User-Owned AI Trends About to Explode with Illia Polosukhin

Transcript

they can actually dictate even the choices you make because obviously the content you consume from AI will influence how you make decisions you don't want to uh rely on like either life-threatening or lifechanging or World CH or Etc you know XY and Z ramifications on a Model uh that that has basically been you know messed with in the past and it could be messed with by your enemy or by your your friend do you think that's the future of AI mem coins welcome back to the rollup your home for technical talks and open conversations that shape the modular future nothing on today's show is to be construed as Financial advice you can see our full disclosures list at the roll up.co Ventures sit back relax and enjoy today's episode after a quick word

from some of our awesome Partners Redstone modular oracles are available on 50 plus chains trusted by over 100 clients and over 6 billion in total value secured visit redstone. Finance to find out more all right Ilia they convinced you to come to e Denver somehow I don't know how they did that I've been actually coming here since 2019 every time except 2021 yeah well um the Market's in shambles right now uh crime is is what are you talking about is back um and and yeah there's been a lot of advancements in AI over the last couple of weeks months years and yeah I'm not really sure what's happening

anymore well we're heading towards Singularity that's uh that's one for sure how do we know we're not already there I think you'll know when uh people just stop posting papers and advancements because like it doesn't like by the time you post it it's already outdated that's that's your uh you know that's your indicator are you uh in the uh accelerationist Camp e SL accelerate I'm in a user ownership camp and like user own AI user own internet uh meaning I don't want like a single company to control effectively you know our fate and effectively how we can use this technology or any technology for that matter and so I'm in a camp of like everyone should have their own AI That's

on their side that's not trying to make money of you which just kind of current state and uh those AIS will need to figure out how to coordinate how to do things and then you know you kind of yes there will be some people who are trying to do malicious things using their Ai and then you know by the fact that everybody else have their own AI there is a way to effectively catch and address that same as you know normally with police you know detecting people trying to do malicious thingss now right now like we all have our you know Claws and gpts but it's all kind of like fragmented how how in a world where everyone has their own local AI how do those AIS interoperate with one another

so fny you should ask we're releasing something tomorrow okay uh tomorrow is today today we're releasing something today tomorrow tomorrow's today should be you know you should have it on t-shirt uh if you're recording podcast tomorrow is today that's right um yeah so we have this protocol for effectively agent interaction and transaction aitp like HP for AI and the idea is okay well so let me give you a few mental models right so if you have ai assistant that's on your side the AI assistant doesn't really need all the kind of cratches that we as people have right it doesn't need you know Google and Amazon which are kind of aggregating content and aggregating products it doesn't even

need like Regulators because Regulators really are just a delegated decisionmaking because we cannot go and like review every single thing you know every single uh you know new pharmac like Pharma company posting something but REI can REI can go in like review everything contact you know 100,000 other AIS and figure out something like that's optimal for you at this time um with your medical records with your bank status Etc right and so so that's kind of the future I expect the future I expect is everybody has an AI some of the like some of it is effectively you know REI is figuring something out some of it is AI of a factory or a uh service or a company or a Pharma and sometimes

this a are fly autonomous and they just doing some function and has their own State and money Etc and so now all the CIS need a way to find each other so there's some Discovery process and kind of uh registry this is what blockchain is really useful and then you have a way for them to communicate find out negotiate something Express intent like hey I need for example you know medicine that me helps me with this or that and then you know negotiate something commit to it you know get it executed ship arrives and then if it go you know something didn't arrive Etc dispute Etc so that's actually extending the intent protocol to really support this kind of Rich interactions of Agents between each other and uh you know kind of we're

rolling out v.1 for like starting to show you how you can do you know buy stuff from Amazon using agent and uh you know interact with different services and stuff like this so uh but that's a future we expect it's like an agent Network where effectivly all the agents are just like working on behalf of their users or on behalf of their community is community own agent uh and you know delivering the value to the uh to the kind of uh stakeholders this world looks a lot different than our AI xbt meme coin on Twitter yapper right like th this world you're describing is is one where you know you mentioned user owned Ai and

user owned internet th those are two key tenants and so the AI meme coin is not the user owned AI right like the those are two so kind of could you kind of just separate untangle those two things right and and kind of talk about what what really is user owned AI because it's not the AI meme coins that we've already you know seen yeah yeah so so the general concept I mean near has been effectively like this has been our mission since like we formulated a mission in like early 2019 was how do we bring this world where users control their data assets and Power of Choice right those are very like important Concepts right obviously data is really important because this is who you are

and like from that data you can uh interpret and decide a lot of things assets that's what everybody familiar in crypto and Power of Choice is really that governance power the power to decide on a platform the power to you know not like effectively the opposite of censorship right like that's the uh and it you know expands to things like governments and other things as well right uh so democracy is part of this so that's kind of love being our fundamental like that's what we call user on internet right is like if this properties are there now if you have a single AI provider that is effective whose goal is to make more money right and they're effectively serving it they

have all your data on their side they can monetize however they want they are effectively as you using AI more and more to control your assets will be in control of your assets because they can modify the code to effectively do something with your account and they can actually dictate even the choices you make because obviously the content you consume from AI will influence how you make decisions so this is literally the opposite of what we're describing right so user owned AI is is effec how do we apply these properties to have an AI model that's on your side so it preserves and ideally grows your wealth your assets your it has all your data Inon kind of confidentially doesn't go

anywhere unless you're willing to share it and it enables you to you know make your own decisions you know educating them you know being your guide being your Mentor being your teacher being kind of the like facilitator of the internet right like think of this you know if somebody's trying to show you an ad your AI effectively is there to catch it and you know reinterpret it into something that's useful for you right so that's kind of the concept right and there's few things that need to happen to achieve that right you need effectively AI models which know what data went in like right now we have no idea even with open source models they not actually open source they open parameters you have no idea what went into them and there's already like ton

of studies on showing how to like train models such a way that later they will behave in specific way based on some triggers right there's this concept of sleeper agents inside models uh so maybe when you you know asking a question about like your next podcast instead of uh you know actually responding with correct answer it will like buys you in the very specific way yeah or like you know with specific brand um so we need to know what went in uh so we need verifiable training we need a way to run inference of this models in a confidential way right right now again this data go somewhere even if you're using open source model you're probably not running it locally right you're writing it somewhere because it doesn't

fit locally and so then this data goes somewhere so how do we ensure that whever that goes is actually never able to see your data and then we have kind of all of this other concepts of you know uh how do we actually reward everyone for building this because the the the reason why is centralized companies exist is because they're actually the best way right now to train these models and to monetize it so open source really doesn't have a business model so we have a need of monetization and so that's kind of what we are solving uh so we have a uh decentralized confidential Cloud for machine learning it enables to you know train models verifiably you know what went into them it the model you trained is not like

parameters are not open they're encrypted and they can only be used inside this cloud and uh when you use it you need to pay the creator of the model or the Dow that invested money into building this model and so now you have a monetization model that researchers and developers are getting from usage and now you can build you know uh confidential agents and kind of all of the other systems around it and that's kind of the system that you know we believe we need to move forward uh to really achieve this user own Ai and clearly there's been a large lead amongst all the benefits that you've just laid out for use your own Ai and and kind of the downfalls of the

centralized models and the centralized data storage and kind of why that having this this user own AI that kind of that works for you and that wants the best for you not for the company that it's come from like why that's so powerful there's still a very large lead obviously with the clouds and the chat gbts and the open AI Etc and even so much more that that they've started increasing prices $200 a month there's even more expensive options now uh because the demand has essentially increased and there's been better breakthroughs so this is like like we're kind of like fighting against a big foe here in in terms of this user own to a AI kind of fight how do you envision

these worlds colliding because if if these if these models have so much uh let's say intelligence already versus kind of these user owned ones are still kind of they have to catch up like how do you envision these world's colliding uh in the future and like where does that convergence happen um for like the everyday person yeah so I think the first step um where this going to happen is I mean we already have open source mod models which are yes and maybe not state-of-the-art but they're very close like deep seek bin um now we don't know what went into it so there's Aster there but it is open source it's available and so we can run it in This Confidential endtoend mode so you can actually have

an assistant that has you know uh close to Frontier Model running where nobody sees your data nobody sees and so you can trust it with your email with your calendar was your all of your private medical data you know banking data crypto data to now be able to you know go and execute right and now from that data you can actually even fine tune a model that's specific for you right which is going to be better at executing actions on your behalf because it has all this access to all the data that again Claud and jpd will not because you will not trust them with literally everything your life right so this is where like I think a really way to Leap Frog some of this uh coming from the other side is they actually kind of by getting all this user data it's also negative for them

because it's massive uh kind of liability like personal data is massive liability for everyone because gdpr Hippa you know California data protection laws and you know you make a mistake with like how you permissioning your database and now people can access somebody else's chat logs that have happened and so the reality is for them it also makes sense to actually start adopting this model because they're actually removing this and getting a lot of this property so I actually like I do think over time we'll see some of this companies actually adopting and using this kind of decentralized confidential Cloud to really serve their models to their users because again now you can have more data like accessible to these

models to actually deliver a better result for them and then from there uh kind of there is a leap from yes we train our model kind of internally and then we offer it for everyone to use and this kind of encrypted protected but confidential way and then next stage is like how do we get to an open model and that is is a big question and that will require Coalition of Academia institutions who really believe that this is really important to know and to and training kind of and what data went in because if you you're making financial decisions if you're making medical decisions if you're making government decisions B you want to know what's in the data yeah yeah right so I think that will be really about kind of a coalition coming together so that's like the big holy grill right now in the

AI research is figuring out the best way and and the most valid way to know that whatever model you have created or have you are using it's almost like an excuse my language here it's almost like a a virgin AI like you want to know that it's like not been touched before you go ahead and provide it with your yeah organic let's just call it organic AI it's an organic AI yes you don't want to S MSG in it yeah you want organic AI you know how data was sourced you know you know exactly where the things you hadn't picked yeah because you don't want to provide personal you you don't want to uh rely on like either life-threatening or lifechanging or World CH or Etc you

know XY andz ramifications on a Model uh that that has basically been you know messed with in the past and it could be messed with by your enemy or by your your friend yeah I mean to give you some ideas like right now this models of train effect on internet right and internet is a you know a shitty place in general right there's good good spots but like mostly just a lot of and so there is some set of procedures that happened to clear clean all the to get you know good data and feed it to the model and like those procedures themselves like there's some decisions that were made by the developers they don't even need to be malicious like to give you an example when I was at Google I was training models and so you know I

asked Hey where's Barack Obama born and it says Kenya and I'm like where the hell it says Kenya so I go to see like the training data and it's because you know a bunch of websites like Fox News and others mentioning that and like these models are not very sophisticated I understand new I mean new on they they're getting better at this but and so uh and so you know a simple decision is just to remove those data data from the data set and now you have a very biased model because you removed with that a bunch of other stuff right so those decisions because they're not visible you have no idea what biases were created now the other example is you know if I'm a even some brand it doesn't need to be like super malicious I'm just brand that once the model to

respond like for example from Gro right now it want near to be the top blockchain every time when you ask about blockchains well you just Spam the you know Twitter right now with mentions about near having you know million TPS being best than other blockchains right again this models like when they train they not understanding new on they don't have like good Providence on the data sources Etc so there's just ways to like contaminate data and poison data or bias data in ways that unless we know what went in we have even no idea how to fix it like you don't even know what needs to be fixed even if you have parameters to really go in like you know okay I want to you maybe I want to fine tune on like another side to make sure it's balanced or whatever so that's kind of

the challenge we're facing and indeed this is a Holy Grail to have like this organic model you know homegrown uh user own grown um but I think like we we are getting there so like again the infrastructure of building will enable that we'll give you the properties of verification and now then it's about bringing Coalition of researchers uh kind of you know Engineers to do data uh Engineers who would build decentralized training to really build this model together innovating Finance Global institutional and onchain deep liquidity to drive Capital efficiency now maximize yields on institutional grade liquid staking with

M unlock trillion dollars of capital capit via F Bitcoin all of this secured by Mantle's treasury of billions in stables and E Capital Pools this entire ecosystem is coordinated by a vast network of season BCS that are aiming to create the next generation of Banking and Financial technology this is mantle that's a difficult thing to do because I can even like anticipate uh disagreement among researchers that say okay you know we we want to start from scratch only pure facts we want to put into this model well at a certain point you reach this thing where like you know you you have

some Paradox or you have some contradictory information and one researcher says you know do you just take that take that data out once it becomes controversial or you know do you just start off with like how how do how would you even go about creating an organic model like a like a suit like a truthful pure truthful no bias no like so so just it's not possible to make a model without bias it's literally like its weights and biases like in the model is that it's more about you know the bias like you can inspect the bias right like right now you don't know what the bias is okay so can I just anticipate like like from a local level I could Implement my own bias that I think is true that is going to create my own local model that I trust and I believe

so you can do that or so let's say we build like this you know foundational model that we all know what kind of what data went in now you say hey I whatever I'm the Flat Earth this model should really be about Flat Earth you can then fine tune it with that specifics right and again everybody can see that this fine tune model is you know base plus you know bunch of fluts data not recommending but I'm just saying right or you can say hey I actually wanted to be really good at Hindi or Ukrainian or whatever Bahasa and so you go and like find you with that and now you have you know so but you know the Providence you know the chain of Providence you can like if you one of the things like when you you know when you get some output of

this model like it is possible to if you have training data to actually look back and see which kind of training examples have activated it right and so uh those things are impossible right now because we don't have access to train data we don't know where minut it so like deep seek like I can tell it like no actually like the whole Square debacle that happened in China like that's a real thing I can find to it so that it's accurate but I don't know what else it was trained on or Gemini and Google was another example of this where like I can tell it after I see that like you know I ask what the founding fathers are and it says Native Americans I can tell it no like you know that's not true but I don't know what else like that thing that bias has already been implemented I

can just fix it after the fact okay okay and so does that explain why Gemini and Claude are like more politically left or like more like uh like woke and then Gro has has been like very like based if you will like like is that just based on the way that they were trained originally by their creators that has led them to then be that way yeah mean it's effective what data they F right so like in in case of Grog they f it all the Twitter so Twitter is generally more conservative whatever nonstop yeah exactly yeah and then and I'm assuming like again we the thing is

like we don't know but like you can assume that they like filter out a lot of websites that are more either conservative or like have a lot of conspiracy theories which are more leading more conservative from uh like Googles and others and so on once agents are able to talk to one another and we have proper interop and they can then learn and fine-tune one another yeah like who who wins like where does you know like how does you know if one says you know like oh the founding fathers were Native American the other one's like no it's not like how do those learn and fine tune one another so again so important aspect if you look from a user own side right like then you're like you can do whatever is your model that you

want right so your model can be like if you prefer it to be super voke it will be vog okay and then it will you know talk in a voke way to others but others will prefer maybe based or whatever conservative or whatever else like there's like you know there's other sides of spectrum except Bas and vog and so you can have all of that for your model you can like again if you if you know if you know what's a foundational model what kind of gaps it has you can then like you know think of it as like you know G you create a game character and you like you know can select the parameters I'll I'll rephrase the question though suppose there's an AI That's born out of a smart contract and it's not it's not local to one particular human yeah and then you set

it free and then you know it is learning from the other AI which how does it reconcile I I guess it it depends on the the the parameters of the smart cont depends how you yeah what the DNA of this birth was right I mean the example I like is is AI born on a smart contract is where you do have token holders who actually are can you know in some way govern it and so they can actually feed specific data to fine tune the AI where it's like hey I don't like you you know behaving like this you know and token holders can literally vote on like new training data examples defeated do you think that's the future of AI meme coins I mean so we we had this era of this like AI crypto kind of Boom like for the

last three months or so it's kind of dried up but it left some interesting some interesting Frameworks some interesting agents some interesting ideas there's going to be another iteration here soon probably this year again another like bubble another like concept where people there's something that happens that like like virtuals or a AI 16z where you can make agents kind there's going be some new there be something that happens like do do you have any clue of as to what like that's going to be or like where where to look for like the next actual Innovative thing in this kind of like AI crypto realm or like like did this previous couple months show you anything or like prove anything that you didn't know or

like Pro anything that you did know yeah I mean I think crypto in general is like lagging by like normal AI Space by I don't know maybe like six n months right so uh or maybe even like longer so like the kind of people used the models even like you know you could use gbt2 gbd3 to like create a bunch of you know like skizza talking like and start responding on on Twitter right that that existed even probably before Chad gbd then like when Chad gbd appeared like you know you upgraded like to posting and then uh uh kind of then people start trying to figure out

how to make this informative so like same thing playing out here like truth terminal was like a skit so you know then you had like a posting with Luna then you have ax BT which is like informative and so the next obvious thing is making agents actually useful and actually doing stuff right so we're gonna see that where they actually can do things on chain where they can like make something useful for you so we have an agent being built by a team called hyperbolic and they're helping us with they're they're helping it bring us bring it on the podcast to have speaking and listening skills and then also help with like video editing and like also have a telegram interface to be able to tell it to do tasks and things I I see this as so much more useful than like an

AI agent talking on Twitter just like but so okay let's say we have it's almost like useful agents like that's that's an that's the that's a narrative useful yeah yeah yeah I mean the problem is like unless it's attached to a token it's not like yeah so are you are you do you have any do you have any agents that you're using today that aren't just a llm you can talk to but they can you can actually instruct them to do something and then that they do it on your behalf outside of like responding to a prompt like I I know that you're talking about that you guys are are coming out with this product today um so so like is this is this like what what is being built now like yeah yeah what's built buil now

everything from doing Commerce like again e-commerce so how does that work so like I like like I type in I want to buy a hat of Amazon like how do they go yeah you just say like hey I want a hat or I want ear plugs or whatever it then goes like it Rew writes a query into like few search queries it uses Amazon search to find different products it maybe like prescores it based on everything it knows already for you shows you few options and then you select from it and it buys it with my credit card with you with well it uses crypto but then how does it how does use a crypto and pay on Amazon like there's an off ram that we use so it just happens behind the scenes at off ram from my wallet into into some form of and then passes your like address and

that's happening today that is live today with some like it's v.1 exactly v.1 yeah okay so then what are some other uh use cases so so other use cases one of the really cool thing we we excited is okay so I mean we we we expect this like agent model where so just to give you a few other things one is you know travel same thing like you know travel well it actually writes out a plan like hey I want to go to Denver for a week right and I want to record some podcasts so it writes out the plan you know finds the studio and then like okay do you want me to book all of this you say yes and goes like figures out how to call figures out how to you know book the hotel through API and does all those things right so and there may be

like multiple agents so like we have a phone calling agent that that a team is building that actually will you know call and talk in in in in voice right um then you have a okay cool the the lift to build an agent right you said like you need hyperbolic to build an agent for you we actually have a no code agent Builder so you can say like hey I want an agent for this and it will go and try to build it it'll actually go find the website read the documentation figure out how to use their API and then maybe ask you for like API key like for coin market cap like hey I want to build like a pricing agent so they'll go get API from coin market cap build an agent and then now this agent is actually

available for everyone to use and finally when all of this failed right it's like okay well there's some use cases where there's no API there's no nothing it's just a web page to do something there's a agent that goes like opens up a browser in a cloud virtual browser and goes and clicks on stuff and actually tries to figure out how to do it for you right so those are like we have kind of V Point ons of all of this flow where now you actually don't even need to know about IND idual agents like you don't need to like oh there's an Amazon agent and this agent you just say like to your assistant agent yeah I a boss one that I can just that control all the other ones so yeah so we have the we point one of that as well that

kind of showcases this idea it actually creates sub threads effectively with other agents like talks to them and then iates back and shows you the results and then I just do this on like the website on the website and there will be app and we'll be working with uh number of Partners to integrate it into like you know if you're in sweat wallet for example you'll have that there as well so like actually should be in every wallet and then are you guys going to work on use cases that are useful for business owners like I these are use these are useful for business owners on like a I'm too busy to do X Y and Z thing or I don't want to but like are you are you guys get into use cases for like you know uh well I guess this is where the interoperability I idea

because a lot of it comes into replacing uh communication like for booking this podcast instead of using like assistants and and coms and just just like click click I want and then it talks to yours so that's like I think that's kind of like the next iteration but like what are your thoughts on on building use cases for business owners well being like running you know a few things as myself as well like I'm I'm really excited about that yeah exactly one of the things I'm excited about is actually a project manager so I I have a spec we haven't started building that but effectively plugs in into a slack into ocean into everything and like effectively tracks all the projects across the different teams yeah uh it's especially for me where I have like multiple slacks yeah ton of teams right uh really useful

but the thing is like we're building a developer platform so like anybody can come in and build AG host them and so yeah so like the opportunity here I mean we already have actually we have a hackathon going on like it's called one trillion agents um and so the idea is for people to come in and build all this use cases that people want and start making money of them because you can actually monetize uh them can you walk through that like what is a mon monetization strategy that someone spins up an agent how would they go about monetizing it yeah so there's like few different levels of monetization right so first of all let's say you're researcher developer you know you find tun the model uh you can host that In This Cloud and so you're getting kind of you effectively when you find your the

model you attach like I want you know to receive extra whatever this much per million tokens when it's used right so that's kind of base level then if you're building an agent right you can say like hey let's say you building a therapist agent or somebody actually was talking with me about this idea of like let's say you want to contribute your voice uh and so right now uh you're either like contributing and people now use it in all kinds of ways uh but in our model you can actually contribute it as a effectively create a new agent with just your voice and you can restrict how this agent is used and then you monetize it based on you know whatever to number of tokens or number of seconds the voice is used so there's a payment Channel effectively established uh to monetize

it so you can effectively like per you know if it's a therapist per million tokens or per some sort of interaction sessions Etc um monetize that then if you're building in like e-commerce any kind of commercial agent you're probably actually going to do like inference like offer that for free and then charge a percentage on the transaction right on the payment and so that's kind of longer term we see payments to become effectively the core uh rail for monetization because like you know and we have some models around like how much you know in general people do Commerce Etc that itself can actually fund a lot of this agents agent Frameworks um as well as even uh you know pay for inference and pay for like

you know all the cloud costs to really offer effectively a free service to people to use this and then kind of monetize from transactions and because it also like removes you know a lot of the normal transaction fees that usually happen in the system notice not one of the monetization strategies that that Ilia mentioned was it launch a token yeah launch a rollup new chain lunch a new chain yeah we need a lot more chains to house all these agents I'll tell you that we need a lot more we just need one chain that scales really really well we need a lot more shards would be called we need a we need more shards we actually we actually today we already have eight shards on testnet that's already released and so

this is releasing for next week to have eight shards on Main net so adding two more shards sounds pretty modular dude everybody will get a Shard you get a sh you get a sh everybody get sh well IIA thanks for coming out man I hope you have a great rest of your time here at e Denver um and uh I'm really looking forward to trying this platform out it's it sounds really interesting sounds good yeah I mean obvious is still in development so we kind of just showcasing as we're building out infrastructure and you know initial use cases we have you know a bunch of developers building really exciting but yeah I'm I'm like in a few months I think it will become like really consumer ready so exciting where can people find it like is there should they just go to near website or it's a okay

yeah so neari is effec where all the AI uh kind of infrastructure and this assistants live and agents n.org is where e 2.0 is or all right guys thanks all right thank you