Redefining AI Trust With Identity, Payment, Governance, and Verification | Scott Shi - Kite

All right, thank you. Thanks for coming. So, my name is Scott. I'm co-founder and CTO of Kite. Um, Kite is build the first agenic payment blockchain.

Um, a little bit about myself. I spent last 15 years building foundational data and AI infrastructure in Bay Area in Silicon Valley. So, yeah, I spent last uh 15 years building that. So I think we are living in a very exciting area. Um what happened in last seven weeks I think is actually exceeded what happened in last 70 years for software engineering.

So half of the material actually present is actually new. If you're not developer you probably going mind-b blown by what's actually happened in last seven weeks. U but if it room out in a very high level I think we are looking into like aic first time. This is actually bigger than internet you know com bubble. This is bigger than mobile area.

This is much much bigger than anything else. Um last year I also come east here. Um I did a demo. This is actually the demo I did last year. So this is actually everything run locally.

So basically say I ask my local agent help me order a burger from McDonald's. I give I give my agent $30 and my agent will do this end to end. Remember this is actually last year you know last year you're already able to do this. It seems that it's got access deny and maybe it's internet issue. Um I yeah so I do have a video recording just in case it got access deny.

So uh you can see that uh the agent are smart enough in this case it will planning for every single step and it will go to the Uber Eats and then click through you as if you as a user here and then notice that it's published like April last year and then underlying what agents say unlike human step. So basically you see the whole information of HTML on page in a in a visual format and then trying to figure out what to click next. So agent will planning say I want to order a burger on McDonald's and then it will go there it will order a burger complete every single step for you order Uber is actually 95% success rate Uber Eats is roughly 70 percentage and then this happens last April and this year we have this cloud boat um I don't know how many of them installed the cloud build I installed them like a few weeks ago I'm really amazed by it performance u this is actually I asked a cloud boat help me do some web coding some change image to you know image to video in this case notice that cloudboard is actually not integrate with any of the image to video automation so what it do is that first that hey I need to do in image to video conversion it will look up on internet figure out there is four vendors here provide image to video conver you know conversion and then it asked me to register in one of the vendor and give a API key to my cloud I just did give a API key to my cloudboard and now cloudboard has this skill which learn on the fly to generate image to video and give me this two image to video generation which I found pretty amazing. So what actually happened here is that you know this all this is revolutionary what actually happens all within last five weeks to six weeks like if you're comparing to early last last year you August last year if you look at the models here it's still can do very small stuff you know it can solve like a 30% your coding problem without going any error but fast forward to December when the chip 5.2 two and Opus 4 4.

5 actually come it becomes a infliction inflection point because now you actually have engineering you know like you have coding agent doing much better engineering than senior engineering itself and then so this is actually a project built by TIDB completely use AI to revamp the front end to instead of use my SQL as a SQL parser they completely rebuild use postgress as SQL parser it's amazing amount of work it's roughly 10 percentage of total work they accumulate over more than 10 years and then completed by one engineer in less than two weeks like original SQL front end pers took like a seven engineer a year and a half to build this is not stand alone if you look at anthropotic posted like a week ago less than two weeks ago they basically have the opus rewrote C compiler use rust less than like less than in one day and spend less than $10k so now you can actually completely rewrote the compiler you know like a like a fundamental infrastructure code use agent with very very little human intervention in this case. So the model become a better developer. So I think I said any software developer you should think about how do I actually design a evaluation system? How do I should I design a test case? What should I have for engineering taste is actually do the concrete coding yourself because agent are coding faster than you do.

you know someone show coding at monday.com in less than one week in this case like a customized monday.com just for your team in this case so um this is a very interesting you know saying AI is hitting a wall but this you know you you if you're growing like exponential speed you're probably not hitting a wall and this me is one of the very meaningful um test um published like mid last year it means that how long the agent can execute with 50% % accuracy rate u when I do the demo last year in stammer I think it's uh 14 minutes and fast forward today it's roughly 7 hours in this case and it's growing still at exponential speed notice that the latest uh metric on this is GPT 5.2 if we look at the next one what happens here is that you know based on the virtual evaluation GPT 5.2 to only has 35 success rate.

It's the last one in this list which released December 18. But if you look at the one released last week, you know, GPT 5.3, it's already reached 90%age. So the performance even for last two months at tripled, you know, comparing to the success rate for this next JS coding uh coding generation tripled in less than two months. So this actually brings a you know fundamental change you know if you're coding if you do the coding now actually generate code is not the hardest thing it's like what do you want to do with the code what do you want to do to with the AI is actually the most important thing so this is actually the model side you know there is tons of the impressive you know breakthrough on the model but there is also tons of the good practice about engineering everyone hear about cloud code what cloud code is actually not improve on the model level but what they did is actually improvement on the engineering level first is that it actually completely give up this MCP in this case because MCP is introduced tons of things agent don't like agent developer agent like execute code agent to like read MD file markdown file because it's intent for developer it's intent to um intent to agent in this case so so what open cloud did is that is actually use the file system and skill.

md in this case It's actually replace rack. You know, if you still have someone doing the rack, you should, you know, re, you know, re-evaluate if rack is the right solution for you because now the foundation model becomes so smart. You don't want to outsmart your foundation model. You should just give your foundation model like MD file, easy to use skills and let your foundation model figure out what to do next. So the rack is probably going to, you know, go to, you know, historical dusting.

Then you have MCP to you know, be replaced by skills. The next you know if I ask me to do make a prediction I predict like prompt pipeline engineering will be history and then lchan will be history and workflow will be history. So everything is moving at exponential speed in this case and then we don't know where is it end at least for next three months everything is changing fundamentally different and then we say another thing when I did a demo um you know when I showed the demo like last year last April demo um the agent is reading screenshot HTML in this case so it has to figure out all those information all those information from this image from this image like what's actually going on here and then You see that the agency much more information than than this information. But um what Claudra um open is that offer is actually provide every single HTML convert to a markdown file. In this case now you actually have a agent native agent native u format you can read to make it much easier to read to make it much easier to read and then uh this will actually reduce the context window for any agent to process it easily by 80% or 95 percentage so make an accuracy level much higher so you know it's already happening you're going to have a internet for agent native you're going to have agent native internet based on markdown file you have agent native browser from chrome in this case and you actually have capture just for agent release this morning you know in this case so nowadays you have to verify your agent in order to do some um posting and do things and then so I think the industry leader are calling out for API first approach as well in this case so if your platform if your service is not API first if your platform service not API first you basically don't exist for the agent because agent can only read API.

For example, when I ask agent to generate this image to video generation, the reason agent was able to complete this because there is API from runway. So agent was able to read runway API doc and they say I just need APK to convert the rest of the conversion and then it happens with oops you know this dashboard generation as well. So every single company going to become a API first company otherwise you don't exist for agent and we probably going to have agent uh marketplace as well in the future in the near future uh this is actually web coding I coded for agent marketplace so you will have ratings for agent you'll have reputation for agent uh this is actually very align with ERC 804 on this proposal for the agent ID and agent reputation okay if you look at very high level a common pattern we see is that user going to delegate the work to agent and agent will actually interact on behalf for you it's going to be a delegation base and then the agent may interact with another agent agent may interact with Uber services over it services agent may call GPO to rental agent may call data API to provide some API services so this is a common pattern you're going to have humanbased internet to agent based internet so what Actually kite play in this case is provide a kite passport to let user fully trust your agent and delegate work to their agent because they if you ask me hey agent help me do something research t research okay you probably you know 100% you know out of 100 probably comfortable to doing it but you say hey agent help me do all the shopping stuff this is my credit card and I'm betting you the agent will spend money in machine speed at machine speed any given time in this is you probably need to replace your credit card every few days in this case because essentially credit card just one private key. If you're exposed to your agent, you basically you don't know what's agent going to do with your credit card and you don't actually know who you're going to spend your credit card at what time in this case. So that's why we need to calling for agentic first internet in this case you need ID for agent you need authorization for agent and you need payment for agent.

Why this is the issue? Because for example for the um the demo use case I run locally Uber Eats don't recognize it's actually my agent scanning through this website it's thought it's me actually doing this serving searching in this case from service provider pro perspective from serving agent perspective they need to figure out it's actually the user's command or agents command without that separation of the ID for agent owner without a ownership delegation between user and the agent it's actually not going to work in this is because the whole you know the whole existing work built based on the credits in this case and then you need to figure out what action belong to user and what action belong to agent. The second thing is that authorization will be a very sim uh important issue as well because think about this if want to use 10 different type of service I need to create API K for every single service. Imagine this with millions of user you have m times an o nightmare issue because whatever agent you build to want to serve your customer you need to ask them to register API key for Uber Eats for Uber for every single platform in this case. So O will be a very particularly hard issue in this case you don't want API key to let agent access each of the services otherwise you're going to have M times a nightmare in this case.

So that's going to be a issue you need to solve. The last part is payment. If you look at about the payment pattern between human payment and agent payment, they're very extremely different. If you look at human payment, the first thing that you need to is looking for the thing you want to shopping and check out and make payment last step. But for agent payment, most likely you're going to give delegation first.

For example, for the Uber Eats demo I did is that I have $30. Buy me a burger. Figure me out something. I want to arrange a birthday party. buy me a birthday cake, you know, figure it out in this case.

So, we want to delegation first. If you look at the price discovery, you know, the the who initialize, who discover the project um product, who discover the price, who make a comparation between different vendors and different service provider, it's all different entity. The payment pattern is different as well. Like for human payment, you're going to have the Swiss bot is $3, you know, $10 to $10,000 because you need to pay credit card company three percentage fee plus 30 cents per transaction. So in that case, it makes like a percent spending almost infeasible.

But for agent most likely going to call one API pay one cents or few cents in this case. So the payment pattern is extremely different. And the last that the data structure going to be different for existing human payment is oneoff payment. I want to buy a birthday cake. You know, I went to this website and buy a birthday cake.

But for agent, it most likely you're going to say, "Hey, help me arrange a birthday park. Invite all 50 my friends, book a venue here." So the agent will try different API service, you know, comparing different birthday cake and then looking for a place to rental. It will do this step by step. At end of day, you have intent to invoice to exclusion to receive.

You may have a millions invoice just because you call you know different types of API service you know looking for weather looking for venue to rent inviting 15 friends in this case interact with their agent to make sure the time work for them so that will happen so kite provide uh so you know actually skip it this is actually what developer can build on top of us so kite provide two two type of service just for agent payment AI payment so underlying kite has kite at layer 1 blockchain. So this provides zero fee in this case. So you don't need to buy Kai token. You know my my design philosophy as a CTO of the company is that even my drunk grand drunk grandma can use it in this case at any day in this case. So so ideal use flow is that my drunk grandma comes here I want to buy a birthday cake.

He deposit a $100 on PayPal and then in this case and say hey u say to the open cloud agency say help me buy a birthday cake I want to deliver tomorrow. So the open cloud agent will go through the uh kite agent passport login to different website could be Amazon website it could be like a birthday kit website it will completely generate this policy envelope in this case is a $100 upper limit uh I need to buy this in 24 hours u this has this limitation in this case you know some restriction on this case it will comparing different vendor make decision make purchase at end of day and then you have intent match to invoice in match to receip and then in this case end to end traceable and audit ball. So underlying um it will it will happen on Kite blockchain with zero gas fee and u you know you can pay in stable coin you don't need to expose to any of the token just PayPal in PayPal out okay so one thing make us fundamentally different than any of the existing solution you know you if you look at our website you found stripe is rolling out a crypto solution in this case you know Coinbase released like a a gentic wallet uh this week earlier this week one thing we found is Fundamental difference is that kite is actually built for a three tiers. So the uh kite passport is fundamental different. Kite is actually built for three tiers of ID system.

So every user has a ID. Every agent will have ID. For example, my cloud agent will have a ID. My CHBD agent will have ID. It has own reputation as well.

But I don't expose this money to my agent directly. Instead, it will generate a one-time session key. The session key is actually delegation first. So it's like a $100 good for 24 hours just make a birthday cake purchase and then in this case kite passport will guarantee you in this case like your spending will not exceed limit and your spending is actually matched to your intent in this case and to end trace to your intent. Okay.

So um we uh this is belowing our launching partners and if you are building an agent if you're building a service for agent come reach out to us and then we have u you know we have several folks here uh in this case if you're looking for a partner come to reach out to us as well. So yeah, so little bit introduction about Kai team. So uh I'm co-founder and CTO and my co-founder Chi was uh previously uh product manager and data brick. He was he was previously AI PhD from uh UC Berkeley. Uh we also just recruited Hydro AI from PayPal joining us.

Um he he actually joining us at this month and then he he leads the integration between PayPal and OpenAI. And we have another machine learning expert uh Yuski lead the machine learning partner. He's previously co-founder of data one the largest autoML company. We also have a blockchain leadership team. And then uh we have lay previously the hydrod near foundation and vehicle previous u ecosystem and near foundation.

We also have u hre uh founded um um sona super team in Malaysia um from Solona ecosystem. Um we have eing um previously uh do doing this manager and investment in VMware. Um we are backed by the best AI and crypto company PayPal venture coinbased general catalyst and then we also listed on the exchange. So anyway so if you have any questions you know looking for partnership looking for more question about agendic payment I think this is the future in less than five years in less than 10 years I believe in less than 10 years 95 percentage of payment will done by AI because payment is last the most tedious step if you're doing a purchase today you need to fill like 14 fields in credit card to make a purchase this not going to be scalable in this case that's the most tedious most boring step this should be fully automated and auto delegated ated to the agent. So I believe in 10 to 15 years every single purchase uh will done by like a aenic payment.

Thank you. Perfect