{"id":"3hQAK2xt7RI","title":"NEAR AI Office Hours #11 w TICLE and SphereOne","description":"Illia & Alex discuss the latest AI research topics and unpack some NEAR.AI focus areas.\r\n\r\nNEAR AI: https://near.ai/\r\nIllia Polosukhin https://x.com/ilblackdragon\r\nAlex Skidanov https://x.com/AlexSkidanov\r\n\r\nGuests:\r\nTICLE\r\nhttps://x.com/TicleOfficial\r\nhttps://www.ticle.io\r\nTrustless low-code developer tool that autonomously plans, build, and deploys cloud-native applications by fetching & integrating crowdsourced APIs.\r\n\r\nSphereOne\r\nhttps://x.com/sphereone_\r\nhttps://sphereone.xyz/\r\nBuilding an AI orchestration and agent building layer to create agents focused on payments, pool re-balancing, AI solvers, automation and wrapper for different protocols, can onramp, bridge, swap, transfer.\n\nReactions to LLAMA 3, 400B - https://ai.meta.com/blog/meta-llama-3/\nOne of the north stars is teaching models how to think well, exposing them to competitive programming\nWorks well with Fireworks https://fireworks.ai/ and TICLE\nFunction calling code? not finetuned for, but no problem\n\nWhat is more important? the models? or what you build with it?\nAn interesting take: using larger, performant models to train smaller models\nThe secret to how GPT-2 continues to improve in quality\n\nDeepminds’ AlphaProof on International Mathematics Olympiad (IMO), analyzing their silver medal\nAI achieves silver-medal standard solving International Mathematical Olympiad problems - Google DeepMind  https://deepmind.google/discover/blog...\n\nHow they did it, Lean + AlphaZero reinforcement learning https://lean-lang.org/about/\nThey took 3 days, humans needed 4 hours\n\nA detour into Lean and model training\nAbout — Lean \nThe need for synthetic data in using Lean for model training\nIn this case, we live in one of two universes\nFrom a business perspective, how viable is this? Lean in real-world applications\nGeneralizing the achievement of DeepMind, what it means for AI development\n\nThe complexity of adding new code to existing codebases\n\nThe challenge of developing an actual AI use cases in/with Web3\nThe pressure is on for Web3 to prove it can play a major role in the AI game\nHow TICLE is taking on this challenge: working alongside, not against big AI models\nHow TICLE tackles the challenge of APIs breaking down when breaking changes: leveraging AI engineer that fetches kubernets clusters + human devs\n\n\nTICLE Deep Dive\nTicle \nLow-code developer tool that uses AI to automatically generate backends for cloud-native applications \nUse community to fetch and crowdsource APIs\nDesigning Backend-As-A-Service system within the past, transitioning into microservice architecture\nBuilding a collaborative as opposed to a penalty-driven developer community\nDetermining incentives for API developers: hit rate (recommendations) and usage\n\n\nSphereOne Deep Dive\nBuilding a one-stop shop orchestration layer for Web3: A SkyNet for Web3 https://sphereone.notion.site/Surface...\nMulti-agent system that creates agents\nHandling agent lifetimes and incentives/payments for agents\nCreating 800 agents in 7 days, done!\nDealing with nefarious actors\n\n\nWhat is the difference between an agent and an API\nAgents: increased tolerance and fault redundancy, flexible\nAPIs: highly specialized and optimized for specific tasks, interoperable\nThe difference is semantic (Agents) vs syntactic (APIs)\nMerging agents and APIs https://websim.ai/\n\n\nWinding down\nExciting possibilities: LLM OS (Andrej Karpathy) https://x.com/karpathy/status/1723140...\nAI in 2017\nHow APIs should be ideally, going forward\nHow do we make agents more like APIs\nMulti-agent systems https://ieeexplore.ieee.org/stamp/sta...\n\nJoin NEAR's community:\nWebsite: https://near.org/\nNEAR AI: https://near.ai/\nTwitter:   / nearprotocol  \nBlog: https://near.org/blog/\nMedium:   / nearprotocol  \nGitHub: https://github.com/near \nDev Docs: https://docs.near.org/\n#NEAR #nearprotocol #NEARAI","channel_name":"NEAR Protocol","published_at":"2024-07-30T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/3hQAK2xt7RI/hqdefault.jpg","transcript_text":"driven Wonderland okay and we are live uh today July 30th uh really excited to have Ian and Ryan uh with me and Alex here for another installment of AI office hours uh I'll pass it to Ian and then Ryan to introduce thems and then we can dive in I know you guys have questions uh we can also discuss some of the uh things has been happening on the news sounds good um hello super excited to be here my name is Ian from tickle co-founder in CSO we are currently building a lowco developer tool where we utilize AI engineer to automatically generate backends for cloud native applications and we do so by fetching and integrating domain specific apis that we crowdsource and manage through our community and a little bit more uh light on the background for the past year we actually have been designing a backend as a service system much in line with the more recent transition into microservice architecture and as in doing so we saw clear limitations in being a centralized backend vender especially looking into Firebase or super base um the apis that they provide has very limited domain coverage and at the same time very fragmented which requires a significant time for integrating and using each one of them so we decid to adopt the centralized ation to source and maintain high quality label code data that we then pass on to our AI engineer to rapidly consume so again super excited to be here and looking forward to the discussion right sounds good Brian hey guys yeah I'm really excited to be here as well I'm the uh co-founder CEO of sphere one essentially what we are doing is we're building a One-Stop shop for web 3 um which is kind of more or less like this orchestration layer which is powered by a self-improving AI um so effectively what we have been kind of calling ourselves acquainting it is more less like a Skynet for web 3 which that sounds kind of a little interesting but the kind of the cool thing and kind of some my background is that I've spent the last two years and our team has spent the last two years essentially integrating different protocols in the space and we kind of ran into this problem where the space is fragmenting fast when what we can actually integrate so instead of us like building out all the Integrations manually which we were doing at about a click RVE about of a a chain and a set of protocols every month we essentially built out a internal tool which now became the product which was a multi-agent system which can create an agent so we can create essentially or leverage about 30 different agents to create One Singular agent pulling in developer docs open API specs or contracts and then U kind of create and place that into to a ecosystem as a whole to be leveraged um kind of by the entire system so you can have access to everything in web 3 that way awesome I I the number of people who building Skynet is a lot and I always recommend to not be building Skynet definitely but yeah um I guess maybe just like to to warm us up um I mean there's been 400b Lama 3 dropped uh last week so I'm kind of curious to hear everybody's reactions you know if you played with it uh kind of how does it compare on what you you work on uh to open source models as well as closed Source models I know we can start with Alex let's yeah I I I app so so I run this uh um yeah so so so one of the one of the uh North Stars we have is to make models reasonable uh and an interesting uh way to approach it is to teach them uh to solve competitive programming which is a very self-contained task which requires a lot of reasoning and I think is nowhere close to being solved yet uh and uh yeah I have this little project where people uh solve them annotating every step and I just plugged in 400b it is not live yet but I played with it and it's uh uh it's it's very um effectively it either solves the problem fully like no human intervention required it just you know like step step step maybe a small mistake fixes it fully solves it submits it's accepted or it's clueless right so like it's just has no idea what to do no matter how much you nudge it to in the right direction it just cannot do anything I I'm yet to find something in between uh yeah so it's a but uh but generally uh compared to previous models it's much better like the number of problems that consol is much much higher and uh uh on the last of his hours we had the Dema from fireworks so so we use fireworks and it's uh uh it's extremely fast with fireworks right so like the experience of just you know pressing the button getting the the output is pretty fast so yeah like a quick way yeah quick way to test it to play with it uh with with pretty fast uh yeah I actually don't know I I I just played with tickle and it seems to be pretty fast but I don't know which model it uses uh and which back end so we're testing out a few different models and for now actually we have been pretty busy testing out Cloud 3.5 Sonic because even though AI code generation is not our main goal actually we are more focused in a low code format but we still look into uh the code generation part in an everyday basis and as we have been doing it we've seen it excel in many different benchmarks of course there has been a little bit of compensations in terms of performance but we've seen it as a very kind of a positive sign for us to actually try out and actually at the same time it really triggered a lot of the discussion internally because sometimes as we are developing our product uh we do consider the ongoing dilemma about is AI Tools in terms of uh the development ecosystem is going to replace developers and that has been an ongoing uh discussion within our company as well and with the new Cloud 3.5 Sonet um it actually triggered a little bit more hype uh internally but if it is to replace developers that will happen regardless right so yeah I mean I we'll take up that topic in a second yeah yeah I mean on on my side we we've tried out um numerous different models so we've done 405b 70b 8B we're predominantly focus in the function calling space um so we've seen really good results from it we've actually also been this isn't live yet but it'll be coming out soon but we've also done some really cool things around um trying to Cascade model uh refence so essentially leveraging a smarter larger model for more planning and reasoning tasks and then use a smaller model that is actually a lot cheaper to run uh like llama 8B for uh like an individual agent so we can have like essentially a series of 30 agents or 10 agents run in parallel and then have the parent for like the group chat manager essentially be like a 405b or 70b and we've seen really good results for that as well and you can run the uh smaller model locally so it's kind of like can we have this world have a local model and a cloud-based model or decentralized model so so with the Llama for function so llama llama is not fine tuned to be doing function cooling right do you do you fine tune it yourself or you like hack it with a PR yeah yeah so uh 405b we have good results for function calling 70b is is not as good grock uh if you guys have used Gro rock is great uh really great TPS um but uh they have a fine tune model for Tool use and we've done some things with that as well that was pretty good um but we've kind of used just a plethora of different models we've also tried centralized models we've had some customers have interesting latency requirements so we've done uh centralized models we've done uh decentralized models we're working with numerous Partners on that side too but yeah I mean all kinds of cool things for function calling there yeah I'm I'm curious because I think obviously seeing an open source model right that you can customize that you can you know fine tune is is definitely being something like that's comparable to kind of close Source models and at same time there's this continuous question if if the models are kind of the important part right or they're just a commodity that you kind of pick up and then you build stuff around right and this is kind of a continuous back and forth where it's like is the next model you know the Next Generation or whatever going to like replace all of the uh kind of wrappers and all of the infrastructure we building around or uh you kind of need to manage all that and indeed if you're optimizing for like cost and performance uh you do need to combine kind of smaller models and uh use big bigger models apparently but yeah just just been curious kind of how people reaction been have you seen um I forget who wrote it but it was a really interesting take I saw the other day which effectively was using 405b to kind of like train or fine-tune smaller models um so you can kind of use a larger model to actually uh kind of look at making more or less like a slm instead of llm so you can actually have these like much much smaller models which are fine-tuned for a particular task but you can use a more generalized model uh to actually create that have you guys kind of looked at any of that stuff I think that's incredibly exciting well I mean the 3.1 is that those models are fine-tuned from the output of the bigger model and so so but I think like karpathy was the one who who made the point the most clear to me at least it was that you know the gpt2 level model is continuously increasing in in quality because we keep training bigger models and then use them to pre-process data for gpt2 so you can imagine that and I think like uh I mean I I I always think you know the current pre-training of predicting next token is let's say rudimentary um or we use the suboptimal word right it's like it's the easiest thing to do it's easiest loss function to use because that's what we have but you can imagine actually if we had a data set that you can train you know pure reasoning question answering you know information retrieval all the tasks code generation you can train way smaller model to be able to do all that right you just need a data set like that and like that data set doesn't exist but if you have a bigger model that is like starting to be pretty good you can actually generate a data set and then train a gp22 load model on that so I mean the I think the the most unclear part is like this new like new set of capabilities that appear right with this different sizes which like as Alex described right it's like it's either it is able to do it even if it takes like multiple tries or it's just purely not able to do it at all right and the part that to me and and I don't think anybody has described very well kind of this boundary like where is that boundary comes from is it literally like it hasn't seen that example enough in data or it just not able to you know compute over it like what what is this boundaries that kind of we see um in in this ability to reason and kind of when it's not not able to grasp it even literally like if you prompt it with you know cor correct direction it's still not able to go from it it's worth mentioning my sample size was pretty small it was like six seven problems right I I actually suspect that there's a there's a frontier it's just relatively small right like if you imagine uh like plotting doesn't have to be competitive programming right like you you take any domain where models can solve like you know subset of it and completely clueless one another like naturally if the if the domain is large enough then you do expect that the FR here is going to be relatively small percentage uh of the current of the total set right because uh uh just because the number how do you explain it just because like you know the the difference between the easiest and the hardest is so is so is so large that uh you know like any you know like small window will only capture a small subset yeah and I think an interesting evidence something I discussed with the Bowen from near one couple days ago is that if you if you if you think about the Deep mindes paper on IMO I don't know if we plan to talk about it later today uh we can shift into that yeah describe it for for those who are not reading all the papers that being published by yeah so so so so I guess it would be pretty hard to live in the modern world and not hear the news that deep mind got the silver on IMO with a couple of asteres uh going both ways one asteris is that they were one point short of gold uh so it's almost gold and moreover they didn't solve at all the last two problems and interestingly on IMO uh when you like when you're in that state and you're a human what you would do is you would write any anything any thoughts you have because that could be worth one point right and deep M could have done that and and like there's a belief among some people in the community that they didn't do that just not to water down the result right like it would be worse if they said hey we have gold that people would be like yeah but like that one point is right so uh but the second as is that it took them three days and the actual participants only had four hours right but but I think that is completely irrelevant like if if there's one thing we are good at as a as as a species it's optimizing something that takes three days to Take 5 Seconds uh right and the way they did it is uh uh so so so so let's put aside Alpha geometry for now so Alpha geometry is older tool they they've had it for more than a year and it solved one problem but the remaining three they solved by uh converting the problem statement into lean that is still done by a human they have a model that kind of can do it but it's uh uh it can hallucinate and if you you know if you make a mistake in the uh translation then uh you know all better of so that is still done by a human there's a human who converts the written statement into lean uh and then they do effectively give or take Alpha zero on Le uh so so they train they train using something you know some sort of reinforcement learning uh with three search and then they do inference using three search until they find the proof and the beauty of Lan is that if you have the statement you know if you have proven it right so unlike human written proofs that are very hard to read and Lan proof uh uh is you know there's there's a system that checks it and it knows if it's correct and so for three problems they found proofs uh and uh those seem to be relatively hard problems so it's not uh it's not an a small achievement right and and something yeah coming back to my conversation with Bowen something he mentioned is that it's interesting that like the way the way they train right the number of lean programs in the world is very small right Lin is a relatively young language and it's very hard language right so the number of people who who who write in Lan is relatively small uh and so if you if you train on all the lean available you're not going to go far right you need to be synthesized ing it and the way they synthesize It Is by um they they get uh that step I don't know fully where they get a lot of problems but they get a lot of problems without proofs right and then you know the the system proves them and when it's proven it goes to the to the data set right and so what is interesting here is that initially of course the system is relatively weak so it can only prove the simplest problems right and so it proves a lot of simple problems it gets better at them uh and uh you would imagine right that there are two possib ities here right we live in one of the two universes in one Universe it will eventually become very very good at solving very simple problems but it will never transition to solving anything harder right so it will get stuck and in another Universe uh it keeps on getting better uh and you know like once it gets better it solves more problems it gets better solves more problems and eventually there's like this continuous transition until it gets to to hard problems right and it seems like we live in the latter Universe because they did eventually get to solving Amo so that is something that is uh that was not necessarily the you know like it wouldn't be obvious that that's the case but it seems to be that's pretty interesting yeah and it actually triggers a different question that in from a business perspective um you mentioned two different uh kind of future that we can expect from this and actually it getting better and better at solving more problems and later on maybe the other side would be doing different uh smaller solving different smaller uh problems and only staying at that level and you would have to find a different method to actually find solutions for solving more complex questions and from a business standpoint um from an economic standpoint what it be a little bit more um what should I say a little bit more closer to what the projects can actually adopt at the current moment and what would be more cost effective for us to actually um EXP and adopt yeah so so I think from this perspective one thing that I still mostly agree with uh is that uh like if you embark on a journey of building something uh that will be ready in a year you should effectively you know be building assuming that the models will be noticeably better right so it's uh like if you try to build something uh that can be built with existing models then you will be a little bit behind uh you know when it's ready um but I think another interesting thought which is uh um so so this like with everything Deep Mind does uh it's extremely mind mind-blowing uh and it's very hard to um it's very hard to generalize to the to the wider um set of applications right because okay we see that uh in this particular application of proving you know uh uh Olympia level mathematics it's very good uh but it's very hard to understand if that will generalize to wider applications of lim right because it wasn't trained on like um like it wasn't trained on practical application on on practical problems right it was synthesized problems and it happened that in this particular case synthesized problems helped to solve limped mathematics right and so we have no idea whether you know similar synthesized problems will help you apply L and other uh means right because lean uh is uh or like generally so I think first of all L is positioned today to be the def facto standard language of formal proofs of tomorrow right because all the research happens in lean the standard library is you know getting bigger and bigger so I think effectively Lin is becoming a synonym uh of formal proving right and if it gets extremely good to interject here the formal proving is needed for actually programming like the AI software engineering that's like foundational piece otherwise yeah yeah so necessary is a strong word right because we've been building code for decades now and I would say throughout those decades it was also necessary but we somehow managed to get by without it and the wor that has happened yet is a complete collapse of the whole world for one day but but but like you can imagine that on on steroid when you also have ai that is like of a level that can break any code and inspect it while inspecting it or yes exactly exactly side effects of it yeah so I think yeah so one thing is that we are probably getting to the state where AI will start breaking code very well so at that point having code which is formally proven would be a massive asset right so like you know it's not going to be broken right no matter how smart you youri is the code is proven to be doing exactly what you want so the at worst uh you can still mess up the actual guarantees right so you have proven that the code does that but that is messed up right it's not what you actually want uh but the second thing is that like you know generally the everything we use today all the website all the apps they all garbage right they it's it's just like uh like you know you open the United App uh and uh in the airport and it like crashes on you or like it doesn't load stuff or like and that's like throughout all the like banking apps notion they all uh right and we s like if if we were to generalize the achievement of Deep Mind we might be moving into the uh phase of humanity where all the code will be formally proven because uh you know like a is getting good at doing it for us right so so so that would be pretty cool and and like uh that would be actually an interesting thing like like you know for a startup to say hey let's say it is the case it might not be the case and then my startup will die but there are you know 99 other reasons it could die right but let's say it is the case let me bet fully uh on you know Lin being fully synthesizable in a year right like can can we do something cool uh you know to be ahead of the curve where all the code is formally verified and you know it doesn't it doesn't get hacked it doesn't fail yeah so so maybe like Ian because you brought it up right like you know with Claude and with others hey the code generation is getting you know it's getting to a level where like it give you a dirty prototype pretty quickly right it may still have some issues or it's not handling some edge cases and uh it it's harder to integrate in a bigger code basis usually but you can see like hey you know if it's a single file thing like it can you know start getting there uh at least like um to to to start but indeed the question is the complexity of integrating kind of a single module into a bigger system is where lots of complexity is like for a person to write a new program it's pretty easy right like you know but actually adding code to existing code Bas is where complexity is because there's so many different conditions and variant and you know usually some some other things that like only in few people's brains potentially that you need to recover from reading this code base and so I think the kind of looking forward we will need some you know combination of the tools to come together to be really both advancing ability to like expand code bases right with was guarantees that you know it's not going to fail again when somebody pushes like one line change and at the same time this also would allow to have kind of more complex transformations to the code base in itself right because semantically you can imagine like I'm looking at some application I'm okay I want to switch database from SQL to uh you know no SQL for example like that's a very simple semantic change which requires a lot of syntactic like reimplementation of in the code base um but like for that you need a lot of like guarantees and understanding of you know what is the differences between SQL and no SQL and how those are being then kind of surfaced in the application in the back end for example because you know of the complexity of the retrieval etc etc right so so all of that is like like models have some of that information just because they statistically saw a bunch of those cases but like making it a lot more uh apparent right by maintaining some of the let's say lean proof around it would allow to to actually have like a pretty good investment I mean that's kind of what we've been you know with Alex uh discussing and and working on some of the pieces of that uh but I mean we do think it's it's coming right although we've been thinking about it that it's coming from 2017 that it's near and before yeah and and before that it was less near before that but yeah I mean I well know that you guys uh started off trying to develop ai ai developer so you guys have an extensive background in this and for us actually uh what we're more focused in here is that we wanted to create an actual use case because it feels like we are at a critical stage where what three has to prove itself that it can prove it can have an important role to play in the AI game and for us how we saw it differently is not from the AI code generation perspective but actually putting it in a low code format so it is a lot easier for us to compose it and for the people as you mentioned like when they have to go back and switch out something it's a lot easier because it is simplified into smaller blocks that they have um pulled through our AI engineer and for us um of course we can't compete against these bigger models or big tech companies that are working on these Technologies especially like IBM just L open sourcing their grandite family and they're trying to receive a lot of feedback and community support from developers as well and for us we are trying to do the same thing but we want it to be in a predefined and a pre-built uh package as an API and which actually gives us a little bit more flexibility of course we won't be able to build super complex uh apps at the moment but the more we actually build out this codebase and build out this API registry we will be able to tailor and cater to more of the custom and unique business needs that the web 2 application Builders actually have and from here on I believe we can actually generate those use cases and based off of that we can actually try to mimic what these bigger models are doing and I don't necessarily believe that we have to work against these bigger um technological innovations but somehow if we can work alongside of it then it would be a better um option for us of course the centralization is um has a lot of value propositions and we do believe in it but for us we're just starting off from the data layer which is the uh bottom layer of theist stack and everything else of course we have it in a traditional setting uh the only difference uh is that we have it all open and publicly available so from there on we're trying to see where this is this can head in building more complex softwares so for the I think one interesting question here is uh um so so the if I understand correctly the the sort of mental model of how a person would build a uh I don't know if you call them tickles if a person builds a tickle uh that uh there's like like I imagine you know like the the end goal is that there's this massive um in in in inventory of of apis right and so I go there and and it's not just like it's not just that I'm relying on an AI to build a whole system from scratch correctly I'm relying that it's writing a small uh effectively gluc code correctly which is a much smaller scope plugging in into all those apis right and like when I was playing I said you know like apis have reputation you know like reputation 10 means that don't touch it reputation 90 means that it's probably legit right so in this model um the the problem I see and I'm curious how you guys think about addressing it is that apis go down apis go down right and so if if we build a whole ecosystem of things that are like interconnected right and like like imagine this completely madeup example that many people need an API to pad a string with spaces to the left let's call it left pad right and then the left pad API goes down right what happens like how how do you how do you think about the problem of API disappearing or like introducing breaking changes and the cascading effect of so as you mentioned API of course one of the apis that can actually go down but what our AI engineer is good at is that it fetches and integrates it as um IIA mentioned of course it's not just the code writing site but it's also the knowledge and expertise that you need from the backend site and from there on we actually have um optimize the kubernets cluster and um allocating resources in an optimal way so whenever one goes down of course at the moment there could be a little bit of delay uh in terms of how the app will run but at the same time we'll have a group of developers that are working on and monitoring the specific API so we'll try to uh push different version updates through it of course um even proprietary uh apis that are provided by local solutions that are that already exist in the web to ecosystem they are not perfect and plus on top of that they are not verifiable so when the users actually receive the apis they don't know if that's the best quality that that they can receive from the market so at least what we can offer is that we actually create this Merit based ecosystem because it needs to be selected search and selected by the AI engineer so at least what we know is that the code itself is as secure and as high code quality as possible if it goes down of course there on it is based on us and it's based on us U being able to quickly ship those problems and Outsource those problems to our developer Community I see but but I think you implicitly answered the the the other question I which is that like for as long as API is deployed by a tickle uh effectively there's no way like like my my bigger question was not just the server going down but the developer like the original developer just completely shutting it down but I guess as long as you can you you own the apis you can just like I guess you can forbid developers shutting them down right it's a right uh so we call user own apis so basically if the user decides to just completely dish the API then of course we will be able to find a followup uh API that we can quickly replace with it so I think it's same with any kind of user owned assets um of course if there are slashing um rules and so on there will be less people who actually dish or whatever asset they're putting out in the marketplace and actually providing people to pay for it but at the same time um for us we we are currently not trying to impose any kind of slashing at this point we just want to make it a little bit more collaborative rather than being penalty driven so for us what when that happens we'll be able to replace it with some other apis within the same domain and within the same domain specific functionalities and there's one more thing which is I'm pretty sure it's on your road map but I I didn't see it there but of course when I deploy my API I want a little bit of money streaming to my account every time it is called so I guess uh if you're a supplier of the API so we basically try to incentivize in a very different way so the first one will be the hit rate which basically means how much recommendation it actually receives and how many times it actually been uh recommended to the end users and the second is actually based on the usage so a lot as we started off as a backend as a service um we still try to um position Us in that way so people will pay monthly based usage fees based on the time that they have instantiated the API stack and based on that we will minus all the infrastructure costs and so on and from whatever is left based on the usage percentage we'll be able to distribute it to the API providers so of course we could ship out um tokens as incentives but we don't see it as a sustainable way to do it of course in the beginning we have different code of recruiting events for different domains and there will be different incentives that people can take off from but the main sustainable uh income for the API providers will be based on the usage actual usage yeah maybe similar question to Ryan around agents right pretty much how do you manage agents Lifetime and payment and uh if like people are able to spun up uh own agents if the agents kind of may disappear yeah yeah absolutely so on our side we actually we explored that space uh quite a bit originally for allowing people to create their own agents to then like support a given protocol if they would like to so maybe like someone wants to create a Unis swap agent or a wormhole agent whatever it may be um what we kind of ran into and what we kind of done is we shifted some things around in our road map for that because we ultimately because we're dealing with money the a concern kind of popped up around uh like how do you keep an nefarious person out so how do you essentially say and and enable a way in which like a nefarious person says okay if I'm doing a swap I'll do the swap normally uh if it's less than a million dollars if it's more than a million dollars then I'm just going to take the money and run so you know you can have reputation systems and what we were looking is like there's reputation systems you can have you can then effectively game those uh for what we're doing uh if you had slashing stuff that was something that we kind of looked at uh there's just some kind of uh concerns that we had around that so what we have actually done is like we built out and we turned our tool loose on our community and we wrapped over three quarters of defi in three days so we ended up doing like 800 different agents uh quite quickly um that was really cool to see but the problem arises where is that if that code touches someone individually then how are you ensuring that they don't have uh a particular like nefarious piece in there and we may do slashing or whatever but if it effectively has a problem before that then I think there's a a larger issue so what we've been focusing on more now is that effectively we can build out a series of Agents ourselves using this tool that's like more self-replicating and then we can actually do more community-owned AI so effectively Implement something that there's an experimental paper that came out called co-learning but effectively it's where uh each individual use can be ran and we can identify we have a knowledge graph that Associates between all different agents so you know how to Traverse and which agent to populate in an event where you're trying to do a given transaction so we can incorporate those directly and then if there's a problem in the overall like flow of that individual agent uh so maybe it gets transaction first before getting a quote we can then uh refine that and through that refinement that person can be rewarded and then the next person has a better experience the next time and then what we want to get to is where users own the agents but we want to have more or less more functionality of using the existing library that we effectively can build out ourselves and then enable people to extend capabilities on top so maybe they want to do uh great payments on uh like an a cost or something we could then say okay you can use these series of Agents up front and then you're effectively composing this into a button so you don't have to actually type out or do anything directly it's all like kind of like through a button interface directly and we can have more and more of these uh buttons or different capabilities be used kind of what we did recently tip link go ahead no go yeah you mentioned you created 800 agents in several days so this is because agents themselves are written by an agent right you actually have humans WR 800 agents so what they did was we had a tool that was deployed in which you could say I would like to create an agent for UNIS Swap and then what we can do is we could pull in the contracts of Unis Swap and the dev veler docs we create a rag agent for like a micro viw over what that agent actually does and what Unis swap is and then we are really what we're doing is we're not creating like a whole generation piece we're more or less like stubbing out the functions associated with that protocol and we're just saying okay the inter agents essentially will interface with those stubbed out functions so like getting a quote or getting a transaction so we're not like replacing their contracts we're just kind of creating a interface layer on on top of that yeah and so what is the um what what is the end goal of an agent right so like let's say I have an agent for a for compound uh well let's not do for compound let's do for Abby for compound is going to be attacking the Govern yeah so go for so so the question is what is the what is the utility of the agent so so first of all is the agent designed to used by another agent or by human if and if by a human what is the like I'm a human right what is the use case for me to use to use the agent um the agent itself can be built by other agents and then the user of it can be other agent or it can be a user directly so we've deployed on we're on a kinetics the number one Dex on Cava chain we're deployed on there we deployed five agents there directly and essentially what you can do there is you can ask it and say like you know what is kinetic who built it you know stuff like that hit the rag agent and then if you wanted to like add in something where it's like I would like to do a swap and then tie it into 100x long position through their perose product you can essentially combine the two together and create one seamless stream um so then it's more about like partner with like you guys and other people to essentially provide a chain abstracted experience so then you populate all the transactions forward and then how do they go about signing that and that's really up to the user they want to go one by one do they want to sign like via a policy engine uh do want something that's chain attracted that kind of thing so so is is it more about like making more more user friendly experiences or is it more about uh like abstracting away the boundaries of uh all applications yeah so uh what we found when we were doing payments was that it just was incredibly difficult to like try and support everything and everyone someone have money on like Cava chain and we don't support Cava chain so then all of a sudden how do add in a bridge from Cava chain over into salana or near or whatever and it's like you what we ended up doing is we incorporate over 130 different pieces of infrastructure so we built out a lot of stuff to support 18 chains and it was just like this is like a NeverEnding problem you know it's like the protocols just kept expanding faster and faster there's more chains now you have l2s and l3s and all that stuff so part of it is to create a userfriendly experience but it's also to extend the cap abilities of that actual like protocol directly so instead of that protocol having to figure out how do I bring in an on-ramp how do I bring in a a bridge to get people's liquidity into my like actual project we can say hey use our like effec our library of agents and then enable your users to pull in funds easily through like kind of like One-Stop shop um kind of system and they don't have to spend the time to do that so I actually want to I want I'm I'm interested to like challenge everybody's perspective on what is the difference between an agent and API um for for me what we've seen is that we when we did payments uh we had a lot of issues with redundancies and rpcs going down and so it's kind of more or less what we continue to see is like there's increased redundancy and fault tolerance so like at least how I'm looking at and I'm biased I think it's like the next step almost because I feel like the unlocks we have had enable a lot more of a uh uh like say I want to hit like a particular RPC I just want to get my balance or I just want to understand like what this wallet is or whatever I I just want to do a transaction RPC goes down we can have an agent then populate a new RPC try it again but instead of it being like written out manually in code um and that would have taken you know a developer maybe a week to implement or go through the whole process we can have effec some with some reasoning capability to provide this fault tolerance in association with what the user is trying to do directly which I think is kind of much more extensible and uh flexible so that's what I get excited about between the two I mean from our perspective it's quite similar as well but what apis actually do is it actually can only serve a specific function and and it can actually only take in um do certain actions based on certain signals and for the agents um where we see it in the future is that they can actually be interoperable and they can actually be working together autonomously and from there on of course uh we can make it to be a bit more fallproof at the same time just providing it as a single more than just providing it as a single API it will have a little B more flexibility in terms of actual uh business applications yeah I think for me the difference is in like semantic versus syntactic right API is a syntactic act like syntactically defined action and kind of fa tolerance as well as you know some intelligent understanding of arguments all of this is more of a semantic like I want the result like I don't care which RPC you use just figure it out um but at the same time you know you can imagine apis becoming like you know open ey API is literally a semantic thing or you're calling an agent API and so I am I am kind of curious where that line lies and and how much more semantic can apis become uh broadly as well I I I want to bring up another thing we we we mentioned it couple office hours ago there's this website called web. which uh it's if if someone who watches it didn't see it yet it's completely mind-blowing even in the Modern Age of AI uh web like web simulation. web. and couple days ago maybe earlier than that someone reverse engineered it right so they found how it works it uses clo which I guess is not very surprising uh and a very interesting aspect of it which is very relevant to elia's question is that effectively they ask ask clae to write the website based on your like small prompt uh you know on the fly but that website might have uh ax queries right to the back end but there's no backend right so so the way they do back end is that whenever there's any ax query they go to CL and say hey clo uh we have this aex query for this website what should we respond with and Claud is you know on the back end is coming up with the response uh and like like if you play with web. you will see that uh like it works you know 99 out of 100 right so so here you know what like that's an agent for sure that's also an API yeah I think somebody did as well an example where the the back end is just llm responding to like exactly right but but but it but but here it is plugged into the to an actual website which is you know built on the Fly and and it all works seemlessly right so like especially uh you can intentionally come up with examples that will have an right like you can have it as part of your prom saying hey it it like calls to an API to do something uh and and it all works uh very well it's also a cool thing you know when you prototype a website before you have the back end you can just St stab it with yeah yeah later you can probably you probably want to replace it with an actual back end but you know for the for the early testing it's it's a nice yeah you know if it does something weird you know you just reboot it but yeah yeah that that is actually not like a new tool that exists in the tool chain of of the developers you imagine the llm that you know has memory that has you know ability to like use a bunch of tools then potentially with like optimizing it you know you have a router in the beginning that you know oh if I if I have if I have already a function like this called a function instead of uh you know unrolling Chain of Thought and figuring out and if not unroll Chain of Thought and write a function so that next time you can call it right and so you kind of cache like while you're doing it you also cache everything and then you have you know some database access let's say so like it will just build the back end under the hood while you're actually doing this I'll go for you I mean it truly is um kind of a crazy world we are jumping into but at the same time uh of course these are newer Technologies and newer kind of like Tri outs with the existing solutions to kind of find ways to make it a little bit more fun and actually just find loopholes into it but at the same time um for extra use cases and Enterprise adoptions when it comes to agents and apis it really depends on how much they can actually rely on it and how much they can actually securely adopted from an Enterprise standpoint and from there on I think um of course where we draw the line between the two is actually important but at the same time just working within the traditional setting and actually trying to provide what they actually need at the moment is also a critical issue that we have right now so of course Drawing the Line we would at some time we will have to draw the line between the two but for now I think we can just work with what we have and actually try to innovate on the usability which is quite similar what to what Ryan is trying to achieve with the project as well the the thing I get excited about is I think Andre carpati uh wrote like a tweet or something like that about like llm OS so he kind of talked about this kind of concept or idea as well and I think that's like uh I mean he's had some very interesting thoughts on like software 1.0 2.0 and this llos I've been kind of following some of that I think that is really an interesting thought or perspective on maybe where software can go um and maybe where the future what the future might hold um but yeah I mean I think it's just exciting times ahead with the existing tools that are available today and even inside of that it's like uh there's lots of really great research out there I mean uh really good products that are like multi-agents type systems like crew Ai and autogen they've done some things around like a I think one was called like agent optimized so effectively if you build out a given system you could then find out a predefined path to do like something in the future so if you do like spin up the backend on the Fly uh in this case you can actually like uh Replay that later so it's almost like you store that in like memory and then you like can cach that and then replay back that same system for the uh a similar like intent or like whatever you're trying to do with that overall like website which I think is awesome and and the post about software 2.0 is from November 11th 2017 yeah which is just wild I mean I I think uh I got into AI in 2017 definitely not nearly as much as you guys uh um that's how I actually got into Bitcoin or crypto was I built out an AI trading bot and then I started a a little uh NLP company that I did for automatically applying people for jobs using like an algorithm I built out but uh I mean even back then I thought okay AI is going to take over the world like today like it's happening today like it's it's you know it's over and then um the llm space though like when it first came out honestly I wasn't as excited like I was like okay this is very cool very interesting um but I wasn't as excited as I probably should have been and it really wasn't until like some of these more recent adaptations um that I found to be quite interesting so more or less like how you add and incorporate things into it maybe Monte Carlo like tree searches for reasoning on top of it it's like the Deep Mind is doing great work on that uh I mean that that's I think is what gets me really excited about what like what is Yet to Come kind of thing yeah yeah I think that's uh generally uh uh when was it I think it was like half a year ago IIA IIA Drew this diagram which I think is the core of how the systems ideally should operate where the user communicates to the agent and the only goal of the agent is to create a symbolic uh representation of what the user wants not the solution yet just you know rephrasing what the user wants in a symbolic space and then it communicates it back using not an llm but a uh deterministic you know like symbolic to natural language conversion right so it tells to the user hey this is uh what I think you want uh translated from my language to your uh language in a deterministic fashion does it sound right and then when the user says yes then we do synthesis in a symbolic space Guided by the model right what Deep Mind does what presumably strawberry is uh yeah and then uh and then when it is synthesized we know that it is correct for as long as the user did properly confirm our understanding right so that's uh uh then you know the API uh that is synthesized is correct and then I actually think that this is in the in the world with lot of apis which I believe myself in uh or or like you know big Global name spaces uh that's that's very important right because uh if you have 10,000 apis built by different people you definitely want to be sure that they actually doing what you want uh and not doing it you know nine times out of 10 yeah so that's that's the exciting uh I I think that's how it should be going forward I mean verifying that these each of these apis do exactly what the user intends to do is very important and completely agree with it without a doubt also spinning off of that is that as you guys open the Asian Marketplace um most of those of course as you talked about there are your Technologies and triy outs but most of those Asian would be connected to certain different apis and for that Marketplace to work seamlessly as a platform um wouldn't you need a protocol to actually integrate those apis seamlessly within your Marketplace as well when the users actually picks it up from there and curious to see if you guys have any thoughts on how the users will pick up those agents and try to adopt them within our ecosystem yeah I mean I was asking this question because I actually think we'll we'll just be living in this API to agent Spectrum right API to to like full llm spectrum Right Where You agents in a way are just constrained llms which you know expect some type of input or can work with some kind of input themselves can use a bunch of apis or other agents right and then you expect the output in certain format right which again API are kind of that just like you don't expect them to be as non-deterministic as agents right now but then agents are more intelligent in result right but it is a spectrum and so you kind of want actually agents to become more like apis where you know they predictably going to deliver results they're not going to fail they're not going to go in like on a side quest of you know building something that you didn't ask them to do right and so I think like how do we make agents more like apis is that's kind of the mindset in a way uh I wanted to kind of challenge as well because I think it will be important to understand yeah like you know livess of this agent's ability to perform the tasks you know how do we Benchmark them um and so similarly to apis you know you want to know all of this information how to use them examples Etc and so yeah to me to me that's kind of a spectrum that uh is probably going to end up in one place like eventually right where all of the species kind of just come together Ryan you wanted to say something oh yeah yeah I was just goingon to say I totally agree I think that uh yeah I mean apis in general um definitely extremely useful into the future and I think like more or less like the agents and how I think about it as well is like it's almost like the glue that kind of sticks it together but you want it to be deterministic somewhat in its output and reliable that it does work and um I think for us like what we have seen really good results from is just adding like a Knowledge Graph associated with it so then that's like more or less like helping the traversal of like different agents um and then incorporating those into a singular conversation which then talk to and communicate to multiple different apis or contracts I think that like those types of like supporting pieces of infrastructure around that system is like very um helpful as far as like what will probably come next and then probably being able to like Traverse like an ideate between different possibilities is also helpful like for us like kind of the end state where we're going to is something like um every Wednesday rebalance my portfolio for the best yield then we can Traverse across the knowledge graph say okay super form ichy and something else come in mind for maybe Moro for good yield protocols we can say hey here's the three results which ones do you like um and then they say Okay I want to go with super form on you know arbitrum or something like that it's like okay you kind of tap that and then we're able to like more or less Traverse and incorporate and and bring that in um but we're really the agent is like the the spectrum of what they're actually looking at and doing and reasoning becomes extremely extremely narrow and we're more or less like I always think about it is like we're almost trying to supply employees with a lot of different bits of information in which they can be successful uh so I think that's a lot better of a end result versus like a just saying hey one two three go then they're just trying to like identify what do I do I have 50 options which one do I grab and then trying to just pick and choose kind of thing so yeah no no I was I was saying we like one and a half minutes away from from the end uh I don't know if we have any final interesting thoughts I don't well I I think the yeah to to wrap it up maybe um I'm curious kind of just in in in a couple sentences uh from Yuan and Ryan um how do you see let's say like your product is fully rolled out you know people are using it um what would be the most like impactful change right uh for for the user like what is the 10x better experience that they cannot get right now uh I I'll go first um actually as you may have noticed from our project is that we actually create a lot of demand from the web two side and actually try to solve what web two has to um has been really thinking about for the past few years which is to really democratize the development so from our end um providing a conver conversational panel and actually being able to just plan build and ship out your backend just within a few seconds and actually be able to just deploy it and instantiate it with just a click of a button I believe it's like an end to-end automation that they haven't really seen uh in the past and at the same time it really brings a lot of freedom and democratization to not just less season developers but at the same time uh non-technical users as well so from there on we're actually excited to see how users actually use our product and how it kind of Fosters a little bit more collaboration within that company yeah and then on my side uh really what we're trying to do from a very high level is just provide like a Visa likee experience to the masses for web 3 so like I I don't think that you should have to know about the given blockchain or even the protocol that you're necessarily implementing or working with in order for a user to actually use the space I think uh like our our old saying when we first started the company was uh we're building crypto for grandmas and I think we've always been kind of true to that ideal which is we're just trying to make this really usable for the masses where people can actually go in and if they want a you know 50% yield on some kind of like LP position um on a deck they can do that but they don't have to understand you know how do I onramp bring this money in how do what's arbitrum versus optimism versus you know whatever like that that just seems kind of ridiculous to me a little bit sometimes it's very cool from a technical side but I think from a user side uh that has always been a huge uh problem I think for the industry so from a high level what I would want to do is is just provide the visel like experience to the user where they don't even have to really think about it it's like a oneclick or a swipe type system yeah we call it chain abstraction for for web three people for for web two people we just don't call it they just should use it and not know about it so yeah great to have you both here hope it was engaging entertaining episode for everyone we'll see you next week and uh have a good week thanks for f","transcript_synced_at":"2026-11-03T05:40:42.167Z","segments":[{"start_seconds":0,"text":"driven Wonderland okay and we are live uh today","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=0s"},{"start_seconds":10,"text":"July 30th uh really excited to have Ian and Ryan uh with me and Alex here for another installment of AI office hours","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=10s"},{"start_seconds":19,"text":"uh I'll pass it to Ian and then Ryan to introduce thems and then we can dive in I know you guys have questions uh we can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=19s"},{"start_seconds":26,"text":"also discuss some of the uh things has been happening on the news sounds good um hello super excited","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=26s"},{"start_seconds":35,"text":"to be here my name is Ian from tickle co-founder in CSO we are currently building a lowco developer tool where we","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=35s"},{"start_seconds":43,"text":"utilize AI engineer to automatically generate backends for cloud native applications and we do so by fetching","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=43s"},{"start_seconds":50,"text":"and integrating domain specific apis that we crowdsource and manage through our community and a little bit more uh light","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=50s"},{"start_seconds":58,"text":"on the background for the past year we actually have been designing a backend as a service system much in line with the more recent transition into","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=58s"},{"start_seconds":66,"text":"microservice architecture and as in doing so we saw clear limitations in being a centralized backend vender","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=66s"},{"start_seconds":74,"text":"especially looking into Firebase or super base um the apis that they provide has very limited domain coverage and at","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=74s"},{"start_seconds":81,"text":"the same time very fragmented which requires a significant time for integrating and using each one of them so we decid to adopt the centralized","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=81s"},{"start_seconds":90,"text":"ation to source and maintain high quality label code data that we then pass on to our AI engineer to rapidly","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=90s"},{"start_seconds":97,"text":"consume so again super excited to be here and looking forward to the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=97s"},{"start_seconds":104,"text":"discussion right sounds good Brian hey guys yeah I'm really excited to be here as well I'm the uh co-founder","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=104s"},{"start_seconds":113,"text":"CEO of sphere one essentially what we are doing is we're building a One-Stop shop for web 3 um which is kind of more","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=113s"},{"start_seconds":120,"text":"or less like this orchestration layer which is powered by a self-improving AI um so effectively what we have been kind","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=120s"},{"start_seconds":128,"text":"of calling ourselves acquainting it is more less like a Skynet for web 3 which that sounds kind of a little interesting","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=128s"},{"start_seconds":135,"text":"but the kind of the cool thing and kind of some my background is that I've spent the last two years and our team has spent the last two years essentially","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=135s"},{"start_seconds":143,"text":"integrating different protocols in the space and we kind of ran into this problem where the space is fragmenting fast when what we can actually integrate","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=143s"},{"start_seconds":151,"text":"so instead of us like building out all the Integrations manually which we were doing at about a click RVE about of a a chain and a set of protocols every month","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=151s"},{"start_seconds":161,"text":"we essentially built out a internal tool which now became the product which was a multi-agent system which can create an agent so we can create essentially or","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=161s"},{"start_seconds":170,"text":"leverage about 30 different agents to create One Singular agent pulling in developer docs open API specs or contracts and then U kind of create and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=170s"},{"start_seconds":179,"text":"place that into to a ecosystem as a whole to be leveraged um kind of by the entire system so you can have access to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=179s"},{"start_seconds":186,"text":"everything in web 3 that way awesome I I the number of people who","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=186s"},{"start_seconds":193,"text":"building Skynet is a lot and I always recommend to not be building","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=193s"},{"start_seconds":204,"text":"Skynet definitely but yeah um I guess maybe just like to to warm us up um I mean","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=204s"},{"start_seconds":212,"text":"there's been 400b Lama 3 dropped uh last week so I'm kind of curious to hear","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=212s"},{"start_seconds":220,"text":"everybody's reactions you know if you played with it uh kind of how does it compare on what you you work on uh to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=220s"},{"start_seconds":229,"text":"open source models as well as closed Source models I know we can start with Alex","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=229s"},{"start_seconds":238,"text":"let's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=238s"},{"start_seconds":242,"text":"yeah I I I app so so I run this uh um yeah so so so one of the one of the uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=242s"},{"start_seconds":251,"text":"North Stars we have is to make models reasonable uh and an interesting uh way to approach it is to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=251s"},{"start_seconds":258,"text":"teach them uh to solve competitive programming which is a very self-contained task which requires a lot of reasoning and I think is nowhere","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=258s"},{"start_seconds":265,"text":"close to being solved yet uh and uh yeah I have this little project where","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=265s"},{"start_seconds":272,"text":"people uh solve them annotating every step and I just plugged in 400b it is not live yet but I played with it and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=272s"},{"start_seconds":280,"text":"it's uh uh it's it's very um effectively it either solves the problem fully like no human intervention","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=280s"},{"start_seconds":289,"text":"required it just you know like step step step maybe a small mistake fixes it fully solves it submits it's accepted or","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=289s"},{"start_seconds":296,"text":"it's clueless right so like it's just has no idea what to do no matter how much you nudge it to in the right direction it just cannot do anything I","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=296s"},{"start_seconds":304,"text":"I'm yet to find something in between uh yeah so it's a but uh but generally uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=304s"},{"start_seconds":312,"text":"compared to previous models it's much better like the number of problems that consol is much much higher and uh uh on","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=312s"},{"start_seconds":320,"text":"the last of his hours we had the Dema from fireworks so so we use fireworks and it's uh uh it's extremely fast with","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=320s"},{"start_seconds":327,"text":"fireworks right so like the experience of just you know pressing the button getting the the output is pretty fast so","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=327s"},{"start_seconds":333,"text":"yeah like a quick way yeah quick way to test it to play with it uh with with pretty","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=333s"},{"start_seconds":341,"text":"fast uh yeah I actually don't know I I I just played with tickle and it seems to be pretty fast but I don't know which","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=341s"},{"start_seconds":348,"text":"model it uses uh and which back end so we're testing out a few different models and for now actually we have been","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=348s"},{"start_seconds":356,"text":"pretty busy testing out Cloud 3.5 Sonic because even though AI code generation is not our main goal actually we are","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=356s"},{"start_seconds":365,"text":"more focused in a low code format but we still look into uh the code generation part in an everyday basis and as we have","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=365s"},{"start_seconds":373,"text":"been doing it we've seen it excel in many different benchmarks of course there has been a little bit of compensations in terms of performance","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=373s"},{"start_seconds":380,"text":"but we've seen it as a very kind of a positive sign for us to actually try out and actually at the same time it really","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=380s"},{"start_seconds":388,"text":"triggered a lot of the discussion internally because sometimes as we are developing our product uh we do consider","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=388s"},{"start_seconds":395,"text":"the ongoing dilemma about is AI Tools in terms of uh the development ecosystem is going to replace developers and that has","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=395s"},{"start_seconds":403,"text":"been an ongoing uh discussion within our company as well and with the new Cloud 3.5 Sonet um it actually triggered a little bit more hype uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=403s"},{"start_seconds":412,"text":"internally but if it is to replace developers that will happen regardless right so yeah I mean I we'll take up that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=412s"},{"start_seconds":421,"text":"topic in a second yeah yeah I mean on on my side we we've","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=421s"},{"start_seconds":428,"text":"tried out um numerous different models so we've done 405b 70b 8B we're predominantly focus in the function","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=428s"},{"start_seconds":436,"text":"calling space um so we've seen really good results from it we've actually also been this isn't live yet but it'll be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=436s"},{"start_seconds":443,"text":"coming out soon but we've also done some really cool things around um trying to Cascade model uh refence so essentially","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=443s"},{"start_seconds":450,"text":"leveraging a smarter larger model for more planning and reasoning tasks and then use a smaller model that is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=450s"},{"start_seconds":458,"text":"actually a lot cheaper to run uh like llama 8B for uh like an individual agent so we can have like essentially a series","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=458s"},{"start_seconds":465,"text":"of 30 agents or 10 agents run in parallel and then have the parent for like the group chat manager essentially be like a 405b or 70b and we've seen","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=465s"},{"start_seconds":473,"text":"really good results for that as well and you can run the uh smaller model locally so it's kind of like can we have this world have a local model and a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=473s"},{"start_seconds":481,"text":"cloud-based model or decentralized model so so with the Llama for function so llama llama is not fine tuned to be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=481s"},{"start_seconds":489,"text":"doing function cooling right do you do you fine tune it yourself or you like hack it with a PR yeah yeah so uh 405b","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=489s"},{"start_seconds":496,"text":"we have good results for function calling 70b is is not as good grock uh if you guys have used Gro rock is great","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=496s"},{"start_seconds":504,"text":"uh really great TPS um but uh they have a fine tune model for Tool use and we've done some","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=504s"},{"start_seconds":512,"text":"things with that as well that was pretty good um but we've kind of used just a plethora of different models we've also","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=512s"},{"start_seconds":519,"text":"tried centralized models we've had some customers have interesting latency requirements so we've done uh centralized models we've done uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=519s"},{"start_seconds":527,"text":"decentralized models we're working with numerous Partners on that side too but yeah I mean all kinds of cool things for function calling there","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=527s"},{"start_seconds":539,"text":"yeah I'm I'm curious because I think obviously seeing an open source model right that you can customize that you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=539s"},{"start_seconds":547,"text":"can you know fine tune is is definitely being something like that's comparable to kind of close Source models and at","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=547s"},{"start_seconds":555,"text":"same time there's this continuous question if if the models are kind of the important part right or they're just","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=555s"},{"start_seconds":564,"text":"a commodity that you kind of pick up and then you build stuff around right and this is kind of a continuous back and forth where it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=564s"},{"start_seconds":571,"text":"like is the next model you know the Next Generation or whatever going to like replace all of the uh kind of wrappers","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=571s"},{"start_seconds":579,"text":"and all of the infrastructure we building around or uh you kind of need to manage all that and indeed if you're","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=579s"},{"start_seconds":586,"text":"optimizing for like cost and performance uh you do need to combine kind of smaller models and uh use big","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=586s"},{"start_seconds":595,"text":"bigger models apparently but yeah just just been curious kind of how people reaction been have you seen um I forget who wrote","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=595s"},{"start_seconds":604,"text":"it but it was a really interesting take I saw the other day which effectively was using 405b to kind of like train or","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=604s"},{"start_seconds":612,"text":"fine-tune smaller models um so you can kind of use a larger model to actually uh kind of look at making more or less","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=612s"},{"start_seconds":618,"text":"like a slm instead of llm so you can actually have these like much much smaller models which are fine-tuned for a particular task but you can use a more","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=618s"},{"start_seconds":627,"text":"generalized model uh to actually create that have you guys kind of looked at any of that stuff I think that's incredibly","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=627s"},{"start_seconds":634,"text":"exciting well I mean the 3.1 is that those models are fine-tuned from the output of the bigger model and so so but","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=634s"},{"start_seconds":643,"text":"I think like karpathy was the one who who made the point the most clear to me at least it was that you know the gpt2","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=643s"},{"start_seconds":652,"text":"level model is continuously increasing in in quality because we keep training bigger models and then use them to pre-process","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=652s"},{"start_seconds":660,"text":"data for gpt2 so you can imagine that and I think like uh I mean I I I always think you know the current pre-training","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=660s"},{"start_seconds":668,"text":"of predicting next token is let's say rudimentary um or we use the suboptimal word right it's like it's the easiest","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=668s"},{"start_seconds":676,"text":"thing to do it's easiest loss function to use because that's what we have but you can imagine actually if we had a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=676s"},{"start_seconds":684,"text":"data set that you can train you know pure reasoning question answering you know information retrieval all the tasks code generation you can train way","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=684s"},{"start_seconds":693,"text":"smaller model to be able to do all that right you just need a data set like that and like that data set doesn't exist but if you have a bigger model that is like","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=693s"},{"start_seconds":702,"text":"starting to be pretty good you can actually generate a data set and then train a gp22 load model on that so I","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=702s"},{"start_seconds":710,"text":"mean the I think the the most unclear part is like this new like new set of capabilities that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=710s"},{"start_seconds":717,"text":"appear right with this different sizes which like as Alex described right it's like it's either it is able to do it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=717s"},{"start_seconds":725,"text":"even if it takes like multiple tries or it's just purely not able to do it at all right and the part that to me and and I don't think anybody has described","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=725s"},{"start_seconds":733,"text":"very well kind of this boundary like where is that boundary comes from is it literally like it hasn't seen that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=733s"},{"start_seconds":741,"text":"example enough in data or it just not able to you know compute over it like what what is this boundaries that kind of we see","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=741s"},{"start_seconds":750,"text":"um in in this ability to reason and kind of when it's not not able to grasp it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=750s"},{"start_seconds":756,"text":"even literally like if you prompt it with you know cor correct direction it's still not able to go from it it's worth","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=756s"},{"start_seconds":764,"text":"mentioning my sample size was pretty small it was like six seven problems right I I actually suspect that there's a there's a frontier it's just","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=764s"},{"start_seconds":772,"text":"relatively small right like if you imagine uh like plotting doesn't have to be competitive programming right like you you take any domain where models can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=772s"},{"start_seconds":782,"text":"solve like you know subset of it and completely clueless one another like naturally if the if the domain is large enough then you do expect that the FR","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=782s"},{"start_seconds":791,"text":"here is going to be relatively small percentage uh of the current of the total set right because","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=791s"},{"start_seconds":801,"text":"uh uh just because the number how do you explain it just because like you know","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=801s"},{"start_seconds":809,"text":"the the difference between the easiest and the hardest is so is so is so large that uh you know like any you know like small window will only","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=809s"},{"start_seconds":818,"text":"capture a small subset yeah and I think an interesting evidence something I discussed with the Bowen from near one couple days ago is that if you if","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=818s"},{"start_seconds":828,"text":"you if you think about the Deep mindes paper on IMO I don't know if we plan to talk about it later today uh we can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=828s"},{"start_seconds":836,"text":"shift into that yeah describe it for for those who are not reading all the papers that being published","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=836s"},{"start_seconds":843,"text":"by yeah so so so so I guess it would be pretty hard to live in the modern world and not hear the news that deep mind got","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=843s"},{"start_seconds":850,"text":"the silver on IMO with a couple of asteres uh going both ways one asteris is that they were one point short of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=850s"},{"start_seconds":858,"text":"gold uh so it's almost gold and moreover they didn't solve at all the last two problems and interestingly on IMO uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=858s"},{"start_seconds":867,"text":"when you like when you're in that state and you're a human what you would do is you would write any anything any thoughts you have because that could be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=867s"},{"start_seconds":875,"text":"worth one point right and deep M could have done that and and like there's a belief among some people in the community that they didn't do that just","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=875s"},{"start_seconds":882,"text":"not to water down the result right like it would be worse if they said hey we have gold that people would be like yeah but like that one point is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=882s"},{"start_seconds":888,"text":"right so uh but the second as is that it took them three days and the actual participants only had four hours","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=888s"},{"start_seconds":898,"text":"right but but I think that is completely irrelevant like if if there's one thing we are good at as a as as a species it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=898s"},{"start_seconds":905,"text":"optimizing something that takes three days to Take 5 Seconds uh right and the way they did it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=905s"},{"start_seconds":913,"text":"is uh uh so so so so let's put aside Alpha geometry for now so Alpha geometry is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=913s"},{"start_seconds":921,"text":"older tool they they've had it for more than a year and it solved one problem but the remaining three they solved by","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=921s"},{"start_seconds":929,"text":"uh converting the problem statement into lean that is still done by a human they have a model that kind of can do it but","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=929s"},{"start_seconds":935,"text":"it's uh uh it can hallucinate and if you you know if you make a mistake in the uh translation then uh you know all better","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=935s"},{"start_seconds":945,"text":"of so that is still done by a human there's a human who converts the written statement into lean uh and then they do","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=945s"},{"start_seconds":952,"text":"effectively give or take Alpha zero on Le uh so so they train they train using something you know some sort of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=952s"},{"start_seconds":959,"text":"reinforcement learning uh with three search and then they do inference using three search until they find the proof and the beauty of Lan is that if you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=959s"},{"start_seconds":968,"text":"have the statement you know if you have proven it right so unlike human written proofs that are very hard to read and Lan proof uh uh is you know there's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=968s"},{"start_seconds":978,"text":"there's a system that checks it and it knows if it's correct and so for three problems they found proofs uh and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=978s"},{"start_seconds":985,"text":"uh those seem to be relatively hard problems so it's not uh it's not an a small","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=985s"},{"start_seconds":993,"text":"achievement right and and something yeah coming back to my conversation with Bowen something he mentioned is that it's interesting that like the way the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=993s"},{"start_seconds":1001,"text":"way they train right the number of lean programs in the world is very small right Lin is a relatively young language and it's very hard language right so the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1001s"},{"start_seconds":1009,"text":"number of people who who who write in Lan is relatively small uh and so if you if you train on all the lean available you're not going to go far right you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1009s"},{"start_seconds":1017,"text":"need to be synthesized ing it and the way they synthesize It Is by um they they get uh that step I don't know fully","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1017s"},{"start_seconds":1027,"text":"where they get a lot of problems but they get a lot of problems without proofs right and then you know the the system proves them and when it's proven","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1027s"},{"start_seconds":1034,"text":"it goes to the to the data set right and so what is interesting here is that initially of course the system is relatively weak so it can only prove the simplest problems right and so it proves","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1034s"},{"start_seconds":1043,"text":"a lot of simple problems it gets better at them uh and uh you would imagine right that there are two possib ities here right we live in one of the two","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1043s"},{"start_seconds":1050,"text":"universes in one Universe it will eventually become very very good at solving very simple problems but it will never transition to solving anything","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1050s"},{"start_seconds":1059,"text":"harder right so it will get stuck and in another Universe uh it keeps on getting better uh and you know like once it gets better it solves more problems it gets","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1059s"},{"start_seconds":1067,"text":"better solves more problems and eventually there's like this continuous transition until it gets to to hard problems right and it seems like we live","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1067s"},{"start_seconds":1074,"text":"in the latter Universe because they did eventually get to solving Amo so that is something that is uh that was not necessarily the you know like it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1074s"},{"start_seconds":1084,"text":"wouldn't be obvious that that's the case but it seems to be that's pretty interesting yeah and it actually","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1084s"},{"start_seconds":1093,"text":"triggers a different question that in from a business perspective um you mentioned two different uh kind of future that we can expect from this and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1093s"},{"start_seconds":1102,"text":"actually it getting better and better at solving more problems and later on maybe the other side would be doing different","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1102s"},{"start_seconds":1109,"text":"uh smaller solving different smaller uh problems and only staying at that level and you would have to find a different method to actually find solutions for","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1109s"},{"start_seconds":1117,"text":"solving more complex questions and from a business standpoint um from an economic standpoint what it be a little","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1117s"},{"start_seconds":1124,"text":"bit more um what should I say a little bit more closer to what the projects can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1124s"},{"start_seconds":1132,"text":"actually adopt at the current moment and what would be more cost effective for us to actually um EXP and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1132s"},{"start_seconds":1140,"text":"adopt yeah so so I think from this perspective one thing that I still mostly agree with uh is that uh like if","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1140s"},{"start_seconds":1149,"text":"you embark on a journey of building something uh that will be ready in a year you should effectively you know be building assuming that the models will","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1149s"},{"start_seconds":1157,"text":"be noticeably better right so it's uh like if you try to build something uh that can be built with existing models","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1157s"},{"start_seconds":1165,"text":"then you will be a little bit behind uh you know when it's ready um but I think another interesting","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1165s"},{"start_seconds":1174,"text":"thought which is uh um so so this like with everything Deep Mind does uh it's extremely mind mind-blowing uh and it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1174s"},{"start_seconds":1183,"text":"very hard to um it's very hard to generalize to the to the wider um set of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1183s"},{"start_seconds":1190,"text":"applications right because okay we see that uh in this particular application of proving you know uh uh Olympia level","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1190s"},{"start_seconds":1198,"text":"mathematics it's very good uh but it's very hard to understand if that will generalize to wider applications of lim","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1198s"},{"start_seconds":1206,"text":"right because it wasn't trained on like um like it wasn't trained on practical application on on practical problems","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1206s"},{"start_seconds":1214,"text":"right it was synthesized problems and it happened that in this particular case synthesized problems helped to solve limped mathematics right and so we have","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1214s"},{"start_seconds":1222,"text":"no idea whether you know similar synthesized problems will help you apply L and other uh means right because lean","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1222s"},{"start_seconds":1231,"text":"uh is uh or like generally so I think first of all L is positioned today to be the def facto standard language of formal proofs of tomorrow right because","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1231s"},{"start_seconds":1239,"text":"all the research happens in lean the standard library is you know getting bigger and bigger so I think effectively Lin is becoming a synonym uh of formal","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1239s"},{"start_seconds":1248,"text":"proving right and if it gets extremely good to interject here the formal proving is needed for actually","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1248s"},{"start_seconds":1255,"text":"programming like the AI software engineering that's like foundational piece","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1255s"},{"start_seconds":1263,"text":"otherwise yeah yeah so necessary is a strong word right because we've been building code for decades now and I would say throughout those decades it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1263s"},{"start_seconds":1271,"text":"was also necessary but we somehow managed to get by without it and the wor that has happened yet is a complete","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1271s"},{"start_seconds":1277,"text":"collapse of the whole world for one day but but but like you can imagine","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1277s"},{"start_seconds":1286,"text":"that on on steroid when you also have ai that is like of a level that can break any code and inspect it while inspecting","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1286s"},{"start_seconds":1294,"text":"it or yes exactly exactly side effects of it yeah so I think yeah so one thing is that we are probably getting to the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1294s"},{"start_seconds":1302,"text":"state where AI will start breaking code very well so at that point having code which is formally proven would be a massive asset right so like you know","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1302s"},{"start_seconds":1310,"text":"it's not going to be broken right no matter how smart you youri is the code is proven to be doing exactly what you want so the at worst uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1310s"},{"start_seconds":1319,"text":"you can still mess up the actual guarantees right so you have proven that the code does that but that is messed up","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1319s"},{"start_seconds":1326,"text":"right it's not what you actually want uh but the second thing is that like you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1326s"},{"start_seconds":1332,"text":"know generally the everything we use today all the website all the apps they all garbage right they it's it's just","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1332s"},{"start_seconds":1340,"text":"like uh like you know you open the United App uh and uh in the airport and it like crashes on you or like it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1340s"},{"start_seconds":1347,"text":"doesn't load stuff or like and that's like throughout all the like banking apps notion they","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1347s"},{"start_seconds":1354,"text":"all uh right and we s like if if we were to generalize the achievement of Deep Mind we might be moving into the uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1354s"},{"start_seconds":1364,"text":"phase of humanity where all the code will be formally proven because uh you know like a is getting good at doing it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1364s"},{"start_seconds":1371,"text":"for us right so so so that would be pretty cool and and like uh that would be actually an interesting thing like like you know for a startup to say hey","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1371s"},{"start_seconds":1379,"text":"let's say it is the case it might not be the case and then my startup will die but there are you know 99 other reasons it could die right but let's say it is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1379s"},{"start_seconds":1387,"text":"the case let me bet fully uh on you know Lin being fully synthesizable in a year right like can can we do something cool","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1387s"},{"start_seconds":1396,"text":"uh you know to be ahead of the curve where all the code is formally verified and you know it doesn't it doesn't get","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1396s"},{"start_seconds":1403,"text":"hacked it doesn't fail yeah so so maybe like Ian because you brought it up right like you know","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1403s"},{"start_seconds":1411,"text":"with Claude and with others hey the code generation is getting you know it's getting to a level where like it give","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1411s"},{"start_seconds":1419,"text":"you a dirty prototype pretty quickly right it may still have some issues or it's not","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1419s"},{"start_seconds":1426,"text":"handling some edge cases and uh it it's harder to integrate in a bigger code basis usually but you can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1426s"},{"start_seconds":1436,"text":"see like hey you know if it's a single file thing like it can you know start getting there uh at least like um to to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1436s"},{"start_seconds":1445,"text":"to start but indeed the question is the complexity of integrating kind of a single module into a bigger system is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1445s"},{"start_seconds":1453,"text":"where lots of complexity is like for a person to write a new program it's pretty easy right like you know but","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1453s"},{"start_seconds":1461,"text":"actually adding code to existing code Bas is where complexity is because there's so many different conditions and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1461s"},{"start_seconds":1469,"text":"variant and you know usually some some other things that like only in few people's brains potentially that you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1469s"},{"start_seconds":1477,"text":"need to recover from reading this code base and so I think the kind of looking forward we will need some you know","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1477s"},{"start_seconds":1485,"text":"combination of the tools to come together to be really both","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1485s"},{"start_seconds":1493,"text":"advancing ability to like expand code bases right with was guarantees that you know it's not going to fail again when","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1493s"},{"start_seconds":1501,"text":"somebody pushes like one line change and at the same time this also would allow to have kind of more complex","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1501s"},{"start_seconds":1510,"text":"transformations to the code base in itself right because semantically you can imagine like I'm looking at some application I'm okay I want to switch","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1510s"},{"start_seconds":1518,"text":"database from SQL to uh you know no SQL for example like that's a very simple semantic change which requires a lot of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1518s"},{"start_seconds":1526,"text":"syntactic like reimplementation of in the code base um but like for that you need a lot of like guarantees and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1526s"},{"start_seconds":1534,"text":"understanding of you know what is the differences between SQL and no SQL and how those are being","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1534s"},{"start_seconds":1543,"text":"then kind of surfaced in the application in the back end for example because you know of the complexity of the retrieval","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1543s"},{"start_seconds":1550,"text":"etc etc right so so all of that is like like models have some of that information just because they statistically saw a bunch of those cases","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1550s"},{"start_seconds":1558,"text":"but like making it a lot more uh apparent right by maintaining some of the let's say lean proof around it would","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1558s"},{"start_seconds":1565,"text":"allow to to actually have like a pretty good investment I mean that's kind of what we've been you know with Alex uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1565s"},{"start_seconds":1572,"text":"discussing and and working on some of the pieces of that uh but I mean we do think it's it's coming right although","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1572s"},{"start_seconds":1580,"text":"we've been thinking about it that it's coming from 2017 that it's near and before yeah and and before that it was","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1580s"},{"start_seconds":1587,"text":"less near before that but yeah I mean I well know that you guys uh started off trying to develop ai","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1587s"},{"start_seconds":1596,"text":"ai developer so you guys have an extensive background in this and for us actually uh what we're more focused in here is that we wanted to create an","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1596s"},{"start_seconds":1605,"text":"actual use case because it feels like we are at a critical stage where what three has to prove itself that it can prove it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1605s"},{"start_seconds":1613,"text":"can have an important role to play in the AI game and for us how we saw it differently is not from the AI code generation perspective but actually","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1613s"},{"start_seconds":1621,"text":"putting it in a low code format so it is a lot easier for us to compose it and for the people as you mentioned like","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1621s"},{"start_seconds":1629,"text":"when they have to go back and switch out something it's a lot easier because it is simplified into smaller blocks that they have um pulled through our AI","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1629s"},{"start_seconds":1637,"text":"engineer and for us um of course we can't compete against these bigger models or big tech companies that are working on these","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1637s"},{"start_seconds":1645,"text":"Technologies especially like IBM just L open sourcing their grandite family and they're trying to receive a lot of feedback and community support from","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1645s"},{"start_seconds":1654,"text":"developers as well and for us we are trying to do the same thing but we want it to be in a predefined and a pre-built","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1654s"},{"start_seconds":1661,"text":"uh package as an API and which actually gives us a little bit more flexibility of course we won't be able to build","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1661s"},{"start_seconds":1668,"text":"super complex uh apps at the moment but the more we actually build out this codebase and build out this API registry","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1668s"},{"start_seconds":1676,"text":"we will be able to tailor and cater to more of the custom and unique business needs that the web 2 application","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1676s"},{"start_seconds":1682,"text":"Builders actually have and from here on I believe we can actually generate those use cases and based off of that we can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1682s"},{"start_seconds":1689,"text":"actually try to mimic what these bigger models are doing and I don't necessarily believe that we have to work against","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1689s"},{"start_seconds":1698,"text":"these bigger um technological innovations but somehow if we can work alongside of it then it would be a better um option for us of course the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1698s"},{"start_seconds":1707,"text":"centralization is um has a lot of value propositions and we do believe in it but for us we're just starting off from the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1707s"},{"start_seconds":1714,"text":"data layer which is the uh bottom layer of theist stack and everything else of course we have it in a traditional setting uh the only difference uh is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1714s"},{"start_seconds":1723,"text":"that we have it all open and publicly available so from there on we're trying to see where this is this can head in","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1723s"},{"start_seconds":1731,"text":"building more complex softwares so for the I think one","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1731s"},{"start_seconds":1737,"text":"interesting question here is uh um so so the if I understand correctly","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1737s"},{"start_seconds":1745,"text":"the the sort of mental model of how a person would build a uh I don't know if you call them","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1745s"},{"start_seconds":1751,"text":"tickles if a person builds a tickle uh that uh there's like like I imagine you know like the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1751s"},{"start_seconds":1759,"text":"the end goal is that there's this massive um in in in inventory of of apis","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1759s"},{"start_seconds":1766,"text":"right and so I go there and and it's not just like it's not just that I'm relying on an AI to build a whole system from scratch correctly I'm relying that it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1766s"},{"start_seconds":1775,"text":"writing a small uh effectively gluc code correctly which is a much smaller scope plugging in into all those apis right and like when I was playing I said you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1775s"},{"start_seconds":1784,"text":"know like apis have reputation you know like reputation 10 means that don't touch it reputation 90 means that it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1784s"},{"start_seconds":1791,"text":"probably legit right so in this model um the the problem I see and I'm","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1791s"},{"start_seconds":1799,"text":"curious how you guys think about addressing it is that apis go down apis go down right and so if if we build a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1799s"},{"start_seconds":1806,"text":"whole ecosystem of things that are like interconnected right and like like imagine this completely madeup example","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1806s"},{"start_seconds":1814,"text":"that many people need an API to pad a string with spaces to the left let's call it left pad right and then the left pad API goes down right what happens","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1814s"},{"start_seconds":1822,"text":"like how how do you how do you think about the problem of API disappearing or like introducing breaking changes and the cascading effect","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1822s"},{"start_seconds":1831,"text":"of so as you mentioned API of course one of the apis that can actually go down but what our AI engineer is good at is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1831s"},{"start_seconds":1839,"text":"that it fetches and integrates it as um IIA mentioned of course it's not just the code writing site but it's also the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1839s"},{"start_seconds":1846,"text":"knowledge and expertise that you need from the backend site and from there on we actually have um optimize the kubernets cluster and um allocating","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1846s"},{"start_seconds":1854,"text":"resources in an optimal way so whenever one goes down of course at the moment there could be a little bit of delay uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1854s"},{"start_seconds":1861,"text":"in terms of how the app will run but at the same time we'll have a group of developers that are working on and monitoring the specific API so we'll try","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1861s"},{"start_seconds":1869,"text":"to uh push different version updates through it of course um even proprietary uh apis that are provided by local","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1869s"},{"start_seconds":1876,"text":"solutions that are that already exist in the web to ecosystem they are not perfect and plus on top of that they are not verifiable so when the users","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1876s"},{"start_seconds":1884,"text":"actually receive the apis they don't know if that's the best quality that that they can receive from the market so at least what we can offer is that we","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1884s"},{"start_seconds":1892,"text":"actually create this Merit based ecosystem because it needs to be selected search and selected by the AI engineer so at least what we know is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1892s"},{"start_seconds":1899,"text":"that the code itself is as secure and as high code quality as possible if it goes down of course there on it is based on","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1899s"},{"start_seconds":1906,"text":"us and it's based on us U being able to quickly ship those problems and Outsource those problems to our developer Community I see but but I think you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1906s"},{"start_seconds":1915,"text":"implicitly answered the the the other question I which is that like for as long as API is deployed by a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1915s"},{"start_seconds":1922,"text":"tickle uh effectively there's no way like like my my bigger question was not just the server going down but the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1922s"},{"start_seconds":1930,"text":"developer like the original developer just completely shutting it down but I guess as long as you can you you own the apis you can just like I guess you can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1930s"},{"start_seconds":1939,"text":"forbid developers shutting them down right it's a right uh so we call user own apis so basically if the user","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1939s"},{"start_seconds":1947,"text":"decides to just completely dish the API then of course we will be able to find a followup uh API that we can quickly","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1947s"},{"start_seconds":1955,"text":"replace with it so I think it's same with any kind of user owned assets um of course if there are slashing um rules","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1955s"},{"start_seconds":1963,"text":"and so on there will be less people who actually dish or whatever asset they're putting out in the marketplace and actually providing people to pay for it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1963s"},{"start_seconds":1972,"text":"but at the same time um for us we we are currently not trying to impose any kind of slashing at this point we just want to make it a little bit more","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1972s"},{"start_seconds":1979,"text":"collaborative rather than being penalty driven so for us what when that happens we'll be able to replace it with some","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1979s"},{"start_seconds":1987,"text":"other apis within the same domain and within the same domain specific functionalities and there's one more","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1987s"},{"start_seconds":1994,"text":"thing which is I'm pretty sure it's on your road map but I I didn't see it there but of course when I deploy my API I want a little bit of money streaming","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=1994s"},{"start_seconds":2003,"text":"to my account every time it is called so I guess uh if you're a supplier of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2003s"},{"start_seconds":2010,"text":"the API so we basically try to incentivize in a very different way so the first one will be the hit rate which basically means how much recommendation","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2010s"},{"start_seconds":2019,"text":"it actually receives and how many times it actually been uh recommended to the end users and the second is actually","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2019s"},{"start_seconds":2025,"text":"based on the usage so a lot as we started off as a backend as a service um we still try to um position Us in that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2025s"},{"start_seconds":2034,"text":"way so people will pay monthly based usage fees based on the time that they have instantiated the API stack and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2034s"},{"start_seconds":2042,"text":"based on that we will minus all the infrastructure costs and so on and from whatever is left based on the usage percentage we'll be able to distribute","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2042s"},{"start_seconds":2049,"text":"it to the API providers so of course we could ship out um tokens as incentives","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2049s"},{"start_seconds":2058,"text":"but we don't see it as a sustainable way to do it of course in the beginning we have different code of recruiting events for different domains and there will be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2058s"},{"start_seconds":2066,"text":"different incentives that people can take off from but the main sustainable uh income for the API providers will be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2066s"},{"start_seconds":2074,"text":"based on the usage actual usage yeah maybe similar question to Ryan","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2074s"},{"start_seconds":2081,"text":"around agents right pretty much how do you manage agents Lifetime and payment and uh if like people are able to spun","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2081s"},{"start_seconds":2091,"text":"up uh own agents if the agents kind of may disappear yeah yeah absolutely so on our","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2091s"},{"start_seconds":2099,"text":"side we actually we explored that space uh quite a bit originally for allowing people to create their own agents to then like support a given protocol if","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2099s"},{"start_seconds":2107,"text":"they would like to so maybe like someone wants to create a Unis swap agent or a wormhole agent whatever it may be um what we kind of ran into and what we","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2107s"},{"start_seconds":2116,"text":"kind of done is we shifted some things around in our road map for that because we ultimately because we're dealing with money the a concern kind of popped up","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2116s"},{"start_seconds":2124,"text":"around uh like how do you keep an nefarious person out so how do you essentially say and and enable a way in","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2124s"},{"start_seconds":2132,"text":"which like a nefarious person says okay if I'm doing a swap I'll do the swap normally uh if it's less than a million dollars if it's more than a million","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2132s"},{"start_seconds":2140,"text":"dollars then I'm just going to take the money and run so you know you can have reputation systems and what we were looking is like there's reputation systems you can have you can then","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2140s"},{"start_seconds":2149,"text":"effectively game those uh for what we're doing uh if you had slashing stuff that was something that we kind of looked at","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2149s"},{"start_seconds":2156,"text":"uh there's just some kind of uh concerns that we had around that so what we have actually done is like we built out and we turned our tool loose on our","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2156s"},{"start_seconds":2165,"text":"community and we wrapped over three quarters of defi in three days so we ended up doing like 800 different agents","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2165s"},{"start_seconds":2172,"text":"uh quite quickly um that was really cool to see but the problem arises where is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2172s"},{"start_seconds":2179,"text":"that if that code touches someone individually then how are you ensuring that they don't have uh a particular like nefarious piece in there and we may","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2179s"},{"start_seconds":2187,"text":"do slashing or whatever but if it effectively has a problem before that then I think there's a a larger issue so what we've been focusing on more now is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2187s"},{"start_seconds":2195,"text":"that effectively we can build out a series of Agents ourselves using this tool that's like more self-replicating and then we can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2195s"},{"start_seconds":2202,"text":"actually do more community-owned AI so effectively Implement something that there's an experimental paper that came out called co-learning but effectively","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2202s"},{"start_seconds":2210,"text":"it's where uh each individual use can be ran and we can identify we have a knowledge graph that Associates between all different agents so you know how to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2210s"},{"start_seconds":2218,"text":"Traverse and which agent to populate in an event where you're trying to do a given transaction so we can incorporate those directly and then if there's a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2218s"},{"start_seconds":2226,"text":"problem in the overall like flow of that individual agent uh so maybe it gets transaction first before getting a quote","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2226s"},{"start_seconds":2233,"text":"we can then uh refine that and through that refinement that person can be rewarded and then the next person has a better experience the next time and then","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2233s"},{"start_seconds":2241,"text":"what we want to get to is where users own the agents but we want to have more or less more functionality of using the existing library that we effectively can","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2241s"},{"start_seconds":2250,"text":"build out ourselves and then enable people to extend capabilities on top so maybe they want to do uh great payments on uh like an a cost or something we","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2250s"},{"start_seconds":2259,"text":"could then say okay you can use these series of Agents up front and then you're effectively composing this into a button so you don't have to actually","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2259s"},{"start_seconds":2266,"text":"type out or do anything directly it's all like kind of like through a button interface directly and we can have more and more of these uh buttons or","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2266s"},{"start_seconds":2275,"text":"different capabilities be used kind of what we did recently tip link go ahead no go yeah you mentioned","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2275s"},{"start_seconds":2286,"text":"you created 800 agents in several days so this is because agents themselves are written by an agent right you actually","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2286s"},{"start_seconds":2294,"text":"have humans WR 800 agents so what they did was we had a tool that was deployed in which you could say I would like to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2294s"},{"start_seconds":2301,"text":"create an agent for UNIS Swap and then what we can do is we could pull in the contracts of Unis Swap and the dev veler docs we create a rag agent for like a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2301s"},{"start_seconds":2310,"text":"micro viw over what that agent actually does and what Unis swap is and then we are really what we're doing is we're not","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2310s"},{"start_seconds":2316,"text":"creating like a whole generation piece we're more or less like stubbing out the functions associated with that protocol","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2316s"},{"start_seconds":2324,"text":"and we're just saying okay the inter agents essentially will interface with those stubbed out functions so like getting a quote or getting a transaction so we're not like replacing their","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2324s"},{"start_seconds":2332,"text":"contracts we're just kind of creating a interface layer on on top of that yeah and so what is the um what what is the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2332s"},{"start_seconds":2341,"text":"end goal of an agent right so like let's say I have an agent for a for compound uh well let's not do for compound let's do for","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2341s"},{"start_seconds":2350,"text":"Abby for compound is going to be attacking the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2350s"},{"start_seconds":2357,"text":"Govern yeah so go for so so the question is what is the what is the utility of the agent so so","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2357s"},{"start_seconds":2365,"text":"first of all is the agent designed to used by another agent or by human if and if by a human what is the like I'm a human right what is the use case for me","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2365s"},{"start_seconds":2373,"text":"to use to use the agent um the agent itself can be built by other agents and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2373s"},{"start_seconds":2380,"text":"then the user of it can be other agent or it can be a user directly so we've deployed on we're on a kinetics the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2380s"},{"start_seconds":2388,"text":"number one Dex on Cava chain we're deployed on there we deployed five agents there directly and essentially what you can do there is you can ask it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2388s"},{"start_seconds":2395,"text":"and say like you know what is kinetic who built it you know stuff like that hit the rag agent and then if you wanted to like add in something where it's like","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2395s"},{"start_seconds":2402,"text":"I would like to do a swap and then tie it into 100x long position through their perose product you can essentially combine the two together and create one seamless stream um so then it's more","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2402s"},{"start_seconds":2411,"text":"about like partner with like you guys and other people to essentially provide a chain abstracted experience so then you populate all the transactions forward and then how do they go about","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2411s"},{"start_seconds":2420,"text":"signing that and that's really up to the user they want to go one by one do they want to sign like via a policy engine uh do want something that's chain attracted","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2420s"},{"start_seconds":2429,"text":"that kind of thing so so is is it more about like making more more user friendly experiences or is it more about","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2429s"},{"start_seconds":2437,"text":"uh like abstracting away the boundaries of uh all applications yeah so uh what we found","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2437s"},{"start_seconds":2445,"text":"when we were doing payments was that it just was incredibly difficult to like try and support everything and everyone someone have money on like Cava chain","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2445s"},{"start_seconds":2454,"text":"and we don't support Cava chain so then all of a sudden how do add in a bridge from Cava chain over into salana or near","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2454s"},{"start_seconds":2461,"text":"or whatever and it's like you what we ended up doing is we incorporate over 130 different pieces of infrastructure so we built out a lot of stuff to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2461s"},{"start_seconds":2470,"text":"support 18 chains and it was just like this is like a NeverEnding problem you know it's like the protocols just kept","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2470s"},{"start_seconds":2478,"text":"expanding faster and faster there's more chains now you have l2s and l3s and all that stuff so part of it is to create a userfriendly experience but it's also to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2478s"},{"start_seconds":2486,"text":"extend the cap abilities of that actual like protocol directly so instead of that protocol having to figure out how do I bring in an on-ramp how do I bring","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2486s"},{"start_seconds":2494,"text":"in a a bridge to get people's liquidity into my like actual project we can say hey use our like effec our library of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2494s"},{"start_seconds":2501,"text":"agents and then enable your users to pull in funds easily through like kind of like One-Stop shop um kind of system","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2501s"},{"start_seconds":2511,"text":"and they don't have to spend the time to do that so I actually want to I want I'm I'm interested to like challenge","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2511s"},{"start_seconds":2517,"text":"everybody's perspective on what is the difference between an agent and API","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2517s"},{"start_seconds":2527,"text":"um for for me what we've seen is that we when we did payments uh we had a lot of issues with","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2527s"},{"start_seconds":2536,"text":"redundancies and rpcs going down and so it's kind of more or less what we continue to see is like there's increased redundancy and fault tolerance","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2536s"},{"start_seconds":2546,"text":"so like at least how I'm looking at and I'm biased I think it's like the next step almost because I feel like the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2546s"},{"start_seconds":2555,"text":"unlocks we have had enable a lot more of a uh uh like say I want to hit like a particular RPC I just want to get my","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2555s"},{"start_seconds":2563,"text":"balance or I just want to understand like what this wallet is or whatever I I just want to do a transaction RPC goes down we can have an agent then populate","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2563s"},{"start_seconds":2570,"text":"a new RPC try it again but instead of it being like written out manually in code um and that would have taken you know a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2570s"},{"start_seconds":2578,"text":"developer maybe a week to implement or go through the whole process we can have effec some with some reasoning capability to provide this fault","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2578s"},{"start_seconds":2586,"text":"tolerance in association with what the user is trying to do directly which I think is kind of much more extensible","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2586s"},{"start_seconds":2592,"text":"and uh flexible so that's what I get excited about between the two I mean from our perspective it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2592s"},{"start_seconds":2601,"text":"quite similar as well but what apis actually do is it actually can only serve a specific function and and it can actually only take in um do certain","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2601s"},{"start_seconds":2609,"text":"actions based on certain signals and for the agents um where we see it in the future is that they can actually be interoperable and they can actually be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2609s"},{"start_seconds":2618,"text":"working together autonomously and from there on of course uh we can make it to be a bit more fallproof at the same time","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2618s"},{"start_seconds":2626,"text":"just providing it as a single more than just providing it as a single API it will have a little B more flexibility in","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2626s"},{"start_seconds":2633,"text":"terms of actual uh business applications yeah I think for me the difference is in","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2633s"},{"start_seconds":2642,"text":"like semantic versus syntactic right API is a syntactic act like syntactically defined action and kind of fa tolerance","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2642s"},{"start_seconds":2651,"text":"as well as you know some intelligent understanding of arguments all of this is more of a semantic like I want the result like I don't care which RPC you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2651s"},{"start_seconds":2660,"text":"use just figure it out um but at the same time you know you can imagine apis becoming","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2660s"},{"start_seconds":2667,"text":"like you know open ey API is literally a semantic thing or you're calling an agent API and so I am I am kind of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2667s"},{"start_seconds":2676,"text":"curious where that line lies and and how much more semantic can apis","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2676s"},{"start_seconds":2684,"text":"become uh broadly as well I I I want to bring up another thing we we we mentioned it couple office hours ago there's this website","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2684s"},{"start_seconds":2693,"text":"called web. which uh it's if if someone who watches it didn't see it yet it's completely mind-blowing even in the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2693s"},{"start_seconds":2702,"text":"Modern Age of AI uh web like web simulation. web. and couple days ago","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2702s"},{"start_seconds":2708,"text":"maybe earlier than that someone reverse engineered it right so they found how it works it uses clo which I guess is not","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2708s"},{"start_seconds":2716,"text":"very surprising uh and a very interesting aspect of it which is very relevant to elia's question is that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2716s"},{"start_seconds":2722,"text":"effectively they ask ask clae to write the website based on your like small prompt uh you know on the fly but that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2722s"},{"start_seconds":2731,"text":"website might have uh ax queries right to the back end but there's no backend right so so the way they do back end is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2731s"},{"start_seconds":2738,"text":"that whenever there's any ax query they go to CL and say hey clo uh we have this aex query for this website what should","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2738s"},{"start_seconds":2745,"text":"we respond with and Claud is you know on the back end is coming up with the response uh and like like if you play","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2745s"},{"start_seconds":2752,"text":"with web. you will see that uh like it works you know 99 out of 100 right so so","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2752s"},{"start_seconds":2759,"text":"here you know what like that's an agent for sure that's also an API yeah I think somebody did as well an","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2759s"},{"start_seconds":2768,"text":"example where the the back end is just llm responding to like exactly right but but but it but but here it is plugged","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2768s"},{"start_seconds":2777,"text":"into the to an actual website which is you know built on the Fly and and it all works seemlessly right so like especially uh you can intentionally come","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2777s"},{"start_seconds":2785,"text":"up with examples that will have an right like you can have it as part of your prom saying hey it it like calls to an API to do something uh and and it all","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2785s"},{"start_seconds":2793,"text":"works uh very well it's also a cool thing you know when you prototype a website before you have the back end you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2793s"},{"start_seconds":2802,"text":"can just St stab it with yeah yeah later you can probably you probably want to replace it with an actual back end but","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2802s"},{"start_seconds":2809,"text":"you know for the for the early testing it's it's a nice yeah you know if it does something weird you know you just reboot it but","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2809s"},{"start_seconds":2817,"text":"yeah yeah that that is actually not like a new tool that exists in the tool chain of of the developers you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2817s"},{"start_seconds":2826,"text":"imagine the llm that you know has memory that has you know ability to like use a bunch of tools","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2826s"},{"start_seconds":2834,"text":"then potentially with like optimizing it you know you have a router in the beginning that you know oh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2834s"},{"start_seconds":2842,"text":"if I if I have if I have already a function like this called a function instead of uh you know unrolling Chain of Thought and figuring out and if not unroll Chain","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2842s"},{"start_seconds":2851,"text":"of Thought and write a function so that next time you can call it right and so you kind of cache like while you're doing it you also cache everything and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2851s"},{"start_seconds":2859,"text":"then you have you know some database access let's say so like it will just build the back end under the hood while","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2859s"},{"start_seconds":2866,"text":"you're actually doing this I'll go for you I mean it truly is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2866s"},{"start_seconds":2874,"text":"um kind of a crazy world we are jumping into but at the same time uh of course","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2874s"},{"start_seconds":2882,"text":"these are newer Technologies and newer kind of like Tri outs with the existing solutions to kind of find ways to make","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2882s"},{"start_seconds":2890,"text":"it a little bit more fun and actually just find loopholes into it but at the same time um for extra use cases and","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2890s"},{"start_seconds":2899,"text":"Enterprise adoptions when it comes to agents and apis it really depends on how much they can actually rely on it and how much they can actually securely","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2899s"},{"start_seconds":2907,"text":"adopted from an Enterprise standpoint and from there on I think um of course where we draw the line between the two","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2907s"},{"start_seconds":2915,"text":"is actually important but at the same time just working within the traditional setting and actually trying to provide what they actually need at the moment is","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2915s"},{"start_seconds":2923,"text":"also a critical issue that we have right now so of course Drawing the Line we would at some time we will have to draw","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2923s"},{"start_seconds":2931,"text":"the line between the two but for now I think we can just work with what we have and actually try to innovate on the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2931s"},{"start_seconds":2939,"text":"usability which is quite similar what to what Ryan is trying to achieve with the project as well the the thing I get excited about","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2939s"},{"start_seconds":2947,"text":"is I think Andre carpati uh wrote like a tweet or something like that about like llm OS so he kind of talked about this","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2947s"},{"start_seconds":2955,"text":"kind of concept or idea as well and I think that's like uh I mean he's had some very interesting thoughts on like","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2955s"},{"start_seconds":2961,"text":"software 1.0 2.0 and this llos I've been kind of following some of that I think that is really an interesting thought or","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2961s"},{"start_seconds":2969,"text":"perspective on maybe where software can go um and maybe where the future what the future might hold","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2969s"},{"start_seconds":2977,"text":"um but yeah I mean I think it's just exciting times ahead with the existing tools that are available today and even inside of that it's like uh there's lots","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2977s"},{"start_seconds":2986,"text":"of really great research out there I mean uh really good products that are like multi-agents type systems like crew Ai and autogen they've done some things","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2986s"},{"start_seconds":2994,"text":"around like a I think one was called like agent optimized so effectively if you build out a given system you could then find out a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=2994s"},{"start_seconds":3001,"text":"predefined path to do like something in the future so if you do like spin up the backend on the Fly uh in this case you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3001s"},{"start_seconds":3008,"text":"can actually like uh Replay that later so it's almost like you store that in like memory and then you like can cach that and then replay back that same","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3008s"},{"start_seconds":3017,"text":"system for the uh a similar like intent or like whatever you're trying to do with that overall like website which I","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3017s"},{"start_seconds":3024,"text":"think is awesome and and the post about software 2.0 is from November 11th","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3024s"},{"start_seconds":3030,"text":"2017 yeah which is just wild I mean I I think","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3030s"},{"start_seconds":3040,"text":"uh I got into AI in 2017 definitely not nearly as much as you guys uh um that's how I actually got into Bitcoin or","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3040s"},{"start_seconds":3048,"text":"crypto was I built out an AI trading bot and then I started a a little uh NLP company that I did for automatically applying people for jobs using like an","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3048s"},{"start_seconds":3056,"text":"algorithm I built out but uh I mean even back then I thought okay AI is going to take over the world like today like it's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3056s"},{"start_seconds":3064,"text":"happening today like it's it's you know it's over and then um the llm space though like when it first came out","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3064s"},{"start_seconds":3071,"text":"honestly I wasn't as excited like I was like okay this is very cool very interesting um but I wasn't as excited","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3071s"},{"start_seconds":3079,"text":"as I probably should have been and it really wasn't until like some of these more recent adaptations um that I found to be quite","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3079s"},{"start_seconds":3087,"text":"interesting so more or less like how you add and incorporate things into it maybe Monte Carlo like tree searches for reasoning on top of it it's like the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3087s"},{"start_seconds":3096,"text":"Deep Mind is doing great work on that uh I mean that that's I think is what gets me really excited about what like what","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3096s"},{"start_seconds":3103,"text":"is Yet to Come kind of thing yeah yeah I think that's","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3103s"},{"start_seconds":3110,"text":"uh generally uh uh when was it I think it was like half a year ago IIA IIA Drew","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3110s"},{"start_seconds":3117,"text":"this diagram which I think is the core of how the systems ideally should operate where the user communicates to the agent and the only goal of the agent","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3117s"},{"start_seconds":3125,"text":"is to create a symbolic uh representation of what the user wants not the solution yet just you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3125s"},{"start_seconds":3133,"text":"know rephrasing what the user wants in a symbolic space and then it communicates it back using not an llm but a uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3133s"},{"start_seconds":3141,"text":"deterministic you know like symbolic to natural language conversion right so it tells to the user hey this is uh what I think you want uh translated","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3141s"},{"start_seconds":3151,"text":"from my language to your uh language in a deterministic fashion does it sound right and then when the user says yes","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3151s"},{"start_seconds":3157,"text":"then we do synthesis in a symbolic space Guided by the model right what Deep Mind","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3157s"},{"start_seconds":3164,"text":"does what presumably strawberry is uh yeah and then uh and then when it is synthesized we know that it is correct","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3164s"},{"start_seconds":3171,"text":"for as long as the user did properly confirm our understanding right so that's uh uh then you know the API uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3171s"},{"start_seconds":3179,"text":"that is synthesized is correct and then I actually think that this is in the in the world with lot of apis which I believe myself in uh or or like you know","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3179s"},{"start_seconds":3188,"text":"big Global name spaces uh that's that's very important right because uh if you have 10,000 apis built by different","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3188s"},{"start_seconds":3196,"text":"people you definitely want to be sure that they actually doing what you want uh and not doing it you know nine times out of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3196s"},{"start_seconds":3205,"text":"10 yeah so that's that's the exciting uh I I think that's how it should be","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3205s"},{"start_seconds":3212,"text":"going forward I mean verifying that these each of these apis do exactly what the user","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3212s"},{"start_seconds":3220,"text":"intends to do is very important and completely agree with it without a doubt also spinning off of that is that as you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3220s"},{"start_seconds":3226,"text":"guys open the Asian Marketplace um most of those of course as you talked about there are your Technologies and triy","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3226s"},{"start_seconds":3234,"text":"outs but most of those Asian would be connected to certain different apis and for that Marketplace to work seamlessly","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3234s"},{"start_seconds":3243,"text":"as a platform um wouldn't you need a protocol to actually integrate those apis seamlessly within your Marketplace","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3243s"},{"start_seconds":3251,"text":"as well when the users actually picks it up from there and curious to see if you guys have any thoughts on how the users will pick up those agents and try to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3251s"},{"start_seconds":3259,"text":"adopt them within our ecosystem yeah I mean I was asking this question","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3259s"},{"start_seconds":3269,"text":"because I actually think we'll we'll just be living in this API to agent Spectrum right API to to like full llm","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3269s"},{"start_seconds":3277,"text":"spectrum Right Where You agents in a way are just constrained llms which you know expect some type of input or can work","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3277s"},{"start_seconds":3286,"text":"with some kind of input themselves can use a bunch of apis or other agents right and then you expect the output in","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3286s"},{"start_seconds":3294,"text":"certain format right which again API are kind of that just like you don't expect them to be as non-deterministic as","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3294s"},{"start_seconds":3303,"text":"agents right now but then agents are more intelligent in result right but it is a spectrum and so you kind of want actually agents to become more like apis","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3303s"},{"start_seconds":3311,"text":"where you know they predictably going to deliver results they're not going to fail they're not going to go in like on a side quest of you know building","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3311s"},{"start_seconds":3320,"text":"something that you didn't ask them to do right and so I think like how do we make agents more like apis is that's kind of","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3320s"},{"start_seconds":3328,"text":"the mindset in a way uh I wanted to kind of challenge as well because I think it will be important to understand yeah","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3328s"},{"start_seconds":3335,"text":"like you know livess of this agent's ability to perform the tasks you know how do we Benchmark them um and so","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3335s"},{"start_seconds":3342,"text":"similarly to apis you know you want to know all of this information how to use them examples Etc and so yeah to me to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3342s"},{"start_seconds":3350,"text":"me that's kind of a spectrum that uh is probably going to end up in one place like eventually right where all of the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3350s"},{"start_seconds":3356,"text":"species kind of just come together Ryan you wanted to say","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3356s"},{"start_seconds":3364,"text":"something oh yeah yeah I was just goingon to say I totally agree I think that uh yeah I mean apis in general um","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3364s"},{"start_seconds":3371,"text":"definitely extremely useful into the future and I think like more or less like the agents and how I think about it as well is like it's almost like the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3371s"},{"start_seconds":3379,"text":"glue that kind of sticks it together but you want it to be deterministic somewhat in its output and reliable that it does work and um I think for us like what we","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3379s"},{"start_seconds":3388,"text":"have seen really good results from is just adding like a Knowledge Graph associated with it so then that's like more or less like helping the traversal","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3388s"},{"start_seconds":3395,"text":"of like different agents um and then incorporating those into a singular conversation which then talk to and communicate to multiple different apis","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3395s"},{"start_seconds":3404,"text":"or contracts I think that like those types of like supporting pieces of infrastructure around that system is like","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3404s"},{"start_seconds":3412,"text":"very um helpful as far as like what will probably come next and then probably being able to like Traverse like an ideate between","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3412s"},{"start_seconds":3421,"text":"different possibilities is also helpful like for us like kind of the end state where we're going to is something like um every Wednesday rebalance my","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3421s"},{"start_seconds":3428,"text":"portfolio for the best yield then we can Traverse across the knowledge graph say okay super form ichy and something else","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3428s"},{"start_seconds":3436,"text":"come in mind for maybe Moro for good yield protocols we can say hey here's the three results which ones do you like","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3436s"},{"start_seconds":3443,"text":"um and then they say Okay I want to go with super form on you know arbitrum or something like that it's like okay you kind of tap that and then we're able to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3443s"},{"start_seconds":3452,"text":"like more or less Traverse and incorporate and and bring that in um but we're really the agent is like the the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3452s"},{"start_seconds":3461,"text":"spectrum of what they're actually looking at and doing and reasoning becomes extremely extremely narrow and we're more or less like I always think about it is like we're almost trying to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3461s"},{"start_seconds":3469,"text":"supply employees with a lot of different bits of information in which they can be successful uh so I think that's a lot","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3469s"},{"start_seconds":3478,"text":"better of a end result versus like a just saying hey one two three go then they're just trying to like identify what do I do I have 50 options which one","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3478s"},{"start_seconds":3487,"text":"do I grab and then trying to just pick and choose kind of thing so","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3487s"},{"start_seconds":3503,"text":"yeah no no I was I was saying we like one and a half minutes away from from the end uh I don't know if we have any","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3503s"},{"start_seconds":3510,"text":"final interesting thoughts I don't well I I think the yeah to to wrap","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3510s"},{"start_seconds":3518,"text":"it up maybe um I'm curious kind of just in in in a couple sentences uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3518s"},{"start_seconds":3525,"text":"from Yuan and Ryan um how do you see let's say like your product is fully rolled out you know people are using it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3525s"},{"start_seconds":3533,"text":"um what would be the most like impactful change right uh for for the user like what is the 10x better experience that","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3533s"},{"start_seconds":3541,"text":"they cannot get right","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3541s"},{"start_seconds":3544,"text":"now uh I I'll go first um actually as you may have noticed from our project is that we actually create a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3544s"},{"start_seconds":3553,"text":"lot of demand from the web two side and actually try to solve what web two has to um has been really thinking about for the past few years which is to really","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3553s"},{"start_seconds":3562,"text":"democratize the development so from our end um providing a conver conversational panel and actually being able to just","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3562s"},{"start_seconds":3570,"text":"plan build and ship out your backend just within a few seconds and actually be able to just deploy it and instantiate it with just a click of a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3570s"},{"start_seconds":3578,"text":"button I believe it's like an end to-end automation that they haven't really seen uh in the past and at the same time it","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3578s"},{"start_seconds":3585,"text":"really brings a lot of freedom and democratization to not just less season developers but at the same time uh non-technical users as well so from","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3585s"},{"start_seconds":3594,"text":"there on we're actually excited to see how users actually use our product and how it kind of Fosters a little bit more","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3594s"},{"start_seconds":3601,"text":"collaboration within that company yeah and then on my side uh really what we're trying to do from a","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3601s"},{"start_seconds":3609,"text":"very high level is just provide like a Visa likee experience to the masses for web 3 so like I I don't think that you","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3609s"},{"start_seconds":3617,"text":"should have to know about the given blockchain or even the protocol that you're necessarily implementing or working with in order for a user to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3617s"},{"start_seconds":3624,"text":"actually use the space I think uh like our our old saying when we first started the company was uh we're building crypto for grandmas and I think","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3624s"},{"start_seconds":3633,"text":"we've always been kind of true to that ideal which is we're just trying to make this really usable for the masses where people can actually go in and if they","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3633s"},{"start_seconds":3641,"text":"want a you know 50% yield on some kind of like LP position um on a deck they can do that but they don't have to","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3641s"},{"start_seconds":3648,"text":"understand you know how do I onramp bring this money in how do what's arbitrum versus optimism versus you know whatever","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3648s"},{"start_seconds":3656,"text":"like that that just seems kind of ridiculous to me a little bit sometimes it's very cool from a technical side but I think from a user side uh that has","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3656s"},{"start_seconds":3664,"text":"always been a huge uh problem I think for the industry so from a high level what I would want to do is is just provide the visel like experience to the","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3664s"},{"start_seconds":3673,"text":"user where they don't even have to really think about it it's like a oneclick or a swipe type system yeah we call it chain abstraction","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3673s"},{"start_seconds":3682,"text":"for for web three people for for web two people we just don't call it they just should use it and not know about it so","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3682s"},{"start_seconds":3688,"text":"yeah great to have you both here hope it was engaging entertaining episode for everyone we'll see you next week and uh","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3688s"},{"start_seconds":3697,"text":"have a good week thanks for","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3697s"},{"start_seconds":3716,"text":"f","url":"https://www.youtube.com/watch?v=3hQAK2xt7RI&t=3716s"}],"youtube_url":"https://youtube.com/watch?v=3hQAK2xt7RI","url":"https://near-pedia.xyz/videos/3hQAK2xt7RI"}