Prof. Angel Marchev: An AI Mentor's Take | ETHSofia

[Music] it's a pleasure for me to introduce Professor Angel marf profess together with the B Hope Foundation he is offering a special Bounty for everybody for every team that is integrating AI elements in their project yeah pay attention uh actually that's a bounty we haven't even announced yet so you're the first to know it was the second on the second s here yeah so uh please you have the floor enjoy thank you uh first of all what an audience what an audience uh you had to collect them from the corridors to come and see me but don't worry this is not the first time today I'm coming from class it happens all the time there so it's uh it's something that I'm used to this is one second of all uh I'm supposed to be the AI Mentor not only me there are two of us we'll see if s comes comes over in the today or tomorrow but he comes on Saturday for sure uh how many of you do think they have some expertise in AI okay one and a half right okay uh this is good uh it seems like I have a niche to to cover them uh if everybody was uh was uh supposed to be uh expert here I'm probably not uh not in the right place anyways so my talk first of all is going to be not so Tech talk uh I have seen some marvelous stuff here a great Workshop uh couple of recordings back um and a lot of technical details I'm going to be more on the creative side trying to sort of put you in the mood to do some something AI uh and of course yes the way to put you in the mood is is like this and by the way bit hope uh you you know V raise your hand who knows V drama okay not not too many okay so actually kudos to him it's by the way V dram is one of the main organizers of the whole whole whole thing not only the hackaton but the conference so kudos to him to try and um put literally effort and and money uh in the idea of uh putting AI uh in the in the hackaton so yes there is a special Bounty that is uh uh for implementing it is splitable Bounty as far as I know so it's not one one off it is splitable uh and um it is for implementing AI by the way what is AI uh it could be so so many things that I'm most probably sure that three4 of the things that you're going to automate are going to pass the criteria for being an AI simply because well one tendency is that uh AI lately is code everything you know a software calculate something outputs you know a number okay my AI calculated something or whatever so this is one and the other thing so sort of two too White and the other tendency is to have it too narrow AI is only a large language model and that's that and and sort of doting in between so um I'm here to actually today to tell you that there are many automations that you can do and many many more that I will not be even thinking about right now that could be AI uh and of course not going into theoretical s here I would like to say that we're most probably going to be a very to go into a very narrow type of error AI anyhow about me so long story short I'm a nerd everything here says basically this uh and I have uh I have uh plenty of things that I'm interested in um mainly probably this is how I I got started I come from Finance basically uh and if you work in finance long enough to end up in a portfolio management or autonomous trading or something like this you all of a sudden become a data scientist so it's only natur natural to go on uh all right so I'm going to skip this I have done plenty of projects and this SL actually is here to tell you that I'm uh I I'm going to show you some ideas today but I'm up for for seeing and discussing with you whatever other ideas there are I mean I've been doing all sorts of things including one of my latest affiliation uh we started an ie block Bulgaria blockchain group uh this year so I'm a co-founder there as well other than this you can see that we've been doing quite a lot with some uh interesting and maybe not so interesting people uh all of these are have been some sort of out ation some sort of not necessar only machine learning projects but some sort of automations recommendations and so on okay now the cool kits this is kind of uh the current state of affairs in uh most it is one cool kit is the AI the other cool kit is the crypto they rarely meet although today I have even the previous presentation I can see that uh uh it is on the way or or it's already you know the meeting or the date is already happening but anyway this is what I'm here for I'm trying to promote ideas to meet these two and when I say meet I like a very how to say it's I I don't want to to use sexy but it is very sexy notion of when ideas have sex so I want you to make your ideas have sex with each other so I want to make these two kids okay uh okay it's modern times after all but you know you got the idea so uh I'm just showing you some examples obviously I just told them like couple of minutes ago when I'm I was making my presentation so trying to looking for uh fruitful combination this is what we are looking for this is what I'm trying to promote here there have been so many great examples of uh two different ideas who seeming which seemingly who okay which seemingly don't have necessar anything in common or not not necessarily too much too many things in common and end it up in a great application or great product so uh this is my pitch for today thank you no I'm just kidding I have four more sides uh okay so when ideas have sex and let's see what ideas okay when uh we started speaking about the conference like uh probably it's not an year but probably good eight months ago and V asked me uh to think about ideas what sort of AI and crypto SL blockchain things we can we can do there are several things that came on of the top of my M uh mind and uh actually this is what I'm going to present in the next couple of slides [Music] but actually the lesson here is that there are plenty many more and you all of us would have your own individual background to result in a different uh um idea sex so idea sex number one and let's I would like to remind you that I'm speaking solely from the AI Mentor point of view so I'm I I know the value here but I'm not sure that there this is the best way to to approach uh uh I mean people coming from the blockchain community might have uh different View and different background and different ideas so this is what I'm promoting so one idea it has already been around at least in a in in papers which not doesn't mean necessary in applications is have some sort of Federated learning over blockchain chain first of all what is Federated learning basically it's a machine learning where you have uh you you keep your decentralized data source with you and you train your local model on your local data and then should things happen or should things require such thing you combine several local models and you get a global model this would be the so-called centralized Federal Federated learning like this or you combine only the ones that are relevant so let's say near neighborhood of models and you get a combined Federated uh model that is relevant for the these subcases which in these data sets that are here so or there on in the neighborhood so this is has already been working uh to tell the truth from my point of view this is a a little bit of a push a push to look for a way to to make sex of ideas I mean it's not necessarily the sort of the biggest application the the the the killer up uh the thing here though because this would be from machine learning point stand point of view this would be very slow I mean this would be impracticably slow and we are striving towards the other the other direction right but um there are there is value here because you keep your data private for your local model so you you you export the model not the data and this uh in itself has um various use cases that are um in many cases you cannot use the data publicly in many cases you don't have license and in many cases um you're only interested in the relations so the model not the data itself so okay this is it and I have I I'm I will find a way to share my presentation at the end I have put some links here so if you're interested in reading the paper for example you can click and see and here is a paper sometimes it's something something else okay so this is ideas sext number one for today uh idea sex number two this has been always been my uh sort of curiosity the first time I heard about smart contracts was back in the year 2013 where boves made okay organized his uh famous currently already now famous conference in Sophia uh and the guys from ethereum actually they were just coming they they didn't have their product yet they just had their concept yet and they came over and present it the ideas and ever since then that moment I had this fascination with the smart contract because from a point of view of you know a computer science this is basically a conditional function but being so such it gives you a lot of useful ways to automate things because it's a trigger right it's a you put you you feed data you feed signal you get result basically it is uh and the result might be nothing at all which is also resol by uh so develop a smart contract functions in an automated way and I call this smart contracts automation it's not only me that I call it of course uh so here the the the barrier the hard part would be that smart contracts of course can cannot AO execute so obviously it will have to be in order to be uh to fulfill the need for an AI and the need for an AI we always have when we have to deal with some sort of a complex problem otherwise we don't need AI to to begin with so in order to do some sort of automation it has to to become a fairly complex system of triggers as an expansion of a service so it cannot it cannot work on its own it I mean only the on its own smart contracts it has to be as an expansion as a part of a a different service by the way um what was the the name of the startup that they just uh presented uh second or third of their challenge here Second Challenge was exactly this smart contract automation okay so I have linked here a a page it's a startup as well but they have fairly um fairly um detailed information on the on this idea so go there read maybe we can discuss it so I'm going to be here today tomorrow on Saturday and if not you can you can find me online for in many many other ways so we can talk uh I have probably some ideas probably you have better okay so this was second idea sex number three this is kind of in my wheelhouse this is what I'm really been doing for long long long years this is autonomous crypto Trader so we have we have prepared this case previously it uh I mean I'm I'm putting the link here for the case not because I expect that some of you would be interested in fulfilling the case exactly but the case itself has a lot of uh I mean it is more of a paper than a case uh it's about in printing P pages is about 20 Pages probably long and most of it is instructions definitions um mathematics how to do it how to achieve it but literally or basically it has actually two sides to the problem uh first one is how do you predict how do you predict your your um cryptocurrency prices actually it doesn't need to be prices it could be um returns or compound returns and which is uh if you take the price and take First uh first difference and you have of course return a compound return uh but anyways you need to have some sort of a viable prediction uh System model if you wish and then once you have this you will have to feed it into some sort of autonomous Trader uh of course this alone is something that has been done already in in many different ways it could be a robot a trading robot it could be uh automated portfolio system it could be uh some sort of a Arbitrage system it could be many many different things here but anyways these are the two the two levels the two steps that you have to go about it and of course don't forget the uh 300 or 3,000 pound gorilla in the house and this is it this is directly comparison of uh the volatility of uh fiat currency and cryptocurrency uh by the way I can see here this should be switched so it is to the to the second the third so basically one and something order of magnitude difference uh so here this is basically euro dollar is the blue and the other is the Bitcoin dollar you can see something that we always all of us know by the way but let's not forget this is kind of the big the big problem here okay idea sex number four it wouldn't be it wouldn't be an AI talk if I don't mention something like this something like a large language model otherwise I would have not followed the trends right uh originally I have put only rack for crypto knowledge so basically I was expecting that you either index uh large uh documentation knowledge base whatever you call it uh on crypto and blockchain and then you can use it and you can use R for uh uh coming up with um generative ideas but since there there used to be a DSX 5 originally now they dropped out of the it was a start of which uh was uh focusing on uh AI agents uh but I like the idea of Agents uh so uh I think it could become a multi-agent r for crypto crypto knowledge uh we just got with my PhD student Sero we just got our paper published on the matter so it's uh it's uh how to say uh very hot topic for for us uh but I have decided to put instead of a paper this time to put a uh a lecture a serus lecture on our summer school up sorry uh uh summer school um which we have every every summer and in this lecture it's a prolonged explanation about uh all the many Tech technical details and application details it's about an hour and something hour and 15 minutes or so so probably this is the source that I would most probably uh refer for you and idx number five and number six and number seven are up to you to F figure out right so this was my contribution today uh don't forget o sorry There Goes My trick right okay don't forget the Bounty and don't forget the $300 3,000 pound gorilla in the room that whatever bount you get in Fiat is worth probably nothing coming in the next several years right okay this is it uh thank you for listening you've been great audience some of you even smiled thank you [Applause] thank you very much do we have questions yeah let's go anybody I guess not okay oh yeah go ahead go ahead H you have one my question is uh you showed us a chart of uh AI trading cryptocurrency yeah yeah and uh I just want to learn more about that because uh I think that if you ask me the only AI strategy for trading cryptocurrency Is MeMe coins you know and you just need to know like you're the first one that buys that you win every other strategy doesn't make too much sense okay well even this even this is not shate strategy you need to figure out which is the right asset right so you will need to have some nonetheless predictive things in your system to predict this and then you nonetheless will have something to execute traes nonetheless but anyways here I have uh this is the the case that we've been running uh and there is plenty of more technical details here which you can uh look around uh but then again this is not uh necessary again this is not a necessary let's say algorithm but this is uh more of an idea if you think you can uh figure out uh other approach let's go and I'm I will be very happy to to to help out because first of all this is really down my wheelhouse this is what I like to do and second of all uh if it works boy if it works then uh we should get rich right okay exactly okay thank you for the question yes one more here there you go uh I have a question about the second idea uh which was related to automations of smart contract so uh if I get that right you mean uh we can combine it with AI in the cases that um maybe the condition is too complex because as much as I know chain link has automations for time based or condition based cases so first of all uh I'm trying to change the the wording that you just asked for me the system that it might be the stupidest way of uh 1,000 uh million if then rules if it is a an automation then we should count it as an automated system right as an AI system this is on theoretical part so this is what I'm trying to say not connect it to but make a AI make an automation but then again you're right uh I mean this is just a remark here uh you're right there there might be different ways to approach you said time uh it could be what I'm most seeing here is more of a uh not time triggered but action triggered this is what I think it should be and it should be a fairly large complex system of many triggers it it it's not going to work with just several uh in order to become um really smart or um expansion to a Smart Service uh thank you and a little bit of a broader question because I'm not as well uh informed um what do you think about the capabilities of watch language models currently are they going to Surge up or they reach the plateau State how much time do we have so uh I'll tell you what large language models uh long story short are not the thing they are not the guy they are something that we like to say in Bulgarian they predict this is what they do they predict the next okay the next token the next part of the word even not the next word and they are excellent at this they are excellent at predicting what would be the next coming part of the next work uh this doesn't mean that they have any conscious of course this doesn't mean that they have any reasoning capabilities which has been proven already uh the paper is just the other day by guys um and one thing that you probably haven't thought about it probably you knew it somehow but I would like to to make it explicit to understand it when you have your dialogue how many of you have used chat something of large language model CH GPT or whatever okay uh all the others are lias of course but uh but when you have this dialogue session you feel like you are putting your question you receive your answer you're putting your question again or whatever additional um uh additional thing to ask you this and you get an answer with the full context and you feel like this guy understands you right uh by the way I purposely put anthropocentric uh uh notion here uh in reality what happen is that the large language model uh gets your prompt reads everything before your prompt uh calculates computes the answer puts it back to the to the web server service to print it out to you and dies and then the next time you push enter it is a totally different large language model iteration which reads everything up to this moment you know computes the answer and then dies and this is totally not the same thing this is why and maybe I have to expand here this is why the agents by Microsoft uh this is out to out to something out to agent I probably was the the paper this is why it's a good idea because what you have here is one AI large one lar language model uh puts as an output task for the other and dies but the task lives on so they die all all the time every everybody does their own job their own thing returns output which in sense triggers or wakes up a new AA and they sort of sort of pass to each other tasks all the time and this is you know you minimum need two of these to to to have uh to have the ability to have a more complex problems to be solved for example s if you're curious to see this talk seru is actually giving um uh a large database and gives uh some sort of a research question and then all these guys work like magic they uh somebody plans so plans the the the the to-do list somebody goes through the to-do list and uh does research somebody else actually reads whatever query we have to the database and and makes an SQL makes it into SQL somebody else is a debugger who tries to check for SQL uh for the for errors in the SQL code and so on and so forth so it when it becomes a system it becomes a little bit more uh smarter but nonetheless it's still very very far off in order to make this uh uh you know a real AI a general Ai and and in this sense for me generative AI will not be the thing that will should not be the the thing uh nobody knows for sure should be should not be the thing to really break the AI barrier if of of sense I think something like evolutionary calculations is the thing that uh should be uh you know as has been predicted by our touring in 1950 so so yes long story short but yeah thank you you're welcome thank you actually we can use multi-agent systems not only for knowledge but also for trading for preventing liquidation stuff like that any more questions oh okay there you go so uh further discuss the question about the information that we feed in AI uh if if I get it so the problem is the large uh set of information that AI is Fe so it cannot properly distinguish answer sometimes or correctly no I don't think this is the we are speaking about fairly not fairly large very large models with uh 117 billion parameters if in terms of neural network this is the biggest neural network that we have made at least the one that we know of of course so it's not about this it is about uh the fact that um [Music] um I mean every large language model finally learns on whatever we feed it right and say the largest has learned from everything on the internet everything on the internet has been created by humans or in even worse some of it has been being created by other large language models which means we we run into another problem with lack of data because it turns out that the large language models do not uh uh train well on Sy synthesized data by the way close of brackets so uh and it turns out into my into a lecture of mine I like it uh so uh um no large language model any large language model knows maximum as much as the Humanity right because it's this is what it has been uh trained on so there is a huge there is a good joke if you remember this uh uh movie AI I Robot I Robot do you remember this one it's 20 years old movie already and uh Will Smith being the sort of anti- AI um character in the movie and he's like oh you cannot speaking to a robot of course oh you cannot make a symphony you cannot draw a picture or whatever and he said yes I cannot but can you this is this is what I mean so a large language model seems like he smarter than you because knows or at least can produce things that you don't know and it you think oh it's smarter than me but then again it's not smarter than than Humanity it cannot surpass the barrier of creativity cannot create new ideas new information it it looks like the gener AI looks like this but in in fact it is generating on on top of whatever it has been trained and by definition this is not creation right uh so uh I'm not sure if should we should we continue in this direction because it's going to take a lot a while but uh I'm I'm here tonight tomorrow next day to speak about even if you're if you're not interested in doing in your projects something in AI I'm here to speak all the this is my passion speaking I mean so thank you for the question thank you thank you very much [Music]