an AI node can you know submit AI jobs to node operators running AI nodes to use their gpus to perform this compute um but all of this can piggyback off of the live pier network which has been running for six years for node Discovery security of compute and payment of settle settlement of payments so you don't have to have some ambitious Pie in the Sky Vision that's you know we have to build everything all at once we're going to use something robust that's been delivering value for for six years so like I said exciting the next couple of weeks to bring these capabilities in the video specific vertical where we think AI is going to change the world lastly I just want to re-emphasize I'm really excited to be working on this this a common theme this week across eth Denver the time is now to actually create these open and decentralized video infrastructures to avoid the regulatory capture which is coming into effect in the next couple of months due to the the Biden executive order and to ensure that there's open models and open access to infrastructure so startups can develop on these things before the opportunity slips away and you have to go through the big Tech gateways so it's been really inspiring to see this really common theme across e Denver this year um I think the industry knows the time is now knows how important it is and everyone is uh you know really excited to run fast to to work on this Mission together so thank you very much for uh listening to The Talk today I'll be around if anyone has any uh any questions but you can find us at live.org and if you want to scan this QR code uh you can kind of sign up for updates on the AI video processing work that's that's coming soon so thank you very much everyone