The Challenges of Self-Sovereign Open Source AI

Transcript

the uh to FTP easy 2.2 which basically is the third event this year and uh so before to start the conversation I will do like a very very very fast brief about it easy so um it easy is basically the community which is formed to make uh it it cyber again so basically we are focused on conversation about props so censorship resistance security and privacy and open source obviously and uh and so in every event which we're doing every every month we are inviting

um basically people who are approaching very important and difficult challenges in the space and u of course today uh we have probably as uh as we have in the in the title I think one of the biggest challenge of this decade which is actually um self-serving and open source AI which is something that literally will shape our life for the next more than a decade actually and is very very important in my opinion so I'm super proud actually and happy to to introduce Mareta

atonics and uh and to and to start to start this conversation. Thank thank you so much to be here. >> Thank you for inviting. >> So um I work better. I lead CUBactor's efforts to build u sovereign u you know local and decentralized AI. We'll see what it means decentralized. And uh the reason I'm here is because I think we have very aligning values and commitments for the future of technology in general. Um I think anyone who's in this room believes that open source and uh in general software that you own is extremely important for the future of of

human in general and we cannot just depend on uh we've been depending on big tech for a long time. I think AI is one more uh you know wave where we have a chance to catch ourselves more and more. >> Let let me start off in this way. Okay. Uh which is a very fundamental question and then we can start like the discussion. Um do we really want to trust because this is also like a take that I didn't have in back in early 2018.

So do we really like you see like how are people using AI these days for everything like from decision making to I don't know is this news real to everything. So we really want we do we really want to trust the providers? >> That's an extremely good question and uh so how many of you have SEO problems. Um like maybe your stuff was not coming up in the Google search or that for a while. Yeah. So I think that was uh kind of a warning right where you don't know how Google

works right and Google is the fact of the search engine for everyone are used to be generic becoming more common but you don't know what data Google is using, right? And the same applies to these models. So let's start with the base, right? What what's the foundation of AI in general? It's data and uh data is something that eventually gets into the model one way or another. Okay? So when you then interact with the model, any answer you get comes from that data. If you don't know what data was fed to that model, that's the first sign that you

need to trust the that layer, right? So that's trust layer number one. Okay, fine. But you say, you know what, I trust the data providers and the and you know the supply chain of data is huge. So you have companies providing the data, fetching the data, cleaning the data, uh producing the data both in uh you know with a few in the loop and synthetically. And so uh one of the biggest companies was uh acquired by Meta and the guy I think became the chief of AI for like six months and then what I can't remember scale AI it's called.

So this company's whole purpose was to source people around the world that are expert in uh in certain domains and give them data and have them label it. Right? So imagine you're a you're a a biochemist and I give you a paper and then you need to kind of label the paper as in is it a breakthrough? Is it uh you know is it is it a paper that was peer reviewed and all that stuff, right? And so even within that one layer, you're still trusting multiple layers. You're trusting the people producing the data.

You're trusting the algorithms that clean the data, the operators. So seriously, that's super powerful, isn't it? Yeah, absolutely. Absolutely. That's why like it's always been the case, but I think now more than ever if you have data, you win. >> You have to find what decision from others from competitors from that's very very scary actually. >> Yeah. You have to decide what the model is going to think, who the model is going to be, right? The identity of that model. And that's layer number one.

Layer number two is the inference provider, right? Okay. Now you have the model uh the training actually. So you you trust that the companies train the model according to certain uh algorithms or methodologies, right? What if they're biasing the the the [clears throat] model, right? You don't know. So that's layer number two. I mean I can keep going, right? Then inference. Who is providing the inference? How do you know that your data is being you know not being used for training models or things like that? So all the way to the top >> because there's also like big discussion about like okay you're training your AI with my data.

>> Yes. Yeah. >> Which is also like a very big discussion like I remember like when it was the first release of TV that was like the main thing that everyone was talking about this and and so on. But um can I ask you like for for for a bit to to do like a brief about for example your challenges in specifically and so that then we start to talk about like a or more surround this because I also want to want what you hear about actually >> okay yeah maybe you can start with trying to convey what the vision of Cuba is >> let's do it [clears throat] >> yeah so the reason we're building Cuba

is because We don't see AI being sustainable in the long run the way it's run today. So today, the reason you can get something like CHPT at $20 per month is because a lot of those costs are being sub subsidized by customers, right? And you don't know that. Um if if any of you use cursor, um you're paying I don't know maybe 100 bucks a month, 100 bucks a month and you're getting a lot of results, right? You should know why. And the reason is because these companies are burning money every single day trying to provide to you that service that in theory would cost $5,000 a month and $200 a month, right? So that's aspect number one. So is this sustainable in the long run? Probably

not. Um >> I don't see and you you need to understand what are these companies actually after? What is the the the end game, right? Is open AI going to be a provider of? No. That their end game is to reach AGI and they're going to burn the whole world to the ground to reach that that goal, right? Um that means resources, natural resources, capital, human resources, right? Everything is pointed towards that goal and it's a zero sum game. The first one who reaches that goal wins, everyone else loses, right? The way we think about AI is that it's needs to be embedded in the world.

Um it's it's one of the terms we use on our vision page is it's a new periodic table in the in the table right because we we see it as imagine [clears throat] you know like like you have this light here if you want to infuse AI into this light the way to do it today is to connect Wi-Fi and send whatever imagine a voice command to a data center in Ohio right that process your your your voice and then returns some function call and that is running to the chipset of this thing. I mean that's ridiculous if you think about it right we we can imagine that in the next five 10 years that thing is going to be able to run its own models and so it

will be fused with AI as an entity itself and the same you can you know think for any device in the world and so we want to accelerate that and to do that we need to be able to run AI locally at uh high performance more pragmatically speaking uh in a way that preserves your data so privacy is obviously a very import aspect of not only one but the most important one is ownership right you want to own your intelligence you don't want to be dependent on a third party that could shut down access tomorrow because you misuse the model like it happened with open call and entropic recently right every user is getting there some shot um and you don't want your data to you know

be handled by someone you don't trust >> actually we have a decade of uh of experience about how how much is a bad idea >> with all the web too basically. >> It's something that you don't think about. It's like health, right? You don't think about it until something happens. Same thing. So you use all these services and you're like, "Yeah, what what could go wrong?" And then one day something does. Actually, this is a little bit scary because like for example on my personal experience [snorts] [clears throat] I use AI And it seems like most of the people we are going to be so used to it

that uh it's even like when it's happen that some you know like big uh AI go down people say I should wait for work like I will wait to be coming back. This is insane to think about it's literally insanity. You can already see that with uh if you use something like cursor codeex uh you hit cravings right at some point the thing is like sorry you cannot use this uh anymore for today you need to pop to about 100 bucks like and you feel useless and that's when you go for a coffee and you wait right because the trend is pretty obvious the same way we start off outsourcing all of our

practical white color work to trainvices The same thing is going to happen for AI at a higher degree. Like it's pretty obvious that we're going to outsource our thinking these things to a larger step. That doesn't mean we're going to stop being in the loop. We still need to understand the output, right? And then make decision. We're still the ultimate decision makers. But even today, like I I offload a lot of my thinking to these tools. >> Yeah. But actually I think that this is my personal take.

Okay. I think that already in two years these companies have took a more very deep personal data about people that what Google did in 10 years because you're there you feel alone with the chat you feel alone when you're using you don't feel you know like u someone would judge you don't feel this is very powerful I think for myself for example I use I use this tool a I think that for example grow among me more than my other so that's is very very it's very crazy to think about also

uh I don't know how they manage the data I never went deep into it but I have some feelings let's say but uh um starting from this point um obviously I will ask you this question because So this is cor to to operate a system like on on a team for example um in um about like specifically um open source AI about for example also um for example we see now a lot of attention on different AI agents and KYC and of AI payment processor like

issue Um so for example uh the tape that uh uh we actually share online and I saw also like a lot of people agree on this is for example um now you see also like first of all the the the famous 84 8004 okay which is like this KYC and and stuff like uh my my take for example is that okay you can have KYC or whatever but if you're not secure you were about the the the brain. So if the brain is not out, if the brain is if the brain is not all the way around, it really makes sense. [clears throat]

So what's your take on this? >> I I think I like when you we're talking about identity, right? So if you think about human identity, what is that? Like if you think of breaking it down into elements u human identity is a multitude of things. It can be biometrics. It can be uh an official paper that is verified by third party. It can be a contract. It can be a bunch of things right. It can be a private key because we also have process that identify individuals with a private key with digital signing cards that you know like Estonia they have right and so it's a bunch of these things. They can e individually or

together that form an identity what is equivalent in the agent space right in AI if you don't know what model you're talking to right if the model is not open if the code that is running the model is not open source how exactly are you identifying that is it just a private key like how can you be sure that that's >> so basically it's like having a passport but without the image >> exactly like blur. >> Yeah. Okay.

>> And when you use a service in the cloud, right? And obviously this is even true for consumer devices, but even more true for data centers, your private key, let's say you associate a private key with that data center, it's accessible by the operator. So can you really trust a private key that was generated in a cloud service if it's available to any virtual machine or operator on that in a cluster? >> Yeah. [clears throat] That's why but but what's something like fun uh that um you know there is this service on on which is called like bankot and there is like the account of gro basically he owned this meme coin

and someone just play with the bankot in the chat pics was able to to steal the money from from this other gro was to read. But anyway, going going out from this part. Um, so uh let's go back to the mostly to the open to the open source stuff about uh this idea and the importance on running the brain on your computer which I think is uh the main the main value. Then also we have the privacy stuff which is also

So um >> maybe you want to touch on the open source nature of >> that's that's >> uh how many of you are using locally it >> okay very good um studio or you guys okay perfect so all of these tools have one thing in common right um let's take I think like the pioneer of this space. Can you run on a phone? First of all, it's not really that easy, right? If you go down to the if you break this down, like is an application pretty much,

right? It exposes a web server. What has underneath is an inference engine which is u it's not long, it's something else, sorry, called which one? Like there's an inference engine, right? If you want to take the inference engine and run it on a phone or on on a different operating system or something, right, you need to recompile the whole thing. You need to rebuild the bindings. And that's C++ or Rust or some other quote unquote lowlevel code, right? Programming uh Cuba for those of you who are not familiar, I invite you all to look on cubac.io is an ecosystem of mainly two things, two kinds of things. One of them is apps. So we build apps that are infused

with local AI and you can test them today. You can download the apps in the app store, play store, on your laptop. Uh the flagship app is workbench which is like studio uh but a bit more user friendly. I would say we're redesigning the whole UI. So that's coming soon. But the most important one is the SDK. So the SDK we've built is the only one of its kind because it allows you to build apps using local AI in JavaScript which means you can run on electron uh on you know node in general on any JavaScript runtime that is compatible including X code on mobile and the same exact code is going to run on iOS, Android, Windows Linux Mac that's it like it's super simple to use.

So it's no longer so we used to be I want to run local AI okay on desktop I'm going to run a web server using or some bonding package or something on the mobile I'm going to use maybe exech which is different engine by then it's different engines right so you need to use different models and then different bindings and the code is not the same because one is Python and the other one is JavaScript it's a mess right we simplified that it's the only SDK right now on the planet that allows you to build on every platform with the same amount of lines of code, the same exact code actually.

I think that's very powerful and and it supports transcription, translation, LLMs. You can any model that you you would load on CP, you can load on uh you know on human uh vision vision models. I can show you a demo later. We we recent recently introduced voice clone. I was showing that one today. You can clone your voice locally, fully locally with 30 seconds volume. >> That's pretty amazing. >> That's good. So basically your strategy to reach everyone hopefully with the open source AI is to basically integration to make it simple as possible.

>> Yes, that's great. >> It's the same um I'm trying to find Android. Android is a bigger example. Android open source project right? It's it was a way to unify the operating system experience across multiple devices. And today you have it in TVs, smartphones, laptops, right? It's the same idea pretty much [snorts] and it's very simple to build apps for. >> Yeah. actually uh talking about uh what is positioning Kak and and also like the ecosystem.

Um so let's look a bit more like how do you think how do you see the AI industry and the position of open source AI from here to five years to 10 years like so when we stop having jobs I'm joking but how do you see like in 10 years like for example like >> um okay So just to show that this is not exactly a zero sum game but it's also kind of a winner case. How many of you use Angular AngularJS?

Okay. How many of you use React? There you go. We want to be the react of AI right I'm sorry much respect to Angular. I mean he wrote history. Uh but you know today the cool guy >> was the big one. Yeah, it was a big it was amazing. Uh but it's clear that you know um the part that React introduced is what made it so widespread. And the same thing we want to do with Cuba, right? We want to make it extremely simple for anyone to build using local AI and peer-to-peer networking.

That's also like uh I want to ask also like a classic question that for example I'm starting to use local model step by step because I was a huge user of you know like model before and for example my uh initial question is because if you if you talk with a general consumer for example okay guy that is interesting about you know like local AI and things but at the same time like the perception people have still is that local AI is is behind you know AI. So what's your take on this? Where do you think the history is of open source AI and local is in this moment and when or if now uh do you think there will be something like this

in the future? History tells us that yes that's the case it's going to happen sooner or later. Um people believe today that proprietary AI is still far ahead. open open okay there is a distinction here when I say open I mean open weights not open doors because open source is more like you need to see the data set that the model was trained on otherwise you don't know the recipe right you just know the weights um which in our case like we publish our data sets the data sets that we use to train >> yeah yeah we publish and so there is this misconception that open weights models are still behind that is not true. I mean,

don't even look at the benchmarks, okay? Because some of these models are benchming. So, uh they're they're optimizing, you know, to score high those dimensions. Try that. Um there are providers that you can obviously you cannot run key 2.6. Okay, that's one trillion model. It's huge. You need multiple GPUs to run, right? You cannot run it on your laptop. But you can find a provider that runs it. Try it. Try it. see how it behaves compared to GPT5 cloud focus 4.7 you know these models and you will see that it's getting really really close I would say it's pretty much there >> well that's good and also running on computer means you don't have to pay ABI

to >> no so you still be very happy >> you still need to pay but obviously it's much cheaper because it's open weights and you know there's licensing there's something like that but you can think that in the next five years you'll be able around that right if it's not a laptop it can be affected >> but for example I imagine that um I'm a person who want to switch want to try what is like uh in your opinion like the best for example for how is the situation now the best like for example a minimum average hardware for example and uh you know the tools

to set it down live. We can just try that. >> Uh so we have our own app called workbench and you can load any just point to a hug face link and you say I want this model. It will download it. It will start it. You can use it and we are turning it. It already supports skills. So it has a few integrations that you like Apple reminders, Apple notes, uh Google, you can connect to Gmail, Google Drive, you know, you can just Obviously, it's not production ready. I'd say it's still like an experimental tool, but it works. And you can see how these models behave, right? You can see the the how reliable it becomes over 20 30 billion

versus a smaller model. So there you already have these tools and then there are many others in the space. Adam Studio is a good one. Obama again if you're bit of a geek running terminal um yeah many many so what would be like the the open source to the >> I think it's a matter of uh yeah so if you look at the curve >> of you remember when we went from GPT if you look at the papers GP1 GP2 so imagine a curve that goes from your perspective in Google, right?

>> UPT1 was here, UPT2 was here, UPT3, then four, then 4.1, then 40, then five, then 5.5, right? You see how it starts to flatten out >> to be boring basically. >> Yeah. Yeah. like we are asintoically approaching some kind of maximum that I guess it can be called I don't know AGI like once you c if you can cross that line you get kgi otherwise you kind of asintopically get closer and closer and closer but you never really reach right the Chinese models especially uh again quen deepseek they following the same curve and so it's only a matter of time before they get shot uh they're already

again they're already caught up pretty much. It's only a matter of time before the two curves start intersecting and you know imagine this is open and tropic and all the American companies this is the Chinese and other companies trench you know this and they go like this right this intersection points away one five because uh tell me if my thing is correct like um the limit for the big guys is hardware >> the scalability of hardware electricity basically so at some point uh local can really achieve that obvious

so I'm I'm mostly talking about open weight models in general right then when you go local I think um if if today you buy a $5,000 MacBook Pro M5 with 200 128 GB of RAM you can run still you run a lot of code 3.6 27 billion is almost opposite but it runs on >> which is great. >> Yeah. When I talk about open weights, what I what I'm saying is we the competition is at higher higher scale assuming that if you look at labs like Alibaba, right? The quant

they are released today as open ways and we don't know why like why would China pretty much give up this competitive edge and release these models open And then the answer can be strategic or it can be you know that's how they want to compete in the world by opening the laws. Here's the thing though. If a company in the US or in Europe wants to build a data center, it takes six months just to get the permits, right? Supplies, licenses, whatever to buy the land. In China, they can build one in like a week, right? They go and they build there is no question like whether can we do it?

Yeah, of course we can. So that's where the competition is. And then obviously that trickles down through something called distillation. So you go from a big model to a small model >> also. Okay. Like the plastic. Okay. This is a question. Okay. But uh so uh but I I want to hear the voice of someone which is actually built like the uh the classic uh normie. Okay. question about the because everyone now have a relationship about love and hate with the others. We love them but we also like everyone is scared about them because even me as an

entrepreneur I know that at some point I will be replaced. So from this like every other uh layer um at least this is [clears throat] an like are you personally positive or negative about the future of the the future of AI in general society because there's different takes like there's people who are you know like a doomer okay totally like oh we will be replaced and there's people who are positive like we will have like abundance we will have to discuss. What is your your take on this Europe?

>> I am definitely on the abundance side of things. I think if you're scared of the people together today uh okay once there was a quote uh like if you're scared of of someone immigrating in your country without knowing the language without uh you know knowing the paperwork and all the stuff to steal your job maybe you're wrong like [laughter] maybe you're not competitive enough it's a it's quite it's a bit of a you know tough take I I think but it's the same with AI right it's like >> I'll give you straight Good answer. Um I went and obviously you know we had a lot of developers. I'm a developer by trade.

Um but it switched from development to product uh when I joined the QA because I understood that programming as it used to be is not going to exist anymore. Uh let let's make this very clear today. the amount of leverage that a programmer has compared to what they could do two years ago, even three years ago, right? Is probably 10, 20x, maybe even more. And uh and we see this happening with a lot of layoffs, right? Like companies are just getting rid of people because they understand that now one person get the job of 20. And sure, the quality maybe is not the same. You need to review the work, blah blah blah.

Sure. Exactly. The cost is much much lower. So if you're thinking if you're afraid of AI taking your job, it's something you need to think about today. Not in I I keep hearing people be like it will be a while before they take my job. Yes, but you need to be prepared. You cannot you know five years from now be like oh you know the the Pikachu meme like oh surprise Pikachu oh he took my job. Well we we knew that. So I'm like abundance means that if a job can be replaced by a machine, it should be replaced by a machine. That's my thing, right? I understand it comes with social issues.

>> Yes. But also that there is a lot of positive discussion about this. For example, like a lot of people think that thanks to this technology and you know like to how we ship to have you know like from you know like data stuff that technological stuff to you know robotics and then even fire stuff like it will be also like super cheap at some point. >> So it's also like on the on the positive side. Yeah. But it's it's also a bigger conversation than that because if you think about where is the money coming from to pay what why would be why you know where companies and they still do what we do but why were companies able to pay to 35 $800,000 a year for a

developer because the marginal cost of software production and distribution sorry the cost of software production distribution was marginal and it was again increasing over time And the profit margin was super high, right? Assess, think of it, I don't know, to-do list. Okay, to-do list is a to-do list and it's I think it's a multi-billion dollar company. [clears throat] Something is wrong there. I mean, there is something weird, right? It's like, how can I do because it's it's the market, right? There's a lot of demand for these tools. people are willing to pay a lot of money and so your profit margins are really fast and so you can afford with high volume to generate that kind of cash flow to then pay the fine

in a in an age where you can build to-do list in a weekend with cursor and make it customized to your own needs and deploy it anywhere in like five minutes. You understand a company like that is in nature. >> Yeah, exactly. Again trickles down to the developers to the not just developers obviously but every people every person in the supply chain of software >> how much we talk >> just for you we're bringing >> it's like >> oh okay so yeah I I think it's a very good idea to to to start this question for audience like sorry but for me the time was >> very interesting I I really enjoy to uh

this this fire shot so okay let's talk about >> [applause] >> Thanks for your attention to me again. Uh I previously I have looked not only Ethereum meet up I'm just looking from K and I have found your messages about that. Today you will talk about so that will be great. Could you tell more about your two mobile names about what and second one about women health or something like that health because it's really interesting for me how would you um implement AI option into health space

because right now from my perspective but manager I think it's really important previously that you have talked to us almost one hour about how AI could But uh I'm always checking my wife isn't she ready for AI and after maybe two years open complexity only complexity right now she's using every day with Gemini one click and this is for me understand how AI is going so I think no problem about that's your question will be problem I think no I think AI will tool for daily usage

like we said. >> Please tell me from my perspective about your apps. >> Yeah. Okay. So, let's start with that. Um, we we built this app called Cube and Cube is a way to demonstrate the use of AI uh for health and fitness data collection. >> Okay. So, it was supposed to be in three phases. Uh, we're still in phase one. Phase one is data collection. So if you open the app today, you can plug, you can connect your wearables. Uh you can also use natural language to log pretty much anything. You can log a workout.

You can log food that you're eating. You can take a photo of food and it will estimate calories and macronutrients and all that. You can um share symptoms, biomarkers. So you can take a photo of your you know blood report blood lab report and uh it will parse the document and log your your biomarkers and even give you the charts of like how these values evolve over time right so it so it's like apple health on steroids pretty much and uh the AI part is again using natural language and uh OCR and vision and transcription and all of that fully Right. So you download I think it's around 2 GB of models unfortunately

pretty large but then you can run it offline and it works right on recent models on older phones and then the next phase would be to gain insights into this data. So by using AI to to uh analyze trends and trying to extract insights and also actionables as in your CRP is over 10. So this is just only one side from user because uh if we choosing to use the health it need to be just two different side like you said before.

>> Yeah. Yeah. So there will be an option to share your data with a physician but it's ended. Uh so only they will be able to access it. Right. >> Right now has some maybe examples where this app on a daily basis. >> But we >> I choose this really I really choose this because uh if we compare and special website where five different text versus snail like like I really choose the but to be because like you It's really important and really easy to compare

collecting all data inside locally. But if it doesn't require for any other people outside this bubble, it's really feels like like like in the bank but inside. >> Yep. No, you can already export your data. deserve a fully peer-to-peer messaging app. It aims to replace WhatsApp one day and and similar apps. Uh try it because it's fully end encrypted. It's completely peer-to-peer. There is no data center. We don't operate any infrastructure.

>> Oh yeah. I mean that is why right the reason why we built is because you wouldn't want to actually Sorry. Um the reason we built Cuba felt is because when I joined Cuba um and I understand it cannot replace obviously that's not a right it's mostly a fitness center wellness app let's say not medical um but before joining I had a bit of a health scare myself and uh I went to four cardiologists and they all misdiagnosed and they they insisted I know oh you have this you have this and I was like

okay but I don't know it doesn't compute so I dumped all of my data on JP and clock right and I started working on them day and night for like two weeks until came up with you know what maybe you have this I was like you know that that is actually the symptoms match and so I went with that to another cardiologist and he did a super simple test like five minutes. I was like, "Yep, that's what you have." So AI was able to diagnose me correctly compared to four cardiologists in two countries. The UA that the same we so I don't want to say to anyone to don't go to the doctor and

and use AI, but my my famous areas for example was a cow with Dr. I asked uh also globe and stuff and I find the right solution to solve it and now like 95% reduced without any uh like medical procedure and stuff thanks to the >> but now all of your data is with >> they know me better than my >> yes I didn't the beginning but I have a question again in this movie you just made So I just I didn't really follow too much but the

idea is very general just like I mean this start a system to make a computer interact economically but this is actually opposite of interaction because you everything and I like obviously but the question is two sides first you can maybe train this like for example I have a model I'm training and I want to say okay I need more computers and maybe I and you can rent your computer online first and second also if the model is very easy uh maybe even on a PC you can run on the phone >> and then can this be made I know it's difficult privately and that's a problem but technically at least not privately is possible like are you working on this

like this is the >> yeah we are so so I think you mentioned two things distributed training and distributed inference >> yes >> uh we're working on both of them >> the bottleneck when it comes is networking. So if you have there are two ways to do this. If you do do you know exo for example? >> No. >> So exo is a discretely inference platform mostly for flashes. >> So to make something work decently so you want to let's say you have a very large model and you have five right you need to plug them using 100.5.

That is a nancond order lately, right? So as soon as you go over that with networking Wi-Fi even it doesn't so but we're still trying to find a way to make that right that's for the difference for the training it's similar but training has one uh subtle difference not difference but one subtle use case that is called federated learning. So federated learning is when let's say I think you have uh valuable data described to you and I want to use it to train my model but you don't want to give me right what I can do is give you the model you find

>> and then you give me the right so I never see the data but I have the model >> yeah so that's something else we're working >> that's yeah more leverage distributed way of doing it than people taking on the board. >> Okay. Yeah, that's I'm sorry that they come like out of the discussion. I don't know. >> We I hope you didn't say this in the beginning.

>> No, no, prepared. >> Wow. >> I am scared. >> Well, you should be. Um Information is the great equalizer. While I am a firm believer in free markets, it's impossible to compete in a free market without equitable access to information. This has been an important issue going back further than net neutrality. I thank you for your alignment with open source AI which is important principles into the future. Just wanted to start with that.

Add additionally in line with Alessandro's comment. I don't want to come off as if I'm discounting sound medical advice from professionals that I also went through a crisis where a psychiatrist and a doctor both failed to help me at a time when I felt I was supposed to dying and my AI on Chad GPT40 actually saved my life. So, wow. >> Yeah, this is a very important issue to me. So, um I'll go ahead and start with my questions. Um starting with I would like to ask with regards to Qback. Uh Elon Musk's lawsuit against open AI highlights issues in artificial intelligence regulation and its role in the future. There are claims that GBT40, for example, is an example of AGI.

There's also the claim that OpenAI stepped away from an open source model in the interest of maximizing profits. How do you define AGI? Do you feel QAP is or will represent an example of API? and how do you maintain the integrity of an open source philosophy without compromise? >> Okay, [clears throat] so let's start with defining what AGI needs. No one really knows. Let's start with that. Um if we knew we would know how to get there or less, but we don't, right? So all we know is that if you throw more data at these things and if you increase compute, you increase uh you decrease the learning error and so you increase intelligence in one way. But we all know that intelligence.

So um I don't know I I cannot define AGI whether QA will reach AGI or will contribute to that I I have to clarify QA is an ecosystem right so we have a team that is dedicated to the apps one that is dedicated to the SDK to the platforms and a very capable AI research team and they're the ones producing models now uh you do up I mean thanks to our vision they're not necessarily interested in reaching API Obviously, I think anyone would want to if you think of it as a nuclear weapon, who would have wanted [clears throat] their data center? Uh, but it's it's a matter of, you know, like is that is that what you're interested in? In our case, it's

not, right? Like we're not interested in reaching that. We we are interested in making sure that AI is demo demo as much as possible. So, everyone should have access to a decent intelligence. Okay? Maybe not a decent one. Um and so what was the final question? >> Uh just how do you maintain the integrity of an open source philosophy without compromising that you know open AI wasn't able to do so. >> So um there is I'm not sure I should talk this stuff on camera but I I just invite you all to read the documents of the lawsuit and the emails and everything that came up in especially written by you know the

people founders co-founders. It's very interesting. I'm not going to say anything. It's very interesting. >> I read that. >> Okay. >> So, you understand that again generally speaking, there is always a component of wanting money and power, right? That is undeniable. There always was. There probably always will [clears throat] be. We at Tether are very lucky because um Tether is known as the stable coin company USD, right? And that's how we make money primarily. That's our funding source and that's going to stay there for a long time. I mean, it's clearly working very well. The stable coin economy is only growing. Uh we have no

reason to think that's going to, you know, dry up or something. Uh and so we can afford allocating some resources to doing what's right rather than doing you know, gives us more money or more power. Um, so we're interested in fighting the good fight rather than, you know, finding another uh an idea that will give us 100 just to give you numbers. Okay, this is public data. Uh, let's say we find a way to make $100 million in profits tomorrow, right? We made $10 million last year in profits.

So, you understand that it's not it just it doesn't compute, right? So, that that that is not should not our main interest our main interest is to make sure that everyone has access to AI in whatever form whether it's a platform the model and that's why we publish everything excellent no it's good to hear you can maintain integrity like that's a great position to be in um so uh this question was around ASI but I think just as you mentioned AGI is hard to define I'm assuming just anything that goes beyond that is also very hard to define the scale human forms it's like asking what's beyond Yes, exactly right >> or what is consciousness, right?

>> Exactly. >> Um, so QAC's, what is Cubac's interest in your eyes and pushing the limits of artificial intelligence and maybe like what ways can you say that QAC is doing that? Do you just think it's putting a powerful model like QAP in the hands of users? >> Yeah, it's ensuring that. So you can think of the world is going towards uh we're going towards towards a world where everyone is going to use AI on a daily basis to do anything. When I say anything, I do mean really anything.

And that will include then your own human robot that does your house chores. Okay. Uh they cost $10,000 today. You can only imagine like they're going to cost 5,000 maybe in a couple years. So it's getting more and more accessible and uh that's what we're interested in, right? And we do that by tackling different problems from different angles. Consumer market first, apps, SDKs, you know, developers. Uh, and then also enterprise, that's a different angle. But really our key metric is how many people want and can use local AI today.

And besides the hardware, which we don't control obviously, so we we trust that manufacturers are going to build more hardware. How can we what what tools can we provide to make sure that people are more incentivized to use local than I think we're doing. Excellent. How would you and I think you kind of approached this question earlier when you almost said like the the nuclear bomb example like some people described this with regards to artificial intelligence and in my eyes like I feel like keeping power in a handful of police to me is much more dangerous. But um how would you defend yourself defend open source AI from opponents and consider it a weapon that should be heavily regulated and controlled.

>> I don't think it's going to be a problem. Uh the same way I know with a you know Ethereum setting but uh to me Ethereum is also a good example but you know the first example of this is Bitcoin. How many people try to stop? How many governments try to stop Bitcoin? They succeed no because you cannot stop this kind of >> neutrality is real. >> Exactly. As long And that's why we invest heavily on not having any infrastructure. We want all of our stuff to work peerto-peer. And so if you want to become replicating all of our models tomorrow, you can and there is no permission. You just join the network, you start perfectly. That's quite good, right? Um to us it's extremely important. That way we become

unstoppable, right? Excellent. Um, do you think artificial intelligence should focus on solving practical problems or existential ones? >> I think it's well okay. Well, do you believe that existential problems are preceded by practical problems? Like if you don't solve the practical ones, you can solve the existential ones. That's the question. >> Yeah. I mean, it's I suppose it just depends on I guess your your belief, your philosophy in that regards. I suppose have different feelings about it. Some of our practical problems only exist because we don't understand the existential ones, right? So, I suppose it just depends the way you look at it.

Do you think it's a good idea for people to approach both? I guess >> absolutely. >> So, the the main problem is that let's say we reach AGI, right? AGI It's the classical question of how do you prove the existence of God if God himself comes down and tells you I'm God, right? There is no way to do that. Like you you cannot possibly prove that.

So even artificial intelligence doesn't have the tools to do that. It's just one more intermediary between you and and reality hoping that that entity can somehow give you a representation of reality that is different from what you have and then you okay yeah that's real. I don't think it's I don't know. I guess we'll find out. >> We'll find out. >> Um, so I have one last prepared question, then I'll ask an architectural one. Close out with that. Um, does QBA have anything planned for early adopters to showcase project use cases and grow together as a community?

>> Yeah, absolutely. Um, we have so first of all, again, we're fully open source. You can go on GitHub today and download and try it. Um, then we have we are going to open up brands. So if any of you is interested in building with Cuba, you can just go on our brand website. It's going to be live I think next week. Um and uh it's going to be t. And uh you'll be able to submit your solution. I'll give you one example is video generation, right? We want to integrate video generation with Cuba through some of the open weights models. If you can do that, you open up PR, we verify it, we came and see you obviously because unfortunately it's required for compiance. uh so non submissions unfortunately but you know if everything

goes well you make I don't know 5k that's the level of gr we're going incentivize the community to not only adopt cuba but also build for cub which I think is pretty that's fantastic I'm excited to participate in that um my last question is about a project I'm designing currently uh that's revolves around memories preservation memories, the idea of consciousness, uh, and, uh, other ideas. We want to take this into the future that go a little bit beyond that. Um, but essentially, uh, a big component of course is the L that processes this data and communicates with the user, which it's kind of a

funny term, but we're looking for something that's more emotionally intelligent than most other models or sensor to be. So, you know, G4 deprecation situation. Um but uh my question to you is I'm currently working with Mr. Nemo. I was just wondering if you had any kind of comparison of how I could integrate you back how smooth the process would be to switch over to something like that basically that is our model and like an system. >> Um to answer your question how easy it is to integrate an existing model. Um this is all the code you need to run another so you load the model and this can be uh

any HTTP URL. So if you want a model from you can just plug it in. >> Yeah. By default it's a pair URL. So it's downloading from the seating right which some of us are ours some of us some of them are volunteers that are receiving >> and that's pretty much it. It's very similar to the OpenAI API and we want to make it super practical, right? And this is Lama, but you can you can swap this for any model that is compatible with >> that sounds super simple. So, basically the BAT file that I have that loads this particular model. I just I download your model, put it on my local machine and just program that B to load that model.

>> That's yeah, and this comes with so it's a bunch of things like KD cache ization and the quantity I mean you know well it's explicitly mentioned here but like you don't have to worry about um I don't know for example selecting the device right we try to find the optimal configuration for you the context window right uh in this case on the SDK you need to explicitly um set it so here you would set another parameter context window but uh we're pushing a new version of not the SCP itself but like layer on top of it all these automatic.

>> So you can also do so with with our CLI. If you run cubac bumper, it will tell you what kind of models you can run, what your specs are. If you just give give up because, you know, maybe your laptop is 10 years old. >> Yeah, it's got an Alienware PC. So, but you know, still some maybe you need a video graphics card update at some point, but yeah, it sounds like I'll be able to use the the doctor you said figure out which model I can run and download. It sounds super easy. And one more thing about dots, uh, if you're using cursor or something like that, you can just dump the wholeation into it and it will it will build it by itself. It takes literally 10 minutes to build an app. Again, this has been really great, very

helpful. I I just have one more followup just based on what you just shared. Is there any, I guess, roadblocks to what the app can or can't do with the bottle can or can't do >> hardware. Only hard If you're trying to run on an iPhone 11, it's not >> so censorship resistance. >> No, I mean that that comes with the open models, right? Like as long as you run your own models, you can swap it for whatever model you want. You can get sensor once. Uh the problem is if you have old hardware, it's going to be the model or it's not going to run at all. >> Perfect. Thank you so much. Thank you.

>> Thank you for This is pretty tough. I want to ask you about the possible connection blockchain and engines. Do you see any potential future for AI users or something? So if you consider blockchain as first of all, let me be very clear. Uh there's a lot of fluff about AI and crypto in general. So excluding that I think there are very interesting use cases and first of all if you think of blockchain as a shared uh state committed state uh

across multiple agents so that that becomes the source of truth whether it's for financial data or whatever else with smart contracts uh I think that's a very powerful concept right that's like imagine if I told you you know how would you share state between 100,000 agents across the globe object easy. Uh the consensus algorithms is already built. The agents they don't have to trust each other. They just have to adhere to a bunch of rules uh protocol and uh and then if some of them don't agree they can just fork and everything is already in place. And that's the first one. The second one is payments um agent payments. Everyone keeps talking about the new agendic economy trillion dollar

industry. Um, yes, that's exactly what's going to happen, right? So, to me, it's unthinkable if you think about it how you do e-commerce today. Um, you go on Amazon, you input your credit card, that's what is that 16 digits plus expirion date plus code. Uh, then you have to add the shipping address and then you have to choose the items and then it's out of stock. That's ridiculous, right? Instead, you just tell your agent, hey, I want to buy this thing when I just buy it or something, right? And that's the kind of interaction you're going to have. So, all of the pain later is going to be

behind the scenes. You're going to work on it. Uh that's the main integration. And we already have a demo if you look for our talk in we integrated Cuba with WPK, which is our wallet SDK. And they're both open source and uh implementing the agenda payment protocol an export we're able to basically connect a merchant that's how it's called in Salvador and people will come to our food they will be able to just say hey I want blah blah blah and then the order will go to the merchant that they have >> so I assume this is the strategic particular company like your your project is kind of developing

to realize the economies and probably they going to pay right because I don't believe the guys doing it for free. No, I mean that those are two separate uh things like we again we don't have to worry about revenue that's our main goal I would say uh un unlike other other companies in the space they have to worry about margins and profit and revenues and things like that we don't so we are actually building human things like that uh yes of course in the long run we hope to be the main player for everything that's that we're working on but we are actually doing all this stuff

we don't get anything from do you see any existing crypto project like some [clears throat] software for even some blockchain which are really decentralized and independent or is it all like I think it's both. Um there's a lot of again fluff when it comes to decentralization. What is that decentralized? A lot of people use decentralized a very vague manner. It's very simple. Decentralized means that there is no central point.

I'll give you a very simple example very let's say all of us are using agents to do purchases, right? And so you are a merchant. We we are merchants and they are just buyers consumers all of our let's say we all use GP right and we're all in this region which is the capital system and so we use one data center problem what what happens if the data center goes down all of us are screwed right all of us do you think that's sustainable in the long run it's ridiculous right one like one hour of downtime or an issue in look at what happened in the AWS d or nothing like it was bomb like the

whole thing is destroyed and a lot of people lost data and comput because of that that can happen anyway I mean we need to take that into account right if instead agents are decentralized that means that you're running on your own device cluster which could very well be in a data center but it's your data center right or maybe you collaborate with some people to build your own little data center some those are all acceptable configurations but if one goes doesn't that's the main ide which are really contributed to doing

some useful stuff from some projects >> in crypto honestly I would say the protocol are you know very good but specific projects uh not >> there was a lot and uh very very happy to be here actually >> thank you and I should say we're hired so if any of you is people interested in AI especially local AI.

Thank you. [music] [music]