G-Alpha AI: How AI Agents Are Transforming Wall Street & Investment Research | Bill Sun | Deconstructed VC

when did you make your first million 6 years old 7 years old first million is through Bitcoin but then I made a big mistake are there any people around who would say don't do this like startups are hard you won't have insurance open ey initially to me is like a very good company that follow similar strategy it's funny like we we going to replace all of the analysts with machines and analysts will just have to bring coffee and your experience is amazing and initially I have bigger dreams oh let me use my math I to prove why def Works can we just accept it as magic I was one of the earliest like Bitcoin like a core researcher stuff it smells of money but yes how do you make it I'm mathematician I was a Stanford math PhD the next step is a leap of face what an impressive story hi we are Nick and Marina Davovs we run DVC which is an early stage Venture Capital firm we invest in early stage startups by repeat Founders revolutionizing the world with AI and in a free time we like to interview pfolio Founders and learn from their experience we try to deconstruct their businesses show you how they're made show what inspired them and how you can start a new company if you like content like this please subscribe and leave a comment this really helps YouTube algorithm notice us so welcome to another episode of deconstructed VC a podcast where two early stage VCS deconstruct businesses that we invest in and we try to explain how they work how they're built how they evolve because it's a process it's not you know set in stone and today we have a very interesting founder Bill

sun and is a repeat founder a very experienced person to run a company who created this out of a very peculiar place and we're going to talk a little bit about AI house because AGI House is you know well maybe not making headlines but definitely there's a lot of Gossip there's a lot of Buzz around Silicon Valley and cerebral Valley I know there's no a separate thing about what's happening inside and it's a it's a select few Elite group of founders building really cool stuff towards AGI and Bill your company G Alpha we really want to try to dissect it and and you know maybe for you that would help in finding more teammates finding more employees finding new partners for us is just general curiosity we love understanding how stuff works and we love understanding journeys of our Founders because like somehow we all in this incredible place to Valley Cal Valley building something but most of the time we could come there from different experiences and your experience is amazing maybe we can start with like what with your journey yeah yeah yeah definitely thanks so much for the invite yeah it's really exciting to be here and yeah I think this is one of the opportunity I got to really share some of the long stories starting from grad school way too like no perfect yeah we can we can start with your childhood okay so usually at this point we hear great stories of success okay early success like when when did you make your first million six years old seven years old no I think my first million is through Bitcoin and yeah it's like I get to so 11 no no yeah like 2015 and all that like a

yeah I get into Bitcoin pretty early I was one of the earliest like a Bitcoin like a core researcher stuff and I I'm mathematician I was a Stanford math PhD so so like I don't know much about coding at that time when I was in Stanford but I know math very well I did like math Olympics I was from like packing Mass legit place to train you to how to like yeah prove serums and in Stanford I was like oh I'm little bit tired of like proving serums let me find something like that really make a d to the world so I look around and apparently deep learning at a time is a rising topic and it's so much more live than traditional Mass Field but I noticed a lot of the tools I learned in the old days can be applied so 2015 that's 2014 when I first jumped to Stanford so I was at like a Princeton as a visiting scholar during last year my undergrad and in I the like the place next to Princeton they were having this statistical physics with bunch of mathematician all that and there are people talking about machine learning with random Matrix I was oh I know this oh this seems much easier than the traditional proving a serum for statistical physics and the impact is much bigger when to SML there is like basically thousand of people squeezing into the same room lessen to Joshua Beno or like Andre and talk while all the classic whole like machine learning Mass is like a yeah a 50 people like a lesson to one person proving serum so I clearly get tracked oh this is clearly the virality and also I say wow it's actually doing something

that shocked me I say at that time it's computer vision it's already like make me feel like black magic that was right after Alex net right yeah yeah yeah yeah it's Alex now is 2012 but like most of the people realize the power of deep learning slowly in like 2013 2014 so many things came out of hinton's lab and funny I was I didn't I I didn't realize for a very long time that apart from Alex Kinski who created Alex net the second name of that paper actually Ilia Suk yeah the famous co-founder of openi yeah yeah of course who nowh to be found what did he see cool so grad school you got in love with machine learning yeah and initially I have bigger dreams oh let me use my mathematics to prove why deep Learning Works of course today we all know it's almost impossible dream and even today we don't understand truly how a countet in 2012 worked like as mathematician I can pretty confidently say we are very far behind the actual practice and I truly basically give up this hope of like oh let's use mathematics to understand deep learning in 2016 I was very fortunate to be the one of the early people that worked at Google brain in its pre Prime Time 2016 as an intern I worked on question answering and I was very fortunate to work on basically attention and the sequence model so at that time it's it's before the Transformer paper and I work in a intern group that basically next to the intern group that basically Ashish is leading which which is about basically trying different architecture for machine translation so they discovered this architecture which do not need lstm a pure self

attention me that here is like oh my God let me try it out since my previous architecture that worked really well I honestly just ripped off the state of art by directional rstm code from Google's production machine translation code I was like hey let me try this code modify run it on question answering while it works and then I basically took the new idea I reimplemented and made a field modification and run self attention and wow it's even like more amazing it's like just simply by swap out the by Direction rstm and swapping the self attention I see like massive performance increase I got very excited to go back tell them like hey if you write paper about this architecture should also add in the question answering stuff but then I made a big mistake because I have like CPT and all that like my summary is ending and I was about to go back to Stanford I didn't really realize oh I should just like basically pay for Stanford to extend my C stay in Google braing be part of that experiment and be part of the paper yeah so kind of regret but like that's life well every every author mentioned on attention is is all you need paper is now a founder of a billion dollar company or new protocol for some reason I think one is is working at open AI but yeah I mean like this this gets this paper get a lot of attention yeah yeah yeah yeah there was this announcement of of a panel at GTC where every person on the attention is all you need paper from Google it was present and I think there was Jensen Juan who was interviewing them and the heading for

that was like Google retention is all you need yeah well yeah I kind of experienced the prime time Google brain but also kind of feel oh it's like a it's a very kind of bottom up research environment it's probably very hard for them to concentrate resource to do one big project at that time at that time it's like the best lab is Google brain and d m so two internal company competing but I already see like de mind move as Army it's before Alpha go right so later they use this Army approach to get to Alpa go like big research groups of people and I think open AI initially to me is like a very good company that follow similar strategy as a the mind they do like data while like open ey do like basically Starcraft and later I think open ey just i r spotted the right direction to put all the resource which is like language model and that took out a lot of people I guess have this question is why do these research organizations that are paid hundreds of millions of dollars to do their research they have enormous budgets what are the like most common application they start with is teach computers how to play video games is this the worthy cause I think it is initially I thought okay it's probably not like well us of like money when open I work on like basically da I was thinking okay they have so much resource and they could have like be more concentrated than do one thing well and that shouldn't be games right like the impact is not as big compared to what you can do elsewhere but later I realized oh it's very hard to find

those perfect simulator that you can generate infant data and just throw computer into a problem and just works and can show you the power of like scaling so I think it's kind of like technically you don't have much other options right so it's really come up with architectures right and then try to train them on real world data to solve real world problems oh I think it's not it's more like how do you get humongous amount of data when you already have humongous amount of like computer resource it's not that many options like go is a great example it has real world impact a perfect simulator it can generate infant data like say I work on finance so I spend so much time thinking can I build a gri simulator in like a finance that I can use infinite computer and then I will get a great algorithm that just work but it's very hard because in finance you cannot get a caal what if simulator say like what if say fat didn't save C value Bank on the weekend how would the work go but in like say gaming you know okay if I didn't kill that hero it's like the hero just walk around in it's a game engine that's the difference that's why game or go is kind of perfect simulator a great place to start doing AI research I'll I'll take it a little bit off course what do you think of Nvidia Isaac in that regard that is effectively simulating the real world to train robots yeah I think it's a very interesting project fun fact before my wife moved to open ey she work on robotics at Google and they were also building simulator so like we play with

Isaac and all that oh by the way she also worked at Nvidia on computer vision in I think 2017 what an awesome family you guys have where did you mean we we met back in undergrad in PAC University she was like physics gold medal and all that and I heard about her for very long and then we went to party for people that about to go to Stanford and then mat was she's not only famous but also like beautiful then just chase after her yeah what an impressive story yeah so you go back from from Google brain to to Stanford what's on your mind you've just seen this you know Sparks of excellence and maybe you can tell a little bit more about your work you've been working on Transformers at that time yes yes so I stumbled upon this great thing I noticed oh okay like it seems we shouldn't just like as mathematicians focusing on try to prove things about one layer linear like MLP or like just like we at that time haven't been able to figure out how to understand like even a two layer like kind of fully connected interet which is like back to the 80s already working so I think the Gap is so large when I observe how Google brain does Empirical research I think okay there is no hope for my generation mathematician to catch at least throughout my PhD I feel like there is no hope and I decide okay maybe if I want to do a startup or something should what why do you think you you were looking for this mathematical way to just explain clearly how multi-layer models work can we just accept it as magic as mathematician we cannot really accept

the Alchemy approach like hey I just throw a bunch of things into a pot and work we want to understand you were a physician yeah cuz Physicians are good at you know assuming things things it's just like this works this works they do not work together like yeah today when I do startup that's kind of my Approach right we do a lot of like Financial agent research like leveraging different large language model and I never really try to think okay mathematically why do this large language model generates token what happen in high dimensional space I was more trating this as a new like operating system where large language model is like my new python interpreter and I try to write something and say what happens it's a computer right it has its own RAM it has this longer memory it's it's interesting so can you give us a two sentence intro of what G Alpha is yes so to me G-Alpha AI is a company that build AI agents that can do a lot of things that traditionally human do in financial investment including like try to analyze like Millions of document as fast as possible and also like being able to combine the structure time series information with the unstructured tax information and Drive summary Drive insight and then drive action for like actual like trades so this is pretty much an artificial analyst like a hedge fund or an IB analyst that is somehow able to read millions of documents yes yes we want to basically but cannot bring coffee no you cannot oh that's a shame yeah but we have other people working on it right okay so it's funny like we we're going to replace all of the analysts with machines and

analysts will just have to bring coffee to the machines yeah yeah maybe they were just or another fuel yeah or another fuel electricity like in the metrix so the idea of generative Alpha was born in AI house right or it was before and then you joined AI house so I think I've been thinking about the idea of how to use the kind of language model style stuff to understand the financial information back when probably 2017 it's like my first real Quant Finance job in 72 and which what's a let let's dive into it what's a Quant Finance job I think a lot of people are super interesting because they it smells of money but how do you make them so basically in 98 8s W St start to have large computers and start have like basically digital information and digital like other books and all that and then people like Renaissance Dall is trying to get a smart mathematician computer scientist physicist to work together to basically automate what the previous generation like Trader in a trading floor do and that is stly involved to basically automate the kind of more simple rule base based and the linear regression based type of decision making systems and those decision making systems run fully automated and then just generate a lot of trades in a day and that becomes itself a big industry today probably occupy a major chunk of like the trading flow of entire like a global market so over time it evolved from rule based systems that basically like this happens then I do this do machine learning and then speed became king so everything started working at at basically instantaneous speeds and what is Quant trading now so I think of Quant trading

today is like a lot of basically smart mathematician statistician physicist still trying to cook up models that are based on this concept of human feature engineering like look at the data and try to understand okay can I write up a feature that describe the market and then do back test and try to find the statistical significance and then evaluate whether if I trade this strategy in history would I make money and then the next step is a Leo pH if it works in history it should work in future that step is like a pure pH the the rest of step is more scientific like proper data science right and then you have this leap phas that things worked in the past work in the future so you're Pro you're just trying to create a model that is complex enough that would try to explain how the actual economics Works in in the world and then you test it going backwards you don't really need to explain essentially your only task is to predict what happens the next minute next day and if it like happen to work and you use some like weird feature that for you it doesn't make sense but it's pretty statistically consist for lot of hat fund it's okay however today like that approach is still having a pretty significant Gap from what a good human investor like Warren Buffett ralo Soros invest right those good human investors working scenarios where there's only like probably like three five examples of similar events happened in history or even in the case like Soros there's like almost no example like happening history you can only use that principal thinking causal understanding of how the marketing mechanisms generalize outside of the training set yes

even beyond that sometimes there's no training set it's like a more reading lot of document try to summarize the information and then try to find the key drivers of things and then figure out which key driver is actually going to drive the next period of Market I think that's how human play the game so G is not a quan system it's exactly the system that is empowering humans to make this more strategic trades yes yes so the initial mission is that I have been working Quan trading for very long and I notice all the Quan firm no matter old on new is basically converging to very similar like high level approach and people are basically only competing on like the small technical details and we joke about Quan Finance as there is a massive roller coaster of like that is running and we just run in front of the roller coaster and trying to pick up some dimes that are spread on the ground that are like kind of deterministic of high statistical significance opportunity but like none of us understand how the roller coaster is moving and I'm trying to say okay for today's AI it's still pretty hard to fully understand how the roller coaster moves but I can work with basically human being and have a human in the loop approach to enable them to be much more efficient and human are in the game of understand roller coaster but they are limited by like their bandwidths of processing information their bandwidths of actually do work right like a good investor only have like limited amount of time he cannot replicate himself a thousand times cannot work like say 10,000 hours a day but in machine you can SK scale everything up

you can just basically replicate 40% of 1% stability but like just scale it up so let's let let's U move a little bit back before the G Alpha so you've been working on Quantum trading this is Quant Quant Quant Quant trading and I'm sorry you've been working on Quant Quant trading and then decided to start your own Journey as a founder yes yes so do you remember how how it happened yes so after I okay so should that Dat Back to the days at Stanford so at Stanford I was actually very interested in do startup and I founded this club called AI plus club together with Tim sh the CTO co-founder of Cresta and we were like exploring different business models and I was already basically that said okay the sector I will choose is AI Plus Finance and spend a lot of time working in different Finance industry firms like a 72 cadl and also like two of my advisor are also very experienced in Quant Finance one advisor Steven Boyd is kind of the gwp of like portfolio optimization and also currently the black rocks aiab another professor David Donal who is like super successful in terms of his research enabled a lot of like say like feature selection in previous generation Quant finance and he's like in one of the most famous Quant training shops that I don't know if I can share in public and throughout this influence I know like what I have been learning is super useful in this industry so it's I give me confidence I can probably do some contribution in that field so I spend a lot of time in that but still thinking okay can I grow a company out of this direction so when

I graduate from PhD or I have like a faculty offers I also thought about startup I was kind of stumbling around trying to figure out what to do for almost a year very different like joining The Faculty or starting a startup is just very different Lifestyles yeah but but on the other hand you you're still like trying to decide long term like do you want to invest next 10 years building a startup or next 10 is doing academic yes so basically then while I was stumble around in 2019 coover hits around and basically everything froze at that time I was all super scared of coover I was like okay I don't want to see anyone I want to just like stay at home only me people zoom and then I find okay it's probably impossible to start a company from scratch now when I cannot meet people and being a faculty while it's a like cover it it seems really not fun I don't want to be like remote teaching every day oh yeah yeah and then I was like okay maybe go back to my old business and I look into Quant Finance industry and because I did okay previously during my PhD in Quant Finance over like internship and all that I'm very fortunate to become like a portfolio manager which basically like could run your own shop in a institutional setting where they provide you the fund and help you to recruit team but you build everything technically from ground up so I got to like like GA a team have my own like basically trading strategy so-called a book in like Millennium War Quant and do you have like how much assets and management do you have at the time I

don't think I can share exact numbers but on the other billion billion yeah and yeah I'm trading mostly like us equity and the the approach is more like using a lot of machine learning to predict how the market move in future and then try to I build a basically a agent that old fashioned fixed agent that like can you share the confidence of these predictions like so I think every single prediction don't need to be super high probably like say if you have a 60% win rate is already very impressive because you accumulate the statistical significance over different stocks and then if you like make decision on 3,000 stocks every day it will give you basically a pretty good end result and stable and I do this for more than three years in the middle I basically observe like Quan Finance as a relatively fixed industry people are just drilling down all the details but like I observe machine learning moving really fast to the point when I say like gbt 3.5 come out I was like oh my God whatever I worked on in 2016 just almost become from like a single to a baby now it can actually do real things and cross the like a gap between a technology that I fun but still in lab to a technology that can directly serve consumer without even any additional like data to train and all that I was super excited and I feel like oh this is one in the lifetime opportunity that I should abandon everything and just do this startup I love this feeling where something that impactful happens that is is just a wave of brilliant people considering not even considering just making the leap of faith quitting Google's opening

eyes like everything well a lot of people quit Google and other companies and to join openi but a lot of people were like well I got to make something with it let's let's continue on your journey and and then you see the opportunity to actually make more Difference by building generative Alpha your company yes yes so I have been thinking about this idea of like hey can I make a basically new technology of AI right like different from classical machine learning to build a company that different from traditional Quant trading but also in the fact that we do Automation and we make human being like like much more efficient and do their work probably like only 70% of their quality but like 10,000 times the effort right like it like a machine in general is this kind of be that if you teach it about one thing it can just do it like endlessly on like every single stock or every single asset so it's really good for like repeated the task and I was just looking around and I think one of the major bottleneck for transitional Quant is we only work with numbers it's very hard for us to work with like a text or even those multi modality content with like a plot with like say even numbers in like say filing that is not super structured like basically we give up because clean those data is so cost ly and we have to have human engineer sitting down like writing pass turn those data into a structure data and then clean all the nons that are created by this and it's like so much work now we have new technology and I think we can literally mimic how human beings analyze

those document and the faster yes just proc it at scale so you have this idea how did you start working on that did you did you find somebody to work with you did you start alone are there any people around who would say don't do this like startups are hard like and yeah come Jo open you won't have insurance and yes better to go to openi yeah I guess about joining openi around the same time I'm thinking about making this move my wife also have a big life event so first we like know we had a baby in like January 200 23 and we're preparing for the baby pretty excited and then like Google basically laid off entire team of my wife in like basically Google robotic they call the team is called Everyday Robots basically like 200 people very strong machine learning talents robotic talents and Google say that project is too long term and doesn't really have a go of reaching business pmf and basically laid off the whole team of course that Google brain gave her offer to continue working on robotics but we were a little bit disappointed at that time and we're looking so as I was considering okay what I should do for startup I was also basically trying very active to say like hey where should my wife go it's like do round of like research on the market so our conclusion is like you either do startup yourself at that time or join openi so basically my wife join open I I yeah I could do startup good like so in the moment of uncertainty you decided to even like bigger put put bigger risk on like uncertainty and build a startup yeah I guess we have

been like working for three four years after PhD we did okay like have enough safety night in the life so it's okay just like to stumble around random things so for a year or two so we don't feel like oh our life is in crisis and it's more like the exciting opportun is coming how should you spend the valuable like time do you think it's important to have safet Tate before like you start like Journey like this when I look back now I think it's less important when I was in PhD I'm a little bit scared so I yeah as I said I was a like a co-founder with team on AI plus club we spent a lot of time doing research on different business model in AI back in the day of 2015 16 17 but Tim jump off the PHD train and basically leaving 30 year to start CA and I was like oh I have such comfortable setup in Stanford my wife is in Stanford same year PhD and I have two probably best advisor I could ever hope for it's like really comfortable setup let me enjoy the Stanford environment a bit more even if it's like a r a kind of luxury consumption time I think it's worth it so when I look back now I feel like oh I might have got too comfortable at Stanford I should have pushed myself harder and forgot about all the safety net and all that that's all like yeah but that's interesting how a lot of people who haven't made the lip yet regard this as like I got to create you know some reserves I going to feel more comfortable I won't have any you know like I won't have dental

insurance when I start my company before I raise any money that's going to be hard and then people who do make the leap always look back and say I should have done it earlier yeah yeah I think that's true maybe we're also looking at survivors and definitely Survivor bias I also think throughout the years that I work at Wall Street I also truly learned a lot about how industry works and how Enterprise work like like what other inefficiencies they face and what other organiz IAL limitation that limit them from doing something new so I can see all this more clearly and that is probably going to be very valuable for me in the long run so I I don't really regret but in terms of safety net I don't feel like I need to accumulate yeah very long let's back to a house and at that time you like met the guys or before that how how how it happened okay so AGI House is a fun subject basically during Co I was working as a Prof manager and I lived in like LA area just to avoid meeting people as yeah it's like I like cannot really meet people before I get the vaccine I was super scared and all that until basically Google calls for everyone to go back and I look at Bay Area I say okay where should we leave now and basic team tell me hey Andrej Karpathy started this fancy Manion that already moved in to on to drawing and I go over I was shocked by how fancy the Mansion is and it's kind of like a very very glorified version of Stanford dorm and so wait but you've been living there with your wife right yes yes

so the only hard part is basically convince my wife hey that's living a co- living style life in this Manion and the pay the same amount of money that we pay for like a pretty nice three bedroom apartment and it's worth it and it's okay to lose some privacy just for the final life it's like a hard sell to my life as you can imagine but I basically yeah tried very hard and get it happen and in the first year with in 2022 we didn't even figure out the house should be about AI so just a bunch of friends leaving together and we thrw a lot of crazy parties even like 400 people party with like Central Park style horses and flying cars yes like a crazy days and then by the end of the year a bunch of our hos mate was in Miami attending art basil and we're just chatting after drink like hey it re like run the house next year what should we do so we get Rocky who is now running AJ house to like say he like interested in really like spend his entire time Runing the house and we we're super excited that someone fully like being the adult and can run so that's like a success factor number one then like we're saying okay we should have a SIM for the house and then yeah basically there is a concept that everyone agree on which is like we should focus on house on AI specific AGI since all of us think okay General AI is coming and we should differentiate on traditional like Ai and disc call a house that's how in 2023 the house change name have a new branding and like stop running any party

with DJ and Fus on becoming like a real AI house that deliver high quality content and basically do good contribution to the AI Community like every two weeks we have like AI hack on that's probably like 200 300 people attend you Chang parties to hakatan yes there's still dinner parties right like every Friday yeah so in the dinner is more like we have like the 30% kind of upscale private Chef style dinner party with like basic industry s leader so in some sense for us it's to replicate the jeonis style like a discussion where one person talks about concept of like say a the future direction of like AI Enterprise how to do go to market for AI star basically with every time select the topic and invite top speakers all like like top VC to come and like discuss this so I don't consider that as like a party party but more like a kind of s yeah like fire chat style that U that is super interesting I have a question I need to know like there were rumors when we last time being ingi that Rocky actually is a Tyran is it is he a Tyran I don't think so I think rocket is the one that Mak sure the entire a house run smoothly and basically if you think about it if you a bunch of kid living in a dorm it be like slowly get messy and people each person want to throw their birthday party here and there there need to be someone who stand out and say no and kind of make people unhappy yeah I think Rock he's not a reant but he's he he's just in he's an adult yes yes I think I'm happy that

Rocky stand out 2013 and yeah he made so much effort in making sure the house runs and getting the community so it's so after Andre left openi again is he more involved with the house or I I think Andre is in a state of life that he want to probably like educate more people enter entering Ai and I'm ready I love he so good at this yeah I think he's like truly the great educator and I think that's how he use his impact and really respect what Andre is doing also I think he's like quite busy even in the first year when he run to the he basically set up the Dum pay for the house but I think we did a calculation together with Andre in 2022 he probably lived in the house for two two days he has like this massive probably like a like a massive master bedroom with like two floors and like a probably yeah like a 3,000 square ft just for a master bedroom suit and there is only one mattress not even a bad frame it's like a very very interesting setup yeah well he's he's brilliant let's let's talk a little bit more about the company so how like how many people are you now and first of all like now it's 9 months baby I would say yeah so less than less than a year and how many people officially have six people and we also have contractors and everything and we try to move fast I constantly inspired by like dami with P able to achieve like great results initially with only four people so trying to say hey let's like move fast and keep them down but also we have so many things want

to try and yeah it's like I'm always like looking for great talent both Technical and business Talent so it's like there's a constant battle in my mind okay should I grow team or should I so I'm more like taking opportunistic approach like if I run into someone ready see fit out just who are the six people we have like a strong technical people from like add cor ml in Facebook now meta right so they are in the team that is basically in charge of the 100 billion Revenue every year and they all engineering yes so majority of our companies engineering we also have business people that like Sero entrepreneur that was in YC in 2014 when Sam was leading YC and they have like a major exits so just try to like mix match like different experience so so how many people are product So currently I do most of the product work and I'm currently very actively looking for like had of product VP product person if you're a great product person come join yeah yeah I I never did product manager work at all so this start out is basically my yeah first yeah official like product work learning experience so initially we just do a lot of things cook and dirty and slowly I realize oh I should find a good designer who have good taste and instead of only a front engineer that's I could work with me to generate a crappy front and I slowly just making this learning experience so and you're still pre-launch yes so currently we are in the pH I consider more like open Beta where we put a product that everyone can come and register but we do not go and publicize and promote our

website since we want to rely on the initial kind organic kind of ear to mouth type of gross approach to test whether people truly use it because we kind of choose a probably 10 20 Sparks of people who are like W Street investors that work in this setting and let them spread to their friends and do the basically slow growth to like have internal beta test to make the product polished and soon when our product is ready I think we'll have a official launch that is for everyone now if you know where to get our product you can try on this almost just yeah out there and it's kind of currently phe so just go there and play with it I loved it I I loved it that was I think we we spoke to a few of our LPS when we were making the decision to invest and I think you met them so Constantin and trading view yes so trading View and their reasoning was actually a little different so you are saying that I'm building this incredible smart product that is going to be the infinite analyst for every person doing trading right and their motivation was different there was like imagine how many analysts are there in investment Banks and that are just spamming all of the rich people people in the world offering them different trades there are like probably hundreds of thousands of people doing just that and this product just can replace all of them yeah yeah yeah yeah yeah yeah that's I think they're basically talking about step one I'm thinking about I think probably step five I don't think we have successfully reached to the point that we can fully replace W even Junior analyst as

human being can still do so many different things we are taking the approach because of the coffee right yeah because of coffee and because like a human can pick up car and actually like ask person and human can basically do lot of indepth thinking that still today's like large language model have not reached to the level of reasoning ability so there's lot of different Bots but we basically try to replace the most boring hours of the analyst which is slowly like moving from 60% of that person's time to 7 % to 90% so the hardest part is from like 95 to 99 then to 100 when you reach 100 your product value will jump in a crazy way from just like a software to to all of a sudden the price of human but like that last mile it's just like Auto driving cars is very hard to bypass the last mile so when you were starting I think Bloomberg GPT was already announced so you were diving had first first into competition with one of the largest players in the field that will inevitably try to compete you out maneuver you was there a lot of people who would tell you like why you doing this yeah yeah yeah I think when I do found raising a lot of people do ask me about Bloomberg gbt and I spent a lot of time reading that paper and all that but I feel like there are so many different approach their approach is basically find one open source model take their C and then spend four5 million dollars pre-t train and then like f turn and do everything themselves to have a large language model to me that approach is really dangerous as a basically

if you think about open ey it's like a they set the Baseline for the entire like field right they're are open source model that are chasing that sea line but they are basically the sea line of intelligence and you can only even Bloomberg with the talent and the money can only like try to match that c line once but open eye have so many talent and the SE line is slowly raising you need to predict where the SE line goes and just build a essentially shape that get ready when the SE line get to you and they would like I think Bloomberg's approach is like try to brly chase the sea line and I think that's I think what what Bloomberg is doing with Bloomberg JP is just ultimately replacing their UI with with a chat and there are there might be prisoners of the in innovators dilemma where they cannot undermine their existing business and the problem for me is that people who professionally use Bloomberg they don't need a chat interface they are good with you know BL Bloomberg keyboard they can do everything in seconds do people still have these Bloomberg big machines I I don't know I I used those like have clients that basically say when they start a new Quant trading firms they can just leave in basement but they need toy to Bloomberg terminal to get working it's basically like a almost a software engineer need to install a terminal to start T coding yeah so it's still the bread and butter and we still have faxes like most of the offices yeah most of the offices still have fa fax machines well Bloomberg's are pretty useful but it's just you know you go two three steps ahead

of that you are automating a person using a Bloomberg terminal not just building a UI for that and I don't think lber can do that because that would undermine their existing business model I think that's one reason but also I think in finance you don't really need to out compete Bloomberg to become like super successful company right as long as in the whole value chain from processing information or like providing like a trustworthy credible data all like having algorithm that can basically give people a little bit of edge all like have a tool that do not give people any predictive Edge but make people much more efficient or like just make people's decision making process slightly more like reliable of making people's risk management slightly more systematic anything you do you can make a lot of money so like a pretty rich mind to start mining which direction you go you might stumble upon some gold and don't really need to go after Bloomberg and out compete them that is that is true any like device you have for aspiring Founders who wants to create anything like in the intersection of AI and Finance are there any like Trends you can see what would be great like products to think about if you weren't working on galpha in but you would do something in finance what would that be what what would that be yeah if I don't work on a current company I would probably work on the intersection of crypto and AI it sounds like a joke right like oh you put the two hottest thing and put them together and you raise a lot of money that's kind of people's joke about Crypt plus AI but if you think more seriously about

what AI lacks today my angle is we really need data especially like data that are not available on the internet and yeah like slowly AI company will realize they run out run out of data to either train or do in context learning right like open a probably already script most of the public open internet so we inherited humans kind of Behavior Trace over the last 20 30 years of like web one internet but when we get to like web two and mobile every data got locked by different company like Facebook by like Tik Tok and by all that and the also the web one company like Google and Amazon also learn to lock their users data so now we do not have any of those data in exposal for like AI companies to touch and that becomes major problem and if you look at business the same right like it's very hard to get touch on Bloomberg terminal data and at Quant company we spend spend millions of dollars just to buy data sets that are are just like U sometime even slightly more refined data than public available data so I think data is truly getting rally the bottom neck for AI to grow of course there's like Al computer and all that but data to me is definitely the number one and I think with the crypto incentive on the future is data or compute I think of to me there are different dimensions Bitcoin is kind of more like the call that based on proof of work which is like energy and compute right yeah and if you think about it there is no such thing as a unit measure for the value of data and I think if I not working

on ji that's probably the direction of spam we've seen lots of people running from crypto to Ai and back and forth so that's the funny thing is the funny thing is that AI does not need crypto crypto needs AI my angle that like a AI company get stuck with getting data is basically like AI need the crypto incentive mechanisms not necessarily crypto but we need a better incentive design yeah to like get people to R share their data so that's something that I think in some sense I would not say AI needs crypto but AI need incentive incentives yeah yeah yeah and in terms of crypto need AI I think today's crypto is still like a small group of people's hobby it hasn't hit the mainstream if you look at how many people have like any wallet it's still a small fraction and also Darkness currency as our friend shaped it it's like fantasy league still yeah it's a fantasy sports investment League yeah so we still need a basically a like major adoption application that make people like get crypto to hit mainstream market and I think AI is probably the one what do you think about worldcoin it's a very interesting concept and friend with some of the co-founders and yeah it's interesting concept that they are able to do this proof of human likelihood yeah the only concern I have which is why personally I did not go and collect the airdrop is I feel like in the future my eyeball retina data might be super valuable if I like use it to lock my safe or something right retina is like such a like unique Biometrics even if everything is super private I might still feel the danger of like sharing

it yeah that makes me also feel quite uncomfortable their their Biometrics wait but I think it's if if it succeeds that might be one of the cases where AI needs crypto because yes but to me I think there are other ways that you can solve this proof of uniqueness of a human humanhood right I I'm more aligned with vitalics Ang go like you need a so bound token that is a like bound to yourself and for me it should be bound to your digital Behavior like human have lived on the internet for the last 20 30 years and every single person that have at least took some educations probably having a phone and have some like internet application so I think there are enough kind of history that makes it super hard to fake a humanhood and that is already something that can be an alternative to the government qyc ID passport system to give people like a unique uniqueness yeah and there's a lot of applications for that let's go back to your company a little bit so tell me the story of your first check how did you how did you fundraise the first the first money for the company okay that's a a hard question and it's been a while and I basically try to talk to a lot of people I know in the old days so back when I was doing Stanford PhD and thinking about startup so I think initially the found reading is mostly like hey I have this idea and we don't have much but like yeah here is a story I interested and yeah got the there is like a multiple checks that probably arrive around same time it's mostly just like hey looks like a

good B and yeah like we have been known for a while and it seems pretty smart go here the chat go figure out so that there was a people that knew you before yeah yeah yeah yeah okay how much roughly did you raise in the like in the first few months so I raised around 4 million in the first few months so that's basically that's a pretty good yeah most of the yeah that's more like a pre pre pre money and yeah that time we don't really have product proed right just to build a proof of concept demo and and how much have you raised total after that I didn't raise much I just occasionally pick up investors that are more like strategic like DVC yeah so what are the plans for you know the next 12 months what do you see as the biggest bottlenecks would you what do you want to work more on and where would you need you know help new hires somebody maybe who's watching us now yeah who might want to join the company yes so I think a few things first a very strong product person a much better product manager than me and also have strong sense of like okay how to disect the potential user groups in more refined ways what are their specific needs what are their major pinpoints that they would tolerate all the like bugs and the same and are willing to even pay to use it and how much are they willing to pay and like what our strategy in terms of pricing this is all the things I'm currently trying to figure out and I want someone that basically has been there done it have deep thoughts about it to drawing

and basically accelerate our learning curves that's one and also like I want strong kind of product/ design person that have also like with a good investment experience understand what's user Journey when people come to a product and what's people's habit how to make it super smooth lless and elegant good taste yeah the third is probably a strong goto Market person that understand okay either like for the sales channels for like a m institution or large institution what are the right way to approach them and also for the like General Market having like give people good impression and like a capture traffic what are the best way and for our product when should we start what type of campaign and when should we start what type of sales process at skill for me it sounds like in spite of the fact you're still early but you have like Journey as a start founder feels pretty smooth like not a lot of like bumps on the road we have P raed quite a lot of money for preed you were able to hire great people on the way to pre-launch and looks like everything is super clear but what was the most challengeable thing for you I think of for I face the same problem that probably most of the early startup founder face which is finding pmf finding product Mark fit understanding okay we have a team that is BAS basically capable of building almost everything with the limited resource right of course so if we say our goal is just train a better model than open eyes probably like not within our resource bucket but technically we can build anything but like what should we BET our chips on which direction should we go and

yeah like what angle will help us accumulate the more we interact with user will become better and then also like different direction some will immediately get cash flow some will like basically grow for a long time until we decide to monetize there different strategies I'm still even today trying to think about this problem so I wouldn't think we have already find the answer to product Market fit and still like learning this every day I think this is probably the hardest part about being a Founder to me like how do you make the bad there's there's a good comparison of of finding product Market fit which I I find pretty accurate in my experience as a Founder so when you're looking for product Market fit it feels like you are Copus and you're pushing the stone up the hill and it's super hard and it's it's hard in every direction and in some direction it's it's even harder but then all of a sudden you reach the summit and this Stone starts just rolling on its own and all of a sudden all of your problems change from trying to you know persuade people to try persuade people to pay persuade people to talk to you to how do we get more resources how do we you know support more users like like wow here's a bottleneck we need more Engineers this is a very interesting thing that one founder explained it to me as like this is like an orgasm like it's very hard to explain but you can definitely not you know like if you had it you 100% sure yeah yeah yeah yeah so yeah that's that's interesting so you're probably going to be experimenting a lot in the in the next like

few months so what what kind of mindset do you approach this with like do you are you certain that you will find U like a great product Market fit in a certain direction or do do you want to maximize number of attempts with the runway that you have what is the what is the idea there so I'm definitely not interested in maximize number of attempts I would hope one attempt get us like a good algorithm right then yeah nice from there I I think for me I already see some very very good sign of early product Market fit so one I I think internal story is we're working with a group of hat fund managers They all have managed around $100 million to $1 billion in assets a and and probably like 20 of them in a group and they are all good friends and I start work with them and we have like basically a weekly meet to like try to deliver them something and they use our public product but they just feel like oh our speed like providing them value is not enough and they do want the product that last week they fly to California I'll leave next to like my office in s till and basically will stay in California for a whole month and this month they will suffer the bad time zone wake up at 500 a.m.

to basically look at Market and also like basically just be there for us and give us daily feedback I think this is something that makes me very confident okay I see early science that smells like a really good early signs what is the number one feature they request I think they want us to help them basically like reduce the number of information they like receive as a human beings they have been really well served as they are like a powerful profol managers so they are having big budgets and they like register for all these different like City Bank government tax research report but they only have limited time so they don't even want to kind of spend time like read even the paid content sometime and they just want us to give them the thing that they needed the most so that they don't just without missing the important stuff that is actually a very challenging task to do yes yes yes how do you like do you do you have any ranking for importance of the facts yeah we're currently trying to build this kind of thing for different so basically you need to understand a different group of people since their ranking is ically different although they all doing investment like they have different Taste of the way they do investment so basically you need to First understanding the landscape to classify groups of investment Styles and for each investment Styles understand their way of working their workflow of information consumption and the workflow of like a analy analytical process and then try to basic help them automate as much of that as possible so that they can save time focusing on basically the most important decisions or the most important games radical

thinking like what are other people thinking and all that that's something that still today's machine cannot really get on and need the best human to participate in that part of game but I think machine can do much more than traditional qu Finance which is only crunchy numbers we should basically help them prepare all the information in the most structured form and digest all the different un structure information and be the filter layer currently we have a great tool that can basically search information or like help you go over the information that you know but we are trying to build in the domain expertise and the knowledge for what other information that we should consume that is that's a lot it sounds like a lot of stuff to code yes yes so we are having a very strong team and the the team is still like just learning how to use the AI large language model as initially we're like okay maybe the AI model is super smart and can do allot things and we test and we immediately see a lot of boundaries that AI today cannot do right even gb4 Cloud three is still in our eyes relatively and it's still pretty far from a well educated dear future Roberts he did not mean to offend you yeah yeah but it's okay right like if you look at the speed of AI on like log skill how fast they grow I still remember working with bird and gbt one so I'm super confident before like we basically figure out all the things AI will also reach the next iteration then we'll have like a new it's like basically operating system upgrade as we use large language model to code our prompts code our agent

stuff and merge that with our software 1.0 like python code and all that like the operating system itself also upgrade and unlock more abilities like say today AI is still not that good at like data science you cannot really hope you give AI the instruction I give to a Quant researcher and that they can give me like a kind of satisfiable answer but I can see the trend where that's something possible on the road and when we reach AGI your Mode's going to be around data first of all we need to have a clear definition what AGI is so I would choose the working definition of like basically the AI need to be General enough and robust enough to do probably like say 95% of like generic kind of knowledge workers work yeah today I think we are not reliable enough as technology there and also for some ability it's still like there's a gap so I kind of agree with the Microsoft paper that in gbt 4 we see spark Sparks of intelligence intelligence for AI but it hasn't got to the point where it's like super reliable and good but GPT 5 from the gossip we're hearing like of course you probably have more information than that but like from the gossip we're hearing it's it's it's a pretty big lip I don't know if it's going to be called 4.5 or five but it's it's a pretty big lip but it's not yet AI I think I have obligation to play poker face in this question okay yeah we're trying to stand up it's this it's funny I think Claude anthropic really is their CLA 3 model which is very impressive and a lot of the world are enjoying it now but

opening I will will retaliate very soon they will they release something very impressive I think large language model is a pretty interesting risk to the bottom as like you need a lot of money to truly train a like a kind of model that will stay in the field for the like even next I think eight months right like if we look eight months ago it's pretty hard to imagine some model as powerful as Gro will be like released and I think this will even accelerate rather than decelerate for the next at least one or two years and in this journey I think for startup we need to buckle up and think ahead basically in a year how good large language model would be and basically build both for the future when AI become better and also build for today to satisfy what what we need to satisfy with our customers there's actually very interesting debate and I think you you'll be able to comment from the point of a researcher with a really really good rep reputation in the field a lot of so we just had this discussion with one of our advisers he told us you guys are investing in thin roppers of companies that are basically building a product on top of somebody else's model and that's not Rel able but effectively you're you're building on top of existing models you treat models as a commodity and you see your product in in a combination of something else how do you like how do you comment this this discussion that a lot of VCS see stuff built on top of foundational models as moldless yeah so initially I also like thought through that thing and there is a period that I also

think oh how can we build technical mode and when I look around entire I think startup space and also big company space I realize oh basically every company that I participated in the game of model have no mode so it's okay it's like no one needs to have mode we are not Nvidia not like a deep chip company so like I think the key is to gain the Insight of like what your user want and how like writing prompts look like such a simple thing but like it's not that far from writing python code like to a very hardcore engineer right like forun writing python is like writing language now like to basically the good like machine learning Engineers writing prompt is like such a stupid job but I think to properly architect A system that mix is like a large language model OS with the transitional like programming OS and make some system that actually works is very n and it will get you will build something much more comp initially when you write Hollow World in Python it's like so triv print Hollow world right and slowly build a very big architect I think in large language model it's the same you were a really good answer and it made me think that maybe we should not be looking at companies in as a you know like as a snapshot because it's really about the culture about the people cuz your small team of six six people is not what you've built but it's what you can build with this team with the culture with you know the motivation how fast you can move so your mode might be in in the dynamism of of the which is essential like in this absolutely

rapid changing B and the change the faster it's changing the more culture matters like if it's a if you're through General Electric probably you don't have enough motivated people and they Department to build you know staff faster and better than Microsoft but maybe but if you're a startup and you have you know a team good enough a culture good enough and maybe your mode is just by being consistently able to outperform and outpace the competitor yeah yeah that's definitely one thing and also I think there are so much things to do with the the new technology I don't really think we are threatened by either the big incumbent or by other startups I think the name of the game is more like exploration game can you find the right format oh yeah that like utilize AI to satisfy your clients and like can you figure out who are the clients that well truly needed to the point that they are willing to suffer all the inconvenience early technology I think just by realizing that is going to be worth the multi-billions if you look at how much Nvidia of Microsoft is work just by unlocking one realization would easily like give them like a few tens of billion dollars more in valuation right so it's like a whole exploration game do you have any advice for expiring finders like what you already learned and think it's essential I I think of startup as more like going on to like a hiking trail so do do not really worry about competitors in the early day before you hit product Market fit of the second phase before you have that organism it's more like you have like other hikers hike with you and you all do exploration

so I think that's one second is it's very hard to basically stand in front of the start of the hike and look forecast the straight and know what the hiking trail looks like so it's extremely hard to like try to Le in a big company and say Okay I want to accumulate enough knowledge that when I do start I know exactly what to build and succeed occasionally that works out in deep technology company where someone assemble a super strong technical team with a well- defined goal and rebuild what is already known but like cannot be built and sell the company for a lot of money but most of the time I think it's just you go on high and it take unexpected turns as it goes and yeah yeah and hopefully not not not meeting a m Mountain line yeah yeah yeah there are some some some risk on the way which is necessary I love the metaphor that's yeah that's a really good metaphor thank you so much bill for joining us today and I hope we made it a little bit more clear for somebody what G Alpha is and what what you guys are building hopefully we inspired somebody maybe we inspired some curiosity this is my biggest hope yeah yeah yeah I would love to have people try our current like open Beta if you go to like G-Alpha AI.

you can like play with our current we released a few different product bot that you can try out and all to get any feedback and maybe you can make some money trading not an investment advice yeah or apply for open positions yeah yeah definitely yeah with hiring as I was saying also like a looking for very technical strong people always thank you very much