Les Amis d’Ethereum France #21 — DePIN avec Aleph et Exaion

Good evening everyone, welcome to a new Ethereum France Meetup. On the agenda tonight is " de pine" (I hope that's the correct pronunciation). We're talking about "de pine," perhaps with Aleppo and Exalon on the program. We opened the doors at 6:30 PM and now we're getting to the Meetup. I'll give you a brief presentation of Ethereum France before handing the floor to Simon Thomas, then Jonathan Chemou.

From 8:00 PM to 9:30 PM, you'll have a small cocktail and some food to share and discuss what we've presented. For those who don't know us, we are Ethereum France, a non-profit organization (under the 1901 French law) whose goal is to popularize blockchain and Ethereum throughout the French-speaking world. We operate in many different regions of France and Navarre, and also in the border region, particularly Brussels, where we will be organizing a major conference soon. The association brings together a dynamic ecosystem, of which Exalon and Alf are a part, and we like to create content so that everyone can make it their own. Ethereum and what the blockchain is, by the way, the eponymous website ethereumfrance.

com is our site and you can find lots of information there. We have values ​​of knowledge sharing here. Most of the time you'll find things that are rather technical, but we also have an open mind, financial interests, dependence, and cooperation. If you want to participate in what Ethereum France does, there's a Discord server, an association, and things you can get involved in and share and exchange with our members. You can join us on the website ettherumfrance.

com. We launched on January 21, 2023. You can discover our Discord server where there's a very active community. You can also get involved in the association's activities. We're always looking to bring more people into our association.

Among the activities of Ethereum France, beyond managing the website etherumfrance.com and organizing these kinds of events, we also organize an annual conference called ETHCC. So you You may have heard about it, and it's happening this year in Brussels. All the THCC videos are available on our YouTube channel, where we're broadcasting the event live. This week, we launched a new website called ethccweek.

fr, a site we maintain every year to support all the side events happening around HCC in Brussels. The THCC conference will take place in July, during the second week of July. And during this conference, beyond the huge HCC event itself, there are tons of side events. Last year we had several hundred, and this year we expect at least as many, even though it's in Brussels. So, if you're planning to organize a side event during this week, you can submit it here.

You can also find information on where to stay and spend time in Brussels, and don't hesitate to contact us through this website. Well, without further ado, I'll now hand it over to the first speaker of this event. Thank you, thank you all. So, to be here, let me introduce myself. I 'm Simon Thomas, and I head the blockchain department at Exayon.

We'll start by introducing the topic of decentralization: what it is, what its overall purpose is. Then we'll talk about Exayon and how Exayon fits into the decentralization topic. So, there's Messari, who defines decentralization very well, I think. Decentralization is an acronym for decentralized physical infrastructure network. These are networks that use crypto-assets to coordinate the construction and management of critical infrastructure.

Today, the decentralization sector has a total market cap of just over $28 billion, and the potential, which is projected over ten years, is over $10 trillion. So, the question that arises is why there's a need to decentralize physical infrastructure. The first answer is the barrier to entry. So, today, building centralized physical infrastructure is very expensive. We already have huge players in place; we're all thinking of AWS and Google Cloud, which come from the FEM (Foreign Exchange Management) sector.

We're building these ultra-centralized infrastructures, which are extremely costly. Regarding resilience, we're all talking about the decentralized infrastructure challenge. In 2021, AWS went down, and DYDX, which was hosted on AWS, was no longer accessible. The front end was down for several hours, so access was only possible via smart contracts. This raises the question of the resilience of what we call decentralized finance and infrastructure in general.

To prove that the cloud market is quite centralized, we have the market shares of the largest cloud providers and the rest of the cloud providers. AWS is first with 28% of the global market share. AWS and Microsoft Azure alone hold half of the global cloud market, making them the top five. Cloud providers today have more than two-thirds of the global cloud market, and if we only consider American clouds, that's almost two-thirds of the global market. So that raises a question about data independence and overall resilience.

If the Americans decide to cut something off, what happens? There's a phrase I really like to say: decentralization starts with the hardware. We all talk about cloud computing, blockchain, and streaming as if it were magic, but in fact, we have huge servers running in different locations that belong to people. So what's outside the scope of decentralization today are real- world assets and blockchain protocols. Indeed, protocols like Bitcoin or Ethereum can be considered decentralized infrastructure, but in that case, this SmartUp would cover almost everything related to blockchain.

So we're not going to consider those sectors and just come back to the... Ethereum decentralization is significant today; more than half of Ethereum nodes run in the cloud, and among these, 40% run on AWS. This is a sector that is becoming increasingly professional. We are seeing more and more companies that no longer intend to host their Ethereum on- premises or at home, but are instead using cloud providers. It's essential to maintain this diversity and independence among Ethereum hosting providers.

That was a brief overview of the Ethereum protocol and decentralization in general. The major sectors of interest—there are six—are: the telecommunications and networking sector; the artificial intelligence sector, which has been booming for the past year; the energy sector; the computing sector in general; the services sector; and the data sector. We will return to each of these sectors later. I will begin by introducing Exaillon, who we are, and how we operate. We operate so that we can then tell you the sectors on which we will focus.

So, for example, we are a wholly-owned subsidiary of the EDF group. We have our own infrastructure; we have two data centers in France and one in Canada. Our goal is to decarbonize data. Our link with EDF is that we refurbish supercomputers that they originally use for nuclear power and electricity. They renew their fleet of supercomputers approximately every three years.

So, we recover them, we buy them back, we erase the disks, and we can then refurbish them. We dismantle the HPC clusters, the processors, and turn them into servers. On these servers, we have built different product lines. The product line I'm in charge of is the blockchain product line. We will have major products, a Node as a Service product, so a Node as a Service platform, a data archive that is indexed on different public protocols.

So, we have the largest ones: Bitcoin, Ethereum, Polygon. Avalanche offers access to a free plan, and of course, paid plans for companies that want to scale. We are also developing a staking platform on Ethereum. This includes non-custodial staking, meaning you retain control of your assets and we only handle the validation, and custodial staking, where we can manage your assets like an exchange with a truly B2B institutional focus. Given our position as a subsidiary of a large group, we have a computing side, which involves performance calculations for AI model training.

This involves HPC clusters— GPU processors and graphics cards—that will be made available to companies, particularly in the finance sector. They will be able to use our HPC for financial simulations like Monte Carlo, and so on. Increasingly, we are also using it in the AI ​​and e-learning sectors. Here, we have be-metal servers and virtual machines available to companies, as well as 3D rendering workstations. So, initially we were aiming for the metaverse, which has really taken off.

Today, most of our clients are animation studios who need, through traditional workstations, access to high computing power and dedicated graphics cards for 3D rendering. They directly access the DAT Center. So now that you know that, we'll move on to the telecommunications, data, and services sector, which isn't directly relevant to us. Telecommunications and networks involve decentralizing 5G and Wi-Fi networks, where we'll incentivize retailers—people like you and me—to have devices at home. Helium is a repeater that can map the entire world to provide network access from anywhere, and each person with a hardware device will receive regular token rewards.

So there's also the data domain, which also applies to retail. There's a well-known company that... It's called iMapper, which is similar; you buy a device and can install it in your car. You can then create a map, an interactive Google Maps that updates every time someone drives over a road. Similarly, you're incentivized, so the more data you provide, the more tokens you earn.

So, the sector then moves into services. This is what all crypto VCs are looking for: the creation of a decentralized Uber, the creation of all these infrastructures, a decentralized Upwork. All these infrastructures, uh, for services that we're trying to decentralize, to give power back to the user. So, these are specialized companies that do this; it doesn't concern Exaillon, the AI ​​sector. We're starting to look at it, so we're already defining how to train models on our infrastructures and then be able to communicate and participate in decentralized AI models.

What's quite interesting about decentralized AI is that it's quite demanding in terms of computing power. We'll be able to connect a lot of infrastructures, which will allow models to communicate with each other, and we'll be able to compare certain models and make them communicate. So this is really something aimed at improving the artificial intelligence sector, and it's a very good example of synergy between AI and blockchain. The two sectors we're going to target, and on which we'll be able to have influence as a provider, are the energy sector and the computing sector. The computing sector was the first sector to appear globally in the digital age with Filecoin, which appeared in 2014.

It's decentralized storage. With Filecoin, we'll have a multitude of providers making storage available. Generally, these are cloud providers that will make several hundred TB, or even several PTAs, available on the network. When we want to store data, we 'll pay with Filecoin tokens, and the data will be replicated across six storage providers. This ensures long-term viability.

Storage, well, it ensures that overall, you shouldn't use storage providers in the same jurisdiction, but it ensures that no jurisdiction comes along and gets its hands on the data, and so on. It's a fairly attractive price, but the big challenge is attracting users. So, when we started talking to Filecoin at Exayon, it was early 2023. So, at the end of 2022, there were 14, or 15,000 PB, and 15 million terabytes available on the network, with 500 PB used. Knowing that each storage location is copied a minimum of 6 times, that means there were 100 PB of effective data on a network of 15 PB.

So, there was a huge amount of storage available, and Filecoin didn't struggle to attract storage providers, but it struggled to attract users. Furthermore, storage providers were only paid based on... of their storage capacity that was being used, so that means that as a storage provider you need tokens and you have to do business development to fill the storage space you've made available on the network. So it's a big effort with relatively little revenue because you might as well sell your storage yourself to companies. And where Fileecin really played well is that we see at the end of 2023 they reduced their storage capacity.

They had 1800 Pa available and they had 1800 Pa used, so we're at x4 of the capacity used overall. And it's in March 2023 that they released the FVM, the Fileecin Virtual Machine, to allow developers to develop on the Fileecin network and use the storage capacity to develop applications. So there we have a real use case for a decentralized storage network and we have real incentives, and that's how they succeeded overall. They made a lot of effort in business development. They notably have the University of Berkeley, which came to them, and so on, but it's really this virtual machine that was a real game changer.

They managed to use 20% of the storage capacity, which was definitely not the case before; it was 3%. Uh, yes, yes, yes, yes, yes, of course, yeah, yeah, of course. There was almost the same number of storage providers, less storage capacity per storage provider, but much more storage capacity actually used, and we saw a doubling in terms of users. And then what happened a little later in the cloud computing field was CPUs first, and then GPUs. So, again, with CPUs, there was a challenge in attracting users and providers as well.

When we started talking with Akash last year, which is—I do n't know if it's the oldest, but one of the oldest and best-known cloud computing providers, especially for CPUs—they hinted to me that there was a big storage provider and supplier. CPUs that were still supplying the network today, and that there was very little usage, and we can clearly see, it was around February/ March that we were talking about €50 per day being spent on the network for a valuation of several hundred million. And what really changed the game for Spine networks in terms of computing power was GPUs. So we see that they launched their GPU network at the end of August, so we see overall in September there is strong usage, a big increase due to several factors, notably the GPUs, and there was also a rather interesting phenomenon with the rise in crypto prices. Proof-of-work tokens increased significantly, and it became profitable, compared to the price of the cash token, to mine using the CPUs of the cash token while dumping the proof-of-work token.

So we see the big increase, it's due to that. So it's a kind of arbitrage of crypto prices using CPUs Decentralized, so it's quite interesting to see that the... the... the... the...

the... can arbitrage the prices of the pro fork tokens. There is, however, a caveat to be made regarding this sharp increase; we're at over $2,500 compared to around $50 a little less than a year ago. The amount spent per day on this network is due to a fad in the use of GPUs because we've had the boom in large language models, so GPT and so on. There are quite a few optimizations being made, notably by Mistral and others who are developing models that are much less demanding in terms of computing power.

So we need to be careful with this enthusiasm and this misconception about GPUs in general, where all companies want to have... and very, very powerful GPUs. Perhaps there will be too many GPUs on the market overall if we manage to reduce the computing power required by artificial intelligence and the energy sector. So, today, decentralizing production... Energy is very difficult, especially in France, so we're a subsidiary of DF.

Nuclear power offers unbeatable prices; it's a very centralized sector. There's a need—well, not a need, but it's very welcome—for state-run nuclear energy production, given the number of jobs required to build a power plant. We operate a validator node for a protocol called Energy Webchain, which will be a forum bringing together energy operators to develop use cases around energy. But the most relevant use case for us today is the free market for tokenized energy certificates. So what we're aiming to do, and what we're studying, is to develop a market between producers and consumers.

We can't tokenize an electron; it's not possible. We can't control it. If you live near a coal-fired power plant, the electron will always take the shortest path. We'll always consume an electron, which is very... Regarding carbon emissions, what's being done is certificates of origin, meaning that a company that wants to show it sponsors green energy, or an energy supplier that guarantees 100% green energy, will provide 100% certificates.

Certificate of origin, meaning I consumed 1 GWh, I buy the equivalent of 1 GWh of green certificate of origin. In other words, I sponsor green energy. This is something that is done today between institutions and large energy producers like EDF, and well, it's a large market between institutions. There's no need for blockchain. Where we might need blockchain is with small electricity producers, meaning people like you and me who have a solar panel on their roof, a wind turbine in their garden, or an electric car (because electric cars can relieve pressure on the grid at night while charging), and so on.

With a sensor, we can tokenize renewable, green, decarbonized electricity production and then feed it back into the grid. The producer can sell this certificate on a global DEX dedicated to electricity, and so we have a free market today. Very centralized, but it creates a free European market between producers and consumers. Thank you very much. So if you want to contact me, I have Telegram and Twitter.

I'm also present on Farcaster. Well, I don't participate much, but you can contact me. The address of our NURPC is note.ex.com.

Hello everyone, I know quite a few people in the room, but not everyone, which is rather good. So, Jonathan Chemou, the founder of Alf M on Twitter, you'll find me under the name Mocher. Why? Because when I started, I was Aon, and then I became DOX. So my real first name is Jonathan.

So, I won't go into too much detail about myself. For your information, I've reused a presentation I gave in one hour at the Bpi in January, so some information isn't entirely up-to-date anymore, but it will give you a good idea. When we talk about DIPin in the world of computing and Lia, and tonight I'll mainly be talking about the part... Computing, and really, it's important to see how it's implemented. We chose to have a network that's a bit like a Layer 2 or Layer 0 of Ethereum.

In fact, all the messages on the alf.im network, which is a decentralized cloud, mean that everything you can do on AWS can be done on the alf.im network. So why do we have Layer 2? Because everything done on the network is a message, and that message must be signed with a private key or come from a trusted blockchain.

All these messages are written once an hour to Ethereum, hashed, and time-stamped on a smart contract on Ethereum, a bit like a Layer 2, if you will. So, I created alf.im in late 2018/early 2019. We raised some funding and so on. Today, the network—by the way, those numbers aren't quite right anymore—has 100 cores.

Channel Notes, which are network controllers, and 600 computing resource nodes, which are the ones that will actually do the work, will be VM supervisors. We work with many companies like Ubisoft, Request, Synapse, Oasis, and so on. The main network we're on is Ethereum. We're compatible with many other networks, including all Layer 2 and, in fact, compatible chains, meaning EVMs, TeSos, Polkadot, Cosmos, and so on, and even C. The question is, why do we need a decentralized network for computing?

We want reliability, meaning we want a network that doesn't go down if a single provider fails, like what happened with AWS, like what's happening with CloudFlare, like what's happening with many others. So how do we achieve that? Well, we don't have a single intermediary doing everything; we have a marketplace where many... Providers will offer computing power, network interconnection, storage, and so on. We also want to ensure that everything we do isn't visible on Web 2.

0. That's somewhat self-evident; we tell ourselves we're not going to resell our information if we use it with Web 3.0. In fact, that's not always true, but generally, the problem with a Web 3.0 network is that everything we do is actually transparent.

On blockchains today, everything is transparent, so how do we manage? We'll have pseudonymity, meaning we'll have addresses that aren't necessarily linked to our identity, and on which we wo n't be able to directly link ourselves to that address. This is somewhat undermined with Mika, in particular, which will arrive in January. We also want to be censorship-resistant because, let's say tomorrow, a government says, " Well, I don't want DeFi anymore," so all the DeFi protocols on IWS will, in practice, be shut down. They'll lose hosting.

Nothing will change if they have their front end on a decentralized network. Maybe there will be a lot of companies that won't want them anymore, but there will inevitably be companies that will say yes, okay, we'll host you. Well, the network will redirect storage and computing to other machines. We want to be able to scale automatically, outside of a single provider. We'll also have lower costs because, in fact, when you have a marketplace, people can compete on price, whereas if you're only on AWS, you're stuck with AWS prices, which are relatively high.

We want decentralized governance for the reasons I mentioned before, and also for performance. I'll skip over all that for a bit, but basically, how does my network work? Acash will work a little differently, FileCoin will work a little differently, and so on. I'm talking about what I know and what I've developed. We have the Core Channel Nodes, which are the network's controller nodes, a bit like validators.

A proof-of-stake network—these nodes will be responsible for storage and the database. They will verify the signatures of all messages, check all the Hs, verify Unchain that the messages were indeed written, and so on. They will replicate them and verify that everything is working correctly. Resource nodes are the nodes that will actually provide the computing power, that is, virtual machines. So, when we talk about virtual machines, when we request computing power, we can request computing power in several forms: microVMs, which is what we call serverless, where we make a request, the network will find somewhere to run a piece of code, execute the code, and return a result; persistent VMs, where we say, "Well, I have a VM with 4 CPUs, 16 GB of RAM, find me where to put it on the network"; or we can even choose to say, "Well, I actually want this provider, and this is the one I..."

So, to host my VM again, it's a marketplace. We can choose to go to a specific location or tell the network, "Well, figure it out yourself." This is what's happening with my VM. And also, what we'll normally have later this year, I hope, so that it could be good for eTHCC, we'll see confidential VMs. That is, confidential VMs that will use what are called trusted execution environments, in this case, a little different from Intel SGX for those who know.

In fact, we'll use next-generation enclaves that will take an entire VM on which we'll encrypt part of the RAM, encrypt the storage as well, and we'll assign it to one or more CPUs. And in fact, nobody can access the memory and what's happening inside. Quite interesting if we want to store private keys, for example, if we're in the defi... uh, if we're going to have Keepers, for example... uh, it's actually this overcatch with Shutter Network, or in fact they They can deploy their Keepers, that is, the clients that will perform actions on the network for users.

These will be stored on a decentralized network, and for that, it's best to use confidential VMs. Well, I won't go into too much detail there. So, the idea when we talk about MICR VMs is that on a host, we'll have a cluster of MICR VMs running on a machine. This machine will be able to schedule many things, and the network will tell it, "You're going to launch this, you're going to launch that, you're going to launch that." Or, what we're also working on is that the user can have a peer-to-peer client in their browser using JavaScript.

They'll be able to find all the hosts and say, "Well, I want the serverless function to run as close to me as possible." And since the person has paid for computing, it doesn't matter where it's launched; they'll be able to go there. So, launching this microcomputer will give it access to pretty much everything, from external sources to computing, specifically IPFS and our blockchain. It will be able to retrieve data from Ethereum for indexing or even write directly to Ethereum from the microVM. Well, that's the roadmap; I won't go into all the details.

But we've been talking about Deepin, really focusing on the computing side. And I know that what's really hyped today is artificial intelligence. When we talk about artificial intelligence, we're talking about many things: text generation, image generation in general, with stable diffusion. When we're using open-source models, we're talking about text-to-speech, speech-to- text, and even multimodal capabilities that can handle all of that. In fact, within a single model, we can pass images and so on over the ALF network.

Currently, we only have CPUs; we don't yet have GPUs, but it works. If you'd like, we can discuss that further. There's a working network inference protocol that's CPU-based and very fast because we put a lot of work into making it work. Again, this one is a little outdated, but it gives you the idea. Before, you absolutely needed Assant and a lot of Veram, etc.

When GPT came out, we estimated that GPT3 and GPT4 have between 200 and 300 billion parameters for an LLM. That's the size of the Large Language Model, or even the Medium or Expert model, so there might be several, even in memory, running simultaneously. This means that on the same machine, you need many large GPUs, 100 of them, connected to each other with what's called an NV Link, for it to work. Needless to say, on a decentralized network, it doesn't work. The only ones with GPUs on decentralized networks today are Akashic Records, iIO.

NET, Bitensor, and others. They don't have CNV Ling re cards for one simple and good reason: if you reserve a single GPU, you don't want it interconnected with others because otherwise, data from one GPU to another, and even for running workloads on a single GPU, doesn't work well. So you can't put in those little things that allow them to communicate together. Therefore, generally, you're tied to a single GPU on a decentralized network. So, to run large Large Language Models, it didn't work.

What changed things a bit was that in February 2023, LAMA was released by Meta, and so we had the open-source communities from all over the world, and in fact, many developers who said, "Yeah, I'm going to be able to play with an LLM." They arrived and said, "Oh yeah, no, but I don't have an LLM, it doesn't work." And so, many people started trying, trying, trying, trying. They did what we call quantization. To say that rather than making it work in 16- bit precision, they managed to make it work in 4-bit, then 8-bit, then 6-bit.

If you look at an image, if the image is 16-bit color, it's perfect; if it's 8-bit, you start to see small dots; in 4-bit, well, it's the same here, except they managed to have something adaptive, which means you don't see the loss as much today. So we were able to use less VAM (Virtual Asset Management). It worked on small 4090s, 4080s, even 4070s. And in June 2023, there was Lama V2 which came out with a list that allowed us to use it legally if we weren't a FEM (Federal Electronic Modeling) user. So that was cool.

Everyone started doing what's called fine-tuning, where you retrain models to add things to them, and so we found ourselves... With models that could be decentralized, it's pretty cool, it's nice. Then we had Mistral, a French company that released new open- source models. There were Mistral fine-tuning versions, so now we're getting very small models that are very efficient, that work very well on a CPU, and they're even quite fast on the CPU. I could give you a demo later on one of the LMs that run on the ALF network.

Then there were many others to remember. This month, Lama V3 was released, and the AMA V3, with 70 billion parameters, well, GPT4 on most benchmarks and on many things. So we're finally reaching GPT4 level on open-source models, or rather, Open Source. So what does a generation of these look like on a 10-pin network? Actually, on the ALF network, we're going to use the Server part that we had earlier on the cache network.

It works a little differently. I'll explain how it works. On their network, we'll have a user who's on the interface. They'll make a request, and it will find a node somewhere on the network that will launch the VM. It will load the VM, load the model directly from IPFS onto the inference layer, which is the text generation layer, and it will start streaming the result to the user.

On networks like Akche, where they only have instances, they do n't have serverless, and they use GPUs, they have to continuously reserve GPUs for their AI because they have an AI chatbot. This also explains the $2500 daily cost they have today, which is largely due to this, because they have to continuously pay their provider to have GPUs on the network. Whereas here, it's on demand, and it will automatically scan. So yeah, well, that's what I just explained. We'll therefore have GPU support on the Alf network which is coming which will add a lot of features on this uh so a use case I'm going to try to see if I can make a demo for you if I have enough network I'll try it huh chatliber.

Okay, yeah, it works. So, actually, I have a model with 7 billion parameters called Neural Beagle. It's basically a tuning of Mistral. I'll say, uh, hi, I 'm giving a presentation. So, while we're talking, it loads a VM, starts loading the widening W model, starts the inference, and it gives us the result.

This runs on the CPU, so it's relatively fast because there's a lot of on- demand optimization. It can use several hosts on the network that are close together. If needed, it will prioritize the hosts with the best RAM speed. That's kind of the crux of the matter. We're talking about the physics infrastructure, and in this case, the crux of the matter for AI is RAM speed.

And when we talk about new processors with DDR5 RAM and many parallel RAM channels, we end up with speeds that are very close to the speed of... The VAM of the GPUs on H1s, even H100s, uh, so there you go. So, if we try it on the Lama 3 Instruct, the one we were talking about, which has the same level as GPT4 TAC, we'll see what it tells us. For once, it's not very well- written in French, so I don't think it's going to answer me in French, even if it will understand what I'm going to say, it's likely. Oh well, never mind, I didn't say anything.

Nice to get to know you, I'm delighted to help you, to discuss with you, blah blah blah. It's a bit slower, it has 10 times more parameters, so it's a bit 10 times slower. [Music] Uh, if you can read that, well, there you go. So basically, it answers the same thing: it's been loaded, and so after that, it's going to keep the VM active for the following users. I think that if we weren't too slow, rather fast, it's because there were still VMs loaded by other users before on a VM.

It can have multiple slots in parallel. We're going to do a lot of things. For those familiar with continuous batching and parallel processing, there are tons of optimizations everywhere. Okay, slideshow, I'll restart it. So, basically, this is another Dimo.

The idea is that what we generally offer, and what many professionals in dipin projects also offer, is storage. For those familiar with IPFS, we'll be able to do decentralized IPFS pinning, computing, document certification, timestamping, document history, and so on. There are many things we'll be able to offer, such as networks like Bitensor. They'll be really specialized in machine learning. The idea is that tensors, in the jargon, are processing steps in machine learning, which we use a lot in AI.

And so, there's a framework that It 's called TensorFlow, which is widely used, and so BitTensor has a bit of a presence on that name. This allows for the decentralization of the execution of these TensorFlows, and for creating subnets, and so on. It's really about computing with instances. We can say, " I want an instance with this many GPUs," and reserve it on ALF. We have much more versatility, so we have many more options on how we can use the compute.

Also, something that can be interesting is that when we have a network with many providers offering services, we can also be compliant with the GDPR. The GDPR will tell you, "Be careful, your data must be stored in the European Union with clearly identified entities," and so on. And above all, it will be necessary to say, "If tomorrow a user has personal data, we must be able to ask the node to delete it." So, the ALF network, and in this respect, certain Other networks, like Harweve, handle deletion requests because it's mandatory; otherwise, we can't succeed. So, okay, I'll come back to Confidential VMs a little bit.

I've already explained that. Now we're going to get to something that's quite interesting. It's a bit of a departure from the topic of Depin, but not entirely. It's the synergy between Web 3.0, blockchain, and so on.

So, it's cool that with blockchain, we'll have something that will be decentralized, with a native payment system, and so forth, versus AI, where we'll have something that's centralized, with large, centralized GPU farms, which will be extremely monopolistic and dependent on payment rails. And so, in the middle, we have a sweet spot that's linked to decentralization on Depin, and so on. On the AI side, we'll be able to say, "Well, when I switch to the model, I'll say, 'Ah yeah, this response'." There, it's good for... it's not good on an open AI, well, nothing happens, we just helped them on a decentralized network.

The decentralized network can say, "Oh, well, I'm going to start paying you if you trained the model, I'll pay you." And since we have tokens, we can do this kind of thing. Or, tomorrow, if we have agents... what is an agent? It's actually an independent AI to which we'll give a goal, a role.

And in fact, this agent will say, "Well, I want to..." Or rather, we'll tell the agent to do this task. The agent will have to consume computing power continuously on a decentralized network. It will be able to pay itself for its computing power. Let's say that we assign an agent to a DAO, that it uses the DAO's funds, that it can earn money for the DAO.

In fact, we have a complete cycle. We can have truly autonomous agents that don't depend on a human's credit card to function. That was the part Vitalque had done. I think that's the right term here. In fact, he saw four kinds of Li as a player in the game, or rather, it's just a participant in the mechanics.

There's an interface where we'll use Lia, which will do things for us. Lia is like the rules of the game. There's an A as a judge, an A as a DAO, which will make decisions based on what people really want in the DAO. And there's an objective there, basically. The idea of ​​the A as an objective.

There are many things we could see, but ultimately, it's the I, it's Lia, like we see in science fiction books. Well, the only way, in my opinion, that could happen is thanks to the Depine. Well, I'll spare you the rest. I won't go into it either. Thank you for taking up so much time, and if you have any questions, please don't hesitate to ask.

[Music]