{"id":"ZFqfT0CGshA","title":"NEAR AI Office Hours #34 Whiteboard Session with NEAR AI team","description":"In this NEAR AI Office Hours, Illia, Alex, and Eric lay out a vision for a new internet powered by user-owned AI. They explain why personal assistants, discovery agents, and service agents will soon mediate online tasks—from buying products to running business processes—under an open protocol. The discussion covers secure compute infrastructure (TEE), payment rails, and how model builders can monetize while preserving data privacy.\n\nThey also explore how decentralized training and encrypted weight models might fuel an ecosystem where data owners, developers, and GPU operators share in rewards. Watch to see how these open protocols and incentives aim to prevent AI from ending up behind walled gardens, paving the way for a more user-focused internet. If you’re developing AI agents, training models, or running web apps, consider plugging into the NEAR AI ecosystem.\n\nTakeaways:\n• Personal & Service Agents: Assistants for users and specialized services will shape a new AI-driven digital economy.\n• Secure Computation: Trusted execution Environment (TEE) ensures private data is protected, even in distributed settings.\n• Open Payment Rails: Crypto-based channels and incentives boost adoption and reward ecosystem participants.\n• Encrypted Model Hosting: Developers can monetize AI models without exposing underlying weights or IP.\n\nTimestamps\n\n0:00 – Introduction & Why AI Shifts Every Interaction\n2:00 – Personal Assistants and Service Agents\n7:00 – Discovery Agents & Matching User Needs\n9:00 – Payment Protocols: Enabling Transactions Across Agents\n14:00 – Trusted Execution Environments (TEEs) for Privacy\n25:00 – Monetizing Encrypted AI Models\n34:00 – Training Open Models & Infrastructure Incentives\n48:00 – How to Get Involved\n\nNEAR is the Blockchain for AI\nWebsite: https://near.org/\nNEAR AI: https://near.ai/\nIntents: https://near.org/intents\nChain Abstraction: https://www.proximity.dev/\nSharding: https://nearone.org/\n\nFor the devs\nGitHub: https://github.com/near \nDev Docs: https://docs.near.org/\nDevhub: https://devhub.near.page/ @NEARDevHub\n\nConnect with socials\nTwitter:   / nearprotocol  \nBlog: https://near.org/blog/\nMedium:   / nearprotocol  \n#NEAR #nearprotocol #NEARAI","channel_name":"NEAR Protocol","published_at":"2025-02-04T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/ZFqfT0CGshA/hqdefault.jpg","transcript_text":"future hand in in this Tech driven Wonder hello everyone welcome back to near AI office hours uh today with me Alex and Eric um all of us are making really good progress in neari and today we want to go over a different topics we're working on as well as uh outline kind of how all of this fits together and obviously invite you to contribute and participate in what we're building so maybe before we go into what we're doing we're going to talk about why we're doing it and kind of how the We Believe future will look like so the first part is we believe that AI is going to be a fundamental shift in how every single interaction with you know computers internet and even economy as a whole is going to work and through that um a lot of the kind of small things you know like you know where what like how you ordering things how you deciding on where to go Etc will change but also big things how you know even governments and industries run uh because AI can not just kind of automate some processes but actually start making decisions and go from kind of this traditional process where we usually delegate decision- making to other people and companies uh now you can have decision- making happening on your side personalized for you given all of the information around the world and all of the other participants around the world so how do we get there right the like this world is you know um pretty complex right now and there will be a lot of kind of transition period before we get to that future and so the first step is what we call assistant so let Eric kind of dive into that sure uh thanks so um really trying to Envision what what your interaction with the internet's going to look like in the future and so we we think basically there's going to be in a same sort of thing you see in sci-fi there's going to be a an assistant sort of works for you knows everything it can about you um knows your preferences knows your general needs knows your information knows your history and when you need something or you want something whether that's um whether that's like hey I want to replace a broken mug in my house or just give me some content to scroll through or tell me today's news or every week make sure like make sure that I wake up tomorrow on time like it will figure out how to do that and give you an answer um and we expect that like the world of individual apps will kind of just fade away over time so uh that assistant can be great but it can't uh won't be able to do everything by itself it's going to know as much as possible about you but it won't know as much as everything about what it can go and accomplish in order to do that it's going to have to talk to other AI um that's what we call Service agents basically your AI assistant represents you it's going out there on the Internet or in the real world there helping things happen in the real world for you uh in order to accomplish those things it has to go communicate with other AI systems that will represent either other people uh other services businesses what have you and those will help you get something done what's what's that a factory oh great so each of these ends up being a service agent which could just be somebody else's assistant really um it's just a matter of uh when you're yeah it's a matter of point of view um in order to figure out how it's going to get something done we have a third category of AI agent we envision that we call a discovery agent um that's something that the assistant can give a more General sort of intent to I'd like to accomplish this thing I'd like this information I don't have it myself can you either go out and find it for Me by understanding the directory of all the service agents out there and just going querying for one or perhaps broadcasting to a to a to to an area of like hey my user needs something done who can help me or just um sort of introducing your assistant to an agent that can go solve the problem maybe just to run a specific example or a few specific examples right um and kind of show the breadth of this right this can be something where hey I want you know to buy uh sleep tracker right and right now you would you know Google then maybe you go to Amazon then you maybe like in my case would Google again because you're like well I don't trust Amazon reviews so I'm going to go check between few different products which were like people reviewed and so you kind of go in that process right and so you can imagine you know the same thing well first of all assistant probably knows that you don't sleep very well so it already predicts that you need a sleep tracker and to improve this uh to get more information about you and recommend you you know different sleeping pattern so it can issu this query okay I need a sleep track uh like who are actually selling this kind of um equipment the discovery agent uh you know knows about two different stores so it either like introduces you to them and so you can connect to them directly say Hey you know I'm looking for sleep tracker uh one of them is responding hey you know this is the one this is um how much it's going to cost maybe you also got a you know some review a agents or maybe you actually just your AI assistant get connected to somebody else's AI assistant who is kind of similar profile uh who you effectively getting like feedback as well about the uh the sleep tracker you considering buying this also works in cases where for example you're actually in professional setting and you're trying to accomplish some task for example run a marketing campaign you know hire someone again you can go you say hey I'm planning to run this marketing campaign I want somebody to do this for me it knows few different marketing agents uh who then give you quotes on how much this marketing campaign would cost uh how would they run it U assistant you know evaluates it either decides directly or asks you to choose from one of the options so it's a pretty versatile kind of system that you know can go from your kind of basic needs you know food and shelter to you know more advanced things to you know business case and kind of uh Global Commerce as well um what connects all these little boxes is that they are all AI agents and that's a word that everybody in the world at least in the AI space is using and nobody has the same definition for us so we're going to give it another shot um what we mean by agents in this context is that they are um they are Services backed by AI inference that uh that that can be talked to um now in this case of the assistant is generally the user that's talking to that agent in the case of Discovery here it's usually your assistant that's talking to it you could also co- talk to it directly but mostly your assistant's going to be doing that same thing goes for these Service agents backed by Services your assistant will talk to them you could skip it but we think in the future you're mostly going to be mediating your interactions with the rest of the internet through this assistant um as opposed to uh what you see a lot of development right now is in things that are called agents but they're really a template for an agent to interact with some existing service some real world um you know some website uh and that's an important component of this system but it doesn't really make an agent that you can go talk to in natural language or whether that's video audio text what have you I mean it the the modality of it doesn't doesn't particularly matter here um nor is this one masterful amazing thing that does everything uh we imagine that there would be multiple competing AI assistants competing for which one is better multiple Discovery agents competing for which is better at matching your intent to uh a solution of course multiple Service agents um I mean this is just the economy of the future internet um what we're trying to build here is um is how we're going to make this happen uh and to be clear this is is it not necessarily that we um we believe that this won't happen if we don't do it we think this is inevitable and we're going to try to make sure that it doesn't suck by um by building this on open protocols uh decentralized infrastructure uh use our own models so that when this happens it's not going to end up as basically uh the AOL ification of the internet um again I think this model is going to be the future uh if we don't do anything about it uh and say open AI builds this whole infrastructure then they get to kind of curate what the whole internet looks like for you it ends up being uh you know they end up being a gatekeeper um they can decide what information the models use they can decide how the discovery is going to work um if we build this as an open protocol and get enough people on it enough users coming in enough service agents um a competitive Discovery system then we can uh build something that's a lot more functional and let the open world sort of swallow the curated World me that yeah so so now the question is okay how do we what is the protocol effectory that connects all those things right because I mean you can imagine this running indeed can be running on open infrastructure which is what we're you know really pursuing it can also still be running on you know Clos infrastructure uh we just want it to be connected into one protocol and so this is where uh we're really kind of proposing this uh protocol called aitp it's agent interaction and transaction protocol to allow us to really interconnect all of the agents in kind of structured format and open up you know ways to pay to transact to share credentials to share information to share kind of state and sessions uh between all of the different participants of the protocols let's so yeah yeah so so uh so what what is required here right well there is kind of a fundamental um uh piece of piece of data that is needed um which uh okay so what do we need to actually achieve this well we need a fundamental kind of object that everybody will share and is able to interact with and so we actually are taking on um an API that already pretty widely used uh for AI assistance U by open Ai and it's effectively a thread API and so thread is right now um effectly a chat his history between you know few different actors right this can can be you know user this can be another agent multiple agents and this can be uh some uh you know some tools calling and a bunch of other stuff and so the idea here is uh we want to add more uh kind of structur uh kind of optionality to allow for various interactions and we want to allow This Thread to be again not a centralized thread where you know you need to go to only one service and only you know through that service you can interact with with it which is kind of current state across the industry but actually a thread that multiple different participants can actually share interact with join leave you know add somebody else and so make it effectively similar to how email works right you can imagine you know right now you're effectively going to open Ai and you're talking to open AI kind of model or agents or cloud or any other system or kind of this new system where you know U System effec creates a you know a thread with a discovery agent Discovery agent can add another agent to the thread and maybe remove itself you know put it in BCC and kind of continue like you now continue conversation with the new agent so very similar to kind of Emil protocol and now we adding a set of properties to it uh to expand it y so I mean it's it's it's a pretty simple idea for a protocol it just sort of standardizes how an agent is going to be able to talk to an agent in the same way that a user can talk to directly to an agent and what are the what are the standard capabilities that um an agent needs when talking to another agent in this format one of the um for a lot of stuff like you can talk over natural language again whether that's text voice image what have you um and that'll work fine but there are places where you want to drop into uh into structured data where you want to make sure that you're getting exactly things correctly a couple big examples of that um one big example of that is payments if if you're going to buy something from this Factory uh you want to make sure you're paying exactly the right amount for exactly the right product that you've um that that it's agreed upon what you're going to get what the terms of that payment are what the return policy is what the payment method is supposed to be and so we're designing um again open standard and protocol for how uh any agent can go request a payment from the agent from from whoever is talking to it uh which essentially is just going to be a tool call um if you're familiar with llm Parlin here uh so on this thread we're going to pass back a tool call uh to say something something like a request pay for payment um now the AI assistant there can uh if it if it is authorized to do so if it has the means to do so um it can do that payment it can respond to that tool call uh go and pay for the thing um it also can sort of pass it further up chain okay maybe I'm not authorized maybe I want maybe I don't have any money um my user hasn't given me any money to work with myself or any credentials for any um for any payment methods and uh and I want to ask my user to pay for you can sort of pass that Upstream that could hop between you know just more than one agent if it needs to until it eventually hits the user who is interacting with this through some sort of UI which we didn't quite draw a box for but I'll do so now and the UI would pop up well a payment box yeah type in your credit card number connect to your crypto wallet whatever method uh is going to work in order to pay for the thing then gets passed all the way back Downstream to execute the transaction um on payments I think uh you know this protocol can support really any payment method that exists whether it's Fiat or crypto um but we think this is going to be huge for adoption of crypto-based payment rails um one main reason like right now if you're uh if you're trying to buy something on the Internet like there are plenty of places where you can spend crypto but it's always like we still haven't come across haven't overcome those those ux problems uh then let logistical problems you know how am I going to get money into my crypto wallet how am I going to do currency conversions um how am I going to on onramp how's the merchant going to offramp uh these things are just getting in the way of of crypto being used for payments dayto day it just like even for me been in the crypto world for 10 years now I still buy things with a credit card um but so when it's the human making the decision like they're reasonably in devised to use Fiat payment methods right now unless they're very principled about it but if this agent is talking to this agent and they're trying to negotiate on what payment method to use that case they really going to want to pass a credit card number over the wire or are they going to be a lot more incentivized to use a cryptobase payment real I think it's the ladder this is the um once agents are making these purchasing decisions on behalf of users or at least recommending them to users really changes the game as far as uh adoption a crypto adoption just for everyday use cases um so we're building out again the open payment protocol we're also building out a crypto-based payment rail on top of that protocol that we think is going to be uh particularly attuned to the needs of AI agents uh to the needs of this system um for instance uh we think you know this discovery agent deserves a little affiliate revenue for introducing you to a service agent we want to bake that into the protocol we believe your UI deserves um some affiliate revenue for getting you to the right assistant if we build this right uh this is going to incentivize more people to build Discovery agents more people to build Service agents more people to build uis and we don't want you to just have to connect to this thing through near. website we want you to be able to connect to this ecosystem through any number of interfaces um and those interfaces should be incentivized to try to get you to connect to this ecosystem uh payments isn't the only thing you can get through here um the the protocol for interagent communication aitp uh would require a few other things um for instance credentials passing but the more and more we put in there the more and more it gets um I mean it gets scary to do on the open internet uh especially with llm sitting behind here um if we're passing all this data you know if if if everything we're doing involves a a a call out to chat gpt's models to uh to to determine what to do next then we've kind of lost the plot and they can still sort of curate the internet for you and adjust what you do um additionally this assistant in order for this assistant to be effective it really has to be inside your brain um if you've ever worked with well the previous definition of an agent in the real world a travel agent a real estate agent you you hit that disconnect where um where it's hard to express it's either really difficult to express your true preferences and plain language to it or um or or you just don't necessarily trust them uh you're not going to tell them your entire life story everything you've already every done everybody you're trying to contact with and your true motivations uh and so in order for this really to be effective you they're going to have to be really inside your brain and you're going to have to trust that you can give them information and it's not going to get it's not going to get compromised it's not going to be stolen and that brings us to why we need trusted execution environments sure yeah go for it I've been talking yeah i' use a mar yeah okay so um as the internet is being revolutionized um people will want to be doing things that depend on their private information so today generally even even in the modern internet the situation which we are where we have to share our information with uh with entities that we have no reason to trust um is already that's already a pretty bleak situation it's going to get even worse where effectively if we envision this future where AI revolutionizes the way we interact with everything it's one of the two Futures either it's fully controlled by the same centralized entities we have today right so like Google open AI you name it right so you go talk to Google and Google has a bunch of services right and so we continue giving all the information to a few to a small number of entities that don't have your interests you know that they optimize for uh and right or it's an open uh network of services uh that's the future that is more desirable but now if you imagine all the services running on random Hardware in the in the wild it's uh it's it's one thing to give your private data to Google it's another thing to give your private data to completely random people whom you don't know at all you definitely have no reason to trust them luckily there's a technology called trusted execution environment or t uh that effectively says you can imagine three entities here interacting right there's a user I will draw him here again that's a user uh the user has some private information uh then there is a service that's supposed to be a little COG um uh the service is a the service is a program right that uh that is something that the user wants to do and it's there's a service provider there's a person who built it who provides the service and then there's a hardware I will draw it bigger because uh IA from behind the scene is hinting me to do that um right so that that is running the service right and the service and the hardware in many cases that would be the same but not always right because we're moving to the to the world where Hardware has to be GPU enabled with expensive gpus not every service provider has to be necessarily infrastructure provider right so ideally the user wants neither the hardware provider nor the service provider to see their private data and te is enable them to do that right so te effectively allows you uh in the environment do we have a green marker no blue somewhere yeah okay so so blue will be green right so red here means um evil right so you can imagine that within this Hardware controlled by an adversary there's a vault uh uh that is such that user can verifiably confirm that everything that executes inside the Vault nobody can see what is happening in there right so so it gives it gives two big guarantees guarantee number one uh it says so user can inspect the the the program that the service provider provid right and the te can guarantee that it is that program that is being executed right so the user knows that their data is processed in a very specific way and the second guarantee is that nobody not the hardware owner not the service provider can see what is happening inside of the te right they cannot see the data right so this way all those agents Discovery you know like shopping agent you name it uh you know all the like your your your psychotherapist the all running travel agent inside of those tees right and this way a lot of people can provide interesting agents which require to see the private data to be useful a lot of Hardware providers can be hosting those agents right and users can be using them being fully certain that you know their data is safe and their data is used in a way that you know they want like like it's not some sort of you know if they sending it to psychotherapist agent they know the psychotherapist agent is actually trying to help them not to upsell uh you know future future sessions right uh and here you know coming back to Eric's point ideally all of those um you know the hardware probably needs a littleit a little cut of the you know of the transaction the service provider right and you would imagine uh it is possible that you know it's running on the hardware like a CPU Hardware but then there's an inference server somewhere Etc right like like thank thanks to the fact that with the payment channel the payment the cost of the payment is effectively zero right you you can have arbitrarily complex interactions of uh multiple of a large number of small components that all you know pay each other on the way you can imagine even the networking part of it uh to have like you know relays and whatnot that guarantee liveness you know and they also get a little cut and you know thanks to payment channels all of those cuts combined still going to be less than 3% that you pay to visa and Master Card right so it's a yeah Beauty yeah so maybe wor drawing the the kind of this idea aide that yeah effectively you have this service agents and uh you know your assistant and so to service this securely uh you have this cloud of uh provider of of compute providers right which which are combined like some of them are GPU enabled some of them are just you know potentially CPU um although we probably move into a world where everything will be you know one where not have uh accelerated mathematical co-processors uh for AI inference but U you can have them effec running the services meaning they're hosted there and you have you know a layer as well on top that allows you to ensure that you know if one of the nodes goes down it kind of rebalances to another one uh the workloads you know there is kind of uh encryption and payments are managed through this so this is kind of the like a concept of a te Cloud uh which requires effectively uh a number of engineering challenges right obviously the te infrastructure and we just uh with Falla released the kind of first version of an SDK to build uh especially for GPU enabled nodes so you can actually have llm inference in a private and verifiable way the CPU uh where you can run you know the agent like business logic and kind kind of payment and routing and everything there's another uh piece of work around kind of this payment channels and liveness guarantees um we've been calling proof of response which enables to uh identify when nodes are not responding you know them to actually lose some uh uh Capital through slashing and through that enable you know rebalancing as well so this's kind of a whole system that's required to power this new internet uh in a private and uh verifiable way now the interesting thing is right as we are actually going toward this um Direction you have the idea that uh right now if somebody you know build a model um yeah um so similarly how with agents right somebody Builds an agent uh it they they want two things one they want it to be discovered and second used and they want to get paid right and so uh Discovery agent is enables the discovery hook up into assistant and kind of this aitp enables the payments uh for uh for the usage now it's the same thing is going about underly models right let's say I have some specialized data and I want to build an agent that uses specialized data so I fine tun the model kind of in a traditional way let's say I'm building you know the most amazing therapist out there and so now I have this model and kind of traditionally you have two routes one is you uh put it in huging face and effectively everybody's using it for free or you are running it yourself and building a whole kind of you know all of this infrastructure yourself to offer a already final product or at least API to somebody else to do it and again we getting into all the same problems which is like if you're running a therapist it has access to your private data now you need to guarantee to me that you're not uh losing it Etc and so this same infrastructure actually allows a few interesting things well first of all you can have the model you trained and you just upload encrypted weights that only decryptable inside te so what this allows you to do is now everyone uh if like if somebody's building a therapist agent or even you doing this you can just Loop and use that model and you have guarantees that only this agent is using this model and you can monetize it or you know the model itself can charge uh for using through the GPU to part of this uh payment Channel now even One Step Beyond that maybe you know if you're therapist yourself you know you have a lot of knowledge about the space but uh you probably don't have a GPU lineing around to F two the model well you can actually find tune the model on private data directly inside te right you can upload private data and say hey fine-tune this you know the result will be kind of this encrypted weight model that then can be used in this system so you're effectively getting to kind of this new way of effectively monetization for models now one step further right you can pre-train a model in the same way right right now you know we have organizations you know meta a deep seek who have done this free training right they spend Fair bit of money and they open source the model and again it it's great it's amazing we can all use it uh only if we use their specific API they're actually getting paid right uh now at the same time it's actually not an open source model it's open parameters model I think we talked about this like in one of the first uh kind of office hours the there's only like few really truly open source models where you actually know what data went in and even then you cannot verify that that actually data and this parameters are like correspond to each other like unless you literally retrain the model and you actually will get the different model anyway so so it's actually really hard to go like hey this is the data that went in right so yeah so you have data logic that you train and you have wait and right now you have fact just getting weights right so we have open weights models there's few examples of smaller models where people have posted data and code to train and weights but you don't have like a prove that this weights are result of this and if you train again literally you will get the different weights because it's randomly initialized and you'll actually get you know different results so what this te infrastructure enables as well is actually the training itself can happen inside te right and you can have like a te cluster or uh there's really interesting advancements around distributor training that's happening and so now you actually have uh a attestation that this weights were trained inside te and that this data went in into this and like and was reused this on this code right so you can actually prove that this weights for example don't contain New York Times articles or that they contain some specific data so it's a pretty kind of powerful way toward this user only the I concept where you actually know what's going on into your model and you actually can be confident that you don't have you know sleeper agents which is kind of a concept of you know finding a model to react on specific uh inputs to give an output which doesn't show up on any benchmarking doesn't show up normally but you know when you pluging it into your ID for example and type in import it will suggest you know a specific Library uh that has you know malicious exploit and so uh that that would be a one way to you know uh through open weights to actually attack a whole host of pro like uh code uh platforms um so so this actually gives you kind of this op like open source model now if you just again open up the weights the problem is you don't have the way to monetize it right everybody can use it for free which is again great but limiting because you cannot reinvest back to get better data train a new model maybe improve the logic right so what you can do now again you can have an encrypted weight model which again only usable inside te and uh this way the whoever trained this can Define the rules on how this model uh it can be used and how much they should be paid to do that so which again this can be done by distributed Community this can be done straight token or can be done by some participant you know who wants to add safety requirements whatever that is right this can now effectively has some additional logic added uh to how this weights can be used and so only usable inside this te uh environment as well just to um accentuate that point uh what right now when you uh deliver an open weight model it's essentially a data file here's the weights um what an encrypted weight model would look like is essentially um a run like it's it's the data in in encrypted form and also the program that you must run in order to run inference you can download and run that locally if you have a GPU farm with a that's te enabled um but even on your own machine you're not a to just decrypt and extract the weights the only way to access it is through the program that wraps it and enforces that um you're running it on the te and just crypt the weights this lets you do and this is useful um if your data set is closed this is also useful if your data set is open um you could have a system where the input data and all the logic used to train it is open but maybe the training is going to take three months and a few million dollars um and you want to recoup that investment you could still have an encrypted weight model and make that make sense yeah or or you training it on you know some uh biochemist tree for example and you don't want it to be used to by weapons so you add the logic to protect against that usage in in this way as well yeah but you still want you know to give it to students and to professionals to enhance their work yeah let's just put you know put restrictions but let you put restrictions on your model uh without making it you have to hide behind a centralized API you have to know to run that all infrastructure yourself you could still run it yourself but it doesn't make it a necessity yeah so so all of this effectively kind of becomes a stat that uh we are working with a lot of Partners to build so obviously this is a lot of infrastructure and a lot of uh kind of logic and and and things to build so near is pretty small team so we are working with a lot of Partners to actually make this happen right so T infrastructures I said working with Fallout to enable this uh we have Partners who are helping with the proof response for the um for the liveness guarantees and uh some of the payment Channel infrastructure uh that uh we're going to be releasing soon we have you know kind of the protocol that uh we're going to be releasing then that we want to welcome other people to build variety of services agents and uh ESP special discovery and Service agents to build on top and obviously this AI assistant we're going to have a prototype kind of a MVP of this but we want to uh open it up for everyone to participate and so kind of if the like we want to welcome everyone to really contribute to this because again as Eric mentioned this is a future that's happening it's you know inevitable kind of how the world was restructure around AI being the facilitator of your kind of uh interactions and it's now really about kind of the open protocols you know and uh open open ecosystem to win over you know one of the closed ecosystems uh to actually you know deliver all this without privacy without kind of uh opportunity for other people to participate um yeah um yeah and we're not just doing this like we're not just asking everybody to contribute to this ecosystem as a favor like there are we're trying to make sure that there's incentive structures in place in order to create this open ecosystem this new internet um and do it in a way that's going to be natural so with um you know building these Service agents that should be a good business that's that's your apps of the future uh building a discovery agent that's that's Google of the future that's App Store of the future building an assistant like these are all things that are going to be um financially rewarding and good for the good for the world good for your users um similarly over on the model training side like with encrypted weights models and with the payment channels you can monetize those models that means that you can have code within the download code within the the model that enforces that you are uh paying for usage and that can go right back to the model developer to anybody who's contributing data anybody's contributing logic um it really makes it so that uh if you feel like open models or even locally run models are important you don't have to just give away all your work um incredibly powerful concept that we think can sort of naturally incentivize the creation of this new open internet um which I think brings us to model training okay I don't know how much deeper we can dive here um the I guess everybody by now knows how we cannot see ourselves anymore but that's chill everybody by now has seen uh you know deep seek we had a reading group yesterday right so we still a week ago yes right a week ago and um uh and so we still live in this beautiful present where entities still have an incentive to throw uh away millions of dollars to train a model and publish it right for the for the secondary effects but that does that will not necessarily last right and so for uh for open source Community to continue uh so to to have a certainty that open source Community will continue being competitive with the Source uh you need first of all we need expertise in the wild how to train those models right so it's great that deeps and meta can do that but it would be great if we have uh open available knowledge how to train those models right so that's something interesting that um we're working towards and as Eric mentioned you do need to be able to to monetize those models right so so it's one thing for us to be able to train a model that is competitive with the with a closed Source it's another thing that we need to have an incentive to do that right so the this whole infrastructure with the being able to train an encrypted model uh and so so I guess there are SE there are several interesting aspects of it here which are get progressively well um more and more interesting right so so so first I guess is imagine you have an entity right so you are you are an entity which has expertise to train the model you want that model to be monetizable but you want it to be open source right I guess it could be because you have value uh you know like it's against your values to publish a model that is not open source or uh it might be that you know you you know you're you're you're you are an entrepreneur you don't care about the world but you know publishing it as an open source gives you uh you know recognition in the community right so in this case you have your own Hardware you don't need to train it in in encrypted way right so you train you have like your cluster of U age 100s uh or or that that cluster could be uh hosted elsewhere like on exhibit or hyperbolic or you name it but they running inside of the kees right so your model is not encrypted um for you meaning that you have the weights but while it is being trained it is trained in an encrypted fashion so so you know the actual providers of the hardware do not see your model as you train it but after it is trained you have the weights you can host it on your API or you can host it inside of the tees with payment channels for everybody right but then you can go one level deeper and say well instead of having that single entity we have a sort of a DA right so if if you if you in Vision training um it is it is hard to tell right now if the models will continue growing right but if you envision that you have to train models which are significantly bigger than the existing ones it could be that no single entity in the open Community is capable of having resources to train it right then you can Envision um a sort of a smart contract which accumulates resources from multiple entities right pull resources from uh from the decentralized providers and then it trains a model that is encrypted meaning that nobody has the weights right so no single entity can run away with the model right you know and screw everybody else over who participated the model exist only in the encrypted uh fashion not a single entity in the world can decrypt it uh right but uh that model comes with a with a virtual machine Runner that can run inside of the tees uh that will be able to decrypt it within the te right so still nobody outside of the te can see it uh right and it would only run inference for you if you pay via the payment Channel and that money gets distributed back to the original smart contract right or or something like this right or like you know um has some other ideas I don't if he wants to share them in terms of how it can be monetized right but the idea here is that you can have now a massive decentralized effort where a large set of people contribute resources ideas and data right and then it trains them model uh you know on on the pulled resources and that model is a community is is a common good you know which pays back everybody who participated yeah I mean maybe to add to this actually because one of the there there's kind of combination of challenges of training you know state-of-the-art model um what I mean the big one is always been is like how to kind of curate data even you know that kind of the base data set is available you know it's just internet and there is there is available craws for people to use but it's mostly low quality so you do need to actually filter down to a high quality data that model uh would be able to train on and that that in itself you know requires kind of a fair bit of uh engineering that uh it needs to be done and right now this engineering is done inside you know meta and openai and Google and deeps but it's not actually done in open source and there's no way to um kind of there's no way to to just you know pick up from the from the kind of shelf a good solution for that and one of the challenges here is that nobody has like a good motivation to do that because if you you know if you do build such a good data set like you're not going to benefit from that uh and it does require a lot of work to you know to carefully toate it potentially run a smaller model onto top of the data to filter it out Etc and so again using this T infrastructure we can actually have you know a uh a system where somebody can curate um kind of somebody can build you know different data filtering uh procedures and submit them to run inside te so the filtering itself happens kind privately the outcome is private and this outcome of a data that then later used again requires a payment right so like whoever then later wants to train on top of this outcome needs to pay to whoever Crea data right so effectively every artifact through the process of training is now monetizable and kind of you can create a system around it now one of the ways to monetize kind of in general in crypto is launch a token right and so even small artifacts like this where you effectively took for example all off all of a common crawl can be you know you can have a token for this specific data filtering and then you require that like you again you attach to this artifact some logic and say if you're going to use my data for training you need to for example you know burn a bunch of my token just as an example right so you have like all this kind of additional logic now that you can attach to any intermediate artifact and the whole process as a as a whole and so indeed you can have a dow that you know uh kind of collects dollars and and gives you know the like model token you can almost imagine this like initial model offering um and then it spends it dollars uh on you know GPU maybe on data processing maybe on uh indeed like knowledge and expertise of Specialists who actually uh you know code it up and then uh it can receive you know the payments right which presumably are in dollars or in near um from this model being used in production through the whole system as well as independently if somebody just wants to use model for their own use case and it can decide you know to distribute it to the uh model token or you know burn it or you know reinvest maybe into next one so this model token just becomes actually a model series so you know this can be you know o can stand for uh so it's a dow AI so you know decentralized autonomous open AI uh system so again this is just examples of what you can build on top of this infrastructure and kind of the way uh the ways to uh to build this and so again the idea here is that we want to partner to enable this kind of use cases and findings both Specialists as well as uh building out some of the learning education uh for kind people who just coming into the space there's more and more people interested in how to do this um to really enable this kind of community built uh and entrepreneur built uh systems with monetization which then flow back into this whole new internet of kind of AI first uh assistants so yeah if you are in model building or interested join us here if you you know already building models and and interested in how to monetize them uh join us in tcloud if you are running a service or planning to build a service uh join us in building effectively agents for the services if you are really interested in how to solve the discovery problem we want talk to you uh if you're really excited about how to build the best assistant in the world uh join us and if you're running an application that it has users and growing and you are interested in how to intr this new internet to your users and effectively enable your application to become a super app this is where all of this infrastructure comes together and really plugs in into your application and you actually get a lot of this for free because it is monetized through the transactions in the process not you know by paying $20 a month you want to Red without s was siren heard I don't know on the mic oh J if you're listening have you heard the S like in the last me okay sounds great right anything else we should add no I think that's a good great finish that's good all right just say thank you thank you all right well uh on this we're gonna finish this session join us uh for you know we have this every week either talking with projects or we partnering on different angles projects that doing really cool stuff or I learn more about what we're working on and follow _ AI for more news","transcript_synced_at":"2026-11-03T05:09:19.483Z","segments":[{"start_seconds":0,"text":"future hand in in this Tech driven Wonder hello everyone welcome back to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=0s"},{"start_seconds":8,"text":"near AI office hours uh today with me Alex and Eric um all of us are making","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=8s"},{"start_seconds":16,"text":"really good progress in neari and today we want to go over a different topics we're working on as well as uh outline","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=16s"},{"start_seconds":23,"text":"kind of how all of this fits together and obviously invite you to contribute and participate in what we're building so maybe before we go into what we're","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=23s"},{"start_seconds":32,"text":"doing we're going to talk about why we're doing it and kind of how the We Believe future will look like so the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=32s"},{"start_seconds":40,"text":"first part is we believe that AI is going to be a fundamental shift in how every single interaction with you know","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=40s"},{"start_seconds":48,"text":"computers internet and even economy as a whole is going to work and through that","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=48s"},{"start_seconds":56,"text":"um a lot of the kind of small things you know like you know where what like how you ordering things how you deciding on","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=56s"},{"start_seconds":64,"text":"where to go Etc will change but also big things how you know even governments and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=64s"},{"start_seconds":70,"text":"industries run uh because AI can not just kind of automate some processes but","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=70s"},{"start_seconds":77,"text":"actually start making decisions and go from kind of this traditional process where we usually delegate decision-","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=77s"},{"start_seconds":84,"text":"making to other people and companies uh now you can have decision- making happening on your side personalized for you given all of the information around","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=84s"},{"start_seconds":94,"text":"the world and all of the other participants around the world so how do we get there right the like this world","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=94s"},{"start_seconds":101,"text":"is you know um pretty complex right now and there will be a lot of kind of transition period before we get to that","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=101s"},{"start_seconds":109,"text":"future and so the first step is what we call assistant so let Eric kind of dive","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=109s"},{"start_seconds":115,"text":"into that sure uh thanks so um really trying to Envision what what your","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=115s"},{"start_seconds":122,"text":"interaction with the internet's going to look like in the future and so we we think basically there's going to be in a same sort of thing you see in sci-fi","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=122s"},{"start_seconds":130,"text":"there's going to be a an assistant sort of works for you knows everything it can about you um knows your preferences knows your general needs knows your","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=130s"},{"start_seconds":139,"text":"information knows your history and when you need something or you want something whether that's um whether that's like hey I want to replace a broken mug in my","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=139s"},{"start_seconds":147,"text":"house or just give me some content to scroll through or tell me today's news or every week make sure like make sure that I wake up tomorrow on time like it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=147s"},{"start_seconds":156,"text":"will figure out how to do that and give you an answer um and we expect that like the world of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=156s"},{"start_seconds":164,"text":"individual apps will kind of just fade away over time so uh that assistant can be great but it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=164s"},{"start_seconds":172,"text":"can't uh won't be able to do everything by itself it's going to know as much as possible about you but it won't know as much as everything about what it can go","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=172s"},{"start_seconds":181,"text":"and accomplish in order to do that it's going to have to talk to other AI um","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=181s"},{"start_seconds":189,"text":"that's what we call Service agents basically your AI assistant represents","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=189s"},{"start_seconds":196,"text":"you it's going out there on the Internet or in the real world there helping things happen in the real world for you uh in order to accomplish those things","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=196s"},{"start_seconds":204,"text":"it has to go communicate with other AI systems that will represent either other people uh other services businesses what","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=204s"},{"start_seconds":212,"text":"have you and those will help you get something","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=212s"},{"start_seconds":221,"text":"done what's what's that a factory oh great so each of these ends up being a","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=221s"},{"start_seconds":228,"text":"service agent which could just be somebody else's assistant really um it's just a matter of uh when you're yeah it's a matter of point of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=228s"},{"start_seconds":237,"text":"view um in order to figure out how it's going to get something done we have a third category of AI agent we envision","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=237s"},{"start_seconds":244,"text":"that we call a discovery agent um that's something that the assistant can give a more General sort of intent to I'd like","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=244s"},{"start_seconds":252,"text":"to accomplish this thing I'd like this information I don't have it myself can you either go out and find it for Me by","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=252s"},{"start_seconds":260,"text":"understanding the directory of all the service agents out there and just going querying for one or perhaps broadcasting to a to a to to an area of like hey my","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=260s"},{"start_seconds":270,"text":"user needs something done who can help me or just um sort of introducing your assistant to an agent that can go solve","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=270s"},{"start_seconds":278,"text":"the problem maybe just to run a specific example or a few specific examples right","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=278s"},{"start_seconds":285,"text":"um and kind of show the breadth of this right this can be something where hey I want you know to buy uh sleep tracker","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=285s"},{"start_seconds":294,"text":"right and right now you would you know Google then maybe you go to Amazon then you maybe like in my case would Google","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=294s"},{"start_seconds":302,"text":"again because you're like well I don't trust Amazon reviews so I'm going to go check between few different products which were like people reviewed and so","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=302s"},{"start_seconds":310,"text":"you kind of go in that process right and so you can imagine you know the same thing well first of all assistant probably knows that you don't sleep very well so it already predicts that you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=310s"},{"start_seconds":319,"text":"need a sleep tracker and to improve this uh to get more information about you and recommend you you know different sleeping pattern so it can issu this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=319s"},{"start_seconds":328,"text":"query okay I need a sleep track uh like who are actually selling this kind of um equipment the discovery","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=328s"},{"start_seconds":337,"text":"agent uh you know knows about two different stores so it either like introduces you to them and so you can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=337s"},{"start_seconds":345,"text":"connect to them directly say Hey you know I'm looking for sleep tracker uh one of them is responding hey you know","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=345s"},{"start_seconds":352,"text":"this is the one this is um how much it's going to cost maybe you also got a you know some review a agents or maybe you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=352s"},{"start_seconds":360,"text":"actually just your AI assistant get connected to somebody else's AI assistant who is kind of similar profile uh who you effectively getting like","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=360s"},{"start_seconds":368,"text":"feedback as well about the uh the sleep tracker you considering buying this also","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=368s"},{"start_seconds":375,"text":"works in cases where for example you're actually in professional setting and you're trying to accomplish some task","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=375s"},{"start_seconds":383,"text":"for example run a marketing campaign you know hire someone again you can go you say hey I'm planning to run this marketing campaign I want somebody to do","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=383s"},{"start_seconds":391,"text":"this for me it knows few different marketing agents uh who then give you quotes on how much this marketing campaign would cost uh how would they","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=391s"},{"start_seconds":400,"text":"run it U assistant you know evaluates it either decides directly or asks you to choose from one of the options so it's a","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=400s"},{"start_seconds":408,"text":"pretty versatile kind of system that you know can go from your kind of basic needs you know food and shelter to you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=408s"},{"start_seconds":417,"text":"know more advanced things to you know business case and kind of uh Global Commerce as","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=417s"},{"start_seconds":423,"text":"well um what connects all these little boxes is that they are all AI agents and that's a word that everybody in the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=423s"},{"start_seconds":431,"text":"world at least in the AI space is using and nobody has the same definition for us so we're going to give it another shot um what we mean by agents in this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=431s"},{"start_seconds":439,"text":"context is that they are um they are Services backed by AI inference that uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=439s"},{"start_seconds":446,"text":"that that can be talked to um now in this case of the assistant is generally the user that's talking to that agent in the case of Discovery here","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=446s"},{"start_seconds":455,"text":"it's usually your assistant that's talking to it you could also co- talk to it directly but mostly your assistant's going to be doing that same thing goes","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=455s"},{"start_seconds":462,"text":"for these Service agents backed by Services your assistant will talk to them you could skip it but we think in the future you're mostly going to be","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=462s"},{"start_seconds":471,"text":"mediating your interactions with the rest of the internet through this assistant um as opposed to uh what you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=471s"},{"start_seconds":480,"text":"see a lot of development right now is in things that are called agents but they're really a template for an agent","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=480s"},{"start_seconds":487,"text":"to interact with some existing service some real world um you know some website uh and that's an important component of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=487s"},{"start_seconds":496,"text":"this system but it doesn't really make an agent that you can go talk to in natural language or whether that's video","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=496s"},{"start_seconds":503,"text":"audio text what have you I mean it the the modality of it doesn't doesn't particularly matter here um","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=503s"},{"start_seconds":510,"text":"nor is this one masterful amazing thing that does everything uh we imagine that there would be multiple competing AI","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=510s"},{"start_seconds":518,"text":"assistants competing for which one is better multiple Discovery agents competing for which is better at matching your intent to uh a solution of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=518s"},{"start_seconds":527,"text":"course multiple Service agents um I mean this is just the economy of the future internet um what we're trying to build","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=527s"},{"start_seconds":534,"text":"here is um is how we're going to make this happen uh and to be clear this is is it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=534s"},{"start_seconds":541,"text":"not necessarily that we um we believe that this won't happen if we don't do it we think this is inevitable and we're","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=541s"},{"start_seconds":549,"text":"going to try to make sure that it doesn't suck by um by building this on open protocols uh decentralized","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=549s"},{"start_seconds":556,"text":"infrastructure uh use our own models so that when this happens it's not going to end up as basically uh the AOL ification","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=556s"},{"start_seconds":565,"text":"of the internet um again I think this model is going to be the future uh if we","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=565s"},{"start_seconds":571,"text":"don't do anything about it uh and say open AI builds this whole infrastructure then they get to kind of curate what the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=571s"},{"start_seconds":580,"text":"whole internet looks like for you it ends up being uh you know they end up being a gatekeeper um they can decide what information the models use they can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=580s"},{"start_seconds":588,"text":"decide how the discovery is going to work um if we build this as an open protocol and get enough people on it enough users coming in enough service","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=588s"},{"start_seconds":595,"text":"agents um a competitive Discovery system then we can uh build something that's a lot more functional and let the open","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=595s"},{"start_seconds":603,"text":"world sort of swallow the curated World me that yeah so so now the question is okay","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=603s"},{"start_seconds":611,"text":"how do we what is the protocol effectory that connects all those things right because I mean you can imagine this running indeed can be running on open","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=611s"},{"start_seconds":619,"text":"infrastructure which is what we're you know really pursuing it can also still be running on you know Clos infrastructure uh we just want it to be","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=619s"},{"start_seconds":628,"text":"connected into one protocol and so this is where uh we're really kind of proposing this uh protocol called aitp","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=628s"},{"start_seconds":636,"text":"it's agent interaction and transaction protocol to allow us to really interconnect all of the agents in kind","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=636s"},{"start_seconds":643,"text":"of structured format and open up you know ways to pay to transact to share credentials to share information to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=643s"},{"start_seconds":650,"text":"share kind of state and sessions uh between all of the different participants of the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=650s"},{"start_seconds":658,"text":"protocols let's so yeah yeah so so uh so what what is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=658s"},{"start_seconds":667,"text":"required here right well there is kind of a fundamental um uh piece of piece of data that is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=667s"},{"start_seconds":676,"text":"needed","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=676s"},{"start_seconds":690,"text":"um which uh okay so what do we need to actually achieve this well we need a fundamental kind of object that","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=690s"},{"start_seconds":697,"text":"everybody will share and is able to interact with and so we actually are taking on um an API that already pretty","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=697s"},{"start_seconds":706,"text":"widely used uh for AI assistance U by open Ai and it's effectively a thread","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=706s"},{"start_seconds":714,"text":"API and so thread is right now um effectly a chat his history between you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=714s"},{"start_seconds":720,"text":"know few different actors right this can can be you know user this can be another agent multiple agents and this can be uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=720s"},{"start_seconds":730,"text":"some uh you know some tools calling and a bunch of other stuff and so the idea","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=730s"},{"start_seconds":737,"text":"here is uh we want to add more uh kind of structur uh kind of optionality to allow","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=737s"},{"start_seconds":747,"text":"for various interactions and we want to allow This Thread to be again not a centralized thread where you know you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=747s"},{"start_seconds":755,"text":"need to go to only one service and only you know through that service you can interact with with it which is kind of current state across the industry but","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=755s"},{"start_seconds":763,"text":"actually a thread that multiple different participants can actually share interact with join leave you know add somebody else and so make it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=763s"},{"start_seconds":772,"text":"effectively similar to how email works right you can imagine you know right now you're effectively going to open Ai and you're talking to open AI kind of model","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=772s"},{"start_seconds":781,"text":"or agents or cloud or any other system or kind of this new system where you know U System effec creates a you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=781s"},{"start_seconds":790,"text":"know a thread with a discovery agent Discovery agent can add another agent to the thread and maybe remove itself you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=790s"},{"start_seconds":797,"text":"know put it in BCC and kind of continue like you now continue conversation with the new agent so very similar to kind of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=797s"},{"start_seconds":804,"text":"Emil protocol and now we adding a set of properties to it uh to expand it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=804s"},{"start_seconds":814,"text":"y so I mean it's it's it's a pretty simple idea for a protocol it just sort of standardizes how an agent is going to be","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=814s"},{"start_seconds":822,"text":"able to talk to an agent in the same way that a user can talk to directly to an agent and what are the what are the standard capabilities that um an agent needs when","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=822s"},{"start_seconds":832,"text":"talking to another agent in this format one of the um for a lot of stuff like you can talk over natural language again","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=832s"},{"start_seconds":839,"text":"whether that's text voice image what have you um and that'll work fine but there are places where you want to drop into uh into structured data where you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=839s"},{"start_seconds":847,"text":"want to make sure that you're getting exactly things correctly a couple big examples of that um one big example of that is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=847s"},{"start_seconds":856,"text":"payments if if you're going to buy something from this Factory uh you want to make sure you're paying exactly the right amount for exactly the right","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=856s"},{"start_seconds":864,"text":"product that you've um that that it's agreed upon what you're going to get what the terms of that payment are what","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=864s"},{"start_seconds":871,"text":"the return policy is what the payment method is supposed to be and so we're designing um again open standard and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=871s"},{"start_seconds":880,"text":"protocol for how uh any agent can go request a payment from the agent from","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=880s"},{"start_seconds":887,"text":"from whoever is talking to it uh which essentially is just going to be a tool call um if you're familiar with llm Parlin here uh so on this thread we're","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=887s"},{"start_seconds":896,"text":"going to pass back a tool call uh to say something something like a request pay for payment um now the AI assistant there can uh if it if it is authorized","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=896s"},{"start_seconds":905,"text":"to do so if it has the means to do so um it can do that payment it can respond to that tool call uh go and pay for the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=905s"},{"start_seconds":913,"text":"thing um it also can sort of pass it further up chain okay maybe I'm not authorized maybe I want maybe I don't have any money um my user hasn't given","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=913s"},{"start_seconds":921,"text":"me any money to work with myself or any credentials for any um for any payment methods and uh and I want to ask my user to pay for you can sort of pass that","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=921s"},{"start_seconds":929,"text":"Upstream that could hop between you know just more than one agent if it needs to until it eventually hits the user who is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=929s"},{"start_seconds":937,"text":"interacting with this through some sort of UI which we didn't quite draw a box for but I'll do so","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=937s"},{"start_seconds":945,"text":"now and the UI would pop up well a payment box yeah type in your credit card number connect to your crypto","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=945s"},{"start_seconds":953,"text":"wallet whatever method uh is going to work in order to pay for the thing then gets passed all the way back Downstream to execute the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=953s"},{"start_seconds":962,"text":"transaction um on payments I think uh you know this protocol can support really any payment","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=962s"},{"start_seconds":969,"text":"method that exists whether it's Fiat or crypto um but we think this is going to be huge for adoption of crypto-based","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=969s"},{"start_seconds":976,"text":"payment rails um one main reason like right now if you're uh if you're trying to buy something on the Internet like there are","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=976s"},{"start_seconds":985,"text":"plenty of places where you can spend crypto but it's always like we still haven't come across haven't overcome those those ux problems uh then let","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=985s"},{"start_seconds":994,"text":"logistical problems you know how am I going to get money into my crypto wallet how am I going to do currency conversions um how am I going to on onramp how's the merchant going to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=994s"},{"start_seconds":1003,"text":"offramp uh these things are just getting in the way of of crypto being used for payments dayto day it just like even for me been in the crypto world for 10 years","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1003s"},{"start_seconds":1011,"text":"now I still buy things with a credit card um but so when it's the human making the decision like they're reasonably in devised to use Fiat","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1011s"},{"start_seconds":1020,"text":"payment methods right now unless they're very principled about it but if this agent is talking to this agent and they're trying to negotiate on what","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1020s"},{"start_seconds":1027,"text":"payment method to use that case they really going to want to pass a credit card number over the wire or are they going to be a lot more incentivized to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1027s"},{"start_seconds":1036,"text":"use a cryptobase payment real I think it's the ladder this is the um once agents are making these purchasing decisions on behalf of users or at least recommending them to users really","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1036s"},{"start_seconds":1045,"text":"changes the game as far as uh adoption a crypto adoption just for everyday use","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1045s"},{"start_seconds":1053,"text":"cases um so we're building out again the open payment protocol we're also building out a crypto-based payment rail","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1053s"},{"start_seconds":1060,"text":"on top of that protocol that we think is going to be uh particularly attuned to the needs of AI agents uh to the needs","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1060s"},{"start_seconds":1069,"text":"of this system um for instance uh we think you know this discovery agent deserves a little","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1069s"},{"start_seconds":1077,"text":"affiliate revenue for introducing you to a service agent we want to bake that into the protocol we believe your UI deserves um some affiliate revenue for","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1077s"},{"start_seconds":1085,"text":"getting you to the right assistant if we build this right uh this is going to incentivize more people to build Discovery agents more people to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1085s"},{"start_seconds":1093,"text":"build Service agents more people to build uis and we don't want you to just have to connect to this thing through near. website we want you to be able to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1093s"},{"start_seconds":1101,"text":"connect to this ecosystem through any number of interfaces um and those interfaces should be incentivized to try to get you to connect to this ecosystem","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1101s"},{"start_seconds":1113,"text":"uh payments isn't the only thing you can get through here um the the protocol for interagent communication aitp uh would","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1113s"},{"start_seconds":1121,"text":"require a few other things um for instance credentials passing but the more and more we put in there the more and more it gets um I mean it gets scary","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1121s"},{"start_seconds":1130,"text":"to do on the open internet uh especially with llm sitting behind here um if we're passing all this data you know if if if","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1130s"},{"start_seconds":1139,"text":"everything we're doing involves a a a call out to chat gpt's models to uh to to determine what to do next then we've","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1139s"},{"start_seconds":1146,"text":"kind of lost the plot and they can still sort of curate the internet for you and adjust what you do um","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1146s"},{"start_seconds":1154,"text":"additionally this assistant in order for this assistant to be effective it really has to be inside your brain um if you've ever worked with well the previous","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1154s"},{"start_seconds":1163,"text":"definition of an agent in the real world a travel agent a real estate agent you you hit that disconnect where um where","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1163s"},{"start_seconds":1171,"text":"it's hard to express it's either really difficult to express your true preferences and plain language to it or","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1171s"},{"start_seconds":1178,"text":"um or or you just don't necessarily trust them uh you're not going to tell them your entire life story everything you've already every done everybody","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1178s"},{"start_seconds":1186,"text":"you're trying to contact with and your true motivations uh and so in order for this really to be effective you they're going to have to be really inside your","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1186s"},{"start_seconds":1194,"text":"brain and you're going to have to trust that you can give them information and it's not going to get it's not going to get compromised it's not going to be stolen and that brings us to why we need","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1194s"},{"start_seconds":1203,"text":"trusted execution environments sure yeah go for it I've been talking","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1203s"},{"start_seconds":1210,"text":"yeah i' use a mar yeah okay so um as the internet is being","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1210s"},{"start_seconds":1219,"text":"revolutionized um people will want to be doing things that depend on their private information so today generally","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1219s"},{"start_seconds":1226,"text":"even even in the modern internet the situation which we are where we have to share our information with uh with entities that we have no reason to trust","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1226s"},{"start_seconds":1234,"text":"um is already that's already a pretty bleak situation it's going to get even worse where effectively if we envision this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1234s"},{"start_seconds":1244,"text":"future where AI revolutionizes the way we interact with everything it's one of the two Futures either it's fully controlled by the same centralized","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1244s"},{"start_seconds":1252,"text":"entities we have today right so like Google open AI you name it right so you go talk to Google and Google has a bunch of services right and so we continue giving all the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1252s"},{"start_seconds":1261,"text":"information to a few to a small number of entities that don't have your interests you know that they optimize","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1261s"},{"start_seconds":1269,"text":"for uh and right or it's an open uh network of services uh that's the future that is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1269s"},{"start_seconds":1276,"text":"more desirable but now if you imagine all the services running on random Hardware in the in the wild it's uh it's it's one thing to give your private data","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1276s"},{"start_seconds":1285,"text":"to Google it's another thing to give your private data to completely random people whom you don't know at all you definitely have no reason to trust them","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1285s"},{"start_seconds":1293,"text":"luckily there's a technology called trusted execution environment or t uh that effectively says you can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1293s"},{"start_seconds":1302,"text":"imagine three entities here interacting right there's a user I will draw him here again that's a user uh the user has some private","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1302s"},{"start_seconds":1310,"text":"information uh then there is a service that's supposed to be a little COG","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1310s"},{"start_seconds":1318,"text":"um uh the service is a the service is a program right that uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1318s"},{"start_seconds":1326,"text":"that is something that the user wants to do and it's there's a service provider there's a person who built it who provides the service and then there's a","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1326s"},{"start_seconds":1333,"text":"hardware I will draw it bigger because uh IA from behind the scene is hinting","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1333s"},{"start_seconds":1340,"text":"me to do that um right so that that is running the service right and the service and the hardware in many cases","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1340s"},{"start_seconds":1347,"text":"that would be the same but not always right because we're moving to the to the world where Hardware has to be GPU enabled with expensive gpus not every","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1347s"},{"start_seconds":1354,"text":"service provider has to be necessarily infrastructure provider right so ideally the user wants neither the hardware","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1354s"},{"start_seconds":1363,"text":"provider nor the service provider to see their private data and te is enable them to do that right so te effectively allows you uh in the environment do we","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1363s"},{"start_seconds":1371,"text":"have a green marker no blue somewhere yeah okay so so blue will be green","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1371s"},{"start_seconds":1379,"text":"right so red here means um evil right so you can imagine that within this Hardware controlled by an adversary","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1379s"},{"start_seconds":1387,"text":"there's a vault uh uh that is such that user can verifiably confirm that everything that","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1387s"},{"start_seconds":1394,"text":"executes inside the Vault nobody can see what is happening in there right so so it gives it gives two big guarantees guarantee number one","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1394s"},{"start_seconds":1403,"text":"uh it says so user can inspect the the the program that the service provider provid right and the te can guarantee","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1403s"},{"start_seconds":1410,"text":"that it is that program that is being executed right so the user knows that their data is processed in a very specific way and the second guarantee is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1410s"},{"start_seconds":1418,"text":"that nobody not the hardware owner not the service provider can see what is happening inside of the te right they cannot see the data right so this way","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1418s"},{"start_seconds":1426,"text":"all those agents Discovery you know like shopping agent you name it uh you know all the like your your your psychotherapist the all running travel","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1426s"},{"start_seconds":1435,"text":"agent inside of those tees right and this way a lot of people can provide interesting agents which require to see","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1435s"},{"start_seconds":1442,"text":"the private data to be useful a lot of Hardware providers can be hosting those agents right and users can be using them being fully certain that you know their","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1442s"},{"start_seconds":1450,"text":"data is safe and their data is used in a way that you know they want like like it's not some sort of you know if they sending it to psychotherapist agent they","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1450s"},{"start_seconds":1458,"text":"know the psychotherapist agent is actually trying to help them not to upsell uh you know future future sessions right uh and here you know","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1458s"},{"start_seconds":1467,"text":"coming back to Eric's point ideally all of those um you know the hardware probably needs a littleit a little cut of the you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1467s"},{"start_seconds":1475,"text":"know of the transaction the service provider right and you would imagine uh it is possible that you know it's running on the hardware like a CPU","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1475s"},{"start_seconds":1483,"text":"Hardware but then there's an inference server somewhere Etc right like like thank thanks to the fact that with the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1483s"},{"start_seconds":1490,"text":"payment channel the payment the cost of the payment is effectively zero right you you can have arbitrarily complex","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1490s"},{"start_seconds":1497,"text":"interactions of uh multiple of a large number of small components that all you know pay each","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1497s"},{"start_seconds":1504,"text":"other on the way you can imagine even the networking part of it uh to have like you know relays and whatnot that guarantee","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1504s"},{"start_seconds":1512,"text":"liveness you know and they also get a little cut and you know thanks to payment channels all of those cuts combined still going to be less than 3%","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1512s"},{"start_seconds":1519,"text":"that you pay to visa and Master Card right so it's a yeah Beauty yeah so maybe wor drawing the the kind of this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1519s"},{"start_seconds":1528,"text":"idea aide that yeah effectively you have this service agents and uh you know your","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1528s"},{"start_seconds":1535,"text":"assistant and so to service this securely uh you have this cloud of uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1535s"},{"start_seconds":1543,"text":"provider of of compute providers right which which are combined like some of them are GPU enabled some of them are","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1543s"},{"start_seconds":1551,"text":"just you know potentially CPU um although we probably move into a world where everything will be you know one where not","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1551s"},{"start_seconds":1558,"text":"have uh accelerated mathematical co-processors uh for AI inference but U","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1558s"},{"start_seconds":1565,"text":"you can have them effec running the services meaning they're hosted there and you have you know a layer as well on","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1565s"},{"start_seconds":1574,"text":"top that allows you to ensure that you know if one of the nodes goes down it kind of rebalances to another one uh the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1574s"},{"start_seconds":1581,"text":"workloads you know there is kind of uh encryption and payments are managed through this so this is kind of the like","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1581s"},{"start_seconds":1588,"text":"a concept of a te Cloud uh which requires effectively uh a number of engineering challenges right obviously","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1588s"},{"start_seconds":1597,"text":"the te infrastructure and we just uh with Falla released the kind of first version of an SDK to build uh especially","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1597s"},{"start_seconds":1605,"text":"for GPU enabled nodes so you can actually have llm inference in a private and verifiable way the CPU uh where you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1605s"},{"start_seconds":1615,"text":"can run you know the agent like business logic and kind kind of payment and routing and everything there's another uh piece of work around kind of this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1615s"},{"start_seconds":1623,"text":"payment channels and liveness guarantees um we've been calling proof of response which enables to uh identify when nodes","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1623s"},{"start_seconds":1633,"text":"are not responding you know them to actually lose some uh uh Capital through","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1633s"},{"start_seconds":1640,"text":"slashing and through that enable you know rebalancing as well so this's kind of a whole system that's required to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1640s"},{"start_seconds":1647,"text":"power this new internet uh in a private and uh verifiable way now the interesting thing is right","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1647s"},{"start_seconds":1655,"text":"as we are actually going toward this um Direction you have the idea that uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1655s"},{"start_seconds":1662,"text":"right now if somebody you know build a model um","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1662s"},{"start_seconds":1670,"text":"yeah um so similarly how with agents right somebody Builds an agent uh it they they","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1670s"},{"start_seconds":1678,"text":"want two things one they want it to be discovered and second used and they want","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1678s"},{"start_seconds":1685,"text":"to get paid right and so uh Discovery agent is enables the discovery hook up","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1685s"},{"start_seconds":1692,"text":"into assistant and kind of this aitp enables the payments uh for uh for the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1692s"},{"start_seconds":1698,"text":"usage now it's the same thing is going about underly models right let's say I have some specialized data and I want to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1698s"},{"start_seconds":1707,"text":"build an agent that uses specialized data so I fine tun the model kind of in a traditional way let's say I'm building","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1707s"},{"start_seconds":1713,"text":"you know the most amazing therapist out there and so now I have this model and kind of traditionally you have","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1713s"},{"start_seconds":1723,"text":"two routes one is you uh put it in huging face and effectively everybody's using it for free or you are running it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1723s"},{"start_seconds":1732,"text":"yourself and building a whole kind of you know all of this infrastructure yourself to offer a already final product or at least API","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1732s"},{"start_seconds":1741,"text":"to somebody else to do it and again we getting into all the same problems which is like if you're running a therapist it has access to your private data now you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1741s"},{"start_seconds":1749,"text":"need to guarantee to me that you're not uh losing it Etc and so this same infrastructure actually allows a few","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1749s"},{"start_seconds":1757,"text":"interesting things well first of all you can have the model you trained and you just upload encrypted weights that only","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1757s"},{"start_seconds":1764,"text":"decryptable inside te so what this allows you to do is now everyone uh if like if somebody's building a therapist","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1764s"},{"start_seconds":1772,"text":"agent or even you doing this you can just Loop and use that model and you have guarantees that only this agent is using this model and you can monetize it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1772s"},{"start_seconds":1781,"text":"or you know the model itself can charge uh for using through the GPU to part of this uh payment Channel now even One","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1781s"},{"start_seconds":1789,"text":"Step Beyond that maybe you know if you're therapist yourself you know you have a lot of knowledge about the space but uh you probably don't have a GPU","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1789s"},{"start_seconds":1796,"text":"lineing around to F two the model well you can actually find tune the model on private data directly inside te right you can upload private","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1796s"},{"start_seconds":1805,"text":"data and say hey fine-tune this you know the result will be kind of this encrypted weight model that then can be used in this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1805s"},{"start_seconds":1813,"text":"system so you're effectively getting to kind of this new way of effectively monetization for models now one step","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1813s"},{"start_seconds":1823,"text":"further right you can pre-train a model in the same way right right now you know we","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1823s"},{"start_seconds":1830,"text":"have organizations you know meta a deep seek who have done this free training right they spend Fair bit of money and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1830s"},{"start_seconds":1839,"text":"they open source the model and again it it's great it's amazing we can all use it uh only if we use their specific API","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1839s"},{"start_seconds":1848,"text":"they're actually getting paid right uh now at the same time it's actually not an open source model it's open parameters model I think we talked about","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1848s"},{"start_seconds":1856,"text":"this like in one of the first uh kind of office hours the there's only like few really truly open source models where","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1856s"},{"start_seconds":1865,"text":"you actually know what data went in and even then you cannot verify that that actually data and this parameters are like correspond to each other like","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1865s"},{"start_seconds":1874,"text":"unless you literally retrain the model and you actually will get the different model anyway so so it's actually really hard to go like hey this is the data","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1874s"},{"start_seconds":1881,"text":"that went in right so yeah so you have data logic that you train and you have wait and right now you have fact just getting weights right so we have open","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1881s"},{"start_seconds":1889,"text":"weights models there's few examples of smaller models where people have posted data and code to train and weights but","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1889s"},{"start_seconds":1898,"text":"you don't have like a prove that this weights are result of this and if you train again literally you will get the different weights because it's randomly","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1898s"},{"start_seconds":1905,"text":"initialized and you'll actually get you know different results so what this te infrastructure enables as well is","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1905s"},{"start_seconds":1913,"text":"actually the training itself can happen inside te right and you can have like a te cluster or uh there's really interesting advancements around","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1913s"},{"start_seconds":1920,"text":"distributor training that's happening and so now you actually have uh a attestation that this weights were","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1920s"},{"start_seconds":1928,"text":"trained inside te and that this data went in into this and like and was reused this on this code right so you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1928s"},{"start_seconds":1936,"text":"can actually prove that this weights for example don't contain New York Times articles or that they contain some specific data so it's a pretty kind of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1936s"},{"start_seconds":1945,"text":"powerful way toward this user only the I concept where you actually know what's going on into your model and you actually can be confident that you don't","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1945s"},{"start_seconds":1954,"text":"have you know sleeper agents which is kind of a concept of you know finding a model to react on specific uh inputs to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1954s"},{"start_seconds":1961,"text":"give an output which doesn't show up on any benchmarking doesn't show up normally but you know when you pluging it into your ID for example and type in","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1961s"},{"start_seconds":1970,"text":"import it will suggest you know a specific Library uh that has you know malicious exploit and so uh that that","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1970s"},{"start_seconds":1978,"text":"would be a one way to you know uh through open weights to actually attack a whole host of pro like uh code uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1978s"},{"start_seconds":1987,"text":"platforms um so so this actually gives you kind of this op like open source","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1987s"},{"start_seconds":1993,"text":"model now if you just again open up the weights the problem is you don't have the way to monetize it right everybody","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=1993s"},{"start_seconds":2000,"text":"can use it for free which is again great but limiting because you cannot reinvest back to get better data train a new","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2000s"},{"start_seconds":2008,"text":"model maybe improve the logic right so what you can do now again you can have an encrypted weight","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2008s"},{"start_seconds":2015,"text":"model which again only usable inside te and uh this way the whoever trained this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2015s"},{"start_seconds":2024,"text":"can Define the rules on how this model uh it can be used and how much they should be paid to do that so which again","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2024s"},{"start_seconds":2033,"text":"this can be done by distributed Community this can be done straight token or can be done by some participant you know who wants to add safety","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2033s"},{"start_seconds":2041,"text":"requirements whatever that is right this can now effectively has some additional logic added uh to how this weights can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2041s"},{"start_seconds":2049,"text":"be used and so only usable inside this te uh environment as well just to um accentuate that point","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2049s"},{"start_seconds":2057,"text":"uh what right now when you uh deliver an open weight model it's essentially a data file here's the weights um what an","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2057s"},{"start_seconds":2066,"text":"encrypted weight model would look like is essentially um a run like it's it's the data in in encrypted form and also the program that you must run in order","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2066s"},{"start_seconds":2074,"text":"to run inference you can download and run that locally if you have a GPU farm with a that's te enabled um but even on","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2074s"},{"start_seconds":2082,"text":"your own machine you're not a to just decrypt and extract the weights the only way to access it is through the program that wraps it and enforces that um","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2082s"},{"start_seconds":2091,"text":"you're running it on the te and just crypt the weights this lets you do and this is useful um if your data set is closed","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2091s"},{"start_seconds":2100,"text":"this is also useful if your data set is open um you could have a system where the input data and all the logic used to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2100s"},{"start_seconds":2108,"text":"train it is open but maybe the training is going to take three months and a few million dollars um and you want to recoup that investment you could still","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2108s"},{"start_seconds":2117,"text":"have an encrypted weight model and make that make sense yeah or or you training it on you know some uh biochemist tree for example","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2117s"},{"start_seconds":2126,"text":"and you don't want it to be used to by weapons so you add the logic to protect against that usage in in this way as well yeah but you still want you know to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2126s"},{"start_seconds":2134,"text":"give it to students and to professionals to enhance their work yeah let's just put you know put restrictions but let","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2134s"},{"start_seconds":2142,"text":"you put restrictions on your model uh without making it you have to hide behind a centralized API you have to know to run that all infrastructure yourself you could still run it yourself","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2142s"},{"start_seconds":2151,"text":"but it doesn't make it a necessity yeah so so all of this effectively kind of becomes a stat that uh we","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2151s"},{"start_seconds":2160,"text":"are working with a lot of Partners to build so obviously this is a lot of infrastructure and a lot of uh kind of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2160s"},{"start_seconds":2168,"text":"logic and and and things to build so near is pretty small team so we are working with a lot of Partners to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2168s"},{"start_seconds":2175,"text":"actually make this happen right so T infrastructures I said working with Fallout to enable this uh we have Partners who are helping with the proof","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2175s"},{"start_seconds":2183,"text":"response for the um for the liveness guarantees and uh some of the payment Channel infrastructure uh that uh we're","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2183s"},{"start_seconds":2191,"text":"going to be releasing soon we have you know kind of the protocol that uh we're going to be releasing then that we want","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2191s"},{"start_seconds":2198,"text":"to welcome other people to build variety of services agents and uh ESP special discovery and Service agents to build on","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2198s"},{"start_seconds":2206,"text":"top and obviously this AI assistant we're going to have a prototype kind of a MVP of this but we want to uh open it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2206s"},{"start_seconds":2213,"text":"up for everyone to participate and so kind of if the like we want to welcome everyone to really contribute to this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2213s"},{"start_seconds":2220,"text":"because again as Eric mentioned this is a future that's happening it's you know inevitable kind of how the world was","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2220s"},{"start_seconds":2227,"text":"restructure around AI being the facilitator of your kind of uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2227s"},{"start_seconds":2234,"text":"interactions and it's now really about kind of the open protocols you know and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2234s"},{"start_seconds":2240,"text":"uh open open ecosystem to win over you know one of the closed ecosystems uh to actually you know deliver all this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2240s"},{"start_seconds":2249,"text":"without privacy without kind of uh opportunity for other people to participate um yeah um yeah and we're","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2249s"},{"start_seconds":2259,"text":"not just doing this like we're not just asking everybody to contribute to this ecosystem as a favor like there are we're trying to make sure that there's incentive structures in place in order","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2259s"},{"start_seconds":2267,"text":"to create this open ecosystem this new internet um and do it in a way that's going to be natural so with um you know","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2267s"},{"start_seconds":2275,"text":"building these Service agents that should be a good business that's that's your apps of the future uh building a discovery agent that's that's Google of the future that's App Store of the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2275s"},{"start_seconds":2283,"text":"future building an assistant like these are all things that are going to be um financially rewarding and good for the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2283s"},{"start_seconds":2290,"text":"good for the world good for your users um similarly over on the model training side like with encrypted weights models","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2290s"},{"start_seconds":2297,"text":"and with the payment channels you can monetize those models that means that you can have code within the download","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2297s"},{"start_seconds":2304,"text":"code within the the model that enforces that you are uh paying for usage and that can go right back to the model developer to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2304s"},{"start_seconds":2312,"text":"anybody who's contributing data anybody's contributing logic um it really makes it so that uh if you feel","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2312s"},{"start_seconds":2320,"text":"like open models or even locally run models are important you don't have to just give away all your work um incredibly powerful concept that we","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2320s"},{"start_seconds":2327,"text":"think can sort of naturally incentivize the creation of this new open internet um which I think brings us to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2327s"},{"start_seconds":2335,"text":"model training okay I don't know how much deeper we can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2335s"},{"start_seconds":2347,"text":"dive here um the I guess everybody by now knows how we cannot see ourselves anymore but that's chill everybody by","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2347s"},{"start_seconds":2355,"text":"now has seen uh you know deep seek we had a reading group yesterday right so we","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2355s"},{"start_seconds":2362,"text":"still a week ago yes right a week ago and um uh and so we still live in this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2362s"},{"start_seconds":2371,"text":"beautiful present where entities still have an incentive to throw uh away millions of dollars to train a model and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2371s"},{"start_seconds":2379,"text":"publish it right for the for the secondary effects but that does that will not necessarily last right and so","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2379s"},{"start_seconds":2385,"text":"for uh for open source Community to continue uh so to to have a certainty","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2385s"},{"start_seconds":2393,"text":"that open source Community will continue being competitive with the Source uh you need first of all we need expertise in","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2393s"},{"start_seconds":2402,"text":"the wild how to train those models right so it's great that deeps and meta can do that but it would be great if we have uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2402s"},{"start_seconds":2409,"text":"open available knowledge how to train those models right so that's something interesting that um we're working towards and as Eric mentioned you do","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2409s"},{"start_seconds":2418,"text":"need to be able to to monetize those models right so so it's one thing for us to be able to train a model that is competitive with the with a closed","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2418s"},{"start_seconds":2426,"text":"Source it's another thing that we need to have an incentive to do that right so the this whole infrastructure with the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2426s"},{"start_seconds":2434,"text":"being able to train an encrypted model uh and so so I guess there are SE there are several interesting aspects of it","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2434s"},{"start_seconds":2440,"text":"here which are get progressively well um more and more interesting right so so so first I guess is imagine you have an entity right so you are you are an","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2440s"},{"start_seconds":2449,"text":"entity which has expertise to train the model you want that model to be monetizable but you want it to be open source right I guess it could be because","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2449s"},{"start_seconds":2456,"text":"you have value uh you know like it's against your values to publish a model that is not open source or uh it might be that you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2456s"},{"start_seconds":2465,"text":"know you you know you're you're you're you are an entrepreneur you don't care about the world but you know publishing it as an open source gives you uh you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2465s"},{"start_seconds":2474,"text":"know recognition in the community right so in this case you have your own Hardware you don't need to train it in in encrypted way right so you train you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2474s"},{"start_seconds":2482,"text":"have like your cluster of U age 100s uh or or that that cluster could be uh hosted elsewhere like on","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2482s"},{"start_seconds":2490,"text":"exhibit or hyperbolic or you name it but they running inside of the kees right so your model is not encrypted um for you meaning that you have the weights but","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2490s"},{"start_seconds":2498,"text":"while it is being trained it is trained in an encrypted fashion so so you know the actual providers of the hardware do not see your model as you train it but","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2498s"},{"start_seconds":2507,"text":"after it is trained you have the weights you can host it on your API or you can host it inside of the tees with payment channels for everybody right but then","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2507s"},{"start_seconds":2517,"text":"you can go one level deeper and say well instead of having that single entity we have a sort of a DA right so if if you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2517s"},{"start_seconds":2523,"text":"if you in Vision training um it is it is hard to tell right now if the models will continue growing right but if you envision that you have to train models","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2523s"},{"start_seconds":2531,"text":"which are significantly bigger than the existing ones it could be that no single entity in the open Community is capable of having resources to train it right","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2531s"},{"start_seconds":2539,"text":"then you can Envision um a sort of a smart contract which accumulates resources from multiple entities right pull resources from uh from the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2539s"},{"start_seconds":2548,"text":"decentralized providers and then it trains a model that is encrypted meaning that nobody has the weights right so no single entity can run away with the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2548s"},{"start_seconds":2556,"text":"model right you know and screw everybody else over who participated the model exist only in the encrypted uh fashion not a single entity in the world can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2556s"},{"start_seconds":2565,"text":"decrypt it uh right but uh that model comes with a with a virtual machine Runner that can run inside of the tees","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2565s"},{"start_seconds":2574,"text":"uh that will be able to decrypt it within the te right so still nobody outside of the te can see it uh right and it would only run inference for you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2574s"},{"start_seconds":2582,"text":"if you pay via the payment Channel and that money gets distributed back to the original smart contract right or or something like this right or like you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2582s"},{"start_seconds":2590,"text":"know um has some other ideas I don't if he wants to share them in terms of how it can be monetized right but the idea here is that you can","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2590s"},{"start_seconds":2598,"text":"have now a massive decentralized effort where a large set of people contribute resources ideas and data right and then","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2598s"},{"start_seconds":2606,"text":"it trains them model uh you know on on the pulled resources and that model is a community is is a common good you know","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2606s"},{"start_seconds":2614,"text":"which pays back everybody who participated yeah I mean maybe to add to this actually because one of the there","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2614s"},{"start_seconds":2622,"text":"there's kind of combination of challenges of training you know state-of-the-art model um what I mean the big one is always been is like how","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2622s"},{"start_seconds":2630,"text":"to kind of curate data even you know that kind of the base data set is available you know it's just internet","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2630s"},{"start_seconds":2639,"text":"and there is there is available craws for people to use but it's mostly low quality so you do need to actually","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2639s"},{"start_seconds":2646,"text":"filter down to a high quality data that model uh would be able to train on and that that in itself you know requires","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2646s"},{"start_seconds":2655,"text":"kind of a fair bit of uh engineering that uh it needs to be done and right now this engineering is done inside you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2655s"},{"start_seconds":2662,"text":"know meta and openai and Google and deeps but it's not actually done in open source and there's no way to um kind of","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2662s"},{"start_seconds":2672,"text":"there's no way to to just you know pick up from the from the kind of shelf a good solution for that and one of the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2672s"},{"start_seconds":2680,"text":"challenges here is that nobody has like a good motivation to do that because if you you know if you do build such a good","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2680s"},{"start_seconds":2688,"text":"data set like you're not going to benefit from that uh and it does require a lot of work to you know to carefully toate it potentially run a smaller model","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2688s"},{"start_seconds":2696,"text":"onto top of the data to filter it out Etc and so again using this T infrastructure we can actually have you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2696s"},{"start_seconds":2704,"text":"know a uh a system where somebody can curate um kind of somebody can build you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2704s"},{"start_seconds":2712,"text":"know different data filtering uh procedures and submit them to run inside te so the filtering itself","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2712s"},{"start_seconds":2720,"text":"happens kind privately the outcome is private and this outcome of a data that then later used","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2720s"},{"start_seconds":2727,"text":"again requires a payment right so like whoever then later wants to train on top of this outcome needs to pay to whoever","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2727s"},{"start_seconds":2735,"text":"Crea data right so effectively every artifact through the process of training is now monetizable and kind of you can create a system around it now one of the","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2735s"},{"start_seconds":2744,"text":"ways to monetize kind of in general in crypto is launch a token right and so even small artifacts like this where you","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2744s"},{"start_seconds":2751,"text":"effectively took for example all off all of a common crawl can be you know you can have a token for","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2751s"},{"start_seconds":2760,"text":"this specific data filtering and then you require that like you again you attach to this artifact some logic and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2760s"},{"start_seconds":2767,"text":"say if you're going to use my data for training you need to for example you know burn a bunch of my token just as an","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2767s"},{"start_seconds":2775,"text":"example right so you have like all this kind of additional logic now that you can attach to any intermediate artifact","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2775s"},{"start_seconds":2782,"text":"and the whole process as a as a whole and so indeed you can have a dow that you know uh kind of collects","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2782s"},{"start_seconds":2791,"text":"dollars and and gives you know the like model token you can almost imagine this like initial model offering um and then","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2791s"},{"start_seconds":2800,"text":"it spends it dollars uh on you know GPU maybe on data processing maybe on uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2800s"},{"start_seconds":2807,"text":"indeed like knowledge and expertise of Specialists who actually uh you know code it up and then uh it can receive","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2807s"},{"start_seconds":2817,"text":"you know the payments right which presumably are in dollars or in near um from this model being used in production","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2817s"},{"start_seconds":2825,"text":"through the whole system as well as independently if somebody just wants to use model for their own use case and it can decide you know to distribute it to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2825s"},{"start_seconds":2833,"text":"the uh model token or you know burn it or you know reinvest maybe into next one so this model token just becomes","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2833s"},{"start_seconds":2840,"text":"actually a model series so you know this can be you know o can stand for","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2840s"},{"start_seconds":2847,"text":"uh so it's a dow AI so you know decentralized autonomous open AI uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2847s"},{"start_seconds":2854,"text":"system so again this is just examples of what you can build on top of this infrastructure and kind of the way uh the ways to uh to build this and so","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2854s"},{"start_seconds":2863,"text":"again the idea here is that we want to partner to enable this kind of use cases and findings both Specialists as well as","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2863s"},{"start_seconds":2872,"text":"uh building out some of the learning education uh for kind people who just coming into the space there's more and more people interested in how to do this","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2872s"},{"start_seconds":2881,"text":"um to really enable this kind of community built uh and entrepreneur built uh systems with monetization which","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2881s"},{"start_seconds":2889,"text":"then flow back into this whole new internet of kind of AI first uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2889s"},{"start_seconds":2897,"text":"assistants so yeah if you are in model building or interested join us here if you you know already building models and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2897s"},{"start_seconds":2907,"text":"and interested in how to monetize them uh join us in tcloud if you are running a service or planning to build a service","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2907s"},{"start_seconds":2915,"text":"uh join us in building effectively agents for the services if you are really interested in how to solve the discovery problem we want talk to you uh","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2915s"},{"start_seconds":2924,"text":"if you're really excited about how to build the best assistant in the world uh join us and if you're running an application that it has users and","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2924s"},{"start_seconds":2933,"text":"growing and you are interested in how to intr this new internet to your users and effectively enable your application to","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2933s"},{"start_seconds":2941,"text":"become a super app this is where all of this infrastructure comes together and really plugs in into your application and you actually get a lot of this for","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2941s"},{"start_seconds":2949,"text":"free because it is monetized through the transactions in the process not you know by paying $20 a month you want to Red","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2949s"},{"start_seconds":2958,"text":"without s was siren heard I don't know on the mic oh J if you're listening","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2958s"},{"start_seconds":2968,"text":"have you heard the S like in the last","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2968s"},{"start_seconds":2976,"text":"me okay sounds great right anything else we should add no I think that's a good great finish that's good all right just","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2976s"},{"start_seconds":2986,"text":"say thank you thank you all right well uh on this we're gonna finish this session join us uh for you know we have","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2986s"},{"start_seconds":2993,"text":"this every week either talking with projects or we partnering on different angles projects that doing really cool stuff or I learn more about what we're","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=2993s"},{"start_seconds":3001,"text":"working on and follow _ AI for more news","url":"https://www.youtube.com/watch?v=ZFqfT0CGshA&t=3001s"}],"youtube_url":"https://youtube.com/watch?v=ZFqfT0CGshA","url":"https://near-pedia.xyz/videos/ZFqfT0CGshA"}