Coin-Operated Agents, or Our New AI Overlords on Avalanche - TOKEN2049 Dubai 2024

[Applause] hi everyone I need a lot more energy in this room than what we've seen before so I need everybody to uh to take a deep breath and think about what it will take for us to go to the moon and as a 19 a Prof someone who did a was a professor for 19 years as someone who's an expert at crypto I will tell you what will not take us to the Moon copying eth will not take us to the Moon doing what we've been doing except on somebody's laptop and calling it an L2 will not take us to the Moon ZK proofs snarks zero knowledge proofs of different kinds um uh verifiable delay functions verifiable random functions accumulators they will not take us to the Moon we need to think outside the box if we want to go somewhere and doing more of the same stuff doing more of the hexad decimal numb sending trying to educate your mom on how to set up her wallet and so forth how to uh to try to write smart contracts they will not take us anywhere I will tell you what I think is one of the interesting paths that can take smart contract programming to the masses and that's what my talk is about today it's about coin operated agents it's a provocative idea it's an outof the-box idea and it's something that we can do on Avalanche because Avalanche is a multi-chain system we are not just a single chain we have multiple different chains each and every one of which can run a different virtual machine and I will describe a virtual machine that we've been building that's at the intersection of AI and blockchains and it's not some bullshitty thing about oh we're going to remember the provenance of stuff we fed into you know it's not about making a putting a little hash on a block chain easy stuff that's not what we do I will tell you something provocative something interesting something that may or may not pan out but I need you to at least think about it and I think it's worth exploring for us to take programming to the next level just to illustrate my point how many people here heard of AI hands up if you heard of AI how many people here interacted with an AI agent almost everyone how many people here know how to write solidity much less how many people here have deployed a solidity contract so what I want to tell you about is a scheme by which the first set of people can do all of the same things as the second set of people okay so let me tell you how this works so uh chat GPT is just the beginning it's it's very exciting very cool and uh but there's so much more we can do with what what what this is allowing us to do so normally how do you write a program you go and learn solidity you then have some idea of what it is that you want to encode right you want to do something like you want to write a a landing service now you need to write a crap ton of code about managing two pools token a versus token B then you deposit collateral of token a so that you can borrow in token B Etc to make it a little simpler think about what you might have to do to write a smart contract about playing chess you would have to encode all of the rules of Chess from scratch into your code you have to write all the all the rules about ANP pasant all about Zang everything all of the niggly little details that make chess a little complicated you have to encode in there and then once you've written that code in solidity um then you have to to see you have to then deploy it it has to get compiled and you have to then check the ultimately deployed code to see if it is what you wanted to encode and if there's any discrepancy between what you wanted to do versus what you ended up encoding then there is the opportunity for a hack and that kind of programming is not easy you got to go get some auditor to go check what you did they have to ensure that uh that the code you wrote is good it captures chess as you know it but there's a much better way of interacting with blockchains these this layered approach of writing code then translating it then encoding it then deploying it that breeds errors that gives rise to all kinds of problems because there's a discrepancy between what you have in your mind versus what you wrote in code versus what got compiled to bite code versus what got executed at runtime there have been many many hacks because of that that uh that problem as you know the Dow hack which I was involved in the parity wall attack the Oracle mispricing attacks unsafe matters these all happened because of these issues in in the different layers of abstraction and because of the mismatch in different layers now how can we eliminate the semantic Gap how can we get rid of the need for programming in the first place I pa it to you that if you could just interact with a blockchain the same way you interact with chat GPT life would be quite different and that's exactly what we're building so um I'll get get i'll skip this stuff but what we're building is a new blockchain and a smart contract architecture that's based on natural language so how does this work on Avalanche we are planning to deploy a separate chain which is programmed in multiple natural languages when you want to for example create uh a chess game you simply say this transaction initiates a chess game where I play the white pieces and the AI already knows the rules of Chess you don't need to encode anything because it already uh digested all of the information that's available on the internet and knows everything there is to know about chess rules suddenly the task of programming is a lot easier because you don't have to start everything from scratch and encode everything from scratch at the same time there's no semantic Gap you're not special you don't have to learn solidity you simply say this is a game of chess or if you want to create a Lending Service you write in English this transaction starts and creates a Lending Service that manages two pools and you say it with those words and you say it in ask key text the same way I just said it you put a period at the end and then you say you know pool number one is convertible to pool number two under the following circumstances there's an oracle etc etc it's going to be a little bit of text right you know if you think about what a does or what banki does it's not that simple Le and yet it's doable and I suspect that all of you every single one of you without any additional instruction know enough English or know enough of some language where you can specify what a Lending Service acts like I bet most of you can't do this properly in solidity I bet most of you will end up having a re-entrance error if you try to do this in solidity I bet all of you will need to go through an audit if you had to do it in solidity but I bet almost all of us can do this in our own native language and uh the nice thing of course is if this ever Works what happened here um my device isn't working anymore um okay next slide please next slide please okay thank you so I want to talk to you about this coin operated agent idea next slide guys uh can we go to the next slide please thank you so it's a new kind of blockchain where there's a Genesis block and there are transactions the transactions are in free form text they need not be in English they can be in multiple languages so I specify the uh Lending Service in English but you can come in and interact with it in Arabic or you can come in and interact with it in Tagalog and you can even say things that normally would be very difficult to say I hereby give you $5,000 this is easy I can write a check but I can do the following thing in this uh particular virtual machine I can say I give you $5,000 if and only if by June 15 you have raised the $5 million you need for the movie you want to make and if you haven't raised the 5 million then the money comes back to me it's as if you took a check and you wrote something in the memo field field that is as descriptive as you want it to be and here is the kicker whatever you write in that field is executed by a trusted agent in the Sky by the AI and you know that the AI is as trustworthy as an executive of a of an estate that's as as as trustworthy as a lawyer with absolutely zero conflicts of interest that's the vision next one please so the way this works is actually fairly straightforward we plan to initiate a new chain a new subnet on Avalanche where each validator runs a standalone llm ah thank you with um where each validator runs a standalone llm there might be more than one llm you might decide well look I like this llm better it speaks more languages I like chat GPT 4 better than chat GPT 3.5 or whatever llm you like you will have a choice of llms to to pick from each validator keeps track of a mapping from addresses to token balances that's all we need to take uh keep track of so you have five tokens I have 15 tokens I hereby give you some of them Etc and as I give these tokens to you there is no bite code format it's just free form text I really literally write I space hereby space give use and I simply specify what it is that I have in my mind in a natural language and uh and there's of course gas tokens that serve as the sort of the the thing for this and um you know as I mention this transaction starts a chess game using standard chess rules where Magnus plays the white pieces the winner will take the pot of $1 million I don't need to specify any of the additional details of a chess game this transaction pays $5,000 Etc I talked about this it's free form English there is no translation to any bite code underneath this is not an agent that sits on the front side it's simply the opposite it's exactly the opposite it's an agent that sits in every validator reads the blockchain and then applies the intent of the person submitting the transaction now there's a bunch of things that need to work of course one of the things that we need to do is we need to to separate different styles of interaction different prompts so to speak so your interaction with me should be separate from my interaction with you and should be separate from the Lending Service that I built with him so there is this concept of a sequence it's kind of like a chat session so this transaction creates a new sequence this transaction Builds on the previous sequence I had where I was uh where I was donating some cash to a movie making effort for example um we need to of course have some congestion control mechanism if you don't have this if you have a gasless system you have a Dos uh a Dos Target a denial of service Target so we need some some kind of gas token and that's actually fairly straightforward we use avax for this fees are charged in full regardless of the outcome this is not like an ethereum style thing where we charge you by the thing you do we simply say look you want your transaction included well you'll have to pay a flat fee they not a flight you'll have to pay a fee and whatever happens with your transaction we will charge you the full fee and um this is a key insight and I need you guys to all understand what we're trying to do and why it's different from people that you might have heard doing the easy thing we at aval Labs with Avalanche we never do the easy stuff we're famous for bringing new science into crypto and one of the things that we're doing here is the exact opposite of translating code into bite code we're not doing that we are writing in English and we are then executing things in the cloud in the in the sky in the on the blockchain by llms no programmers what does this mean this is kind of crazy you don't have to speak any kind or understand any kind of solidity any kind of programming you don't have to know what a for Loop is you don't have to know what a while loop is you don't have to know what Boolean algebra is you simply write things the way you've always been saying things to people that you know so that means that everyone now is a programmer if you speak a language if you can hear me now and understand what I'm saying then you're quite capable of programming this system now yeah that's kind of crazy I don't know if it's good or bad so I'll be fair and I'll talk about the downsides of this approach as well there are quite a few one of the things of course is who's good at this right you know I would like to think I'm kind of okay at specifying things in English but that's because I'm a programmer U so programmers are going to be good at this I think but um but I also think that a bunch of lawyers are going to be fantastic because they're really good at using natural language to specify contracts to specify what ought to happen and making sure that the language is so bulletproof that an executive going through the same agreement will always come to the same conclusion there are cool things you can do with because there is a there is this learned agent in the sky you can appeal to its sense of common sense you can say look you know I'm not going to specify all the crazy conditions uh Whenever there are three or more parties that are that are in disagreement resolve it fairly now I don't know what fair is so I I taught classes for 19 years and uh students would come up to me and said that's not fair and I would say look you know when you get older you realize the word fair is completely undefined right it's I don't know what fair is I couldn't tell you like this is there's a rule for fair but uh but I do have some gut feeling for what fair is and I think the AI agents are getting to the point where they are realizing the Sim similar concept of of these abstract ideas so they will be able to say okay well look if there are three people who are in this agreement in this particular configuration this would be a fair outcome and uh and as I said there is no need to specify minusa there is no need to go into every single detail the way you would have to with solidity because you can appeal to the common sense to the learn knowledge of the llm agent there are some issues you need determinism so the current AI agents are not deterministic they don't always yield to the same outcome and so if you have a setup where um let's say the landing service resolves in a in a in a different state on this validator versus that validator we will have an issue so one of the core ideas or one of the technical challenges here is making sure that the llm that we use on the validators is deterministic that all sources of determin non-determinism have been exterminated this seems hard at first but it's not actually that hopeless because where there is non-determinism we can appeal to a deterministic random number generator for the source of non-determinism and have the same outcome happen even though there's some non-determinism underneath um there might be cases where it's actually difficult to understand what's going on when people interact with the same service in multiple languages I've defined a landing Service uh he came in and and interacted with it using Tagalog what's the state of the Lending Service now it might actually be difficult so that's an issue and uh that's something that we've encountered that could be a problem um is this even a good idea right are there downsides with this so I actually don't know okay I wouldn't say that this is the next thing that's going to revolutionize everything all I will say as a scientist is it's worth exploring and this is exactly what we're doing it could be the next thing that allows a billion people to use blockchains and without realizing that they're using a blockchain at the same time these agents are prone to hallucinations they might suddenly say you know I take all the money and I use it to finance the next I don't know AI robot army that could that happen it's actually kind of possible so there are some downsides with this approach um they could also say things like I award all the money to that person and why because you know they make up some hallucinated idea and that those would all be absolutely horrible outcomes and that's one of the reasons why you don't see us turning this technology and throwing over the fence and making it public yet it it's just not there but in about a year or so we expect this to be actually realizable and and the hallucination problems controlled well enough that we could begin to do basic transactions with them and uh and then of course the final thing how hard is it or how easy is it to use a natural language to define a contract I'd like to think it's actually within the C capabilities of most people but it might not be it might just be a specialized thing maybe we always need a lawyer to write out our contracts I'm not sure about that so it's but it's something that worth trying it's much easier to try this approach than for us to try teaching solidity to the masses so that's why I think um I'm really excited about this billions may be lost I I wrote here billions will be lost people will end up writing something in English the comma will be in the wrong spot the AI will interpret it you know literally and suddenly we go down a very very dangerous and bad path that can happen with this approach and yet I would pause it to you it's still far more promising than most of the other mumbo jumbo that people are feeding you about what's going to allow these systems to scale so um so I think that the llm subnet is an experiment that we simply cannot afford not to perform that's why we're squarely in the middle of this that's why we're really excited about it and uh I will leave you with the the simple message that avalanche's multi-chain platform the multi-chain architecture allows us to EXP expent with groundbreaking ideas and I hope you agree with me that this is at least if not groundbreaking it's at least worth exploring of an idea that can allow blockchains to be accessible to a much wider audience thank you so much