{"id":"USclcdfj-mY","title":"[REDACTED] 2024 | Alex Skidanov - NEAR.AI Hub for Researchers","description":"Alex Skidanov on the latest developments at NEAR AI.\n\nTakeaways:\n• Historical Pattern in Tech Innovation: Technology often starts as a promising advancement, only to later be dominated by a few corporations that control access and value capture, ultimately sidelining user benefits. AI is now at a similar juncture, with a few major players like OpenAI, Google, and Meta holding much of the power.\n\n• NEAR’s Unique Position for Decentralized AI: NEAR Protocol aims to create an ecosystem for decentralized AI, leveraging its Web3 infrastructure and collective expertise to offer an alternative to centralized AI models. This decentralized approach could provide access to tools, data, and computing power for AI development without relying on a single entity.\n\n• Trusted Execution Environments (TEEs) for Private AI: TEEs are crucial for ensuring privacy in decentralized AI. They allow computations to be performed on external hardware without exposing sensitive data to the hardware owner. This makes it possible to run models and manage data securely, preventing misuse by third parties.\n\n• Frontier Models via Community Collaboration: Through TEEs and decentralized infrastructure, NEAR plans to enable the creation of competitive large language models that are community-funded and open-source. These models would remain privately monetizable, allowing backers to recoup their investments while providing competitive, accessible AI.\n\n• Competitions to Democratize AI Development: NEAR is launching a series of competitions to engage researchers globally, allowing them to build and improve AI models. Starting with small models and gradually increasing in complexity, this initiative provides a way for talented individuals outside major corporations to contribute to advanced AI development in a decentralized ecosystem.\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":"2024-11-11T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/USclcdfj-mY/hqdefault.jpg","transcript_text":"how you guys doing okay let's talk about uh artificial intelligence um if you if you look back at the history of Technology uh if you look at what's been happening with u search Social Mobile it always Begins by something new and exciting uh something that changes Our Lives to the better and then one or two companies fully capture uh the whole you know attention whole value from the from the new market and then you know a decade later that market is entirely you know used to capture every little every single dollar of our um of value without producing value for users anymore right so with search if you go to Google today it's almost impossible to find anything with social it's all optimized for your attention so that you spend you know more and more time on in there and artificial intelligence is in a very similar space right now in a very similar moment in history where we are still at the moment where we're very excited about something new uh but we already can see that you know one two maybe three companies are fully capturing value uh and that will create a uh a data whe uh that will effectively make it less uh more and more hard for uh for others to get throw them in the future uh and as long as you know if you like me believe that AI will be a very big part of our interactions with the internet with the world uh with uh practically everything in the future that means no user owned internet uh no user Centric future is possible as long as large language models are controlled by large corporations and if you look at those large corporations they have like like if you think of open AI entropic Google Facebook they have data experiments compute the best researchers all in the same place are all very nice centrally orchestrated right so that allows them to iterate very quickly while outside of them if we wanted to replicate it in a decentralized fashion first of all we don't have access to uh resources that are of the same quality today uh second of all even the resources that exist they all spread out right so a person who tries to to um to push it Forward they have they spend a lot of time you know looking around finding things that they need and we want near near is in a sort of unique position to try and go ahead and solve it because uh relative to web two companies right such as you know like any any Web Two compan company that would try to to Dethrone you know hypothetical open AI we have years of experience of pulling resources together of orchestrating a large sets of people in a decentralized fashion something that is absolutely impossible in web to um uh uh you know sending microp payments annotating data Etc while relative to web three companies near is is in unique position because near internally has a lot of experience building artificial intelligence uh one and only of my co-founders is literally the creator of Transformers so it's a uh you know nobody in the world is positioned better than near to make this change and so correspondingly it is effectively our responsibility to drive this uh this Innovation forward and effectively the way we want to do it the way we're doing it today is we want near to be a uh central decentralized place uh for all the things related to AI so that all the data sources all the models all the experiments are are discoverable uh agents so that you have access to compute you have access to data annotations in one place uh and yet each of those is coming from various ecosystem projects from different uh decentralized entities right and none of that you know nobody disappearing tomorrow because they ran out of money or turned out to be you know from day one affect in any way uh the whole effort so before I get into uh the fun part let's address the most important part of uh decentralized AI the decentralized AI one way or another needs to be private when you interact with chat GPT you know that your data goes to open AI but at least you can assume that you know um that open AI has at least some effort to protect the data it's it's a bad place to be in but it's not as bad as if you were using some decentralized servers where you know some random person is running a machine with the uh a h100s and your data just goes to them and they sell it to the lowest bidder highest bider uh and uh interestingly private AI is possible uh and the the way it is possible is all the modern CPUs and gpus allow uh a technology called trusted execution environment something IIA mentioned during his talk uh which effectively says that if someone let's say me has a computer and I'm a malicious actor you should never trust me uh and you have some computation you want to do but you don't have access to compute you can send it to me and I can execute it in a trusted execution environment and you will be absolutely certain that I had no idea what was happening on my computer uh I could not see your data I could not see uh the code you're executing uh all I know that something was executed on my machine right and so that allows you to build all sort of interesting use cases it allows you to do private inference meaning that I could be running a model on my machine and you could come to me run inference on my machine I will not see your data uh I will not see what you were doing with the in there right you can you can send any medical records any private data and you will have 100% confidence that uh that was not in any way seen by me uh it allows you to run agents you know like an agent provider says this is my agent it's open source it's you know it's beautiful it does everything you want uh here is the end point there is no way for me without trusted execution environments to know know that the end point I'm hitting is actually the agent that they see on GitHub with TR that execution environment I can do that that allows you to to have a network of Agents running uh in a on a decentralized network where you know for every agent that it is doing exactly what is promised you can use any secrets private Keys everything uh and be certain at every point that the system behaves exactly the way you want and uh none of your data is is going to any malicious or any other actors uh but the most exciting uh the most exciting thing that trusted execution environments unlock is that they allow us to create Frontier open source Lang large language models which are monetizable right so today if you think about Frontier AI you know like uh the the the models like you know Lama 405b or gp4 or you know GPT 5 is probably around the corner they are massive training such a model requires hundreds of millions of dollars and there are very few entities who can for spending hundreds of millions of dollars to create such a model and then just drop the weights and say hey guys use it uh it's all yours right and uh you know if if the entirety if the whole future of Open Source AI is fully in the hands of meta it's a very Bleak future we have and with te is what you can do is you can pull together resources from people uh you can train a model entirely inside of the tees using open source code uh open source data pipelines uh and then get the model that is open source meaning that the source is open but not open weights meaning that the weights are not seen by anyone uh and yet anyone can pull the the uh the encrypted weights to their machine and run it inside the KE but the KE would only serve the model to you uh if you establish a payment Channel and pay for the usage right and the and the price of the usage could be extremely competitive right so it could be competitive with any inference providers that exist today but you know that the model once it is trained will be monetizable every usage of the model will be paid for and therefore you can actually pull together those hundreds of millions of dollars and have those people have a reasonable way to expect to repay their their contribution uh from the usage of the model so that is extremely exciting because that technology actually gives us you know gives us a path towards the future in which the frontier models are created by the community for the community and served in a decentralized fashion um without without sacrificing e omal viability so that is probably the most exciting uh area that we are exploring right now this is uh effectively the way this is a little diagram which highlights what I was just describing right Bob who you can clearly see is not a good uh good guy uh has a machine uh and the Machine is fully controlled by Bob right so like he can he can open it up he can use a microscope to like inspect every little transistor in the processor uh but there's a trusted execution environment within his machine and everything that happens in The Trusted execution environment Ellis can trusted but most importantly the protocol as a whole can trust it right so the protocol as a whole can know that the model once deployed to any of the machines with the te will not be used without uh proper payment back to the uh to the to the Dow that initially created it so it's a yeah and payment channels obviously something that we know works very well uh there's there's a way to to build this technology such such that payment channels are open between the trusted execution environment and the and the user so it's a it's an exciting exciting technology and exciting Way Forward for user owned AI the only problem is it's not enough to train a very large language model ideally you want to train a a large language model that is good right so so for example if you think if you think about lama 405b lama 405b in terms of size is comparable to gp4 uh or CLA 3.5 uh but in no way is L 45 be competitive with those two models right so even Facebook with all its resources cannot train a model that is competitive with an open AI model so how could we without having access to researchers that Facebook has today train a model that will be competitive because if the model is not competitive you cannot monetize it people will not use it right now the way we approach it uh is we we start uh a series of competitions where we're just going to bring all the researchers who today uh don't work for open AI right there are students uh all over the world who are um getting into AI today a lot of them are extremely talented and extremely knowledgeable who do not have a way uh to get access to compute to get access to uh to good data to get access to other resources uh and we will provide access to those researchers and a way for them to compete and show you know that they are the best and they can contribute to the best mod model in existence and you can think of it of course we cannot have a competition you know we cannot train a one trillion parameter model in in such a way because you know even training one of them cost hundreds of millions of dollars but you can try train a small model and you can begin with like a half a billion parameter model which is small and then you can have thousands of researchers simultaneously competing for training the best half a billion parameter model and then you take the best of them those of those who who show that they are promising and and can do exciting things and all of them go to the next step and now we train the best 2 billion parameter model right and then the best of those Advance further uh and at each step you have fewer people but those are the best people and uh but you need more compute you know fewer people more compute it's the same amount of resources you need to spend until you get the very best uh you know 10 to 20 people working on the very large language model that will be competing with the gp4s and cloes of the world and that's uh that's a that's an exciting way to to enable first of all to enable people who are very talented but don't have a way uh to um to work for those uh large companies either because of uh uh their you know because of lack of opportunity or because of their values right not every person wants to contribute to the evil Corp uh you know to get more and more of uh of user of user data and user attention so for many of them that will be the way uh to apply their talent to show what they can do to contribute to open source AI uh and to get rewarded for that cool uh so that the first competition is starting effectively now that will be a competition to train a very small half a billion parameter model so let's together train the uh the small one and then uh we have a plan uh we will publish it very soon uh effectively you know in a in like six seven steps we will get to uh to very large numbers which are way more than half a billion you know 400 billion to a trillion ET so every everything I was talking about today uh a lot of it is uh live at app. it's a hub for all things AI data sets models compute uh it's a place where the competitions uh will be hosted uh so the you know the future of decentralized AI has never been brighter uh so yeah let's build the future together","transcript_synced_at":"2026-11-03T05:25:36.917Z","segments":[{"start_seconds":0,"text":"how you guys doing okay let's talk about uh artificial intelligence um if you if you","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=0s"},{"start_seconds":9,"text":"look back at the history of Technology uh if you look at what's been happening with u search Social Mobile it always","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=9s"},{"start_seconds":18,"text":"Begins by something new and exciting uh something that changes Our Lives to the better and then one or two companies fully capture uh the whole you","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=18s"},{"start_seconds":27,"text":"know attention whole value from the from the new market and then you know a decade later that market is","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=27s"},{"start_seconds":34,"text":"entirely you know used to capture every little every single dollar of our um of value without producing value for users","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=34s"},{"start_seconds":43,"text":"anymore right so with search if you go to Google today it's almost impossible to find anything with social it's all optimized for your attention so that you","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=43s"},{"start_seconds":50,"text":"spend you know more and more time on in there and artificial intelligence is in a very similar space right now in a very similar moment in history where we are","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=50s"},{"start_seconds":60,"text":"still at the moment where we're very excited about something new uh but we already can see that you know one two maybe three companies are fully","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=60s"},{"start_seconds":67,"text":"capturing value uh and that will create a uh a data whe uh that will effectively","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=67s"},{"start_seconds":74,"text":"make it less uh more and more hard for uh for others to get throw them in the future uh and as long as you know if you","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=74s"},{"start_seconds":84,"text":"like me believe that AI will be a very big part of our interactions with the internet with the world uh with uh practically everything in the future","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=84s"},{"start_seconds":92,"text":"that means no user owned internet uh no user Centric future is possible as long as large language models are controlled","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=92s"},{"start_seconds":100,"text":"by large corporations and if you look at those large corporations they have like like if you think of open AI entropic Google","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=100s"},{"start_seconds":109,"text":"Facebook they have data experiments compute the best researchers all in the same place are all very nice centrally orchestrated right so that allows them","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=109s"},{"start_seconds":118,"text":"to iterate very quickly while outside of them if we wanted to replicate it in a decentralized fashion first of all we don't have access to uh resources that","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=118s"},{"start_seconds":127,"text":"are of the same quality today uh second of all even the resources that exist they all spread out right so a person who tries to to um to push it Forward","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=127s"},{"start_seconds":136,"text":"they have they spend a lot of time you know looking around finding things that they need and we want near near is in a sort","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=136s"},{"start_seconds":145,"text":"of unique position to try and go ahead and solve it because uh relative to web two companies right such as you know","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=145s"},{"start_seconds":153,"text":"like any any Web Two compan company that would try to to Dethrone you know hypothetical open AI","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=153s"},{"start_seconds":160,"text":"we have years of experience of pulling resources together of orchestrating a large sets of people in a decentralized fashion something that is absolutely","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=160s"},{"start_seconds":169,"text":"impossible in web to um uh uh you know sending microp payments annotating data","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=169s"},{"start_seconds":176,"text":"Etc while relative to web three companies near is is in unique position because near internally has a lot of experience building artificial","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=176s"},{"start_seconds":185,"text":"intelligence uh one and only of my co-founders is literally the creator of Transformers so it's a uh you know","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=185s"},{"start_seconds":192,"text":"nobody in the world is positioned better than near to make this change and so correspondingly it is effectively our responsibility to drive","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=192s"},{"start_seconds":201,"text":"this uh this Innovation forward and effectively the way we want to do it the way we're doing it today is","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=201s"},{"start_seconds":208,"text":"we want near to be a uh central decentralized place uh for all the things related to AI so that all the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=208s"},{"start_seconds":217,"text":"data sources all the models all the experiments are are discoverable uh agents so that you have access to compute you have access to data","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=217s"},{"start_seconds":225,"text":"annotations in one place uh and yet each of those is coming from various ecosystem projects from different uh decentralized entities right and none of","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=225s"},{"start_seconds":234,"text":"that you know nobody disappearing tomorrow because they ran out of money or turned out to be you know from day one affect in any way uh the whole","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=234s"},{"start_seconds":243,"text":"effort so before I get into uh the fun part let's address the most important part of uh decentralized AI the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=243s"},{"start_seconds":251,"text":"decentralized AI one way or another needs to be private when you interact with chat GPT you know that your data goes to open AI but at least you can","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=251s"},{"start_seconds":258,"text":"assume that you know um that open AI has at least some effort to protect the data it's it's a bad place to be in but it's","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=258s"},{"start_seconds":266,"text":"not as bad as if you were using some decentralized servers where you know some random person is running a machine","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=266s"},{"start_seconds":273,"text":"with the uh a h100s and your data just goes to them and they sell it to the lowest bidder highest","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=273s"},{"start_seconds":279,"text":"bider uh and uh interestingly private AI is possible uh and the the way it is","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=279s"},{"start_seconds":287,"text":"possible is all the modern CPUs and gpus allow uh a technology called trusted","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=287s"},{"start_seconds":295,"text":"execution environment something IIA mentioned during his talk uh which effectively says that if someone let's say me has a computer and I'm a","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=295s"},{"start_seconds":302,"text":"malicious actor you should never trust me uh and you have some computation you want to do but you don't have access to compute you can send it to me and I can execute it in a trusted execution","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=302s"},{"start_seconds":311,"text":"environment and you will be absolutely certain that I had no idea what was happening on my computer uh I could not see your data I could not see uh the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=311s"},{"start_seconds":319,"text":"code you're executing uh all I know that something was executed on my machine right and so that allows you to build all sort of interesting use cases it","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=319s"},{"start_seconds":327,"text":"allows you to do private inference meaning that I could be running a model on my machine and you could come to me run inference on my machine I will not see your data uh I will not see what you","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=327s"},{"start_seconds":336,"text":"were doing with the in there right you can you can send any medical records any private data and you will have 100% confidence that uh that was not in any","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=336s"},{"start_seconds":345,"text":"way seen by me uh it allows you to run agents you know like an agent provider says this is my agent it's open source it's you know it's beautiful it does","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=345s"},{"start_seconds":353,"text":"everything you want uh here is the end point there is no way for me without trusted execution environments to know know that the end point I'm hitting is","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=353s"},{"start_seconds":361,"text":"actually the agent that they see on GitHub with TR that execution environment I can do that that allows you to to have a network of Agents","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=361s"},{"start_seconds":370,"text":"running uh in a on a decentralized network where you know for every agent that it is doing exactly what is promised you can use any secrets private","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=370s"},{"start_seconds":377,"text":"Keys everything uh and be certain at every point that the system behaves exactly the way you want and uh none of","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=377s"},{"start_seconds":386,"text":"your data is is going to any malicious or any other actors uh but the most exciting uh the most","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=386s"},{"start_seconds":394,"text":"exciting thing that trusted execution environments unlock is that they allow us to create Frontier open source Lang large language models which are","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=394s"},{"start_seconds":402,"text":"monetizable right so today if you think about Frontier AI you know like uh the the the models like you know Lama 405b","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=402s"},{"start_seconds":410,"text":"or gp4 or you know GPT 5 is probably around the corner they are massive training such a model requires hundreds of millions of dollars and there are","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=410s"},{"start_seconds":418,"text":"very few entities who can for spending hundreds of millions of dollars to create such a model and then just drop the weights and say hey guys use it uh","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=418s"},{"start_seconds":425,"text":"it's all yours right and uh you know if if the entirety if the whole future of Open Source AI is fully in the hands of meta it's a very Bleak future we have","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=425s"},{"start_seconds":434,"text":"and with te is what you can do is you can pull together resources from people uh you can train a model entirely inside of the tees using open source code uh","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=434s"},{"start_seconds":443,"text":"open source data pipelines uh and then get the model that is open source meaning that the source is open but not open weights meaning that the weights","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=443s"},{"start_seconds":451,"text":"are not seen by anyone uh and yet anyone can pull the the uh the encrypted weights to their machine and run it inside the KE but the KE would only","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=451s"},{"start_seconds":460,"text":"serve the model to you uh if you establish a payment Channel and pay for the usage right and the and the price of the usage could be extremely competitive","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=460s"},{"start_seconds":467,"text":"right so it could be competitive with any inference providers that exist today but you know that the model once it is trained will be monetizable every usage","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=467s"},{"start_seconds":475,"text":"of the model will be paid for and therefore you can actually pull together those hundreds of millions of dollars and have those people have a reasonable","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=475s"},{"start_seconds":484,"text":"way to expect to repay their their contribution uh from the usage of the model so that is extremely exciting","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=484s"},{"start_seconds":491,"text":"because that technology actually gives us you know gives us a path towards the future in which the frontier models are","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=491s"},{"start_seconds":501,"text":"created by the community for the community and served in a decentralized fashion um without without sacrificing e","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=501s"},{"start_seconds":509,"text":"omal viability so that is probably the most exciting uh area that we are exploring right now this is uh","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=509s"},{"start_seconds":518,"text":"effectively the way this is a little diagram which highlights what I was just describing right Bob who you can clearly","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=518s"},{"start_seconds":524,"text":"see is not a good uh good guy uh has a machine uh and the Machine is fully","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=524s"},{"start_seconds":531,"text":"controlled by Bob right so like he can he can open it up he can use a microscope to like inspect every little transistor in the processor uh but","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=531s"},{"start_seconds":539,"text":"there's a trusted execution environment within his machine and everything that happens in The Trusted execution environment Ellis can trusted but most","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=539s"},{"start_seconds":547,"text":"importantly the protocol as a whole can trust it right so the protocol as a whole can know that the model once deployed to any of the machines with the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=547s"},{"start_seconds":555,"text":"te will not be used without uh proper payment back to the uh to the to the Dow","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=555s"},{"start_seconds":562,"text":"that initially created it so it's a yeah and payment channels obviously something that we know works very well","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=562s"},{"start_seconds":570,"text":"uh there's there's a way to to build this technology such such that payment channels are open between the trusted execution environment and the and the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=570s"},{"start_seconds":578,"text":"user so it's a it's an exciting exciting technology and exciting Way Forward for user owned","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=578s"},{"start_seconds":586,"text":"AI the only problem is it's not enough to train a very large language model ideally you want to train a a large","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=586s"},{"start_seconds":593,"text":"language model that is good right so so for example if you think if you think about lama 405b lama 405b in terms of size is comparable to","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=593s"},{"start_seconds":601,"text":"gp4 uh or CLA 3.5 uh but in no way is L 45 be competitive with those two models right","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=601s"},{"start_seconds":609,"text":"so even Facebook with all its resources cannot train a model that is competitive with an open AI model so how could we without having access to researchers","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=609s"},{"start_seconds":617,"text":"that Facebook has today train a model that will be competitive because if the model is not competitive you cannot monetize it people will not use it right","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=617s"},{"start_seconds":624,"text":"now the way we approach it uh is we we start uh a series of competitions where we're just going to bring all the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=624s"},{"start_seconds":632,"text":"researchers who today uh don't work for open AI right there are students uh all over the world who are um getting into","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=632s"},{"start_seconds":639,"text":"AI today a lot of them are extremely talented and extremely knowledgeable who do not have a way uh to get access to compute to get access to uh to good data","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=639s"},{"start_seconds":648,"text":"to get access to other resources uh and we will provide access to those researchers and a way for them to compete and show you know that they are","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=648s"},{"start_seconds":657,"text":"the best and they can contribute to the best mod model in existence and you can think of it of course we cannot have a competition you know we cannot train a one trillion parameter model in in such","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=657s"},{"start_seconds":666,"text":"a way because you know even training one of them cost hundreds of millions of dollars but you can try train a small model and you can begin with like a half a billion parameter model which is small","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=666s"},{"start_seconds":675,"text":"and then you can have thousands of researchers simultaneously competing for training the best half a billion parameter model and then you take the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=675s"},{"start_seconds":682,"text":"best of them those of those who who show that they are promising and and can do exciting things and all of them go to the next step and now we train the best","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=682s"},{"start_seconds":689,"text":"2 billion parameter model right and then the best of those Advance further uh and at each step you have fewer people but","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=689s"},{"start_seconds":696,"text":"those are the best people and uh but you need more compute you know fewer people more compute it's the same amount of resources you need to spend until you","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=696s"},{"start_seconds":704,"text":"get the very best uh you know 10 to 20 people working on the very large language model that will be competing","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=704s"},{"start_seconds":710,"text":"with the gp4s and cloes of the world and that's uh that's a that's an exciting way to to enable first of all to enable","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=710s"},{"start_seconds":719,"text":"people who are very talented but don't have a way uh to um to work for those uh large companies either because of uh uh","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=719s"},{"start_seconds":727,"text":"their you know because of lack of opportunity or because of their values right not every person wants to contribute to the evil Corp uh you know","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=727s"},{"start_seconds":736,"text":"to get more and more of uh of user of user data and user attention so for many of them that will be the way uh to apply","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=736s"},{"start_seconds":744,"text":"their talent to show what they can do to contribute to open source AI uh and to get rewarded for","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=744s"},{"start_seconds":752,"text":"that cool uh so that the first competition is starting effectively now that will be a competition to train a very small half a billion parameter","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=752s"},{"start_seconds":760,"text":"model so let's together train the uh the small one and then uh we have a plan uh","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=760s"},{"start_seconds":766,"text":"we will publish it very soon uh effectively you know in a in like six seven steps we will get to uh to very large numbers which are way more than","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=766s"},{"start_seconds":775,"text":"half a billion you know 400 billion to a trillion ET","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=775s"},{"start_seconds":787,"text":"so every everything I was talking about today uh a lot of it is uh live at app.","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=787s"},{"start_seconds":793,"text":"it's a hub for all things AI data sets models compute uh it's a place where the competitions uh will be hosted uh so the","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=793s"},{"start_seconds":802,"text":"you know the future of decentralized AI has never been brighter uh so yeah let's build the future together","url":"https://www.youtube.com/watch?v=USclcdfj-mY&t=802s"}],"youtube_url":"https://youtube.com/watch?v=USclcdfj-mY","url":"https://near-pedia.xyz/videos/USclcdfj-mY"}