David and Goliath How Web3 Can Compete in AI | Gurnoor Narula - Placeholder VC

[Music] Analyst at placeholder Ventures uh which is a venture Crypt native Venture fund based out in New York City and today I wanted to talk a little bit and provide some clarity on the whole crypto AI vertical and really break it down to its barebones technical level there's been a lot of talk about an excitement uh about crypto and Ai and how the two intersect and I think today I'm going to put a little bit more of a somber tone and break it down to its technicals on where the crypto stack helps Ai and where the AI stack helps crypto so where I want to start is by defining what I believe the AI stack consists of right on your bottom you have your compute your data layers right above that you have your training layers which is essentially taking in the Raw goods and processing it into something useful on top of that you have your Dev tools your inferencing providers and at the top you have developers creating applications and agents that leverage these training and INF training and inference infrastructures to create something valuable for us everyday users and I'd like to focus on the bottom two the compute data as well as the training layers now a lot of folks have probably mentioned the fact that we can decentralize some of these training these compute these data collection mechanisms but the question is why it seems that we're centralizing for the sake of decentralizing without considering the performance and efficiency tradeoffs for doing so and my thesis here is that centralized providers currently that collect data that house compute resources and therefore translate those into training Foundation models currently do and will continue to have an advantage as compared to our decentralized counterparts and the reason for that just pure economics and scarcity if we think about it today demand for Topline Hardware data and compute resources will always outpace Supply and since those are the primary ingredients in creating training resources what we'll see is that centralized providers will continue to outcompete and outperform decentralized counterparts at least for the foreseeable future what This lends itself to is an observation that scarcity for Top Line Hardware data and capital equates to centralization within our most performant Foundation models and this is going to continue happening you have your Nvidia chips now in their current V1 you have V2 V3 V 100 versions of these coming out and the Topline providers will always have a higher demand as compared to supply for some of these compute resources thereby leading to centralization at the most fundamental training compute and data uh data locations now this kind of puts a bleak uh outlook for the the future of decentralized compute and training layers but I'm here to tell you that it's not all bad what we'll find is that when you start generating these different versions of Topline Hardware it's not like the older Hardware is now suddenly moot or useless right the the the Topline Hardware is only supporting the most performant Topline models but what happens to the older Hardware that works perfectly fine but just for older use cases well I believe that it finds allocative efficiency so as you can see here we see old Hardware that's supporting mid-tier models as compared to decentral and then Services decentralized compute networks therefore servicing on Prem and Edge models personal data sets and maybe smaller tasks now on the other side what we see is Topline models these Topline models sorry let me make sure I'm presenting this properly there we go so these Topline models therefore will service more complex models therefore will be centralized and might be centrally hosted complex performing tasks now an easy way to walk yall through this is let's say I'm a personal user and I'm using my iPhone to execute a simple task maybe I want to book an airline do I really need nvidia's h100 chip for me to execute a task on Expedia probably not and so what I'm probably doing is okay we have old Hardware on my Edge device serviced from mid-tier models that can fulfill said task and are good enough to do so they're able to run on decentralized compute and training networks because it's cheaper more affordable and more flexible than its centralized counterparts and it can exist on our Edge and can service our personal data on the other hand let's say I'm executing an incredibly complex task maybe in this case my iPhone that's running some Edge model can't service said task what we can imag imagine then is my iPhone or my Edge device offcat my personal data my personal identity and my personal request sends it to a complex model maybe hosted in a centralized way because that's what their advantage is it's run on a centralized training model and therefore can perform more complex performing tasks now moving on from here what we see is this thesis that maybe the factors that we're competing on for decentralized training and decentralized compute are incorrect if you think about what people are marketing themselves as today they're saying we're going to become the next open AI we're going to essentially replace what folks are building on the open AIS the microsofts Googles of the world maybe the factors that we want to compete on instead is look those guys have the most performant Hardware they have the most performant data that's great good good for them and that's where they can operate but what we on the decentralized area really compete on is verifiability flexibility and censorship resistance and to a certain extent affordability and I think if we continue harping down on those areas that's where we'll see our advantages now moving on we can see that This lends itself to an eventual transition in the training model currently our decentralized training models are censorship resistance and that's great for them on the centralized part we have potentially biased and centrally run models and we can see that that's a relative market share split most of the decentralized models are run by hobbyists independent researchers and startups whereas most of us use the open AIS of the world what we could see though is a transition to this future where decentralized models end up capturing most of the Mind share for our individual users and then centralized models are potentially biased and centrally run but are reserved for Topline performant and complex tasks and so it doesn't have to be a centralized versus decentralized it could be that they exist together now we move on to entrancing so let's move one layer above on the stack right now it's it's the same thing with with inferencing models that run on top of decentralized or centralized training models they offer zero downtime public evaluation and reputation scores and a set of diverse outputs on the centralized side though there's bias vendor locking a lack of verifiability we're placing trust in these actors because we have no other option but again if we focus on competing about censorship resistance we focus on competing on the factor of affordability and flexibility we can exist in this future where most user requests for inferencing are completed by decentralized inferencing providers and only the most complex requests are completed by your centralized counterparts again this whole thesis of coexisting rather than one against the other now what This lends me to is a thesis that I express in a blog post which yall can read online as well that decentralized compute networks and inferencing networks present a unique value proposition around more personalized affordable openly competitive and censorship resistant Ai and I think these are the factors that we have to compete against to eventually gain market share against our centralized counterparts as compared to butting heads with them rather treating them as equals and therefore coexisting with them for certain tasks now before I end my speech I would like to say that there's a lot of innovation that I've seen at this eth Denver specifically on the agentic Apper and a little bit of a thesis there is you could imagine as I said before you operating with your iPhone or another Edge device and fulfilling requests in natural language and therefore not having to deal with the complexity of finding resources yourself but rather just speaking it into existence now how does that coordination at that agentic L happen I think that's a really really cool question and answer anything around agent Discovery agentic coordination and communication so if yall are building in that space researching in that space feel free to get in touch uh that's my Twitter over there um and more than happy to talk about it afterwards as well appreciate y'all's time appreciate e Denver for having me out here thanks again