{"id":"0uvxudxSsuk","title":"NEAR AI Office Hours #7 with Bitte and Exabits","description":"Navigating the Future of AI Open Models and Compute Power\n\nIllia & Alex discuss the latest AI research topics and unpack some NEAR.AI focus areas\n\nNEAR AI: https://near.ai/\nIllia Polosukhin https://x.com/ilblackdragon\nAlex Skidanov https://x.com/AlexSkidanov\n\nBitte: https://x.com/BitteProtocol\nExabits: https://x.com/exa_bits\n\nThe intersection of decentralized AI, blockchain, and compute power with leading experts from Bitte Protocol and Exabits. Discover how open-source AI can counteract a Black Mirror-esque future, the role of Web3 in democratizing AI, and the potential for cross-chain interactions and agent marketplaces.\n\nTakeaways:\n• Decentralized AI as a Counterbalance: Open-source AI and blockchain technology can help prevent a centralized, dystopian future as depicted in Black Mirror.\n• Importance of Compute Power: The future of AI hinges on access to compute power, making it a critical component of the value chain in AI development.\n• Convergence of AI Models: As AI models scale, they may converge in functionality, suggesting that foundational models could become commoditized.\n• Web3 and Public Good Funding: Web3 can offer innovative funding mechanisms for public good projects, including decentralized AI research and development.\n• Security in Agent Marketplaces: Building secure and scalable agent marketplaces is crucial for enabling seamless, secure interactions across different blockchain environments.\n\nTimestamps:\n00:03:54 - Black Mirror Scenario: Discussing how decentralized AI and blockchain can prevent a dystopian future.\n00:06:13 - AI Models and Centralization: The risks of AI being controlled by a small group of entities.\n00:11:17 - Convergence of AI Models: Exploring the Platonic Representation Hypothesis and its implications for AI.\n00:15:07 - NVIDIA's Role in AI: The significance of compute power and NVIDIA’s impact on the AI landscape.\n00:19:18 - Public Good Funding and Web3: How Web3 can fund decentralized AI projects and research.\n00:24:15 - Government Involvement in AI: Debating the role of governments in funding and controlling AI development.\n00:28:01 - Chain Abstraction and Multi-Chain Interactions: How chain abstraction can simplify blockchain interactions for users.\n00:37:16 - Consumer Applications for AI: Potential consumer applications that can integrate AI and blockchain technologies.\n00:45:44 - Prompt Engineering and UI Generation: The future of prompt engineering and its role in AI interactions.\n00:54:02 - Building Secure Agent Marketplaces: The importance of security in developing agent marketplaces for decentralized applications.\n01:00:02 - Future of Decentralized AI: Final thoughts on the future of decentralized AI and the need for open-source solutions.\n\nJoin NEAR's ecosystem:\nWebsite: https://near.org/\nTwitter: https://x.com/NEARProtocol\nBlog: https://near.org/blog/","channel_name":"NEAR Protocol","published_at":"2024-06-25T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/0uvxudxSsuk/hqdefault.jpg","transcript_text":"light feel the custom feel hack the world and make it tight blockchain flows near to one we're close to won't stop us now together we'll Rine game when near we'll find the flame empowered Minds rise and sore use your own AI we'll explore hold the future hand in hand in this Tech driven Wonder okay hello hello to the new episode of AI office hours we have a great group here today uh let's uh start with uh Nate do you wna and M do you want to guys introduce yourself sure I'm Nate uh co- founder of bits of protocol previously uh mint base the old nft Marketplace uh We've recently discovered in the last year of cranking well in the last five years of building an nft Marketplace within the last year of integrating with chat TBT and some other bits we've learned we can do anything on our nft platform through prompts and so that's led us to believe we can now do defi swaps and now we're in this magical position of cross chain uh signatures so we're super excited to have our new name fits a protocol means please in German uh please AI build and execute this transaction for me and I'll hand it off to our other co-founder mikro who's been driving a lot of our AI initiatives also the CTO of mbase yeah I mean I think you pretty much summed it up uh but yeah I've been in the near ecosystem for about three years uh with mid-base as long as as as long as I have been with in the near ecosystem uh and super excited for what's to come and uh what we've been building um in in the DI space do you want to take it away uh wansu oh yes my name is hansu I'm co-founder of exhibits we're in the business of building the layer zero for decentralized compute um and we service some of the biggest and most renowned um compute marketplaces in the world nice to meet you good to meet you too good to have you guys so uh I know you guys had some questions uh from before so maybe maybe we can launch into some of that uh I guess I'm sure Nate you have a lot of questions so sure so I was recently watching uh Black Mirror rewatching it so the uh episode two to me was the most like intense one I think this was like from like over a decade ago maybe even more uh where everything was a number right whether you brush your teeth you know you get a number reduced whether you're watching a horrific ad uh and if you're not watching it you have to reduce your number uh and basically the people are just spinning their bike all day to acrw more number right this is like the evil world that we're probably actually heading towards especially with AI um so the the big question is uh how is connecting near protocol uh and the open- source world going to help sort of not have that world because there the ideal world is we all don't work anymore right um we all have these efficient machines doing efficient things but I mean how do we get how do we get to that uh and what's your sort of uh open open take on on the uh the savior of the Black anti-black Mirror ification yeah I I also remember that the person who who who really wanted to to to fix the system in that episode ended up being a big part of it um yeah but uh well well I guess we the I actually don't know I didn't see that episode when it came out so it's hard for me to tell how relatable it was back then but yeah it is getting more and more relatable that's that's certainly uh and uh I mentioned it I think like couple episodes again but the uh like every breakthrough that happened uh since like uh the Advent of internet has been you know have has has been improving Our Lives in some ways but was also making it more and more black mirror is and then like you know like social networks uh generally uh web two right and uh and I think the big part of it is precisely because um so so I think there are two parts here right I think both this episode and there's another one called nose dive which is sort of with a similar uh similar message I think I think they they sort of present two different aspects of so one of them is is this Society right so so like there's a big part which which has nothing to do with centralization that Society is building uh you know other uh so so that I don't think we will address in anyway that needs to be addressed uh you know collectively as humans but the another one is that yeah uh as centralization or rather as every breakthrough happens uh it it uh so far hisor ially it inevitably becomes controlled by a small group of people who the moment they fully capture the the market they optimize for Revenue not for Value right and obviously optimizing for Revenue sometimes is aligned with optimizing for Value but not always right and example I like using is that if social graph was a public good we would have way more usages of it uh than we actually have today right primarily today it is used you know for for engagement uh farming and it's a it optimizes for Revenue doesn't optimize for value at all and so in this case AI could be a very big uh change in the way we uh do everything uh and if it is again controlled by small set of people whatever values they uh you know state today those values can change similarly to how Google was don't do evil uh and became what it is today right and and like I think there's a good chance that Larry and Sergey had that value back in the day when they started right so like whatever values those companies stay today they will they will change right once once you become you know a trillion dollar company you you naturally become this you know profit making Machinery so it's very important that for AI there is a uh um well like I think it is early enough that a sufficiently strong effort to make it free and open from the GetGo will win uh and then we will at least avoid that Pitfall and we'll have you know will have a strong technology which works for the good of people not for the uh for the small set of investors yeah I would add to figuring out how do we how do we have a more Collective good optimization right versus single centralized optimization which is just a natural way like if a centralized company trains a model whatever kind of individual pieces are trying to do at the end it's going to be for the benefit of that centralized company right that's just how does things work and so we I mean I think blockchain have been struggling with this like hey this is a way to organize people at scale for Collective good and it's easy to do when it's something that's like very simple to describe right you know let's all just hash you know and find find an non starts with zero and that actually protects Network and it gets harder and harder right with you know things like gitcoin been trying to do this at at like in the more kind of abstract ways and so I think like here we have something that's like extremely demanding figuring this out where we need to um globally have a model that's optimized for successive individuals and communities versus uh optimized for any single Company Success and and that's indeed very challenging I'm you're you're muted Nate yeah and it's it's already doing its job right a little bit even though you're working with you know private llms we're also dealing with private uh Financial systems with their own closed systems whereas blockchain is already open source right so if we have that bridge we're already sort of doing the good like we don't have to do much more because we're now bridging that working with open source contracts working with a much more open world even if we're still working with closed Source uh big models we're kind of already starting that that step for forward into the the the non- eval or being able to interact with different contracts that can be replaced by you know other folks that can fork and do something a little bit better do you want to say something um sorry I wasn't too sure about the formatting um no I think that's an important question um but but um can we switch gears uh and um perhaps discuss another topic yeah go ahead you know I was just following what's what's been happening with Nvidia in the marketplace um I don't know if you any of you guys are the the big news is NVIDIA briefly last year um you know surpassed the market capitalization of Microsoft and Apple and since then it's it's sort of LW some steam um but uh you know one of the big questions that I'm grappling with these days is sort of the long run vision of where things are headed with AI you know so when you see this big rise and the spotlight of Nvidia in the public markets you know the first question is where where is the value chain you know is where is a lot of this value in AI going to be captured and it appears right now Nvidia is sort of the kicking butt in doing that um and then you know when you look at the competitive landscape of of AI models you know you have you have Claw from anthropic you have mistro B from Google you know meta you know everyone's building their own AI um and then I recently came across this research paper uh by MIT researchers it's called the platonic representation hypothesis okay and um so these authors they found that representations in AI models um particularly deep networks are converging so as models get larger different neural networks representing data are becoming more aligned um and the implication is that there's going to be this in the long run a convergence of all of the models you know the more and more you train them with with more data and more parameters they they will essentially become all identical um uh you know so I'd like to get your thoughts on this and I mean like an immediate uh sort of a followup that comes to mind is you know if these mods are converging um you know open AI anthropic Mistral you know um what what are their modes in the long run right you know or or do they just become completely indistinguishable I want to open up the floor and get your thoughts I think interesting question is do they already have like do they have a mode today like let's say models do not converge right I think they like at the given size with the given amount of data the capabilities are sort of similar so in the absence of a breakthrough uh you do not get much of a mode right so like M and llama uh they kind of uh like like you know already somewhat indistinguishable From perspective of the downstream tasks I think the fact if the models actually convert right I didn't I skimmed the paper uh before before the this call but I didn't look very deeply but if they did converge that's actually good news that effectively means that at some point uh there is no benefit in in creating more and more models so once you have one uh model that you know very close to that convergence it's as good as it gets so so that that is potentially good news especially if it also if it also implies I don't know if it implies it but if it implies that there is the plateau at which point you cannot go you know any further that's even better that means that the model will not be getting better and better uh and will not eventually replace us right it's even better if that Plateau is is our level of intelligence that would be outstanding uh yeah but I think I mean there there there's few interesting points here right one is you asked about the value chain and indeed right now we see massive investment in kind of compute because even if training itself is you know takes 10 20 50,000 gpus the thing is that the inference St can take any amount of gpus right afterwards because if now you're doing inference kind of you know one token at a time you can imagine as this like instrumentation gets more sophisticated you will want to run you know you know Chain of Thought you'll want to run multiple models at the same time to like on different different subtasks so you can imagine that the amount of compute you can you know spend an inference will not be bounded by any kind of even magnitude so even if we imagine like all let's imagine like at we trained you know indeed what Al said the gold model that's like this is a final model we don't need to train anything new though I mean obviously there's like lots of asteres on discussion there but even then like at an inference time you'll need you can use as much computer as want right to actually produce the result you want right and we already see this with you know you run multiple agents they themselves refining results they read you know the read the context and may be like oh let me write that let me run it with tests let me do this or that um and so indeed I mean clearly the market you know reacts to that that there's like this amazing massive investment in the hardware that happens um will that stop I mean probably not anytime soon right just because as I said like there's just so so much more you can do if you have that compute at INF time and if you multiply by a billion people well you need like massive amount of amount of this to to serve it now indeed on the training side I mean I've been saying this for over a year now that I think like model fun foundational models will be a commodity and kind of indeed may not they may not be like completely converging uh but it makes sense that at the end you know you're training it on somewhat the same data you're using the same method it kind of makes sense you get the same results at the end uh so I don't know yeah I haven't read the paper outside of abstract so I'm actually curious what are they comparing because obviously you cannot go compare CHP to CLA um outside of maybe the embeddings of the surface um right um and it appears U I mean you know if you're using the same data the the underlying mathematics is roughly equivalent then then right just comes down to who has access to most compute so they can have the biggest and the most complicated model um now so for me when I think about this I I also kind of uh you know I think there are some really important implications about ownership so let's say um uh scaling we're sort of nowhere near the plateau you know so we get uh you know you can hear guys think Mark Zuckerberg and his podcast you know admit in in their Facebook in The Meta Labs um they have yet to see um any signs of diminishing uh Mar marginal returns to scaling you know they're seeing like you know the more you pump up the models with more compute more data um it's it's doing more and more fascinating things um so let's then assume that for examp assume that for a moment that you know we can continue to scale up the models become bigger and bigger um and sort of just seeing and going going back to the whole Invidia story of where things are with them you know since compute at scale is just so Capital intensive right you know this is the reason why only big tech companies are really able to even have carry out the ambition of training these Large Scale Models um it seems like a scenario in which several competitors you know spending hundreds of billions or even trillions of dollars to replicate essentially the same AI just seems too inefficient to be viable right you know so so if one is convinced that AI brings immense public value um I certainly I certainly am then what one can apply the reasoning that um large Investments with really challenging economics that provide significant public public benefits are typically financed you know by a Central Public Authority right or the government so think of you know think of utilities infrastructure or other Investments that require you know really complex multiple stakeholder alignments the highways the power grid gas pipelines you can name them so you know like is there a future in which the US government will attempt to fund and subsequently hold ownership of an AGI um you know you know I guess we're already seeing this is the case in China and kind of kind of thinking ahead um you know we just as Elon Musk had a had to innovate and engineer reusable Rockets to bring down the cost of launches you know demand for scaled compute we need to uh discover some dramatic efficiencies to remain privatized right and um um I guess or or maybe the need for a some sort of creative ownership structure is where we may get substantial input or you a solution from web 3 um and this was sort of where I started to connect the dots and see um you know it where to find that intersection between decentralized computer or Ai and where the the scaling needs are but you love to get your further input on this oh I would really hope to never have the government run the full AI system it's so I mean if you been to a California DMV have you ever been to like it's just so slow and brutally and that's why NASA is basically getting overrun by space I'm not aent of a central ownership right but but just looking at historical examples you know when you need a ton of resource to build out something that everyone agrees has tremendous value you know is AI going to enhance human productivity absolutely yes right um so it's it's sort of inevitable reality that someone some somewhere is going to build this AGI right with enough resources uh but typically when you get uh you know historically when you have something that's incredibly Capital intensive and if what everyone is competing to build out essentially the same thing and it costs trillions of dollars to build um you know in equilibrium it's most likely um a much bigger uh socially neutral agency is going to build that you know so you know if if if we to avoid that fate uh you know what are some things that we would have to see develop um and one idea is know web 3 brings in this very creative uh decentralized ownership model and I think that's one of the aspects that people haven't really thought very hard as much as you want to see this intersection in web 3 and AI yeah so I mean as I mentioned like we need to figure out I mean this is literally the problem of the public goods funding right just at at a larger scale that so far has been tested out in crypto and sadly Beyond I mean like the biggest working public goods funding is Bitcoin where um you know everybody runs a minor and they're getting funded for doing that um and so there's a clear protocol how you get rewarded for providing public goods service while you have a self-interest in participating in it and so in a way like we kind of need very similar mechanism but with a lot fine grain more fine grain research Focus right where um like I actually don't think like you know going in massively training like training massive modeling at this point is is even that interesting because like we've seen you know some of the new papers showcasing that even with smaller models if you Leverage search or like Monte Carlo search or you know using to reinforce doing thought uh tokens internally as you predict things you're able to re you know kind of increase the capacity of the model um just again because you're moving the compute pretty much out of the weights into the kind of that inference time to actually into activations and is also just I mean the fact there also this new paper on inductive bias to that propagation is also crazy right I don't know if you guys seen but they pretty much fine tune on examples and the model was able to pretty much infer the function from fine tuning so we clearly have like new methodologies that are I think emerging around using optimizers using search and kind of intern like intermittent thought process inside model to actually improve the results and so what we're now missing is really an ability to coordinate researchers to really swarm on this ideas and test them out and and see exactly where the ins and out of this are because the challenge is like somebody published the paper you know there's some source code they trained it on they did it on like a 7B model with a little bit of fine-tuning um and like you you don't really know how much more you can do you don't know if it works in 70b so I think this is where we have right now an opportunity where you know you guys as exhibits for example can be like hey we actually going to give a grant to this researchers to for this amount of compute and this is actually something we're trying to work out how to do effectively um but let's try to run this model like train this model on 70b now and like again F tuning from existing model and see you know if this this properties hold if they do then potentially is a state-of-the-art across a bunch of parameters um and now this is still all in open source because we kind of gave a grant for this to be in open source and you get you know promotion pretty much from doing this as well uh and so people want to use it for inference because they know it runs on your Hardware in the first place so I think like that's kind of where what what we're missing like some of this mechanism of collaborative funding research where we have researchers right now making progress not not on like let's train a bigger model right I mean the the amount of research there is like go into config add a zero right start and then babysit that gigantic model that you know crashes all the time versus like people are actually making progress on you know structure and and and uh you know ways to optimize things and then like how do we fund that how do we leverage that and and again how do we like replicate it in such a way that then people know okay this is actually works let's incorporated into next technique so yeah that I pretty much I agree with you and I think web three is like have the methods and and actually funding to do that now and actually showcase not on kind of some like you know in theory you can train a bigger model right as you collect a lot of clusters but actually like we can get interesting research out of the existing Academia researchers by just being able to provide them access and kind of manage that a little bit more freely and fluidly right and and uh you know our our business mission is very much align uh with that with that uh with that view um and we we take it very serious seriously upon ourselves U to kind of provide that sort of democratization and access to compute you know for various like these these other applications that are currently under research and to make it Equitable for anyone to access compute you know want to build out models um but neat I got I got a question so if uh since I have you guys I'm excited to ask it's fun to talk about tomorrow and what's what's coming up in the future but it's also really fun to talk about what what do we have now what can we play with today on near and where things are heading so I'm sure you've sort of seen uh we've built that AI agent allows you to generate images uh it builds a transaction that you can execute you can create drop links we've now built a ref Finance agent that allows you to do defi uh Swap this for that uh and we're going to be opening up that agent Marketplace for other folks to start plugging in potluck uh just took them three days to to build basically their core pits of their agents you'd be able to say fund this company for that uh where do you want to see this heading uh where should we start uh plugging away today uh and what do you see value in basically AI building transactions uh first and B uh what what interesting use cases should we really tackle if if you think that there's value there because again I think we're going we're building the Black Mirror uh kind of theory it's being able to communicate with the financial system through human readable conversations uh and I'm just curious to get your input on on what we've built so far and and where where we can head to help uh push the near ecosystem further I mean I think the biggest opportunity right now is to find the consumer applications which can integrate this right I mean there's both there's few interesting consumer applications in crypto that are that can leverage this right away usually they are in already existing chat uh environment so this is Telegram and Discord Bots um where if you plug that in uh and create economic incentives around for example Community managers and kind of channel uh server owners to benefit from originating transactions from from these channels uh you you know make it really easy to create your accounts either through existing wallets or embedded wallet directly there you can pretty much like supercharge the both web three adoption and kind of amount of accounts that are transacting through this because you know to just give you an example like cab land right has over I think on their website over 10 uh users across Discord and telegram across 50,000 different communities and so you know you can imagine that bot becoming also originating transactions through natural language same you know with hot with the telegram Bots as well so I think and then obviously going into web 2 as well and seeing how to connect with their existing ecosystems uh and you know turning on pretty much the um the financial component of what right now a lot of this assistants are just like there is a bunch of stuff you can do and then you still need to go and like book it and figure it out yourself right the real CR of that is that you're you know incentivizing folks so how much is it uh folks that are actually wanting to get value out of what you're doing or how much of it is folks that just want to get paid with the tokens so it's it's a it's something that we're looking into deep uh and it's kind of a weird line that we're trying to to ride but oh go ahead go ahead I'm I'm curious your your answer no well I think I mean the general expectation should be you you like same as credit card right you should not give you know 100% money back that just doesn't work um you should get you know if you're building out a loyalty P program which is what all of these token you know user distributions are they should be percentage of the value that is being transacted um and so and at the same time it should be something that's like below the margin that you're capturing then it's a sustainable program that you can run I think crypto indeed kind of went into like Hey we're going to just you know we need to distribute this token so we're going to spend all of it uh kind of at once for everyone who's even you know barely tried our product that obviously is not a good idea um although you know for a while that's been the way people want to approach it but you can imagine you know this is your gross engine so like how do you create a loyalty program that you know benefits from that and potentially yeah you overpay in the beginning with the idea that your margins will improve over time but I think the the benefits you potentially can have is transitioning away from like it should be valuable for the users to urgin transactions uh but the people who may be teaching and encouraging and providing a UI Etc like you can you can pretty much uh reward them which will be a lot I mean a lot less like again if you integrate it into some bot you're either rewarding the bot creators or the community that installed this body again hot there's Villages and collab land and all of this like there servers that are actually you know there's Community managers behind them and so they can be the ones educating the user base about what they can do through this and so you can reward them versus individual users there and so you're kind of preventing a lot of the uh negative of usage farming right engagement farming that happens transaction building real world use cases you got you got any Alex uh I think anything which is more complex than a uh like a simple transfer uh could benefit like like uh like I think as an example as simple as staking is uh causing a lot of confusion so so so if that was possible like you know I go there and like uh I have like a simple task of I guess especially uh uh like even with the example of staking right even staking on itself could be confusing uh with like you know hey there's like withdrawal you know interval Etc then you have liquid staking reaking Etc I think like going there and communicate communicating with something that will tell you your options uh and uh executing it from there is quite helpful yeah from a single prompt as well you're not going from one UI piece to another you're not going to ref to something else you're in the single prompt where it's like Swap this token for this token now stake it stake half of it um and then send the other half to my ledger wallet right all within a single prompt and these things we can start doing with like bulk prompts right being able to send multiple transactions through a single prompt right mint an nft generate the image load it to R weave listed for five bucks could be one single prompt and that's or and do one on eth and do one on near it sounds crazy but this is this is uh this is where we're at it's uh I I just think it's it's uh working and focusing more on that the agent side is uh and incentivizing those agents to build better agents is is probably where we're we're going to be touching in micro you got anything sure um yeah I was thinking since you're talking about the multiple prompts or you know submit a prompt that's composition of multiple prompts I was wondering to hear your guys thoughts on uh the future of prompt engineering is it you know a real thing will it become like even more um of a real thing in the future or you know as models evolve and like uh kind of understand better and better like what users want uh it will kind of Fade Away over over time UI generation yeah I mean I think it it's going to be a combination of things right because you still unless it's like a completely autonomous agent that you know goes and figures out everything from a A to Z and you just gave it like a super high level task the problem is like you still want something that's like so I mean the example we're doing on AI developer right now the it's not even about the prompt it's like you still want to have a conversation with it because as part of that conversation which refines your understanding and kind of you answer questions you have tradeoffs you're like figuring out what needs to be done and so even if the model itself is kind of conditioned on that uh you still need that kind of like I wouldn't call that prom engineering per se but like you need that convers system that actually figures out from the user what you want right um and so that can come in the form of a one shot prompt where like I just know exactly what I'm doing but it's more probably going to become like you can imagine a a prompt bot right which you go to it like asks you what you want it figures out like exactly all the tradeoffs ETC and then it creates a prompt for another AI to then do it repetitively already now knowing exactly what you know the process and thinking you've done um and so I'm imagining something like this will happen where uh AIS actually will write kind of the the programmatic like what what we right now have for programmatic prompts right that are like behind an agent right where user doesn't interact with that and then on the user side it will be more of like a conversational plus UI plus like we actually use a spec as a format where it like writes out stuff and then you can read that and like correct it what you want right and that's actually really useful because you kind of see the whole context you're like oh I want this user stor or like actually I want this new journey I want this feature and so it gives you like a lot more context on what's kind of in the model on the other side so I I'm imagining more of that and then kind of generated programmatic prompts that go into the backend of the engine I of people disagree with me it's kind of a a tricky conversation because uh Hano is very back in hardcore uh hardware and all those pieces and we're front in app layer so it's interesting Bridging the Gap I don't want to take up too much time onto so go for it no that's okay um well so I mean since we're getting into the some of these technical questions um so as a crypto compute project U so exhibits is trying to find methods to provide users with more flexible ways to use Compu you know so um think of the the following examples so for in compute defi being swap e on ethereum for a compute voucher and salana um for compute lending uh use so is collateral to borrow compute registered on near so we we think n's um chain abstraction can help help us abstract complexity of blockchain away from the user which I think is a a very valuable and and critical um barrier uh for for wider adoption and is so you know first off is that understanding correct and how long do you think it would take to realize such a technical system yeah I mean well chain abstraction for sure is designed to well kind of as the idea to remove the need to think about which chain things are on and instead move into user space where you just indeed as as Nate said like just describing with natural language what you want right um and so I mean the the use cases you described are exactly the type of things that we actually have projects that are using chain signatures to are and building right kind of a multi-chain defi multi-chain lending that uh that are working where yeah you can you know borrow against a combination of Bitcoin Dodge Coin you know some some M lady and a uh you know some locked vcv on ethereum and then you can borrow usbd on to and similarly you can swap you know any combination of assets across different chains so that is indeed kind of the direction we're going as well as on the wallet side right kind of what beta is doing what H is doing where you can kind of have like a singular view of assets across all chains you don't actually need to think about like where they are and you can navigate it you know either UI or natural language where you say hey you know I just want a position like I mean I just typing an example like hey I want an index of top five coins across all chains right and you know buy $ five dollar worth of it right like I don't want to figure it out like how to do that it should just go you know fetch from coin market cap build me an index and then you know buy kind of across all of them and say me like oh I want you know I want to be able to have my AI run for an hour thinking about this problem and you know like I'll pay is whatever eum and so it you know buys a kind of compute uh coin in the way and then pays for it when it actually runs the inference on the back end right for an agent that goes and like navigates and thinks through and writes Out The Chain of Thought and maybe writes some code and does all the things and then comes back with you with an answer and saying like hey I spent You Know This Much computer pretty much on it so like I can imagine all of that is kind of coming together pretty quickly actually because um I know you know folks with wallets already doing s this the applications kind of are building this like multichain def5 multichain Ling are as well have like their kind of alpha versions that come together so I would I would expect within like three months we have some of that kind ofun that's where we're also heading to with the UI uh building on the Fly being able to not only do swaps and these things but say give me a report of the charts of this and this and then it just builds a nice pretty graph and then zoom in a bit more on that okay execute my swap on this based on these forms that I've I've processed so yeah I uh yeah I'm very on boarded yeah turn the table back to you Nate or Alex do you have more to add now we can move on if if people have other fun questions uh okay what would be your ideal scenario for an open Agent Market connected with near at the core so I know you guys are working on an agent agent Marketplace as well I don't know how much you want to reveal of that land I know it was revealed in a couple open uh office hours before but just curious because we're we're also working on probably something that we could work together deeply on I'm curious how we can work together to really build just an epic system that allows not just ecosystem Partners but with this cross chain you know world that we're we're jumping into how do we basically start taking over other chains and I think we can do it pretty quick uh if if we execute this right together right um so the biggest question is what is what is your ideal agent Marketplace that is really focused on sort of either the execution layer or so on I mean I can mention maybe a few components we're looking at right so the way we're thinking about it is kind of agent is anything that um Can can take an environment and we can talk about what the environment is and perform set of actions right and so it has access to llm it has access to you know tools it has access to any other stuff including submitting transactions so you can model that as just a python you know file python script that executes that the the challenge obviously is like how to make it secure how to do per missioning how to do all of this kind of n greedy details and so so that's where I think a lot of thinking and kind of Designing is to go in while also still iterating on you know some useful agents to build as kind of you guys are building um so the I think the big important part is kind of how do we go to a situation where you can have you know a very simple prompt that then opens up an environment right in a kind of private in some private container uh you know potentially goes like into a browser opens up coin market cap you know fetches the required information like for example for this index of top five tokens fetches that you know brings it to kind the agent that constract the for example the buy transaction right and then passes it to back to the user to valid it but this is exactly what you want and so kind of how do you construct do containers where actions can be executed for example they can navigate browsers they can write code they can execute some of the code as well within that environment and so then this is also usable for developer this is usable for process automation where you know near foundation for example somebody can just say like hey you know run this process it goes like opens up browser and does that or somebody in the ecosystem is like hey you know I want to find all of the all of the things that's going on in because then you know and give it to me and in this format and then potentially get me to contact people of of the folks who are doing so that's kind of the you know high high level structure that we kind of thinking and obviously the the system itself the code is all going to be open source and so um as well as the agents are just stored somewhere on chain for example over time to uh to you know for anyone to be able to pick them up and use it um and then the questions are yeah like if there are additional use cases around like you want to have agents that are paying each other because they need special data uh you know how do you model data how do you model payments stuff like this right we have never mind which do doing payments for example there's few projects that doing kind of data coordination and like fetching and you can add payments to all that so kind of there's like somewhat of a framework that uh we're trying to con conceptualize and obviously I mean for context we did talk with Nate about this on Friday so um so we I mean we're gonna have this kind of contining this conversation and also hopefully can open source what we've been working on pretty soon yeah our world is trying to tackle both ends it's not just the user typing in things to go do swaps and do things I mean blockchain has such had has had such a horrendous onboarding uh excruciating onboarding hurdles right not only you have to get users to figure out how to like buy nfts but you have to get Brands to figure out how to create accounts and do all that sort of stuff so being able to also enable Brands to create agents super quick like imagine lvmh being able to hook up their inventory uh sheets and being able to do a mass mint of nfts pegging every single asset to maybe some of their items with highly detailed explanation of what that asset is and then people being able to search for those assets while they're watching a Netflix movie you know I'd like to buy that red handbag from that's that's really where I see us evolving over the number of years uh it's not only helping reduce the onboarding for people interacting but also for the other side the asset creation because if we look at the traditional Finance system it's just not going to keep up with AI there's just no way that AI is going to be able to evolve with traditional Finance uh AI is such a perfect match being able to boot up accounts within seconds do trades create assets uh on the Fly ask its inventory sheets like so to me the the agent Marketplace is really kind of this space where we can open it up where we're not doing all the work we can focus a bit more on the security right being able to say this is a level one user who is just messing around don't let them do trades over 10 bucks right and then once they messed around with a little bit more maybe they can go higher these sort of like security measures we just want to focus in that that land um curious to get your thoughts Alex it does sound a little bit like an episode of Black Mirror when you can like you know while watching a TV show you can like buy stuff from it right yeah yeah that's where we're heading yeah I I I think generally uh it's the definition of an agent still uh like I think a lot depends on the definition of an agent because you know an agent could be a very generic thing right like arbitrary python script running in a Docker container is an agent um to some extent right um but I do think um so so so a couple things that definitely necessary uh so you mentioned security that's an interesting one right so not only like hey this is $10 but also ideal in the agent Marketplace it's two sides like there there are consumers there are providers right so ideally you want uh to limit quite a bit what providers can do and that's something that um like I think s similar to how what would be a good example like generally uh in many aspects of our life security evolved as the system was building right like Windows 95 you can run any process and sure sure thing very quickly all the computers were you know covered in viruses because right and then there was like Windows NT which like patched it right but the security Mo modes are not adequate uh for what we use systems today right like for example Linux security model is not quite adequate for like you know when you use Pap or uh npm install you arbitrary code from some random person and and like really your security comes from uh from you believing that uh whomever is moderating that package manager is not uh evil first of all malicious second of all is actually paying attention right similar to hugging face like unknown to many a model in hugging face is an arbitrary script right so whenever you install a model from hugging face you're actually running someone's code uh from from under your user which is extremely dangerous right and so I think uh like the moral is that if you first build a cool system and then start doing your own Patchwork of uh of security you end up in a system where security is is non present right so I think like laying laying out a good security model ahead of time is is actually a very good idea uh and like putting a lot of effort into it and maybe like involving Security Experts like that would be something that is extremely valuable for the community as a whole mik micro actually built a uh uh script that just starts running I don't know how amazing it was but it was a good uh proof of concept that basically we we show before you run the transaction all the function calls that you're doing in the contracts and all that and below that we had a little thing that says in plain English this is what you're doing you're interacting with this instead of you know so yeah it's a it's going to be a crazy world we're going to right right because I think uh there are many uh so so so whether or not we want it uh at some point agents will have access obviously uh you know to the whole internet uh to money Etc right so it's a um yeah it's a I I think I have a kind of follow-up question here um so we're seeing for example like Google just um announced recently that they're enabling Gemini Nano to to run on on Chrome right so I think started in version 128 so sometime soon I think under some Flags you can you can enable like the model to run locally also we're seeing like um Apple partnering with open AI to like run natively sort of uh in in in the user's device uh or at least have access to a bunch more data and access to multitude of applications so I'm just really wondering what what's your guys's thoughts on this and um and I know I know we running quite late here but uh yeah just quite curious to see you know like near is building on the open web like we're all like we really want the open web to succeed and we want these open models and decentralized execution everything to to succeed but these are giant players that will have access to a lot of users a lot of devices and um what are you guys' thoughts on on this and uh is is this potentially harmful uh in the short term I I I think in case of open AI like you can think of open ai's decision in the in the From perspective of the fact that they have had an integration with Google for a while now right like if you go to Safari and type something in the address bar uh it goes to Google right so so uh it it sort of makes sense because Apple needs to meet expectations of the users and users are expecting to have access to uh to the cool stuff it is obviously harmful because uh open AI gets a massive FL uh data whe right so so they already they already quite ahead of everybody else in terms of how much data they get from the users once it becomes the standard uh on Apple devices they get significantly more data I I would expect uh I actually don't know the numbers but in my head what you get from integration with apple is order of magnitude bigger than what you get from chat GPT interface right so it is definitely harmful for the uh for the effort of building an acttion ual uh open um artificial intelligence right and it's sort of I think it's a um it sort of points at the fact that you know time is running out right so if we started as a community a year ago right and maybe if we had some good uh options to to offer today it would be different right I'm pretty sure apple is not happy with their you know privacy Center approach to integrate chat GPT just that open Solutions were not like I'm pretty sure they did some user studies right because they had an option to run llama but they went with chat GPT so like clearly they you know did some research and concluded that uh chat GPT turned out to be better it's it's unlikely that you know chat GPT is open a is bribing them just like from the uh you know relative sizes of the companies and the resources they have right but if there was a something competitive in the open I'm pretty sure Apple would have went with it uh and so I think uh that sort of but there will be uh like I think as time passes more and more uh the you know the status like the current market leader gets uh their Roots into the ecosystem and it's harder and harder to Dethrone them so obviously integration with apple is one of them and the faster we get to the state where where open uh uh open models compete like agents and infrastructure in general with the proprietary uh the the faster we will start turning the tide uh backward yeah maybe just to add I think the the challenge is that we're we're living in the world where we have two kind of operating systems as well right which what one of them is developed by the company that is building AI models as well um and the other one is indeed I mean the the the actually interesting thing that they they build the small model themselves the ROM device uh and then they are asking explicitly every time if you w to go at least from their demos to go to open AI um which yeah I think it kind of shows that they're not particularly happy with that approach but kind of need to need to get the best results possible so indeed if there is a way to kind of provide a privacy first way right where for example you run it inside secure enclaves on H 100s right in a distributed Network that would be something that may more preferred now the there is more the question is the data wheel actually that important going forward right if if the if like at some point there's a situation of amount of data you can use right and it's maybe becomes more about synthetic data and like in general other methods but uh we over time I I want to thank everyone for joining uh we had indeed a very wide range conversation hopefully you guys enjoyed it um and yeah excited to work on this together and actually build something that is better product and more do their own going forward thank you awesome","transcript_synced_at":"2026-11-03T05:47:22.123Z","segments":[{"start_seconds":0,"text":"light feel the custom feel hack the world and make it","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=0s"},{"start_seconds":12,"text":"tight blockchain","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=12s"},{"start_seconds":19,"text":"flows near to one we're close to won't stop us","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=19s"},{"start_seconds":26,"text":"now together we'll Rine game when near we'll find the flame","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=26s"},{"start_seconds":34,"text":"empowered Minds rise and sore use your own AI we'll","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=34s"},{"start_seconds":48,"text":"explore hold the future hand in hand in this Tech driven","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=48s"},{"start_seconds":57,"text":"Wonder okay hello hello to the new episode of AI office hours we have a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=57s"},{"start_seconds":64,"text":"great group here today uh let's uh start with uh Nate do you wna and M do you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=64s"},{"start_seconds":71,"text":"want to guys introduce yourself sure I'm Nate uh co- founder of bits of protocol previously uh mint base","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=71s"},{"start_seconds":79,"text":"the old nft Marketplace uh We've recently discovered in the last year of cranking well in the last five years of building an nft Marketplace within the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=79s"},{"start_seconds":88,"text":"last year of integrating with chat TBT and some other bits we've learned we can do anything on our nft platform through prompts and so that's","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=88s"},{"start_seconds":97,"text":"led us to believe we can now do defi swaps and now we're in this magical position of cross chain uh signatures so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=97s"},{"start_seconds":105,"text":"we're super excited to have our new name fits a protocol means please in German uh please AI build and execute this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=105s"},{"start_seconds":114,"text":"transaction for me and I'll hand it off to our other co-founder mikro who's been driving a lot of our AI initiatives also","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=114s"},{"start_seconds":122,"text":"the CTO of mbase yeah I mean I think you pretty much summed it up uh but yeah I've been in the near ecosystem for about three","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=122s"},{"start_seconds":130,"text":"years uh with mid-base as long as as as long as I have been with in the near ecosystem uh and super excited for","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=130s"},{"start_seconds":139,"text":"what's to come and uh what we've been building um in in the DI","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=139s"},{"start_seconds":147,"text":"space do you want to take it away uh wansu oh yes my name is hansu I'm co-founder of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=147s"},{"start_seconds":155,"text":"exhibits we're in the business of building the layer zero for decentralized compute um and we service","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=155s"},{"start_seconds":161,"text":"some of the biggest and most renowned um compute marketplaces in the world nice to meet","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=161s"},{"start_seconds":170,"text":"you good to meet you too good to have you guys so uh I know you guys had some questions uh from before so maybe maybe","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=170s"},{"start_seconds":180,"text":"we can launch into some of that uh I guess I'm sure Nate you have a lot of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=180s"},{"start_seconds":186,"text":"questions so sure so I was recently watching uh Black Mirror rewatching it","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=186s"},{"start_seconds":192,"text":"so the uh episode two to me was the most like intense one I think this was like from like over a decade ago maybe even","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=192s"},{"start_seconds":201,"text":"more uh where everything was a number right whether you brush your teeth you know you get a number reduced whether you're watching a horrific ad uh and if","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=201s"},{"start_seconds":210,"text":"you're not watching it you have to reduce your number uh and basically the people are just spinning their bike all day to acrw more number right this is","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=210s"},{"start_seconds":218,"text":"like the evil world that we're probably actually heading towards especially with AI um so the the big question is uh how","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=218s"},{"start_seconds":227,"text":"is connecting near protocol uh and the open- source world going to help sort of not have that world because there the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=227s"},{"start_seconds":235,"text":"ideal world is we all don't work anymore right um we all have these efficient machines doing efficient things but I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=235s"},{"start_seconds":242,"text":"mean how do we get how do we get to that uh and what's your sort of uh open open take on on the uh the savior of the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=242s"},{"start_seconds":251,"text":"Black anti-black Mirror ification yeah I I also remember that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=251s"},{"start_seconds":261,"text":"the person who who who really wanted to to to fix the system in that episode ended up being a big part of it um yeah","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=261s"},{"start_seconds":269,"text":"but uh well well I guess we the I actually don't know I didn't see that episode when it came out so it's hard for me to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=269s"},{"start_seconds":278,"text":"tell how relatable it was back then but yeah it is getting more and more relatable that's that's certainly uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=278s"},{"start_seconds":284,"text":"and uh I mentioned it I think like couple episodes again but the uh like every breakthrough that happened uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=284s"},{"start_seconds":292,"text":"since like uh the Advent of internet has been you know have has has been improving Our Lives in some ways but was","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=292s"},{"start_seconds":301,"text":"also making it more and more black mirror is and then like you know like social networks","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=301s"},{"start_seconds":309,"text":"uh generally uh web two right and uh and I think the big part of it is precisely","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=309s"},{"start_seconds":317,"text":"because um so so I think there are two parts here right I think both this episode and there's another one called nose dive","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=317s"},{"start_seconds":324,"text":"which is sort of with a similar uh similar message I think I think they they sort of present two different","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=324s"},{"start_seconds":333,"text":"aspects of so one of them is is this Society right so so like there's a big part which which has nothing to do with centralization that Society is building","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=333s"},{"start_seconds":341,"text":"uh you know other uh so so that I don't think we will address in anyway that needs to be addressed uh you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=341s"},{"start_seconds":348,"text":"collectively as humans but the another one is that yeah uh as centralization or rather as every breakthrough happens uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=348s"},{"start_seconds":357,"text":"it it uh so far hisor ially it inevitably becomes controlled by a small group of people who the moment they","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=357s"},{"start_seconds":364,"text":"fully capture the the market they optimize for Revenue not for Value right and obviously optimizing for Revenue sometimes is aligned with optimizing for","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=364s"},{"start_seconds":372,"text":"Value but not always right and example I like using is that if social graph was a public good we would have way more usages of it uh than we actually have","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=372s"},{"start_seconds":382,"text":"today right primarily today it is used you know for for engagement uh farming and it's a it","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=382s"},{"start_seconds":390,"text":"optimizes for Revenue doesn't optimize for value at all and so in this case AI could be a very big uh change in the way","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=390s"},{"start_seconds":400,"text":"we uh do everything uh and if it is again controlled by small set of people whatever values they uh you know state","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=400s"},{"start_seconds":408,"text":"today those values can change similarly to how Google was don't do evil uh and became what it is today right and and like I think there's a good chance that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=408s"},{"start_seconds":416,"text":"Larry and Sergey had that value back in the day when they started right so like whatever values those companies stay today they will they will change right","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=416s"},{"start_seconds":425,"text":"once once you become you know a trillion dollar company you you naturally become this you know profit making Machinery so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=425s"},{"start_seconds":432,"text":"it's very important that for AI there is a uh um well like I think it is early","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=432s"},{"start_seconds":440,"text":"enough that a sufficiently strong effort to make it free and open from the GetGo will win uh and then we will at least avoid that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=440s"},{"start_seconds":448,"text":"Pitfall and we'll have you know will have a strong technology which works for the good of people not","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=448s"},{"start_seconds":454,"text":"for the uh for the small set of investors yeah I would add to figuring","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=454s"},{"start_seconds":465,"text":"out how do we how do we have a more Collective good optimization right","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=465s"},{"start_seconds":474,"text":"versus single centralized optimization which is just a natural way like if a centralized company trains a model","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=474s"},{"start_seconds":482,"text":"whatever kind of individual pieces are trying to do at the end it's going to be for the benefit of that centralized company right that's just how does","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=482s"},{"start_seconds":490,"text":"things work and so we I mean I think blockchain have been struggling with this like hey this is a way to organize","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=490s"},{"start_seconds":498,"text":"people at scale for Collective good and it's easy to do when it's something","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=498s"},{"start_seconds":506,"text":"that's like very simple to describe right you know let's all just hash you know and find find an non starts with","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=506s"},{"start_seconds":514,"text":"zero and that actually protects Network and it gets harder and harder right with you know things like gitcoin been trying to do this at at like in the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=514s"},{"start_seconds":524,"text":"more kind of abstract ways and so I think like here we have something that's like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=524s"},{"start_seconds":531,"text":"extremely demanding figuring this out where we need to um","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=531s"},{"start_seconds":539,"text":"globally have a model that's optimized for successive individuals and communities","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=539s"},{"start_seconds":547,"text":"versus uh optimized for any single Company Success and and that's indeed very","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=547s"},{"start_seconds":560,"text":"challenging I'm you're you're muted Nate yeah and it's it's already doing its job right a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=560s"},{"start_seconds":568,"text":"little bit even though you're working with you know private llms we're also dealing with private uh Financial systems with their own closed systems","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=568s"},{"start_seconds":577,"text":"whereas blockchain is already open source right so if we have that bridge we're already sort of doing the good like we don't have to do much more","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=577s"},{"start_seconds":585,"text":"because we're now bridging that working with open source contracts working with a much more open world even if we're","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=585s"},{"start_seconds":592,"text":"still working with closed Source uh big models we're kind of already starting that that step for forward into the the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=592s"},{"start_seconds":600,"text":"the non- eval or being able to interact with different contracts that can be replaced by you know other folks that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=600s"},{"start_seconds":607,"text":"can fork and do something a little bit","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=607s"},{"start_seconds":619,"text":"better do you want to say something um sorry I wasn't too sure about the formatting um no I think","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=619s"},{"start_seconds":627,"text":"that's an important question um but but um can we switch gears uh and um perhaps","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=627s"},{"start_seconds":633,"text":"discuss another topic yeah go ahead you know I was just following what's what's been happening with Nvidia in the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=633s"},{"start_seconds":640,"text":"marketplace um I don't know if you any of you guys are the the big news is NVIDIA briefly last year um you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=640s"},{"start_seconds":648,"text":"surpassed the market capitalization of Microsoft and Apple and since then it's","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=648s"},{"start_seconds":654,"text":"it's sort of LW some steam um but uh you know one of the big questions that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=654s"},{"start_seconds":662,"text":"I'm grappling with these days is sort of the long run vision of where things are headed with AI you know so when you see","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=662s"},{"start_seconds":671,"text":"this big rise and the spotlight of Nvidia in the public markets you know the first question is where where is the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=671s"},{"start_seconds":678,"text":"value chain you know is where is a lot of this value in AI going to be captured and it appears right now Nvidia is sort","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=678s"},{"start_seconds":686,"text":"of the kicking butt in doing that um and then you know when you look at the competitive landscape of of AI models","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=686s"},{"start_seconds":694,"text":"you know you have you have Claw from anthropic you have mistro B from Google you know meta you know everyone's building their own","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=694s"},{"start_seconds":702,"text":"AI um and then I recently came across this research paper uh by MIT researchers it's called the platonic","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=702s"},{"start_seconds":710,"text":"representation hypothesis okay and um so these authors they found that representations in AI models um","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=710s"},{"start_seconds":718,"text":"particularly deep networks are converging so as models get larger different neural networks representing","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=718s"},{"start_seconds":725,"text":"data are becoming more aligned um and the implication is that there's going to be this in the long run a convergence of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=725s"},{"start_seconds":733,"text":"all of the models you know the more and more you train them with with more data and more parameters they they will essentially become all","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=733s"},{"start_seconds":741,"text":"identical um uh you know so I'd like to get your thoughts on this and I mean like an immediate","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=741s"},{"start_seconds":749,"text":"uh sort of a followup that comes to mind is you know if these mods are converging um you know open AI anthropic Mistral","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=749s"},{"start_seconds":757,"text":"you know um what what are their modes in the long run right you know or or do they just become completely","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=757s"},{"start_seconds":765,"text":"indistinguishable I want to open up the floor and get your thoughts I think interesting question is do they already have like do they have a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=765s"},{"start_seconds":774,"text":"mode today like let's say models do not converge right I think they like at the given size with the given amount of data the capabilities are sort of similar so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=774s"},{"start_seconds":783,"text":"in the absence of a breakthrough uh you do not get much of a mode right so like M and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=783s"},{"start_seconds":790,"text":"llama uh they kind of uh like like you know already somewhat indistinguishable","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=790s"},{"start_seconds":798,"text":"From perspective of the downstream tasks I think the fact if the models actually convert right I didn't I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=798s"},{"start_seconds":805,"text":"skimmed the paper uh before before the this call but I didn't look very deeply but if they did converge that's actually","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=805s"},{"start_seconds":813,"text":"good news that effectively means that at some point uh there is no benefit in in creating more and more models so once","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=813s"},{"start_seconds":819,"text":"you have one uh model that you know very close to that convergence it's as good as it gets so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=819s"},{"start_seconds":827,"text":"so that that is potentially good news especially if it also if it also implies I don't know if it implies it but if it implies that there is the plateau at","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=827s"},{"start_seconds":835,"text":"which point you cannot go you know any further that's even better that means that the model will not be getting better and better uh and will not","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=835s"},{"start_seconds":841,"text":"eventually replace us right it's even better if that Plateau is is our level of intelligence that would be outstanding uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=841s"},{"start_seconds":852,"text":"yeah but I think I mean there there there's few interesting points here right one is you asked about the value chain and indeed right now we see","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=852s"},{"start_seconds":862,"text":"massive investment in kind of compute because even if training itself is you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=862s"},{"start_seconds":871,"text":"know takes 10 20 50,000 gpus the thing is that the inference St can take any","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=871s"},{"start_seconds":878,"text":"amount of gpus right afterwards because if now you're doing inference kind of you know one token at a time you can","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=878s"},{"start_seconds":885,"text":"imagine as this like instrumentation gets more sophisticated you will want to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=885s"},{"start_seconds":892,"text":"run you know you know Chain of Thought you'll want to run multiple models at the same time to like on different different subtasks so you can imagine","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=892s"},{"start_seconds":902,"text":"that the amount of compute you can you know spend an inference will not be bounded by any kind of even magnitude so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=902s"},{"start_seconds":911,"text":"even if we imagine like all let's imagine like at we trained you know indeed what Al said the gold model that's like this is a final model we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=911s"},{"start_seconds":918,"text":"don't need to train anything new though I mean obviously there's like lots of asteres on discussion there but even then like at an inference time you'll","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=918s"},{"start_seconds":926,"text":"need you can use as much computer as want right to actually produce the result you want right and we already see this with you know you run multiple","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=926s"},{"start_seconds":934,"text":"agents they themselves refining results they read you know the read the context and may be like oh let me write that let","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=934s"},{"start_seconds":941,"text":"me run it with tests let me do this or that um and so indeed I mean clearly the market you know reacts to that that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=941s"},{"start_seconds":950,"text":"there's like this amazing massive investment in the hardware that happens um will that stop I mean","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=950s"},{"start_seconds":959,"text":"probably not anytime soon right just because as I said like there's just so so much more you can do if you have that compute at INF time and if you multiply","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=959s"},{"start_seconds":968,"text":"by a billion people well you need like massive amount of amount of this to to serve it now","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=968s"},{"start_seconds":978,"text":"indeed on the training side I mean I've been saying this for over a year now that I think like model fun foundational models will be a commodity and kind of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=978s"},{"start_seconds":986,"text":"indeed may not they may not be like completely converging uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=986s"},{"start_seconds":993,"text":"but it makes sense that at the end you know you're training it on somewhat the same data you're using the same method","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=993s"},{"start_seconds":1000,"text":"it kind of makes sense you get the same results at the end uh so I don't know yeah I haven't read the paper outside of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1000s"},{"start_seconds":1008,"text":"abstract so I'm actually curious what are they comparing because obviously you cannot go compare CHP to CLA um outside","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1008s"},{"start_seconds":1016,"text":"of maybe the embeddings of the surface um right um and it appears U I mean you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1016s"},{"start_seconds":1024,"text":"know if you're using the same data the the underlying mathematics is roughly equivalent then then right just comes","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1024s"},{"start_seconds":1031,"text":"down to who has access to most compute so they can have the biggest and the most complicated model um now so for me","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1031s"},{"start_seconds":1039,"text":"when I think about this I I also kind of uh you know I think there are some really important implications about","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1039s"},{"start_seconds":1046,"text":"ownership so let's say um uh scaling we're sort of nowhere near the plateau you know so we get uh you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1046s"},{"start_seconds":1055,"text":"you can hear guys think Mark Zuckerberg and his podcast you know admit in in their Facebook in The Meta Labs um they","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1055s"},{"start_seconds":1062,"text":"have yet to see um any signs of diminishing uh Mar marginal returns to scaling you know they're seeing like you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1062s"},{"start_seconds":1070,"text":"know the more you pump up the models with more compute more data um it's it's doing more and more fascinating things","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1070s"},{"start_seconds":1078,"text":"um so let's then assume that for examp assume that for a moment that you know we can continue to scale up the models","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1078s"},{"start_seconds":1085,"text":"become bigger and bigger um and sort of just seeing and going going back to the whole Invidia story of where things are","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1085s"},{"start_seconds":1094,"text":"with them you know since compute at scale is just so Capital intensive right you know this is the reason why only big","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1094s"},{"start_seconds":1101,"text":"tech companies are really able to even have carry out the ambition of training these Large Scale Models um it seems like a scenario in which several","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1101s"},{"start_seconds":1110,"text":"competitors you know spending hundreds of billions or even trillions of dollars to replicate essentially the same AI","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1110s"},{"start_seconds":1119,"text":"just seems too inefficient to be viable right you know so so if one is convinced that AI brings immense public value um I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1119s"},{"start_seconds":1127,"text":"certainly I certainly am then what one can apply the reasoning that um large Investments with really challenging","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1127s"},{"start_seconds":1136,"text":"economics that provide significant public public benefits are typically financed you know by a Central Public Authority right or the government so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1136s"},{"start_seconds":1144,"text":"think of you know think of utilities infrastructure or other Investments that require you know really complex multiple stakeholder alignments the highways the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1144s"},{"start_seconds":1153,"text":"power grid gas pipelines you can name them so you know like is there a future in which the US government will attempt","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1153s"},{"start_seconds":1161,"text":"to fund and subsequently hold ownership of an AGI um you know you know I guess we're already seeing this is the case in China","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1161s"},{"start_seconds":1169,"text":"and kind of kind of thinking ahead um you know we just as Elon Musk had a had to innovate and engineer reusable","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1169s"},{"start_seconds":1178,"text":"Rockets to bring down the cost of launches you know demand for scaled compute we need to uh discover some","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1178s"},{"start_seconds":1185,"text":"dramatic efficiencies to remain privatized right and um um I guess or or maybe the need for a some sort of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1185s"},{"start_seconds":1194,"text":"creative ownership structure is where we may get substantial input or you a solution from web 3 um and this was sort","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1194s"},{"start_seconds":1201,"text":"of where I started to connect the dots and see um you know it where to find that intersection between decentralized","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1201s"},{"start_seconds":1210,"text":"computer or Ai and where the the scaling needs are but you love to get your further input on this oh I would really","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1210s"},{"start_seconds":1219,"text":"hope to never have the government run the full AI system it's so I mean if you been to a California DMV have you ever","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1219s"},{"start_seconds":1226,"text":"been to like it's just so slow and brutally and that's why NASA is basically getting overrun by space I'm","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1226s"},{"start_seconds":1234,"text":"not aent of a central ownership right but but just looking at historical examples you know when you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1234s"},{"start_seconds":1241,"text":"need a ton of resource to build out something that everyone agrees has tremendous value you know is AI going to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1241s"},{"start_seconds":1250,"text":"enhance human productivity absolutely yes right um so it's it's sort of inevitable reality that someone some","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1250s"},{"start_seconds":1258,"text":"somewhere is going to build this AGI right with enough resources uh but typically when you get uh you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1258s"},{"start_seconds":1266,"text":"historically when you have something that's incredibly Capital intensive and if what everyone is competing to build out essentially the same thing and it","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1266s"},{"start_seconds":1274,"text":"costs trillions of dollars to build um you know in equilibrium it's most likely um a much","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1274s"},{"start_seconds":1282,"text":"bigger uh socially neutral agency is going to build that you know so you know if if if we to avoid that fate uh you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1282s"},{"start_seconds":1290,"text":"know what are some things that we would have to see develop um and one idea is know web 3 brings in this very creative","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1290s"},{"start_seconds":1299,"text":"uh decentralized ownership model and I think that's one of the aspects that people haven't really thought very hard","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1299s"},{"start_seconds":1308,"text":"as much as you want to see this intersection in web 3 and AI yeah so I mean as I mentioned like we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1308s"},{"start_seconds":1316,"text":"need to figure out I mean this is literally the problem of the public goods funding right just at at a larger scale that so far has been tested out in","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1316s"},{"start_seconds":1325,"text":"crypto and sadly Beyond I mean like the biggest working public goods funding is Bitcoin where um","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1325s"},{"start_seconds":1333,"text":"you know everybody runs a minor and they're getting funded for doing that um","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1333s"},{"start_seconds":1340,"text":"and so there's a clear protocol how you get rewarded for providing public goods service while you have a self-interest","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1340s"},{"start_seconds":1347,"text":"in participating in it and so in a way like we kind of need very similar mechanism but with a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1347s"},{"start_seconds":1356,"text":"lot fine grain more fine grain research Focus right where um like I actually don't think","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1356s"},{"start_seconds":1364,"text":"like you know going in massively training like training massive modeling at this point is is even that interesting because like we've seen you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1364s"},{"start_seconds":1372,"text":"know some of the new papers showcasing that even with smaller models if you Leverage search or like Monte Carlo","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1372s"},{"start_seconds":1380,"text":"search or you know using to reinforce doing thought uh tokens internally as you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1380s"},{"start_seconds":1388,"text":"predict things you're able to re you know kind of increase the capacity of the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1388s"},{"start_seconds":1395,"text":"model um just again because you're moving the compute pretty much out of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1395s"},{"start_seconds":1402,"text":"the weights into the kind of that inference time to actually into activations and is also just I mean the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1402s"},{"start_seconds":1409,"text":"fact there also this new paper on inductive bias to that propagation is also crazy right I don't know if you guys seen but they pretty much fine tune","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1409s"},{"start_seconds":1417,"text":"on examples and the model was able to pretty much infer the function from fine","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1417s"},{"start_seconds":1423,"text":"tuning so we clearly have like new methodologies that are I think emerging","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1423s"},{"start_seconds":1430,"text":"around using optimizers using search and kind of intern like intermittent thought","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1430s"},{"start_seconds":1437,"text":"process inside model to actually improve the results and so what we're now missing is really an ability to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1437s"},{"start_seconds":1446,"text":"coordinate researchers to really swarm on this ideas and test them out and and see exactly where the ins and out of this are because the challenge is like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1446s"},{"start_seconds":1455,"text":"somebody published the paper you know there's some source code they trained it on they did it on like a 7B model with a little bit of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1455s"},{"start_seconds":1462,"text":"fine-tuning um and like you you don't really know how much more you can do you don't know if it works in 70b so I think this is where we have right now","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1462s"},{"start_seconds":1470,"text":"an opportunity where you know you guys as exhibits for example can be like hey we actually going to give a grant to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1470s"},{"start_seconds":1477,"text":"this researchers to for this amount of compute and this is actually something we're trying to work out how to do","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1477s"},{"start_seconds":1483,"text":"effectively um but let's try to run this model like train this model on 70b now","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1483s"},{"start_seconds":1490,"text":"and like again F tuning from existing model and see you know if this this properties hold if they do then","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1490s"},{"start_seconds":1498,"text":"potentially is a state-of-the-art across a bunch of parameters um and now this is still all in open source because we kind","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1498s"},{"start_seconds":1506,"text":"of gave a grant for this to be in open source and you get you know promotion pretty much from doing this as well uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1506s"},{"start_seconds":1514,"text":"and so people want to use it for inference because they know it runs on your Hardware in the first place so I think like that's kind of where what what we're missing like some of this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1514s"},{"start_seconds":1521,"text":"mechanism of collaborative funding research where we have researchers right now making progress not not on like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1521s"},{"start_seconds":1529,"text":"let's train a bigger model right I mean the the amount of research there is like go into config add a zero right start and then babysit that gigantic model","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1529s"},{"start_seconds":1537,"text":"that you know crashes all the time versus like people are actually making progress on you know structure and and and uh you know ways to optimize things","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1537s"},{"start_seconds":1546,"text":"and then like how do we fund that how do we leverage that and and again how do we like replicate it in such a way that then people know okay this is actually","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1546s"},{"start_seconds":1554,"text":"works let's incorporated into next technique so yeah that I pretty much I agree with you and I think web three is","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1554s"},{"start_seconds":1561,"text":"like have the methods and and actually funding to do that now and actually showcase not on kind of some like you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1561s"},{"start_seconds":1570,"text":"know in theory you can train a bigger model right as you collect a lot of clusters but actually like we can get interesting research out of the existing","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1570s"},{"start_seconds":1577,"text":"Academia researchers by just being able to provide them access and kind of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1577s"},{"start_seconds":1584,"text":"manage that a little bit more freely and fluidly right and and uh you know our our","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1584s"},{"start_seconds":1592,"text":"business mission is very much align uh with that with that uh with that view um and we we take it very serious seriously","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1592s"},{"start_seconds":1600,"text":"upon ourselves U to kind of provide that sort of democratization and access to compute you know for various like these","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1600s"},{"start_seconds":1608,"text":"these other applications that are currently under research and to make it Equitable for anyone to access compute you know want to build out models um","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1608s"},{"start_seconds":1619,"text":"but neat I got I got a question so if uh since I have you guys I'm excited to ask","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1619s"},{"start_seconds":1627,"text":"it's fun to talk about tomorrow and what's what's coming up in the future but it's also really fun to talk about what what do we have now what can we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1627s"},{"start_seconds":1634,"text":"play with today on near and where things are heading so I'm sure you've sort of seen uh we've built that AI agent allows","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1634s"},{"start_seconds":1642,"text":"you to generate images uh it builds a transaction that you can execute you can create drop links we've now built a ref","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1642s"},{"start_seconds":1649,"text":"Finance agent that allows you to do defi uh Swap this for that uh and we're going to be opening up that agent Marketplace","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1649s"},{"start_seconds":1658,"text":"for other folks to start plugging in potluck uh just took them three days to to build basically their core pits of their agents you'd be able to say fund","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1658s"},{"start_seconds":1666,"text":"this company for that uh where do you want to see this heading uh where should we start uh plugging away today uh and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1666s"},{"start_seconds":1674,"text":"what do you see value in basically AI building transactions uh first and B uh what what interesting use cases should","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1674s"},{"start_seconds":1683,"text":"we really tackle if if you think that there's value there because again I think we're going we're building the Black Mirror uh kind of theory it's","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1683s"},{"start_seconds":1692,"text":"being able to communicate with the financial system through human readable conversations uh and I'm just curious to get your input on on what we've built so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1692s"},{"start_seconds":1701,"text":"far and and where where we can head to help uh push the near ecosystem further","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1701s"},{"start_seconds":1710,"text":"I mean I think the biggest opportunity right now is to find the consumer applications which can integrate this right I mean there's both there's few","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1710s"},{"start_seconds":1718,"text":"interesting consumer applications in crypto that are that can leverage this right away usually they are in already","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1718s"},{"start_seconds":1727,"text":"existing chat uh environment so this is Telegram and Discord Bots um where if you plug that in uh and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1727s"},{"start_seconds":1737,"text":"create economic incentives around for example Community managers and kind of channel uh server owners to benefit from","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1737s"},{"start_seconds":1747,"text":"originating transactions from from these channels uh you you know make it really easy to create your accounts either through existing wallets or embedded","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1747s"},{"start_seconds":1756,"text":"wallet directly there you can pretty much like supercharge the both web three adoption and kind of amount of accounts","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1756s"},{"start_seconds":1765,"text":"that are transacting through this because you know to just give you an example like cab land right has over I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1765s"},{"start_seconds":1772,"text":"think on their website over 10 uh users across Discord and telegram across 50,000 different communities and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1772s"},{"start_seconds":1781,"text":"so you know you can imagine that bot becoming also originating transactions through natural language same you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1781s"},{"start_seconds":1788,"text":"with hot with the telegram Bots as well so I think and then obviously going into web 2 as well and seeing how to connect","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1788s"},{"start_seconds":1796,"text":"with their existing ecosystems uh and you know turning on pretty much the um the financial","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1796s"},{"start_seconds":1805,"text":"component of what right now a lot of this assistants are just like there is a bunch of stuff you can do and then you still need to go and like book it and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1805s"},{"start_seconds":1813,"text":"figure it out yourself right the real CR of that is that you're you know incentivizing folks so how much is it uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1813s"},{"start_seconds":1820,"text":"folks that are actually wanting to get value out of what you're doing or how much of it is folks that just want to get paid with the tokens so it's it's a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1820s"},{"start_seconds":1829,"text":"it's something that we're looking into deep uh and it's kind of a weird line that we're trying to to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1829s"},{"start_seconds":1836,"text":"ride but oh go","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1836s"},{"start_seconds":1842,"text":"ahead go ahead I'm I'm curious your your answer no well I think I mean the general expectation should be you you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1842s"},{"start_seconds":1852,"text":"like same as credit card right you should not give you know 100% money back that just doesn't work um you should get","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1852s"},{"start_seconds":1861,"text":"you know if you're building out a loyalty P program which is what all of these token you know user distributions are they should be percentage of the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1861s"},{"start_seconds":1869,"text":"value that is being transacted um and so and at the same time it should be something that's like below the margin","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1869s"},{"start_seconds":1879,"text":"that you're capturing then it's a sustainable program that you can run I think crypto indeed kind of went into like Hey we're going to just you know we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1879s"},{"start_seconds":1887,"text":"need to distribute this token so we're going to spend all of it uh kind of at once for everyone who's even you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1887s"},{"start_seconds":1894,"text":"barely tried our product that obviously is not a good idea um although you know for a while that's been the way people","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1894s"},{"start_seconds":1901,"text":"want to approach it but you can imagine you know this is your gross engine so like how do you create a loyalty program that you know benefits from that and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1901s"},{"start_seconds":1910,"text":"potentially yeah you overpay in the beginning with the idea that your margins will improve over time but I think the the benefits","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1910s"},{"start_seconds":1918,"text":"you potentially can have is transitioning away from like it should be valuable for the users to urgin","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1918s"},{"start_seconds":1926,"text":"transactions uh but the people who may be teaching and encouraging and providing a UI Etc like you can you can","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1926s"},{"start_seconds":1935,"text":"pretty much uh reward them which will be a lot I mean a lot less like again if you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1935s"},{"start_seconds":1942,"text":"integrate it into some bot you're either rewarding the bot creators or the community that installed this body again","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1942s"},{"start_seconds":1949,"text":"hot there's Villages and collab land and all of this like there servers that are actually you know there's Community managers behind them and so they can be","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1949s"},{"start_seconds":1958,"text":"the ones educating the user base about what they can do through this and so you can reward them versus individual users","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1958s"},{"start_seconds":1967,"text":"there and so you're kind of preventing a lot of the uh negative of usage farming","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1967s"},{"start_seconds":1974,"text":"right engagement farming that happens transaction building real world use","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1974s"},{"start_seconds":1982,"text":"cases you got you got any Alex uh I think anything which is more complex than","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1982s"},{"start_seconds":1991,"text":"a uh like a simple transfer uh could benefit like like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1991s"},{"start_seconds":1998,"text":"uh like I think as an example as simple as staking is uh causing a lot of confusion so so so if that was possible","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=1998s"},{"start_seconds":2006,"text":"like you know I go there and like uh I have like a simple task of I guess","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2006s"},{"start_seconds":2013,"text":"especially uh uh like even with the example of staking right even staking on itself could be confusing uh with like you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2013s"},{"start_seconds":2021,"text":"hey there's like withdrawal you know interval Etc then you have liquid staking reaking Etc I think like going","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2021s"},{"start_seconds":2028,"text":"there and communicate communicating with something that will tell you your options uh","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2028s"},{"start_seconds":2034,"text":"and uh executing it from there is quite helpful yeah from a single prompt as well you're not going from one UI piece","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2034s"},{"start_seconds":2043,"text":"to another you're not going to ref to something else you're in the single prompt where it's like Swap this token for this token now stake it stake half","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2043s"},{"start_seconds":2050,"text":"of it um and then send the other half to my ledger wallet right all within a single prompt and these things we can start doing with like bulk prompts right","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2050s"},{"start_seconds":2059,"text":"being able to send multiple transactions through a single prompt right mint an nft generate the image load it to R","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2059s"},{"start_seconds":2066,"text":"weave listed for five bucks could be one single prompt and that's or and do one on eth and do one on near it sounds","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2066s"},{"start_seconds":2074,"text":"crazy but this is this is uh this is where we're at it's uh I I just think it's it's uh working and focusing more","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2074s"},{"start_seconds":2083,"text":"on that the agent side is uh and incentivizing those agents to build better agents is is probably where we're we're going to be","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2083s"},{"start_seconds":2092,"text":"touching in micro you got anything sure um yeah I was thinking","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2092s"},{"start_seconds":2099,"text":"since you're talking about the multiple prompts or you know submit a prompt that's composition of multiple prompts I was wondering to hear your guys thoughts","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2099s"},{"start_seconds":2108,"text":"on uh the future of prompt engineering is it you know a real thing will it become like even more um of a real thing","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2108s"},{"start_seconds":2116,"text":"in the future or you know as models evolve and like uh kind of understand better and better like what users want","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2116s"},{"start_seconds":2124,"text":"uh it will kind of Fade Away over over time UI generation","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2124s"},{"start_seconds":2133,"text":"yeah I mean I think it it's going to be a combination of things right because you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2133s"},{"start_seconds":2140,"text":"still unless it's like a completely autonomous agent that you know goes and figures out everything from a A to Z and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2140s"},{"start_seconds":2148,"text":"you just gave it like a super high level task the problem is like you still want something that's like so I mean the example we're doing on AI developer","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2148s"},{"start_seconds":2156,"text":"right now the it's not even about the prompt it's like you still want to have a conversation with it because as part of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2156s"},{"start_seconds":2163,"text":"that conversation which refines your understanding and kind of you answer questions you have tradeoffs you're like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2163s"},{"start_seconds":2170,"text":"figuring out what needs to be done and so even if the model itself is kind","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2170s"},{"start_seconds":2178,"text":"of conditioned on that uh you still need that kind of like I wouldn't call that prom engineering","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2178s"},{"start_seconds":2185,"text":"per se but like you need that convers system that actually figures out from the user what you want right um and so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2185s"},{"start_seconds":2193,"text":"that can come in the form of a one shot prompt where like I just know exactly what I'm doing but it's more probably","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2193s"},{"start_seconds":2200,"text":"going to become like you can imagine a a prompt bot right which you go to it like asks you what you want it figures out","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2200s"},{"start_seconds":2210,"text":"like exactly all the tradeoffs ETC and then it creates a prompt for another AI to then do it repetitively already now","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2210s"},{"start_seconds":2217,"text":"knowing exactly what you know the process and thinking you've done um and so I'm imagining something like this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2217s"},{"start_seconds":2225,"text":"will happen where uh AIS actually will write kind of the the programmatic like what what we right now have for programmatic prompts right that are like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2225s"},{"start_seconds":2233,"text":"behind an agent right where user doesn't interact with that and then on the user side it will be more of like a conversational plus UI","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2233s"},{"start_seconds":2241,"text":"plus like we actually use a spec as a format where it like writes out stuff and then you can read that and like correct it what you want right and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2241s"},{"start_seconds":2250,"text":"that's actually really useful because you kind of see the whole context you're like oh I want this user stor or like actually I want this new journey I want this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2250s"},{"start_seconds":2258,"text":"feature and so it gives you like a lot more context on what's kind of in the model on the other side so I I'm","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2258s"},{"start_seconds":2265,"text":"imagining more of that and then kind of generated programmatic prompts that go into the backend of the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2265s"},{"start_seconds":2275,"text":"engine I of people disagree with","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2275s"},{"start_seconds":2288,"text":"me it's kind of a a tricky conversation because uh Hano is very back in hardcore","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2288s"},{"start_seconds":2295,"text":"uh hardware and all those pieces and we're front in app layer so it's interesting Bridging the Gap I don't want to take up too much time onto so go","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2295s"},{"start_seconds":2303,"text":"for it no that's okay um well so I mean since we're getting into","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2303s"},{"start_seconds":2310,"text":"the some of these technical questions um so as a crypto compute project U so exhibits is trying to find methods to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2310s"},{"start_seconds":2320,"text":"provide users with more flexible ways to use Compu you know so um think of the the following examples so for in compute","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2320s"},{"start_seconds":2327,"text":"defi being swap e on ethereum for a compute voucher and salana um for compute","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2327s"},{"start_seconds":2335,"text":"lending uh use so is collateral to borrow compute registered on near so we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2335s"},{"start_seconds":2342,"text":"we think n's um chain abstraction can help help us abstract complexity of blockchain away from the user which I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2342s"},{"start_seconds":2350,"text":"think is a a very valuable and and critical um barrier uh for for wider","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2350s"},{"start_seconds":2358,"text":"adoption and is so you know first off is that understanding correct and how long do you think it would take to realize","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2358s"},{"start_seconds":2366,"text":"such a technical system yeah I mean well chain abstraction for","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2366s"},{"start_seconds":2372,"text":"sure is designed to well kind of as the idea to remove the need to think about","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2372s"},{"start_seconds":2381,"text":"which chain things are on and instead move into user space where you just","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2381s"},{"start_seconds":2388,"text":"indeed as as Nate said like just describing with natural language what you want right um and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2388s"},{"start_seconds":2395,"text":"so I mean the the use cases you described are exactly the type of things that we actually have projects that are","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2395s"},{"start_seconds":2403,"text":"using chain signatures to are and building right kind of a multi-chain defi multi-chain lending that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2403s"},{"start_seconds":2412,"text":"uh that are working where yeah you can you know borrow against a combination of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2412s"},{"start_seconds":2418,"text":"Bitcoin Dodge Coin you know some some M lady and a uh you know some locked vcv","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2418s"},{"start_seconds":2427,"text":"on ethereum and then you can borrow usbd on to and similarly you can swap you know any combination of assets across","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2427s"},{"start_seconds":2436,"text":"different chains so that is indeed kind of the direction we're going as well as on the wallet side right kind of what","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2436s"},{"start_seconds":2443,"text":"beta is doing what H is doing where you can kind of have like a singular view of assets across all chains you don't actually need to think about like where","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2443s"},{"start_seconds":2452,"text":"they are and you can navigate it you know either UI or natural language where you say hey you know I just want a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2452s"},{"start_seconds":2459,"text":"position like I mean I just typing an example like hey I want an index of top five coins across all chains right and you know buy $ five dollar worth of it","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2459s"},{"start_seconds":2467,"text":"right like I don't want to figure it out like how to do that it should just go you know fetch from coin market cap","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2467s"},{"start_seconds":2474,"text":"build me an index and then you know buy kind of across all of them and say me like oh I want you know I want to be able","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2474s"},{"start_seconds":2482,"text":"to have my AI run for an hour thinking about this problem and you know like I'll pay is whatever","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2482s"},{"start_seconds":2491,"text":"eum and so it you know buys a kind of compute uh coin in the way and then pays","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2491s"},{"start_seconds":2499,"text":"for it when it actually runs the inference on the back end right for an agent that goes and like navigates and thinks through and writes Out The Chain","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2499s"},{"start_seconds":2508,"text":"of Thought and maybe writes some code and does all the things and then comes back with you with an answer and saying like hey I spent You Know This Much","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2508s"},{"start_seconds":2516,"text":"computer pretty much on it so like I can imagine all of that is kind of coming together pretty quickly actually because","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2516s"},{"start_seconds":2524,"text":"um I know you know folks with wallets already doing s this the applications","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2524s"},{"start_seconds":2531,"text":"kind of are building this like multichain def5 multichain Ling are as well have like their kind of alpha","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2531s"},{"start_seconds":2538,"text":"versions that come together so I would I would expect within like three months we have some of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2538s"},{"start_seconds":2545,"text":"that kind ofun that's where we're also heading to with the UI uh building on the Fly being able","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2545s"},{"start_seconds":2554,"text":"to not only do swaps and these things but say give me a report of the charts of this and this and then it just builds","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2554s"},{"start_seconds":2561,"text":"a nice pretty graph and then zoom in a bit more on that okay execute my swap on this based on these forms that I've I've","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2561s"},{"start_seconds":2569,"text":"processed so yeah I uh yeah I'm very on boarded","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2569s"},{"start_seconds":2581,"text":"yeah turn the table back to you Nate or Alex do you have more to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2581s"},{"start_seconds":2589,"text":"add now we can move on if if people have other fun questions uh okay what would be your","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2589s"},{"start_seconds":2599,"text":"ideal scenario for an open Agent Market connected with near at the core so I know you guys are working on an agent agent Marketplace as well I don't know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2599s"},{"start_seconds":2607,"text":"how much you want to reveal of that land I know it was revealed in a couple open uh office hours before but just","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2607s"},{"start_seconds":2616,"text":"curious because we're we're also working on probably something that we could work","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2616s"},{"start_seconds":2623,"text":"together deeply on I'm curious how we can work together to really build just an epic system that allows not just","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2623s"},{"start_seconds":2631,"text":"ecosystem Partners but with this cross chain you know world that we're we're jumping into how do we basically start taking over other chains and I think we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2631s"},{"start_seconds":2639,"text":"can do it pretty quick uh if if we execute this right together right um so the biggest question is what is what is your ideal agent Marketplace that is","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2639s"},{"start_seconds":2648,"text":"really focused on sort of either the execution layer or so","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2648s"},{"start_seconds":2657,"text":"on I mean I can mention maybe a few components we're looking at right so the way we're thinking about it is kind of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2657s"},{"start_seconds":2664,"text":"agent is anything that um Can can take an environment and we can talk about what","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2664s"},{"start_seconds":2672,"text":"the environment is and perform set of actions right and so it has access to llm it has access to you know tools it","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2672s"},{"start_seconds":2680,"text":"has access to any other stuff including submitting transactions so you can model that as just a python you know file","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2680s"},{"start_seconds":2689,"text":"python script that executes that the the challenge obviously is like how to make it secure how to do per missioning how","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2689s"},{"start_seconds":2697,"text":"to do all of this kind of n greedy details and so so that's where I think a lot of thinking and kind of Designing is","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2697s"},{"start_seconds":2704,"text":"to go in while also still iterating on you know some useful agents to build as kind of you guys are building um so the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2704s"},{"start_seconds":2714,"text":"I think the big important part is kind of how do we go to a situation where you can have you know a very simple","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2714s"},{"start_seconds":2724,"text":"prompt that then opens up an environment right in a kind of private in some private container uh you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2724s"},{"start_seconds":2731,"text":"potentially goes like into a browser opens up coin market cap you know fetches the required information like for example for this index of top five","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2731s"},{"start_seconds":2739,"text":"tokens fetches that you know brings it to kind the agent that constract the for example the buy transaction right and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2739s"},{"start_seconds":2747,"text":"then passes it to back to the user to valid it but this is exactly what you want and so kind of how do you construct","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2747s"},{"start_seconds":2755,"text":"do containers where actions can be executed for example they can navigate browsers they can write code they can","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2755s"},{"start_seconds":2762,"text":"execute some of the code as well within that environment and so then this is also usable for developer this is usable","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2762s"},{"start_seconds":2770,"text":"for process automation where you know near foundation for example somebody can just say like hey you know run this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2770s"},{"start_seconds":2777,"text":"process it goes like opens up browser and does that or somebody in the ecosystem is like hey you know I want to find all of the all of the things that's","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2777s"},{"start_seconds":2786,"text":"going on in because then you know and give it to me and in this format and then potentially get me to contact","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2786s"},{"start_seconds":2794,"text":"people of of the folks who are doing so that's kind of the you know high high level structure that we kind of thinking","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2794s"},{"start_seconds":2802,"text":"and obviously the the system itself the code is all going to be open source and so um as well as the agents are just","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2802s"},{"start_seconds":2812,"text":"stored somewhere on chain for example over time to uh to you know for anyone to be able","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2812s"},{"start_seconds":2821,"text":"to pick them up and use it um and then the questions are yeah like if there are additional use cases around like you want to have agents that are paying each","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2821s"},{"start_seconds":2830,"text":"other because they need special data uh you know how do you model data how do you model payments stuff like this right we have never mind which do doing","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2830s"},{"start_seconds":2839,"text":"payments for example there's few projects that doing kind of data coordination and like fetching and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2839s"},{"start_seconds":2846,"text":"you can add payments to all that so kind of there's like somewhat of a framework that uh we're trying to con conceptualize and obviously I mean for","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2846s"},{"start_seconds":2855,"text":"context we did talk with Nate about this on Friday so um so we I mean we're gonna have this kind of contining this conversation and","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2855s"},{"start_seconds":2863,"text":"also hopefully can open source what we've been working on pretty soon yeah our world is trying to tackle","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2863s"},{"start_seconds":2872,"text":"both ends it's not just the user typing in things to go do swaps and do things I mean blockchain has such had has had","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2872s"},{"start_seconds":2880,"text":"such a horrendous onboarding uh excruciating onboarding hurdles right not only you have to get users to figure out how to like buy nfts but you have to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2880s"},{"start_seconds":2888,"text":"get Brands to figure out how to create accounts and do all that sort of stuff so being able to also enable Brands to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2888s"},{"start_seconds":2895,"text":"create agents super quick like imagine lvmh being able to hook up their inventory uh sheets and being able to do","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2895s"},{"start_seconds":2903,"text":"a mass mint of nfts pegging every single asset to maybe some of their items with highly detailed explanation of what that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2903s"},{"start_seconds":2911,"text":"asset is and then people being able to search for those assets while they're watching a Netflix movie you know I'd like to buy that red handbag from that's","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2911s"},{"start_seconds":2919,"text":"that's really where I see us evolving over the number of years uh it's not only helping reduce the onboarding for","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2919s"},{"start_seconds":2927,"text":"people interacting but also for the other side the asset creation because if we look at the traditional Finance system it's just not going to keep up","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2927s"},{"start_seconds":2935,"text":"with AI there's just no way that AI is going to be able to evolve with traditional Finance uh AI is such a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2935s"},{"start_seconds":2944,"text":"perfect match being able to boot up accounts within seconds do trades create assets uh on the Fly ask its inventory","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2944s"},{"start_seconds":2951,"text":"sheets like so to me the the agent Marketplace is really kind of this space where we can open it up where we're not doing all the work we can focus a bit","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2951s"},{"start_seconds":2960,"text":"more on the security right being able to say this is a level one user who is just messing around don't let them do trades over 10 bucks right and then once they","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2960s"},{"start_seconds":2968,"text":"messed around with a little bit more maybe they can go higher these sort of like security measures we just want to focus in that that land um curious to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2968s"},{"start_seconds":2977,"text":"get your thoughts Alex it does sound a little bit like an episode of Black","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2977s"},{"start_seconds":2985,"text":"Mirror when you can like you know while watching a TV show you can like buy stuff from it right yeah yeah that's","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2985s"},{"start_seconds":2995,"text":"where we're heading yeah I I I think generally","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=2995s"},{"start_seconds":3002,"text":"uh it's the definition of an agent still uh like I think a lot depends on the definition of an agent because you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3002s"},{"start_seconds":3010,"text":"an agent could be a very generic thing right like arbitrary python script running in a Docker container is an agent um to some extent","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3010s"},{"start_seconds":3020,"text":"right um but I do think um so so so a couple things that definitely necessary uh so you mentioned","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3020s"},{"start_seconds":3029,"text":"security that's an interesting one right so not only like hey this is $10 but also ideal in the agent Marketplace it's two sides like there there are consumers","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3029s"},{"start_seconds":3036,"text":"there are providers right so ideally you want uh to limit quite a bit what providers can do and that's something","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3036s"},{"start_seconds":3043,"text":"that um like I think s similar to","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3043s"},{"start_seconds":3051,"text":"how what would be a good example like generally uh in many aspects of our life security evolved as the system was","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3051s"},{"start_seconds":3059,"text":"building right like Windows 95 you can run any process and sure sure thing very quickly all the computers were you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3059s"},{"start_seconds":3067,"text":"covered in viruses because right and then there was like Windows NT which like patched it right but the security Mo modes are not adequate uh for what we","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3067s"},{"start_seconds":3076,"text":"use systems today right like for example Linux security model is not quite adequate for like you know when you use","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3076s"},{"start_seconds":3083,"text":"Pap or uh npm install you arbitrary code from some random person and and like really your security comes from uh from","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3083s"},{"start_seconds":3091,"text":"you believing that uh whomever is moderating that package manager is not uh evil first of all malicious second of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3091s"},{"start_seconds":3099,"text":"all is actually paying attention right similar to hugging face like unknown to many a model in hugging face is an arbitrary script right so whenever you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3099s"},{"start_seconds":3107,"text":"install a model from hugging face you're actually running someone's code uh from from under your user which is extremely dangerous right and so I think uh like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3107s"},{"start_seconds":3116,"text":"the moral is that if you first build a cool system and then start doing your own Patchwork of uh of security you end up in a system where security is is non","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3116s"},{"start_seconds":3125,"text":"present right so I think like laying laying out a good security model ahead of time is is actually a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3125s"},{"start_seconds":3133,"text":"very good idea uh and like putting a lot of effort into it and maybe like involving Security Experts like that would be something that is extremely","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3133s"},{"start_seconds":3141,"text":"valuable for the community as a whole mik micro actually built a uh uh script that just starts running I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3141s"},{"start_seconds":3148,"text":"don't know how amazing it was but it was a good uh proof of concept that basically we we show before you run the transaction all the function calls that","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3148s"},{"start_seconds":3156,"text":"you're doing in the contracts and all that and below that we had a little thing that says in plain English this is what you're doing you're interacting","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3156s"},{"start_seconds":3163,"text":"with this instead of you know so yeah it's a it's going to be a crazy world we're going to right right because I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3163s"},{"start_seconds":3172,"text":"think uh there are many uh so so so whether or not we want it uh at some point agents will have","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3172s"},{"start_seconds":3179,"text":"access obviously uh you know to the whole internet uh to money Etc right so it's a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3179s"},{"start_seconds":3189,"text":"um yeah it's a I I think I have a kind of follow-up question here um so we're","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3189s"},{"start_seconds":3196,"text":"seeing for example like Google just um announced recently that they're enabling Gemini Nano to to run on on Chrome right","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3196s"},{"start_seconds":3205,"text":"so I think started in version 128 so sometime soon I think under some Flags you can you can enable","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3205s"},{"start_seconds":3213,"text":"like the model to run locally also we're seeing like um Apple partnering with","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3213s"},{"start_seconds":3219,"text":"open AI to like run natively sort of uh in in in the user's device uh or at least have access to a bunch more data","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3219s"},{"start_seconds":3227,"text":"and access to multitude of applications so I'm just really wondering what what's your guys's thoughts on this and um and I know I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3227s"},{"start_seconds":3236,"text":"know we running quite late here but uh yeah just quite curious to see you know like near is building on the open web","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3236s"},{"start_seconds":3243,"text":"like we're all like we really want the open web to succeed and we want these open models and decentralized execution everything to to succeed but these are","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3243s"},{"start_seconds":3251,"text":"giant players that will have access to a lot of users a lot of devices and um what are you guys' thoughts on on this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3251s"},{"start_seconds":3259,"text":"and uh is is this potentially harmful uh in the short term I I I think in case of open AI like","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3259s"},{"start_seconds":3267,"text":"you can think of open ai's decision in the in the From perspective of the fact that they have had an integration with Google for a while now right like if you","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3267s"},{"start_seconds":3275,"text":"go to Safari and type something in the address bar uh it goes to Google right so so uh it it sort of makes sense","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3275s"},{"start_seconds":3283,"text":"because Apple needs to meet expectations of the users and users are expecting to have access to uh to the cool stuff it is obviously","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3283s"},{"start_seconds":3290,"text":"harmful because uh open AI gets a massive FL uh data whe right so so they already they already","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3290s"},{"start_seconds":3299,"text":"quite ahead of everybody else in terms of how much data they get from the users once it becomes the standard uh on Apple devices they get","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3299s"},{"start_seconds":3309,"text":"significantly more data I I would expect uh I actually don't know the numbers but in my head what you get from integration with apple is order of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3309s"},{"start_seconds":3317,"text":"magnitude bigger than what you get from chat GPT interface right so it is definitely harmful for the uh for the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3317s"},{"start_seconds":3324,"text":"effort of building an acttion ual uh open um artificial intelligence right","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3324s"},{"start_seconds":3332,"text":"and it's sort of I think it's a","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3332s"},{"start_seconds":3338,"text":"um it sort of points at the fact that you know time is running out right so if we started as a community a year ago","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3338s"},{"start_seconds":3346,"text":"right and maybe if we had some good uh options to to offer today it would be different right I'm pretty sure apple is","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3346s"},{"start_seconds":3353,"text":"not happy with their you know privacy Center approach to integrate chat GPT just that open Solutions were","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3353s"},{"start_seconds":3361,"text":"not like I'm pretty sure they did some user studies right because they had an option to run llama but they went with chat GPT so like clearly they you know","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3361s"},{"start_seconds":3368,"text":"did some research and concluded that uh chat GPT turned out to be better it's it's unlikely that you know chat GPT is open a is bribing them just like from","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3368s"},{"start_seconds":3377,"text":"the uh you know relative sizes of the companies and the resources they have right but if there was a something competitive in the open I'm pretty sure","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3377s"},{"start_seconds":3385,"text":"Apple would have went with it uh and so I think uh that sort of but there will be uh like I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3385s"},{"start_seconds":3393,"text":"think as time passes more and more uh the you know the status like the current market leader gets uh their Roots into","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3393s"},{"start_seconds":3403,"text":"the ecosystem and it's harder and harder to Dethrone them so obviously integration with apple is one of them and the faster we get to the state where where","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3403s"},{"start_seconds":3411,"text":"open uh uh open models compete like agents and infrastructure in general with the proprietary uh the the faster","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3411s"},{"start_seconds":3420,"text":"we will start turning the tide uh backward yeah maybe just to add I think the the challenge is that we're we're","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3420s"},{"start_seconds":3429,"text":"living in the world where we have two kind of operating systems as well right which what one of them is developed by","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3429s"},{"start_seconds":3437,"text":"the company that is building AI models as well um and the other one is indeed I","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3437s"},{"start_seconds":3444,"text":"mean the the the actually interesting thing that they they build the small model themselves the ROM","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3444s"},{"start_seconds":3451,"text":"device uh and then they are asking explicitly every time if you w to go at","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3451s"},{"start_seconds":3458,"text":"least from their demos to go to open AI um which yeah I think it kind of shows that they're not particularly happy with","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3458s"},{"start_seconds":3466,"text":"that approach but kind of need to need to get the best results possible so indeed if there is a way to kind of","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3466s"},{"start_seconds":3475,"text":"provide a privacy first way right where for example you run it inside secure enclaves on H 100s","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3475s"},{"start_seconds":3482,"text":"right in a distributed Network that would be something that may more preferred now the there is more the","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3482s"},{"start_seconds":3489,"text":"question is the data wheel actually that important going forward right if if the if like at some point there's a situation of amount of data you can use","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3489s"},{"start_seconds":3497,"text":"right and it's maybe becomes more about synthetic data and like in general other methods but uh we over time","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3497s"},{"start_seconds":3506,"text":"I I want to thank everyone for joining uh we had indeed a very wide range conversation hopefully you guys enjoyed it um and yeah excited to work on this","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3506s"},{"start_seconds":3515,"text":"together and actually build something that is better product and more do their own going","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3515s"},{"start_seconds":3523,"text":"forward thank you awesome","url":"https://www.youtube.com/watch?v=0uvxudxSsuk&t=3523s"}],"youtube_url":"https://youtube.com/watch?v=0uvxudxSsuk","url":"https://near-pedia.xyz/videos/0uvxudxSsuk"}