{"id":"dDXTOPRKyjg","title":"NEAR AI Office Hours #28 w Matthew Wang founder of OpenGradient","description":"Illia & Alex discuss the latest AI research topics and unpack some NEAR.AI focus areas.\r\n\r\nGuest:\r\nMatthew Wang\r\nOpenGradient\r\nhttps://x.com/0xdeltahedged\r\n\r\nhttps://www.opengradient.ai\r\nhttps://x.com/OpenGradient\n\nIn this episode of NEAR AI Office Hours, Alex speaks with Matthew, co-founder of Open Gradient, about building a decentralized AI platform for hosting models, secure inference, and application deployment. They explore how Open Gradient bridges Web3 and AI with features like trusted execution environments, zkML, and a decentralized model hub.","channel_name":"NEAR Protocol","published_at":"2024-12-10T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/dDXTOPRKyjg/hqdefault.jpg","transcript_text":"near to one we're close to won't stop us now together we'll redefine the game when near we'll find the flame empowered Minds rise and sore use your own AI we'll explore how the future hand in hand in this TX driven Wonderland okay hello everyone uh welcome to the next episode of near office hours uh today we have mat you from open gradient um interestingly a while ago I and I were in a little room with a whiteboard brainstorming uh on What U you know user own AI should look like and a lot of things that open gradient are doing is something we had on the on that whiteboard so it's going to be an exciting exciting conversation I might you do you want to take a couple words about yourself and what open gradient is doing yeah absolutely uh first of all I wanted to thank you and then your team for putting this together very excited to be here um nice to meet everyone my name is Matthew one of the co-founders of open gradient so um open gradient is uh an open intelligence Network and we're building a decentralized platform for hosting models secur C inference agentic reasoning and application deployment so uh we have you know full stack of a variety of different products that I can get into later but from a high level we're excited to be building this you know decentralized secure infrastructure that can help not just make AI more accessible from anyone you know leaning into the whole uh web 3 and blockchain Ethos but also build solutions that can allow blockchain developers to easily integrate ARL into their applications and build more optimized or more intelligent ation so yeah super excited to be here personally for me my background was in ECE I'd done stints of ml research as well and know back when I was in in school um had some experience doing things like data science at Nasa uh building ml models at Google um engineering at Facebook and I spent most of my professional career uh as a research engineer at two Sigma which is a Quant fund based in New York where I built quantitative models as well as a lot of modeling infrastructure for an equity options Market making desk and yeah I met um Adam uh my co-founder in New York who had spent uh six to seven years at Palante tier where he was a tech lead on AIP which is their AI platforms and he led um notably the development efforts on palente teers patented AI agent execution framework which I found to be very very fascinating so yeah and uh we've been building um since we met uh at the end of 2023 and we're very excited to be you know continuing to show the world what open grent is about yeah so so open gradient it consists of several things right so I I guess we can unpack each of them separately or or maybe we can we can talk about how they all fit together but uh so so one of them is the model Hub uh which is my understanding is it's a decentralized take on uh on hiding face right so so and couple things that are interesting here one is what is the what is the infrastructure behind the modelhub what is it built upon and what is the I guess another well let's be let's begin with a more why right so what's wrong with hiking face why do we need a decentralized modle hub and then what is what is powering it behind the scenes yeah I mean I would definitely not I I would definitely not say that anything's wrong with hugging face it's a product that has found tremendous product Market fit um and they've built bu a very very rich Suite of different tools and technologies that I think many tech companies today are now kind of relying on I would say um you know what the open gradient AI model Hub aims to do is to kind of fit that Niche within um the the web 3 space and make it easy for web 3 application developers to integrate ml into their applications so what that actually means in practice is we have a lot of these outof the- boox Solutions um that come with you know as an example example running AI models in tees generating zkl proofs for different model inferences that you know are necessary for a lot of applications building the blockchain space so it's kind of this um taking hugging face and what they've already built and building kind of a version of that that is you know extremely well tailored towards web 3 audience and on top of that of course leaning into the ethos of decentralization from the architecture um you can think about the model storage um you know that we have as a fork of ipfs so anyone can run a storage node and you know be able to contribute to that Network and the blockchain infrastructure that I can get into down line uh essentially interacts with this storage layer and uh We've also created some abstractions that are available on the AI model Hub uh you know as an example running you know inference through our playground or if you want to use our SDK you can interact with these models that are stored in our decentralized file store um you know utilizing our blockchain in without having to deal with the whole blockchain complexities do you don't need to build your own smart contract you don't need to kind of like access these notes so you can just from a very familiar developer kind of CLI or or python environment be able to run all these things yeah and I I went to the model Hub I saw there's a couple of uh models STL along the lines of price prediction so I see your background from uh from the past life is uh uh uh contributing to the yeah um so a little bit of uh kind of you know context on that I I think uh you know back when I was at two Sigma there was this term that we threw around a lot that we called like research velocity and the entire idea behind that term is that uh you know models especially things for like you know trading or finance a lot of the time individual models just aren't going to be that good right uh models uh their Half Lives just aren't very long some times they're overfit on data sometimes your Alpha gets crowded out sometimes model loses its Edge um so a lot of the times we have to build technologies that you know as an example do model ensembles that help reduce you know prediction uh kind of loss and and error rate of of the models also on top of that you know people are constantly building new models and iterating on them so I guess like from a high level you think about the Hub as a tool that allows people to you know take other models that people have developed and rapidly iterate on them and and that's just because we want to give people the the power to kind of continue to iterate and build better models over time you know there's not going to be one model in our opinion that trumps everything we kind of just built some sample models but um we're very excited to kind of see you know ml developers and web 3 developers alike kind of build these technologies that will gradually improve uh and the quality will get better over time and you mentioned a couple times by now the uh the blockchain so is it is it just like a set of connectors uh of the of the model Hub to the blockchain or is there anything specific uh that that like so so so it mentions here that it enables performance AI so I wonder what is the performance side of it yeah so typically in in regular blockchains which are replicated State machines every single node re-executes every single transaction and then they attest to hey you know is this the current state of the blockchain you know consensus is reached and we move on to the next block of transactions what we've done here is we've kind of fundamentally reimagined what kind of a blockchain you know what blockchain architecture might look like and we've kind of leaned into this idea of node specialization or or we call it like heterogenous compute which is creating different types of clusters of of nodes that do different things so we have as an example regular blockchain notes that execute stateful evm transactions we have um a new of nodes called uh inference nodes that uh compute or you know powerful gpus or tees designed to compute inference and generate any Associated proofs or add stations and we also have data access nodes which are tees that use TLS certificates for secure data access and then the inference noes then uh kind of uh then they kind of interface with our storage nodes which is the the uh the model store so what happens uh when someone wants to run let's say a simple inference is uh that transaction submitted the sequencer sequences the transaction and it goes into the specialized kind of simulation pipeline in which the inference piece is load balance to the cluster of inference nodes that then computes the the inference and you know generates any Associated proofs sends that back um and we embed in the state transition function the validation of those proofs and AD stations um and so this is kind of I guess you can think about it like it's done offchain except these nodes are all registered on our Network as well and participate in cryptoeconomic you know security via an igen layer ABS so if they generate garbage proofs they can get slashed and on top of that you know the validation of these proofs are natively embedded um in the St transition function as well um so this allows anyone who wants to validate uh you know uh you know inference transactions on our Network as an example to run a node and all the precom Sor all the um you know zkl or t enabled proofs will will go through their node and they can ensure that they get trust minimized access to AI so I have a couple of questions about inference but before we jump into inference guess there's a more fundamental question uh so this the the nodes you mentioned is is it is it a a brand new layer one or is it built on top of existing infrastructure and it's only the imprint noes that are unique topen gradient yeah so technically um using I guess like mainstream blockchain kind of terminology and semantic definitions we would be classified as a layer to because we use near da um so so all the you know proofs um and an ad stations in addition to the block data are stored in in that blob space um and we have like a sequencer um but you know we kind of see ourselves as kind of this Sovereign layer where uh you know we want to kind of have native AI driven applications built on top of us um and when I say AI driven appc can come this can come in many ways shapes and forms which I can elaborate you know later when we talk about use cases yeah so for inference right you mentioned that it's running you mentioned two things you mentioned trusted execution environments and you mentioned zml so I'm curious right so so my prompt like prompt goes into the system right we need we need to get the response what happens to the prompt from there why is it multiple nodes why not a single one right and uh how what sort of what is the nature of the proof that ends up on chain and is validated yeah so right now we support um I I guess I can maybe talk about what the user flow is initially so with directly you know within the the kind of precompile libraries there are options for you to choose from either um we have pessimistic zkl which is a zkl proof is generated on every single inference we have optimistic zkl which is where we only run inference um but we don't generate the proof to save on cost and you know users can then submit challenges within a Time window to have to see that proof be generated and you know we can decide to arbitrate if that proof is valid Challenger losers stake vice versa the inference don't get slashed and then we also have te at a stations um and right now we have two different types of of tees and adest stations on our Network one is um Nitro enclaves and the other one is like HW h100s their confidential compute enclaves um that were recently on boarding um so that's kind of like I guess the different levels of security that we uh we have and we allow kind of users to choose from what type of you know security they want for their inference directly from the the function call and kind of directly answering your question of why we have multiple nodes and multiple kind of like Security Solutions for this um it is ultimately allowing the users to be very very or developers in this case to be very very flexible with what best suits their intended use case as an example if they want to inference just a small you know like a gradient boost a tree model or regression model zkl works just fine and can you know be quite powerful for that right whereas for things like llm inference you know it zkl is just not a particularly practical solution due to compute costs and we want to rely on um tees and the multiple node situation comes in where we can have many different models cached on different nodes so the load balancer embedded directly in the sequencer uh you know pings these nodes every couple seconds and is able to say hey if I want to route a model inference for let's say um you know llama I know which nodes are capable of producing which kind of proofs and I also know which nodes have that model cached so this is purely just for performance optimization I see but but the particular PRP will go to a single node or will it go to multiple it would go to a single node I see and there are two things so when it comes to te te help with two things one is that uh you can actually get a proof that the output you're looking at is in fact uh a response of a model to your prompt right like you can go to the GitHub and load the virtual machine that is used compute the whatever the hashes and get the quote but it also gives you private infint gtml only gives you the form right so so if I choose to use gml my promp will be visible to the no right so for zkl you can actually um so we use Ezekiel uh their library and you can actually choose what you want to hide and choose what you want to show um there actually are different visibility settings you can either choose to to hide the model and make the inputs and outputs visible or you can choose to hide um let's say like some like output or input instead uh that you can actually there is some variability there I think right now currently we're focused on you know just you know we're not focused on privacy preserving compute just yet uh kind of one of the benefits that we see of putting this on the blockchain to begin with is the verifiability aspect but we are kind of working on some kind of privacy preserving Solutions involving both zkl and tees for use cases in which someone develops a valuable model or you know someone wants to run compute on some private data and they don't want you know certain parts of it to be exposed to the world but I think that's definitely more of a a a an ad hoc kind of um you know use case that that we have a lot of our uh kind of existing clients are kind of build you utilizing us as opposed to you know centralized solutions for that verifiability aspect and what are the what are the primary use cases for for verifiability yeah no I'm I'm glad you asked we kind of put out like this like write up on all the different use cases uh that we have I guess I can just like maybe quickly walk through a couple of them um so the biggest use case I think that we're excited about is um is Defi and we think that you know for example uh there are a lot of lwh hang fruit like Unis swap amm fee optimization research that we had put out a while ago um where we can see that hey if we use you know ml models to compute Dynamic fees we can reduce net impermanent loss for LPS and increase fee collection for LPS by up to 18% um so this is kind of preventing or mitigating effects of lvr um things such as risk models um to help improve you know yield yield aggregators as well as Vault protocols as risk adjusted returns um as well as kind of you know know risk models for Lending protocols to optimize collateralized like LTV ratios in these pools um so outside of defi there's also you know reputation models for deepend networks so we can you know outline a couple different features like you know liveness uptime things like that uh and we allow these models to dynamically score you know the the performance of different minor nodes on dpin networks and actually if you check our model Hub filecoin um we have a couple file coins this models on on our model Hub and this basically allows us to say hey you want to be a good dpin node you should get a good reputation score and we'll allocate rewards you know proportionally to your reputation score um you know we also have some interesting use cases around using ml models to price different prediction markets um things like that I can maybe dive into uh you know some of these if you're curious but a lot of these kind of use cases are high impact kind of use cases where if the model inference goes wrong and the AI or ml work is running correctly could could lead to kind of um you know implications with users's funds and as a result we want to ensure that everything is entend you know verifiable and that users get this guarantee around the compute that we provide yeah and what is the so as a user because there's a model Hub I can download the model right the models are public right so what is the um like what would be my like what are the tradeoffs between running it on top of open gradient as opposed to you know for example just running it locally on on on my own machine yeah so I mean this kind of leans into the whole uh you know we I would say two things number one is is verifiability right uh you know people want to ensure that the quality of the model is good they might you know even if they can guarantee the compute is is done correctly they want to be able to maybe download the model and run their own simulations or audit make sure that the model like you know looks reasonable um so similar to like kaggle or hungy face we provide a big platform that allows people to go and download and access these models as well and the second one is kind of like model versioning um which allows developers to constantly iterate on these models allows people to come you know upload their own versions of these models Cod so it kind of Fosters this kind of open- source Community around development of these models and on top of that you know people can build workflows that involve say models I've developed other people have developed prediction models as well we can kind of aggregate them and do like a model stacking you know Ensemble approach and ultimately create a workflow that is even more you or that leads to more accurate forecast than just a singular model yeah and right now this is primarily used uh like has there been any work towards autonomous agents that use those models uh to uh to take action on chain or is it still primarily operated by humans yeah so that's actually one of the things I'm I'm very bullish about and I I had talked to um a couple folks from near about this as well which is uh you know I kind of see AI agents uh and I would love to hear your opinion on this as well I think they're currently in a stage where um initially it's kind of a mimic right these AI agents can put out tweets on Twitter it's particular it's just interesting and the next step step for agents is they can gradually replace what humans can do um so you know I don't need a and place a trade on Unis swap if I can just tell an agent hey can you swap out you know one eth for usdc on Unis swap for me just does that for me and I think the next step for agents in this case is to be able to allow agents to complete complex tasks better than humans as an example how do I manage a portfolio right how do I take in these spot and these volatility forecasts and perform Marco width Style mean variance analysis to optimize my portfolio allocations and holdings right these are all kind of complex tasks that that uh you know a simple llm might struggle to to kind of do effectively and as a result we kind of want to build these powerful ml workflows that allows agents to leverage them to do cooler things right if you tell an agent what token or if tell an LM you know if asking what token should I buy to maximize my risk adjusted returns it'll probably say something generic like Bitcoin or like you know I don't want to provide Financial advice but in this case um you know these AI agents will then be able to leverage these ML workflows and models we've built to kind of do all these things which which is now something we're diving into and we're actually putting out um something or releasing an an agent with a double up which is this uh you know gamey slash you know gamble uh protocol onui we're building this AI agent that leverages ouru spot forecasting model to set kind of uh money uh money line lines on their protocol and allows users to bet against it um so that's going to be our initial kind of foray into the agent space and we're kind of talking to some other AI agent protocols and partners like exponent AI as an example on how they can integrate uh you know our ml models into their agents to help users you know you know do do bigger better cooler things yeah yeah that's that sounds pretty cool we have a we have this toy example called Shield GPT where people can Shield their meme coins to a to a model to A system that is powered by uh by uh I don't remember which model but uh it's not it's not particularly resilient that's exactly the point of the name right if if you sh well enough it will it will get convinced and it will buy it and it's uh the performance of its portfolio was suboptimal to say the least right so so in this case I think it's an interesting yeah if if the models so so factly it's interesting how far we are from the state where there is a AI system sufficiently sophisticated that people cannot take advantage of it right because you know on the market you play against others right if people can take advantage of your system they will right so so once the model can be sufficiently reliable that people cannot take advantage of it easily that or like an sample of models yeah I think in this case like as an example you could build in a simple like you know like a market risk or Market forecasting model into Shield GPT so as an example if you're bullish on the crypto Market at large like it's likely that a lot of these meme coins will trade as levered beta to the market and you know you can as an example say Hey you know if I am underexposed to the market I can buy some more meme coins if I'm Overexposed I can opt to be more risk averse in buying these um tokens being shied to the the agent um you know I mean I I can dive deeper into these many like kind of uh Quant Finance upgrades that you know the agent can benefit from but I'm particularly excited about kind of these more advanced um agents that can can leverage these Technologies and uh yeah um it's kind of like what we're building towards um at open gradian so so on the model Hub what are the what are the incentives for people to publish good models do do they uh is there is there an incentive structure or is it mostly just a place for people to share for free yeah so we're we're kicking off a hackathon kind of later this uh this month as well so initially it's going to be kind of like bootstrap by hey if you're a data scientist and you want to earn some some uh you know cash prizes come and build some cool models on on top of us um but I I think down the line we want to leverage uh you know both potential token incentives as well as kind of just like you know create an organic Community uh that is kickstarted by you know application and protocol developers wanting to benefit you know each other and the ecosystem at large to to kind of contribute to these models and and what I say by that is uh practically speaking you know an amm might say Hey you know we want to put out this this new model that we're working on that you know does you know dynamic V optimization some someone else might be able to take that model you know and and also put out a new one to not just better their protocol but to kind of Advance the space from a research standpoint overall um and yeah a lot of it will be uh you know both kind of inorganic incentives as well as you know organic growth from kind of just like research mind share yeah um so um going back to the to the blockchain part of things uh so you me mention that it's running its own Sovereign chain so I'm curious if it's a uh if there exists some consensus algorithm that is that it's using or if it's a single sequencer and if there is a in either case what what are the what is the incentive structure there for the for those entities to do their work properly is is there a like is there staking or are they elected in a way yeah so currently we're a single sequencer um we're kind of like working off of of celestial roll kit framework uh as of right now and we're kind of working with them they kind of have a lot of initiatives to essentially down the line move it into kind of like a a shared sequencer um kind of you know approach but right now it is a sequencer um and when it comes to incentives for the uh you know inference nodes as an example or or data access nodes to do their work Faithfully um so these nodes as I mentioned um you know are registered nodes on the network and they uh earn a portion of the inference transaction fee for uh doing compute um so you know I guess for them it is uh you know wanting to to do good work in order to get you know monetarily rewarded and also avoiding not avoiding doing any bad work as to not get slashed um you know by the I layer AVS us because they have to stake some tokens to participate in the network and the same same for data data nodes they get like a key from the instance yeah I think like we're still kind of deciding on what the right amount of stake is for these different types of nodes um like it is likely that for these data access nodes um you know the stake will be lower um but yes uh for these nodes to participate they will have to participate in the ABS Y and there's another component called neuroml so is neuroml just away from uh for you to invoke those inference end points from a solidity or or is there more to the library yeah so uh that that's defin a part of it um but I think we're you know we realize that building AR ml driven workflows is a lot of the time not just about inference I mean for for like llms like perhaps so but a lot of the time you might need to pre-process data like normalize or standardize features you might need to um you know if you want to run some simple statistical tests like Co integration tests or run simple regressions a lot of the time like you know you want these tools and building them in solidity is just uh you know kind of clunky not particularly easy to do um and for us to you know incentivize people to kind of or or to help actually sorry people build these AML workflows in their smart contracts uh these tools are necessary for for them to kind of build with um and that you know includes inference but also yeah all these things that I spoke about yeah and if I do any of those like like let's say I'm doing data processing right from using this solidity Library uh will the data processing happen on chain or is it also going to be sent to some other uh kind of node under the network yeah no that's a great question so right now we use kind of like a mixture of both um so for things like inference I mention we're kind of um you know sending that over to these these cluster of inference nodes for things such as you know data pre-processing or if it's like more light work we're actually going to embed it directly um within kind of like the the blockchain notes itself um because it's not you know a big workload um so so there there is a mixture of both just depending on on kind of how how heavy the compute is I see but but if like like like let's say I have a massive data set I want to preprocess I will just use the the the inference so so this is an API for me to uh and like like custom workload to inference nodes or is it more rigid than that it's uh it's handled behind the background I think right now um that it is true like we have like the main consideration is mostly just to move inference onto the inference nodes and essentially the rest you know including like processing data that's going to be done kind of on the uh like within the blockchain nodes within like uh go ethereum um so so uh that would be done in in gol but we are thinking about um creating a kind of like a special path for transactions as an example if you want to run like convex optimization or if it's like like you mentioned maybe a huge data set uh we don't obviously want to replicate that compute uh many times over so that might be kind of like load balance to these less inference notes as well um that's kind of currently still like in development we don't have it live on our Network as of right now um but I think it's also important to kind of consider the fact that a lot of our workloads are kind of more geared towards like inference um or running kind of statistical tests and less so around training so these workloads might be large but it's not going to be kind of on on the magnitude of what's necessary for training yeah awesome okay so we we're getting close to the end of the of this U session so I guess my my last question is what is the uh what is sort of your your overall vision of the future with the you know Ai and the web 3 together and what is the what's next exciting for open gradient yeah I I think a lot of the times I I like to kind of use this mental experiment to kind of think about what might uh kind of lie in the future for for what 3 and that is if you look at kind of existing web 3 apps many of them kind of don't really leverage arml to to you know at all but if you look at web 2 uh you know I couldn't name you know a couple apps that don't use arml in any capacity at all that I you know interact with on a daily basis and I think for web 3 to really gain Mass adoption and to be competitive with either trafi or web2 it you know is inevitable that a lot of these applications will integrate AR ml um to some extent to optimize kind of the features or the user experience or the liquidity provider experience um and so you know my question is okay what's necessary for these applications to kind of utilize these Technologies in a way that is easy and also secure um and to me that's what kind of open grad represents we're kind of filling that Gap that allows web three applications whether you're defi or social fi or whether you're building a pure AI application to really like bridge that Gap Gap and be able to utilize secure and you know easy to use out of the box solutions that gets you to that next level and also while fostering kind of this open source Community around these you know exploring of new design spaces enabled uniquely in web 3 through you know this ARL kind of product that we have so yeah you know the vision is hopefully you know five years down the line most applications are are smarter more intelligent more optimized because they have access to this technology and you know we will continue to build not just infrastructure but also kind of showcase applications continue pushing on research and building models that can help people you know eventually brid that Gap yeah yeah ex exciting future ahead of us thank you mat you for uh for joining us today h yeah and uh see you all guys on the next episode awesome thank you so much for having me uh you know had a great conversation and uh really excited to kind of continue pushing on this thanks so much Alex","transcript_synced_at":"2026-11-03T05:12:14.814Z","segments":[{"start_seconds":1,"text":"near to one we're close to won't stop us","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1s"},{"start_seconds":8,"text":"now together we'll redefine the game when near we'll find the flame","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=8s"},{"start_seconds":16,"text":"empowered Minds rise and sore use your own AI we'll explore","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=16s"},{"start_seconds":30,"text":"how the future hand in hand in this TX driven","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=30s"},{"start_seconds":37,"text":"Wonderland okay hello everyone uh welcome to the next episode of near office hours uh today we have mat you","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=37s"},{"start_seconds":46,"text":"from open gradient um interestingly a while ago I and I were in a little room with a whiteboard","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=46s"},{"start_seconds":53,"text":"brainstorming uh on What U you know user own AI should look like and a lot of things that open gradient are doing is","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=53s"},{"start_seconds":61,"text":"something we had on the on that whiteboard so it's going to be an exciting exciting conversation I might you do you want to take a couple words","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=61s"},{"start_seconds":68,"text":"about yourself and what open gradient is doing yeah absolutely uh first of all I wanted to thank you and then your team for putting this together very excited","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=68s"},{"start_seconds":76,"text":"to be here um nice to meet everyone my name is Matthew one of the co-founders of open gradient so um open gradient is","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=76s"},{"start_seconds":84,"text":"uh an open intelligence Network and we're building a decentralized platform for hosting models secur C inference agentic reasoning and application","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=84s"},{"start_seconds":92,"text":"deployment so uh we have you know full stack of a variety of different products that I can get into later but from a high level we're excited to be building","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=92s"},{"start_seconds":100,"text":"this you know decentralized secure infrastructure that can help not just make AI more accessible from anyone you know leaning into the whole uh web 3 and","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=100s"},{"start_seconds":109,"text":"blockchain Ethos but also build solutions that can allow blockchain developers to easily integrate ARL into their applications and build more","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=109s"},{"start_seconds":118,"text":"optimized or more intelligent ation so yeah super excited to be here personally for me my background was in ECE I'd done","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=118s"},{"start_seconds":125,"text":"stints of ml research as well and know back when I was in in school um had some experience doing things like data","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=125s"},{"start_seconds":132,"text":"science at Nasa uh building ml models at Google um engineering at Facebook and I spent most of my professional career uh","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=132s"},{"start_seconds":140,"text":"as a research engineer at two Sigma which is a Quant fund based in New York where I built quantitative models as well as a lot of modeling infrastructure","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=140s"},{"start_seconds":147,"text":"for an equity options Market making desk and yeah I met um Adam uh my co-founder in New York who had spent uh six to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=147s"},{"start_seconds":156,"text":"seven years at Palante tier where he was a tech lead on AIP which is their AI platforms and he led um notably the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=156s"},{"start_seconds":163,"text":"development efforts on palente teers patented AI agent execution framework which I found to be very very fascinating so yeah and uh we've been","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=163s"},{"start_seconds":172,"text":"building um since we met uh at the end of 2023 and we're very excited to be you know continuing to show the world what open grent is","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=172s"},{"start_seconds":180,"text":"about yeah so so open gradient it consists of several things right so I I guess we can unpack each of them","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=180s"},{"start_seconds":188,"text":"separately or or maybe we can we can talk about how they all fit together but uh so so one of them is the model Hub uh","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=188s"},{"start_seconds":197,"text":"which is my understanding is it's a decentralized take on uh on hiding face","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=197s"},{"start_seconds":204,"text":"right so so and couple things that are interesting here one is what is","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=204s"},{"start_seconds":213,"text":"the what is the infrastructure behind the modelhub what is it built upon and what is the I guess another well let's","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=213s"},{"start_seconds":220,"text":"be let's begin with a more why right so what's wrong with hiking face why do we need a decentralized modle hub and then what is what is powering it behind the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=220s"},{"start_seconds":229,"text":"scenes yeah I mean I would definitely not I I would definitely not say that anything's wrong with hugging face it's a product that has found tremendous","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=229s"},{"start_seconds":237,"text":"product Market fit um and they've built bu a very very rich Suite of different tools and technologies that I think many tech companies today are now kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=237s"},{"start_seconds":246,"text":"relying on I would say um you know what the open gradient AI model Hub aims to do is to kind of fit that Niche within","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=246s"},{"start_seconds":254,"text":"um the the web 3 space and make it easy for web 3 application developers to integrate ml into their applications so","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=254s"},{"start_seconds":262,"text":"what that actually means in practice is we have a lot of these outof the- boox Solutions um that come with you know as","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=262s"},{"start_seconds":269,"text":"an example example running AI models in tees generating zkl proofs for different model inferences that you know are","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=269s"},{"start_seconds":276,"text":"necessary for a lot of applications building the blockchain space so it's kind of this um taking hugging face and what they've already built and building","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=276s"},{"start_seconds":284,"text":"kind of a version of that that is you know extremely well tailored towards web 3 audience and on top of that of course leaning into the ethos of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=284s"},{"start_seconds":292,"text":"decentralization from the architecture um you can think about the model storage um you know that we have as a fork of ipfs so anyone can run a storage node","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=292s"},{"start_seconds":301,"text":"and you know be able to contribute to that Network and the blockchain infrastructure that I can get into down line uh essentially interacts with this","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=301s"},{"start_seconds":311,"text":"storage layer and uh We've also created some abstractions that are available on the AI model Hub uh you know as an example running you know inference","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=311s"},{"start_seconds":319,"text":"through our playground or if you want to use our SDK you can interact with these models that are stored in our decentralized file store um you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=319s"},{"start_seconds":328,"text":"utilizing our blockchain in without having to deal with the whole blockchain complexities do you don't need to build your own smart contract you don't need to kind of like access","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=328s"},{"start_seconds":336,"text":"these notes so you can just from a very familiar developer kind of CLI or or python environment be able to run all these","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=336s"},{"start_seconds":344,"text":"things yeah and I I went to the model Hub I saw there's a couple of uh models STL along the lines of price prediction","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=344s"},{"start_seconds":353,"text":"so I see your background from uh from the past life is uh uh uh contributing to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=353s"},{"start_seconds":361,"text":"the yeah um so a little bit of uh kind of you know context on that I I think uh","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=361s"},{"start_seconds":369,"text":"you know back when I was at two Sigma there was this term that we threw around a lot that we called like research velocity and the entire idea behind that","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=369s"},{"start_seconds":376,"text":"term is that uh you know models especially things for like you know trading or finance a lot of the time individual models just aren't going to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=376s"},{"start_seconds":385,"text":"be that good right uh models uh their Half Lives just aren't very long some times they're overfit on data sometimes your Alpha gets crowded out sometimes","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=385s"},{"start_seconds":394,"text":"model loses its Edge um so a lot of the times we have to build technologies that you know as an example do model ensembles that help reduce you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=394s"},{"start_seconds":402,"text":"prediction uh kind of loss and and error rate of of the models also on top of that you know people are constantly building new models and iterating on","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=402s"},{"start_seconds":410,"text":"them so I guess like from a high level you think about the Hub as a tool that allows people to you know take other models that people have developed and","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=410s"},{"start_seconds":418,"text":"rapidly iterate on them and and that's just because we want to give people the the power to kind of continue to iterate and build better models over time you","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=418s"},{"start_seconds":427,"text":"know there's not going to be one model in our opinion that trumps everything we kind of just built some sample models but um we're very excited to kind of see","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=427s"},{"start_seconds":434,"text":"you know ml developers and web 3 developers alike kind of build these technologies that will gradually improve","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=434s"},{"start_seconds":441,"text":"uh and the quality will get better over time and you mentioned a couple times by","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=441s"},{"start_seconds":446,"text":"now the uh the blockchain so is it is it just like a set of connectors","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=446s"},{"start_seconds":454,"text":"uh of the of the model Hub to the blockchain or is there anything specific uh that that like so so so it","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=454s"},{"start_seconds":463,"text":"mentions here that it enables performance AI so I wonder what is the performance side of it yeah so typically in in regular","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=463s"},{"start_seconds":473,"text":"blockchains which are replicated State machines every single node re-executes every single transaction and then they attest to hey you know is this the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=473s"},{"start_seconds":480,"text":"current state of the blockchain you know consensus is reached and we move on to the next block of transactions what we've done here is we've kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=480s"},{"start_seconds":487,"text":"fundamentally reimagined what kind of a blockchain you know what blockchain architecture might look like and we've kind of leaned into this idea of node","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=487s"},{"start_seconds":494,"text":"specialization or or we call it like heterogenous compute which is creating different types of clusters of of nodes","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=494s"},{"start_seconds":501,"text":"that do different things so we have as an example regular blockchain notes that execute stateful evm transactions we have um a new of nodes called uh","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=501s"},{"start_seconds":510,"text":"inference nodes that uh compute or you know powerful gpus or tees designed to compute inference and generate any Associated proofs or add stations and we","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=510s"},{"start_seconds":518,"text":"also have data access nodes which are tees that use TLS certificates for secure data access and then the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=518s"},{"start_seconds":525,"text":"inference noes then uh kind of uh then they kind of interface with our storage nodes which is the the uh the model","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=525s"},{"start_seconds":532,"text":"store so what happens uh when someone wants to run let's say a simple inference is uh that transaction submitted the sequencer sequences the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=532s"},{"start_seconds":541,"text":"transaction and it goes into the specialized kind of simulation pipeline in which the inference piece is load balance to the cluster of inference","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=541s"},{"start_seconds":550,"text":"nodes that then computes the the inference and you know generates any Associated proofs sends that back um and we embed in the state transition","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=550s"},{"start_seconds":558,"text":"function the validation of those proofs and AD stations um and so this is kind of I guess you can think about it like","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=558s"},{"start_seconds":566,"text":"it's done offchain except these nodes are all registered on our Network as well and participate in cryptoeconomic","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=566s"},{"start_seconds":573,"text":"you know security via an igen layer ABS so if they generate garbage proofs they can get slashed and on top of that you know the validation of these proofs are","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=573s"},{"start_seconds":582,"text":"natively embedded um in the St transition function as well um so this allows anyone who wants to validate uh","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=582s"},{"start_seconds":589,"text":"you know uh you know inference transactions on our Network as an example to run a node and all the precom","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=589s"},{"start_seconds":595,"text":"Sor all the um you know zkl or t enabled proofs will will go through their node and they can ensure that they get trust","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=595s"},{"start_seconds":603,"text":"minimized access to AI so I have a couple of questions about inference but before we jump into inference guess there's a more","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=603s"},{"start_seconds":610,"text":"fundamental question uh so this the the nodes you mentioned is is","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=610s"},{"start_seconds":617,"text":"it is it a a brand new layer one or is it built on top of existing infrastructure and it's only the imprint","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=617s"},{"start_seconds":626,"text":"noes that are unique topen gradient yeah so technically um using I guess like mainstream blockchain kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=626s"},{"start_seconds":634,"text":"terminology and semantic definitions we would be classified as a layer to because we use near da um so so all the you know proofs um and an ad stations in","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=634s"},{"start_seconds":643,"text":"addition to the block data are stored in in that blob space um and we have like a sequencer um but you know we kind of see","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=643s"},{"start_seconds":650,"text":"ourselves as kind of this Sovereign layer where uh you know we want to kind of have native AI driven applications built on top of us um and when I say AI","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=650s"},{"start_seconds":658,"text":"driven appc can come this can come in many ways shapes and forms which I can elaborate you know later when we talk about use","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=658s"},{"start_seconds":665,"text":"cases yeah so for inference right you mentioned that it's running you mentioned two things you mentioned trusted execution environments and you","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=665s"},{"start_seconds":673,"text":"mentioned zml so I'm curious right so so my prompt like prompt goes into the system right we need we need to get the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=673s"},{"start_seconds":681,"text":"response what happens to the prompt from there why is it multiple nodes why not a single one right and uh how what sort of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=681s"},{"start_seconds":689,"text":"what is the nature of the proof that ends up on chain and is validated yeah so right now we support","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=689s"},{"start_seconds":698,"text":"um I I guess I can maybe talk about what the user flow is initially so with directly you know within the the kind of precompile libraries there are options","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=698s"},{"start_seconds":707,"text":"for you to choose from either um we have pessimistic zkl which is a zkl proof is generated on every single inference we have optimistic zkl which is where we","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=707s"},{"start_seconds":716,"text":"only run inference um but we don't generate the proof to save on cost and you know users can then submit challenges within a Time window to have","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=716s"},{"start_seconds":725,"text":"to see that proof be generated and you know we can decide to arbitrate if that proof is valid Challenger losers stake vice versa the inference don't get","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=725s"},{"start_seconds":732,"text":"slashed and then we also have te at a stations um and right now we have two different types of of tees and adest","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=732s"},{"start_seconds":739,"text":"stations on our Network one is um Nitro enclaves and the other one is like HW h100s their confidential compute enclaves um that were recently on","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=739s"},{"start_seconds":748,"text":"boarding um so that's kind of like I guess the different levels of security that we uh we have and we allow kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=748s"},{"start_seconds":756,"text":"users to choose from what type of you know security they want for their inference directly from the the function call and kind of directly answering your","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=756s"},{"start_seconds":763,"text":"question of why we have multiple nodes and multiple kind of like Security Solutions for this um it is ultimately allowing the users to be very very or","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=763s"},{"start_seconds":771,"text":"developers in this case to be very very flexible with what best suits their intended use case as an example if they want to inference just a small you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=771s"},{"start_seconds":779,"text":"like a gradient boost a tree model or regression model zkl works just fine and can you know be quite powerful for that right whereas for things like llm","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=779s"},{"start_seconds":788,"text":"inference you know it zkl is just not a particularly practical solution due to compute costs and we want to rely on um","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=788s"},{"start_seconds":795,"text":"tees and the multiple node situation comes in where we can have many different models cached on different","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=795s"},{"start_seconds":803,"text":"nodes so the load balancer embedded directly in the sequencer uh you know pings these nodes every couple seconds and is able to say hey if I want to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=803s"},{"start_seconds":812,"text":"route a model inference for let's say um you know llama I know which nodes are capable of producing which kind of proofs and I also know which nodes have","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=812s"},{"start_seconds":821,"text":"that model cached so this is purely just for performance optimization I see but but the particular PRP will go to a single node","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=821s"},{"start_seconds":829,"text":"or will it go to multiple it would go to a single node I see and there are two","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=829s"},{"start_seconds":835,"text":"things so when it comes to te te help with two things one is that uh you can actually get a proof that the output","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=835s"},{"start_seconds":845,"text":"you're looking at is in fact uh a response of a model to your prompt right like you can go to the GitHub and load the virtual machine that","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=845s"},{"start_seconds":854,"text":"is used compute the whatever the hashes and get the quote but it also gives you private infint gtml only gives you the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=854s"},{"start_seconds":863,"text":"form right so so if I choose to use gml my promp will be visible to the no","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=863s"},{"start_seconds":871,"text":"right so for zkl you can actually um so we use Ezekiel uh their library and you can actually choose what you want to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=871s"},{"start_seconds":878,"text":"hide and choose what you want to show um there actually are different visibility settings you can either choose to to hide the model and make the inputs and","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=878s"},{"start_seconds":885,"text":"outputs visible or you can choose to hide um let's say like some like output or input instead uh that you can actually there is some variability there","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=885s"},{"start_seconds":894,"text":"I think right now currently we're focused on you know just you know we're not focused on privacy preserving compute just yet uh kind of one of the benefits that we see of putting this on","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=894s"},{"start_seconds":902,"text":"the blockchain to begin with is the verifiability aspect but we are kind of working on some kind of privacy preserving Solutions involving both zkl","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=902s"},{"start_seconds":909,"text":"and tees for use cases in which someone develops a valuable model or you know someone wants to run compute on some private data and they don't want you","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=909s"},{"start_seconds":917,"text":"know certain parts of it to be exposed to the world but I think that's definitely more of a a a an ad hoc kind of um you know use case that that we","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=917s"},{"start_seconds":925,"text":"have a lot of our uh kind of existing clients are kind of build you utilizing us as opposed to you know centralized solutions for that","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=925s"},{"start_seconds":933,"text":"verifiability aspect and what are the what are the primary use cases for for","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=933s"},{"start_seconds":940,"text":"verifiability yeah no I'm I'm glad you asked we kind of put out like this like write up on all the different use cases uh that we have I guess I can just like","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=940s"},{"start_seconds":948,"text":"maybe quickly walk through a couple of them um so the biggest use case I think that we're excited about is um is Defi","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=948s"},{"start_seconds":955,"text":"and we think that you know for example uh there are a lot of lwh hang fruit like Unis swap amm fee optimization research that we had put out a while ago","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=955s"},{"start_seconds":963,"text":"um where we can see that hey if we use you know ml models to compute Dynamic fees we can reduce net impermanent loss for LPS and increase fee collection for","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=963s"},{"start_seconds":971,"text":"LPS by up to 18% um so this is kind of preventing or mitigating effects of lvr um things such as risk models um to help","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=971s"},{"start_seconds":981,"text":"improve you know yield yield aggregators as well as Vault protocols as risk adjusted returns um as well as kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=981s"},{"start_seconds":988,"text":"you know know risk models for Lending protocols to optimize collateralized like LTV ratios in these pools um so","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=988s"},{"start_seconds":996,"text":"outside of defi there's also you know reputation models for deepend networks so we can you know outline a couple different features like you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=996s"},{"start_seconds":1003,"text":"liveness uptime things like that uh and we allow these models to dynamically score you know the the performance of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1003s"},{"start_seconds":1010,"text":"different minor nodes on dpin networks and actually if you check our model Hub filecoin um we have a couple file coins","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1010s"},{"start_seconds":1018,"text":"this models on on our model Hub and this basically allows us to say hey you want to be a good dpin node you should get a good reputation score and we'll allocate","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1018s"},{"start_seconds":1027,"text":"rewards you know proportionally to your reputation score um you know we also have some interesting use cases around using ml models to price different","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1027s"},{"start_seconds":1035,"text":"prediction markets um things like that I can maybe dive into uh you know some of these if you're curious but a lot of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1035s"},{"start_seconds":1041,"text":"these kind of use cases are high impact kind of use cases where if the model inference goes wrong and the AI or ml work is running correctly could could","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1041s"},{"start_seconds":1050,"text":"lead to kind of um you know implications with users's funds and as a result we want to ensure that everything is entend you know verifiable and that users get","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1050s"},{"start_seconds":1059,"text":"this guarantee around the compute that we provide yeah and what is the so as a user because there's a model Hub I can","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1059s"},{"start_seconds":1068,"text":"download the model right the models are public right so what is the um like what","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1068s"},{"start_seconds":1076,"text":"would be my like what are the tradeoffs between running it on top of open gradient as opposed to you know for example just running it locally on on on","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1076s"},{"start_seconds":1085,"text":"my own machine yeah so I mean this kind of leans into the whole uh you know we I","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1085s"},{"start_seconds":1092,"text":"would say two things number one is is verifiability right uh you know people want to ensure that the quality of the model is good they might you know even","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1092s"},{"start_seconds":1101,"text":"if they can guarantee the compute is is done correctly they want to be able to maybe download the model and run their own simulations or audit make sure that","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1101s"},{"start_seconds":1108,"text":"the model like you know looks reasonable um so similar to like kaggle or hungy face we provide a big platform that","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1108s"},{"start_seconds":1116,"text":"allows people to go and download and access these models as well and the second one is kind of like model versioning um which allows developers to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1116s"},{"start_seconds":1124,"text":"constantly iterate on these models allows people to come you know upload their own versions of these models Cod so it kind of Fosters this kind of open- source Community around development of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1124s"},{"start_seconds":1133,"text":"these models and on top of that you know people can build workflows that involve say models I've developed other people","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1133s"},{"start_seconds":1141,"text":"have developed prediction models as well we can kind of aggregate them and do like a model stacking you know Ensemble approach and ultimately create a","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1141s"},{"start_seconds":1148,"text":"workflow that is even more you or that leads to more accurate forecast than just a singular model yeah and right now this is","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1148s"},{"start_seconds":1156,"text":"primarily used uh like has there been any work towards autonomous agents that use those models","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1156s"},{"start_seconds":1164,"text":"uh to uh to take action on chain or is it still primarily operated by humans yeah so that's actually one of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1164s"},{"start_seconds":1172,"text":"the things I'm I'm very bullish about and I I had talked to um a couple folks from near about this as well which is uh you know I kind of see AI agents uh and","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1172s"},{"start_seconds":1181,"text":"I would love to hear your opinion on this as well I think they're currently in a stage where um initially it's kind of a mimic right these AI agents can put","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1181s"},{"start_seconds":1189,"text":"out tweets on Twitter it's particular it's just interesting and the next step step for agents is they can gradually replace what humans can do um so you","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1189s"},{"start_seconds":1197,"text":"know I don't need a and place a trade on Unis swap if I can just tell an agent hey can you swap out you know one eth for usdc on Unis swap for me just does","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1197s"},{"start_seconds":1206,"text":"that for me and I think the next step for agents in this case is to be able to allow agents to complete complex tasks better than humans as an example how do","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1206s"},{"start_seconds":1215,"text":"I manage a portfolio right how do I take in these spot and these volatility forecasts and perform Marco width Style mean variance analysis to optimize my","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1215s"},{"start_seconds":1223,"text":"portfolio allocations and holdings right these are all kind of complex tasks that that uh you know a simple llm might","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1223s"},{"start_seconds":1231,"text":"struggle to to kind of do effectively and as a result we kind of want to build these powerful ml workflows that allows agents to leverage them to do cooler","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1231s"},{"start_seconds":1240,"text":"things right if you tell an agent what token or if tell an LM you know if asking what token should I buy to maximize my risk adjusted returns it'll probably say something generic like","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1240s"},{"start_seconds":1248,"text":"Bitcoin or like you know I don't want to provide Financial advice but in this case um you know these AI agents will then be able to leverage these ML","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1248s"},{"start_seconds":1255,"text":"workflows and models we've built to kind of do all these things which which is now something we're diving into and we're actually putting out um something","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1255s"},{"start_seconds":1263,"text":"or releasing an an agent with a double up which is this uh you know gamey slash you know gamble uh protocol onui we're","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1263s"},{"start_seconds":1272,"text":"building this AI agent that leverages ouru spot forecasting model to set kind of uh money uh money line lines on their","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1272s"},{"start_seconds":1280,"text":"protocol and allows users to bet against it um so that's going to be our initial kind of foray into the agent space and we're kind of talking to some other AI","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1280s"},{"start_seconds":1287,"text":"agent protocols and partners like exponent AI as an example on how they can integrate uh you know our ml models","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1287s"},{"start_seconds":1294,"text":"into their agents to help users you know you know do do bigger better cooler things yeah yeah that's that sounds","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1294s"},{"start_seconds":1303,"text":"pretty cool we have a we have this toy example called Shield GPT where people can Shield their meme coins to a to a","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1303s"},{"start_seconds":1311,"text":"model to A system that is powered by uh by uh I don't remember which model but uh it's not it's not particularly","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1311s"},{"start_seconds":1319,"text":"resilient that's exactly the point of the name right if if you sh well enough it will it will get convinced and it will buy it and it's uh the performance","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1319s"},{"start_seconds":1327,"text":"of its portfolio was suboptimal to say the least right so so in this case I think it's an","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1327s"},{"start_seconds":1334,"text":"interesting yeah if if the models so so factly it's interesting how far we are from the state where there is a AI","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1334s"},{"start_seconds":1343,"text":"system sufficiently sophisticated that people cannot take advantage of it right because you know on the market you play against others right if people can take","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1343s"},{"start_seconds":1350,"text":"advantage of your system they will right so so once the model can be sufficiently reliable that people cannot take","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1350s"},{"start_seconds":1358,"text":"advantage of it easily that or like an sample of models yeah I think in this case like as an example you could build in a simple","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1358s"},{"start_seconds":1366,"text":"like you know like a market risk or Market forecasting model into Shield GPT so as an example if you're bullish on the crypto Market at large like it's","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1366s"},{"start_seconds":1374,"text":"likely that a lot of these meme coins will trade as levered beta to the market and you know you can as an example say Hey you know if I am underexposed to the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1374s"},{"start_seconds":1383,"text":"market I can buy some more meme coins if I'm Overexposed I can opt to be more risk averse in buying these um tokens","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1383s"},{"start_seconds":1390,"text":"being shied to the the agent um you know I mean I I can dive deeper into these many like kind of uh Quant Finance upgrades that you know the agent can","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1390s"},{"start_seconds":1399,"text":"benefit from but I'm particularly excited about kind of these more advanced um agents that can can leverage these Technologies and uh yeah um it's","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1399s"},{"start_seconds":1407,"text":"kind of like what we're building towards um at open gradian so so on the model Hub what are the what are the incentives for people","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1407s"},{"start_seconds":1416,"text":"to publish good models do do they uh is there is there an incentive structure or is it mostly just a place for people to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1416s"},{"start_seconds":1423,"text":"share for free yeah so we're we're kicking off a hackathon kind of later this uh this month as well so initially","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1423s"},{"start_seconds":1430,"text":"it's going to be kind of like bootstrap by hey if you're a data scientist and you want to earn some some uh you know cash prizes come and build some cool","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1430s"},{"start_seconds":1438,"text":"models on on top of us um but I I think down the line we want to leverage uh you know both potential token incentives as","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1438s"},{"start_seconds":1446,"text":"well as kind of just like you know create an organic Community uh that is kickstarted by you know application and","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1446s"},{"start_seconds":1453,"text":"protocol developers wanting to benefit you know each other and the ecosystem at large to to kind of contribute to these models and and what I say by that is uh","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1453s"},{"start_seconds":1461,"text":"practically speaking you know an amm might say Hey you know we want to put out this this new model that we're working on that you know does you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1461s"},{"start_seconds":1469,"text":"dynamic V optimization some someone else might be able to take that model you know and and also put out a new one to not just better their protocol but to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1469s"},{"start_seconds":1476,"text":"kind of Advance the space from a research standpoint overall um and yeah a lot of it will be uh you know both kind of inorganic incentives as well as","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1476s"},{"start_seconds":1485,"text":"you know organic growth from kind of just like research mind share yeah um","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1485s"},{"start_seconds":1492,"text":"so um going back to the to the blockchain part of things uh so you me mention that it's running its own","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1492s"},{"start_seconds":1499,"text":"Sovereign chain so I'm curious if it's a uh if there exists some consensus algorithm that is that it's using or if","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1499s"},{"start_seconds":1507,"text":"it's a single sequencer and if there is a in either case what what are the what is the incentive structure there for the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1507s"},{"start_seconds":1516,"text":"for those entities to do their work properly is is there a like is there staking or are they elected in a","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1516s"},{"start_seconds":1523,"text":"way yeah so currently we're a single sequencer um we're kind of like working off of of celestial roll kit framework uh as of right now and we're kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1523s"},{"start_seconds":1531,"text":"working with them they kind of have a lot of initiatives to essentially down the line move it into kind of like a a shared sequencer um kind of you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1531s"},{"start_seconds":1540,"text":"approach but right now it is a sequencer um and when it comes to incentives for the uh you know inference nodes as an example or or data access nodes to do","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1540s"},{"start_seconds":1549,"text":"their work Faithfully um so these nodes as I mentioned um you know are registered nodes on the network and they uh earn a portion of the inference","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1549s"},{"start_seconds":1557,"text":"transaction fee for uh doing compute um so you know I guess for them it is uh you know wanting to to do good work in","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1557s"},{"start_seconds":1566,"text":"order to get you know monetarily rewarded and also avoiding not avoiding doing any bad work as to not get slashed","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1566s"},{"start_seconds":1574,"text":"um you know by the I layer AVS us because they have to stake some tokens to participate in the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1574s"},{"start_seconds":1580,"text":"network and the same same for data data nodes they get like a key from the instance yeah I think like we're still","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1580s"},{"start_seconds":1589,"text":"kind of deciding on what the right amount of stake is for these different types of nodes um like it is likely that for these data access nodes um you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1589s"},{"start_seconds":1596,"text":"the stake will be lower um but yes uh for these nodes to participate they will have to participate in the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1596s"},{"start_seconds":1602,"text":"ABS Y and there's another component called neuroml so is neuroml just away from uh for you to invoke those","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1602s"},{"start_seconds":1611,"text":"inference end points from a solidity or or is there more to the library yeah so uh that that's defin a part of it um but","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1611s"},{"start_seconds":1619,"text":"I think we're you know we realize that building AR ml driven workflows is a lot of the time not just about inference I mean for for like llms like perhaps so","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1619s"},{"start_seconds":1628,"text":"but a lot of the time you might need to pre-process data like normalize or standardize features you might need to um you know if you want to run some","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1628s"},{"start_seconds":1636,"text":"simple statistical tests like Co integration tests or run simple regressions a lot of the time like you know you want these tools and building","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1636s"},{"start_seconds":1643,"text":"them in solidity is just uh you know kind of clunky not particularly easy to do um and for us to you know incentivize","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1643s"},{"start_seconds":1651,"text":"people to kind of or or to help actually sorry people build these AML workflows in their smart contracts uh these tools are necessary for for them to kind of","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1651s"},{"start_seconds":1660,"text":"build with um and that you know includes inference but also yeah all these things that I spoke about yeah and if I do any of those like","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1660s"},{"start_seconds":1667,"text":"like let's say I'm doing data processing right from using this solidity Library uh will the data processing happen on chain or is it also going to be sent to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1667s"},{"start_seconds":1676,"text":"some other uh kind of node under the network yeah no that's a great question so right now we use kind of like a","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1676s"},{"start_seconds":1683,"text":"mixture of both um so for things like inference I mention we're kind of um you know sending that over to these these cluster of inference nodes for things","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1683s"},{"start_seconds":1691,"text":"such as you know data pre-processing or if it's like more light work we're actually going to embed it directly um within kind of like the the blockchain","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1691s"},{"start_seconds":1699,"text":"notes itself um because it's not you know a big workload um so so there there is a mixture of both just depending on","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1699s"},{"start_seconds":1708,"text":"on kind of how how heavy the compute is I see but but if like like like let's say I have a massive data set I want to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1708s"},{"start_seconds":1715,"text":"preprocess I will just use the the the inference so so this is an API for me to uh and like like custom workload to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1715s"},{"start_seconds":1723,"text":"inference nodes or is it more rigid than that it's uh it's handled behind the background I think right now um that it is true like we have like the main","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1723s"},{"start_seconds":1732,"text":"consideration is mostly just to move inference onto the inference nodes and essentially the rest you know including like processing data that's going to be","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1732s"},{"start_seconds":1740,"text":"done kind of on the uh like within the blockchain nodes within like uh go ethereum um so so uh that would be done","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1740s"},{"start_seconds":1748,"text":"in in gol but we are thinking about um creating a kind of like a special path for transactions as an example if you","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1748s"},{"start_seconds":1756,"text":"want to run like convex optimization or if it's like like you mentioned maybe a huge data set uh we don't obviously want to replicate that compute uh many times","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1756s"},{"start_seconds":1765,"text":"over so that might be kind of like load balance to these less inference notes as well um that's kind of currently still like in development we don't have it","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1765s"},{"start_seconds":1773,"text":"live on our Network as of right now um but I think it's also important to kind of consider the fact that a lot of our workloads are kind of more geared","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1773s"},{"start_seconds":1781,"text":"towards like inference um or running kind of statistical tests and less so around training so these workloads might","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1781s"},{"start_seconds":1787,"text":"be large but it's not going to be kind of on on the magnitude of what's necessary for training yeah awesome okay so we we're","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1787s"},{"start_seconds":1797,"text":"getting close to the end of the of this U session so I guess my my last question","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1797s"},{"start_seconds":1804,"text":"is what is the uh what is sort of your your overall vision of the future with the you know","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1804s"},{"start_seconds":1812,"text":"Ai and the web 3 together and what is the what's next exciting for open gradient yeah I I think a lot of the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1812s"},{"start_seconds":1821,"text":"times I I like to kind of use this mental experiment to kind of think about what might uh kind of lie in the future for for what 3 and that is if you look","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1821s"},{"start_seconds":1829,"text":"at kind of existing web 3 apps many of them kind of don't really leverage arml","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1829s"},{"start_seconds":1835,"text":"to to you know at all but if you look at web 2 uh you know I couldn't name you know a couple apps that don't use arml","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1835s"},{"start_seconds":1843,"text":"in any capacity at all that I you know interact with on a daily basis and I think for web 3 to really gain Mass adoption and to be competitive with","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1843s"},{"start_seconds":1852,"text":"either trafi or web2 it you know is inevitable that a lot of these applications will integrate AR ml um to some extent to optimize kind of the","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1852s"},{"start_seconds":1861,"text":"features or the user experience or the liquidity provider experience um and so you know my question is okay what's","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1861s"},{"start_seconds":1868,"text":"necessary for these applications to kind of utilize these Technologies in a way that is easy and also secure um and to","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1868s"},{"start_seconds":1876,"text":"me that's what kind of open grad represents we're kind of filling that Gap that allows web three applications whether you're defi or social fi or","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1876s"},{"start_seconds":1884,"text":"whether you're building a pure AI application to really like bridge that Gap Gap and be able to utilize secure and you know easy to use out of the box","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1884s"},{"start_seconds":1892,"text":"solutions that gets you to that next level and also while fostering kind of this open source Community around these","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1892s"},{"start_seconds":1899,"text":"you know exploring of new design spaces enabled uniquely in web 3 through you know this ARL kind of product that we","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1899s"},{"start_seconds":1906,"text":"have so yeah you know the vision is hopefully you know five years down the line most applications are are smarter more intelligent more optimized because","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1906s"},{"start_seconds":1915,"text":"they have access to this technology and you know we will continue to build not just infrastructure but also kind of showcase applications continue pushing","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1915s"},{"start_seconds":1922,"text":"on research and building models that can help people you know eventually brid that Gap yeah yeah ex exciting future","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1922s"},{"start_seconds":1931,"text":"ahead of us thank you mat you for uh for joining us today h yeah and uh see you all guys on the next episode awesome","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1931s"},{"start_seconds":1940,"text":"thank you so much for having me uh you know had a great conversation and uh really excited to kind of continue pushing on this thanks so much Alex","url":"https://www.youtube.com/watch?v=dDXTOPRKyjg&t=1940s"}],"youtube_url":"https://youtube.com/watch?v=dDXTOPRKyjg","url":"https://near-pedia.xyz/videos/dDXTOPRKyjg"}