AI’s Appetite for Compute—and the Rise of Decentralized Mobile Compute - Alessandro De Carli

Transcript

[Music] thank you so much um yeah I'm very happy to be back here in University uh I also was kind of studying on on these very um tables uh a couple of years back uh I see that we have currently the wrong screen on um let me see oh yeah perfect thank you all right so today I'm going to be talking about the topic that I'm very passionate about that I'm actually quite obsessed uh about it's the uh topic of compute and more specifically how uh mobiles uh play a big role my background is in Mobile security I uh

used to be working with some of the largest uh financial institutions and defense companies to basically bring in mobile devices into uh their setups so that they um Can optimize workflows and basically Embrace this this innovation I do have a couple of slides so let me Jump Right In um when we're looking at the demand and appetite for compute of humans we will show well we will see two things right one is the hardware obviously gets uh better and better um at an exponential scale here you see uh basically how uh we are printing more and more

transistors into um CPUs um this does not talk about the density of transistors uh it's just the number of transistors in in that case um and another metric that we can see as well to to basically showcase uh this appetite is um the power of uh supercomputers uh which ultimately is measured in teraflops uh you see that this is as well a a logarithmic scale um that um shows you how year after year we're basically doubling uh if not more the um the amount of compute that we that we have at our uh

disposal now um here is however a breaking change um with now ai uh we we changed a little bit the course of action okay so so you saw the linear uh line well you saw the line before that was on a logarithmic scale um and that uh you you can basically see here in the uh in in how many data points uh there are uh being consumed by these uh machine learning and AI labeled uh things but since a couple of years uh this even um accelerated quite a lot okay so what does this tell

us well um this acceleration also comes with a big centralization uh because a lot of this Compu actually lands in uh centralized uh data centers and uh as you know there is no Cloud it's just somebody else's uh computer and the big problem with this centralization is that uh it started off with the users's best interest in mind uh offering them the best possible service but um shifted at some point to a um more kind of um pushing for for shareholder value uh approach um long story short what it really means is that you now have a huge

concentration of compute a huge concentration of power um into the hands of people that don't have your best interest in mind and this becomes very um clear now even more with uh AI you saw before this uh acceleration of of data points um which already on a logarithmic scale changed um to to an exponential uh graph what that really means is up to a moment we had machine learning models already around since many many years but they were all prone to a um phenomena called overfitting so at some point even though

you were giving it more data it would just get worse uh the Precision uh but clearly uh this is not the case with llms so either we haven't reached that stage yet so we haven't reached the moment where it's too much data or llms are simply not prone uh to that right and if you now think about yourself as an individual this means that all of these uh big hyperscalers are super hungry for data because they're able to extract more and more value so up to this moment it was not so relevant um that uh you were sharing a lot of data with them because they couldn't really process or extract value out of them and

now they sure can um for all sorts of um yeah use cases that may also go against you one of the uh things that uh make this very very visible is that if you have access to these big uh models um that are centralized basically able to influence uh the perception of um the general public in this specific case uh we see how um yeah Google being a company that is pushing a lot for diversification and diversity um ended up creating a model that would

uh reflect that those values um in in their answers right and here basically and actually then uh returning uh incorrect uh data um because as you know um the chances of um female or or black um or Asian uh soldiers um on the German side in World War II are are very very slim right but it's not only Google um with deep seek uh this becomes also very clear um here as asking uh a critical question uh that the CCP is not um yeah interested in in Sharing uh you see how

basically the model itself even though it has all of the knowledge because you see the thinking process I don't know if uh I can you see the thinking process we're we're asking about um this iconic picture of of a man holding backs in front of uh this line of tanks right the model starts answering the thing correctly but the minute um the T men um or tman Square um keyword shows up it's being uh censored so a lot of of power is in this uh computer in this um AI because it can shape the way we we

perceive and and uh basically act in the digital world long story short you need to to take control back um into your hands um and what we really need is a the centralized system that puts you basically first now decentralized systems are not uh immune to anch ification that I described before which is the process of uh moving from something that had value to basically shifting uh the the alignment to something else um but if you're able to to come up with something that is truly global scale um and decentralized by the people for the

people you're able to uh offer an alternative that actually has again your best interest in mind this is nothing new during the um Ico times this narrative has been um handed out couple of times uh already today there's a a couple of um tokens that are trading actively uh in the space of the centralized compute but all of these tokens and all of these projects are basically missing one big point and that point is why would you rely on centralized data centers uh when you're trying to build a decentralized compute Network because if you do if you basically just use the same old tools

and wrap them uh with a new wrapping you end up with a system that we have today which is this pseudo decentralization where people people say we have um thousands or even Millions uh of notes uh meanwhile those notes actually live on the AWS Cloud right um or also um the the problem that a lot of these networks are basically um just working with self-reporting where people just say yeah yeah I do have that Hardware trust me uh which obviously in an adversarial space such as blockchain never works uh because people will take advantage of that and they actually

already took advantage of that and last but not least all of these systems require trust in whoever is providing you with the Computing infrastructure which obviously is a problem because if they can tamper with your execution or can tamper with the data you share with them this will never take off this will never reach Mass adoption and now we come to my obsession uh because basically we don't have to look that far for the answer and the answer lies in your pocket um everybody has a mobile device um if you remember the the the short uh and brief um chart that I had before when it comes to compute uh these

chips actually uh lead the way you see here many mobile chips as being uh the ones chosen uh to to prove this this more slow if you now would even start um dividing the number of transistors by the actual area that you have uh you will see that mobile chips already are working on 3 nanometer Technologies are way more efficient than than servers um and uh are basically already out there the way more modern uh Hardware now I will quickly try to um bring my my

passion and Obsession for mobile devices to you with a couple of with a couple of um examples and why I believe um this really matters so one thing is that mobile devices are secure by Design and the reason is mobile devices don't have the privilege uh to have physical security like a server in a data center in a data center you don't have Bad actors physically accessing your Hardware with mobile devices you cannot rule out this attack Vector this means that if some of you now uh needs to go to the bathroom and leaves um a phone here on the table

somebody else could just pick up that phone now why would you leave your phone here on the table um and and go to the bathroom well because you know that that device device is actually secure right you're using it for all of the mobile banking managing your wallets managing your identity and so on and so forth um a lot of effort has been put into creating the most advanced Hardware security modules out there that offer both verifiable and confidential compute and we'll see why this is important in a bit the other aspect is that mobile devices are the most centralized form of compute out there already today we have 8 billion humans and more than seven billion Mobile smartphones actually uh

in circulation so the vast majority of compute is actually in mobile devices um we sell 13 million mobile devices annually while we uh sell as well 1.2 billion sorry servers annually while we sell 1.2 uh billion new phones each and every year what you have to add on top of this is that the lifespan of a server is usually 10 years while a mobile device is about two two and a half years that means that you have also more law uh working in your favor in this kind in this uh case but it does not end there so you have apple that basically invests

the entire market capitalization of all of these compute coins um each and every year in just R&D the iPhone is Big Business the iPhone alone makes up for more than 60% of Apple's entire Revenue so it's clear that this is also not only one of the most secure scalable and decentralized devices but also one of the devices where a lot of R&D money is actually being poured in so who would have guessed um this is basically the project um accurist that uh we're building upon these uh core values and these ideas which is a real

decentralized compute Network that allows you to um onboard mobile devices how that looks like we can see here um with one of our users uh who onboarded about 200 devices all of these devices are interconnected with um USB 3.2 so they have a 40 gbit um interconnection offer eight cores 8 GB of RAM and what you can see here at the end is actually the the killer use case which is um you're able to load uh or upcycle devices that are broken um you are able also to tap into uh vas Community uh already today so if

you have a mobile device a spare phone or something like that uh you're welcome to to actually onboard them and even though it's just on test net this concept has already been battle tested uh at scale um and uh we already have um quite a distribution it's not perfect yet I think we still have a lot of work to do um on the Africa side of things now looking at time um I see that my time is almost up um that's why for whoever wants to see uh the demo that I prepared um I'm very happy to uh do this uh right after the call but basically everything is uh completely

permissionless and it's not that you get slapped in the face with a contact us form uh whenever you try to deploy something uh use cases uh for the developers uh in here is you can think of it as a general compute platform so so you interact with it uh using just um wasm workloads or or no Jazz workloads um and the best uh use cases are actually the ones that rely on these um three main uh USPS that that the um platform offers you um concrete use cases actually lie in the uh verifiable secure onchain

automation um as well as then web scraping um and uh distributed simulations um because yeah you just have a lot of uh compute out there uh already on the network with more than 100,000 um high performance performance course so that's it from my side I would love to talk longer about uh all of these uh topics um and as I said if you want to see a demo I'll I'll gladly uh show you outside um yes uh if you have a spare smartphone or you want to onboard your everyday device please do so uh the the mobile apps um are already in the stores for

you to grab or even on GitHub thank you so much thank you Alandra are there any questions from the room three two one okay thank you very much thank [Music]