Data will be AI's bottleneck - Matthijs de Vries

hello everyone how's the energy here today still good yay or are we all a little bit tired from all the talks oh I try to keep it uh easy uh not really technical uh today uh to help easy easy for exactly I'm um what I'm what what am I going to do today I'm going to say the word data a lot just be warned um I'm going to take you through a journey of What's blocking the next generation of AI in my opinion and um that's going to be data that's the first time I said data um just uh let's let's do something first like um uh by raise of hands who here is actually from Poland okay not that much okay and who has been here since morning raise of hands okay no okay so what I just did is collecting essentially data I know now something about audience could be useful for me could be useful for the organizers um but I have now data it's not maybe super useful data but would ask more questions uh what is data and uh what I'm trying to say is data is all around us it's everywhere and people say data is the new oil and I would like to argue that data is really like the new oxygen because in this Society without data we cannot function like data is is really it's everywhere you know it's generated and consumed when we use public transport when we are on our phones when we order food um is used really everywhere at the problem with data it's very hard to obtain the data that's already generated it's not freely available or easily available and um there's a problem that uh I want to discuss today because let's let's take um an easy example for example from an insurance company that does car insurance and they uh they give out uh small devices to uh the people the drivers that they plug into their OBD port in the car and then they generate data about how they're driving whether they drive carefully or a little bit Wilder and that data is going to be used to calculate the premium of their insurance so if they drive really carefully they get less uh they pay less right um that data then served served its purpose meaning it's reached a goal and the data is then somewhere on a server siloed way not available for anyone anymore which is maybe a little bit wild because that data could be useful for other business startups maybe even be useful for conducting research um and it's it's a hassle to sell that data and then to also reward the original uh creator of the data the driver for a part of that revenue and there is another problem with the current reality is the data that is available you get issues like this you know uh just recently maybe not so recently anymore but uh the New York Times They sued open AI because they scraped millions of Articles from the new New York Times in order to train their uh uh large language model right and uh we can go back and forth whether who is in the right who is in the wrong doesn't really matter the fact is is that there is an increasing amount of companies that generate content data that uh are trying to make it harder to get access to the data um data is for example locked behind paywalls uh they are applying anti-c scraping measures stuff like that and because of that all all all different mechanisms it makes it harder for uh uh creators of uh models like this like open AI to have access to the data which makes the next generation of AI more difficult and yeah we can say it's simply going to get harder to get access to more data but there is another problem a couple of other problems but another major problem is that uh besides the fact that it's harder to obtain access to data there is if we look at uh if we look specifically zoom in on large language models generative AI specifically um the the the large language models the chat ones um they have a problem that they hallucinate so they're trained on massive amounts of uh unstructured data and there is a big lack of access to structure data so let's uh stop a little what is structure data was unstructured Data uh unstructured data uh sorry no structured data it's uh for example what you have in a spreadsheet is predictable has a structure single format and uh you have a bunch of records and if it's a record about objects that you can buy uh one of the columns is uh price for example and that price can be different between records could be $70 in one it could be 200 in the other but it can never be uh purple or amam it will always be a dollar amount that's structured unstructured data is everything else it can be a news article conversation between people it can be an image a video uh an audio clip it's data that's not very predictable it's not very structured structured in a sense but not structured as typical structured data and having more structured data available to large language models will help uh large language large language models to come up with uh factual information meaning that will hallucinate less because all this unstructured data means that they can have excellent conversations like they're excellent conversation list so it's very nice uh uh piece of text but if you going to um any logical reasoning with an llm it falls apart and it's not as easy to say okay we give it a bunch of structured data sets and then they will know what to do it's not enough because structured data needs to be contextualized because without contextualization you have no idea how to interpret the data so to zoom in a little bit on contextualization what that is um here's an example of Wikipedia is maybe not a perfect example but makes it absolutely clear what how powerful contextualization can be on the if we pick apart a Wikipedia page you have two big elements you have at the bottom you have a list of sources references these are important because through this references you essentially prove that what you state in the article above is true can be verified that list of references is structured data because every reference it's a number it's a URL it's a description and the next reference is the same so it's a predictable structure format if you just have a Wikipedia page with only structured data only these references nobody understands the topic they will have to go through all this these uh these links read it and they know that all of this combined is this topic but how the context Behind These combinations is unclear so you have the page above with the paragraphs that's unstructured data by the way can contain images can contain videos uh tables um but through reading that page you know the context of all these sources essentially so the same goes for any kind of structured data set if we give it to an llm we need to add that context so we need to have mechanisms in place in order to have contextualization through metadata so that the data can be fed through llms and llms knows to uh how to interpret it so for llms we already have kind of a solution through retrieve augmented generation bra and what it means is that you don't have to train a model on your business data your structure data for it to use it in its reasoning when it's answering you so let's pick that apart a little bit to make this uh digestible if um a business creates a custom llm it can uh give it access to its own data his business data about clients about uh products stuff like that so that when uh for example llm is used for customer support it can um while generating the response check the documents the databases for the factual information and take that into account for its answer which is cool which means everything I said doesn't matter there is no problem we already have the solution uh not true because this is very a very individual approach this works for a single B business a single use case so if we want to have truly a next generation of AI where we bring all the world's data together not necessarily for free but then at least with Fair rewarding uh mechanisms Revenue distribution mechanisms uh data governance and stuff like that we need to go a step F further than having just a rack for individual use cases where we can unlock all the world's data into an llm and we can also solve issues like New York Times we can say if article from The New York Times is used as a reference then you have to pay a little bit for as an example let's say I agree but it could be done so what I'm essentially talking about what I see in the future is a completely new data economy which is essentially forming an ontology of of the world's data onology is an interesting word what does it mean onology it's a combination of properties about something so here it's an example of a pizza pizza contains topics you have different types of pizzas you see all the uh hierarchy between those elements so we can do the same with structured data and unstructured data from the whole world so if we bring all of these data together and we are able to form the connections and combinations between this data we can essentially form an antology of the entire world the entire universe of course this kind of utopian what I'm saying here it's very ambitious but every step towards that will be an improvement of the current situation we probably never really fully realize the world's or universe onology but we can go really far and something very underrated what is also needed in order to get that onology is reference data and reference data is the boring type of data not worth anything does it belong to a business it's just data that is it's virtually non-changing or very slow changing data but it's useful for uh us as reference like we can talk about the flag of Poland but with reference data you would include information as that has a red part and it has a white part and which order they are and uh as an example here in the presentation is the periodic table periodic table table is perfect example of what reference data is it's not owned by anyone it's just is and it can change but it's very slow to change and if you have a data set a business data set that reference an element it could be further extended with information from a reference data set that includes this so we need to have um we need to accommodate a bunch of reference data sets um to allow to on board them to allow free contributors like on on Wikipedia that uh on board these data sets that they're rewarded in some way uh because only then we can have a full onology because now I mentioned all these problems with what I see what's data what's still needed and then there is synthetic data and um there's really interesting story here this tweet was very account intuitive so let's first uh come to what's synthetic data um synthetic data is essentially an extrapolation of existing data it's filling in the gaps if you want more of that data you're going to make that up by algorithms averages using an AI model to come up with that data and you come to synthetic data uh so that you have more of the data but the data isn't real uh there was a discussion uh well there are several discussions about what happens to large language models when they are going to be trained on their own data that they have generated because there is a a lot of concern that with this huge influx of llm generated content is it going to be of any high quality in the future because it's trained on its own data going forward and I think that situation is very similar to a situation where we uh speak about synthetic data because that's also AI generated data it's also alori uh generated data so it's very similar and you would expect from that discussion and from intuition that having more fake data will make an generative AI worse or Flatline will hit is ceiling and this shows that having uh a lot of synthetic data actually makes the model perform better than the current models which is insane and in a way it maybe makes some sense because if you think about it um I think maybe and and this is not based on anything but just just follow my thoughts for a moment I think maybe as humans for Humanity we kind of already wrote everything it's already all out there it's just not all the different combinations between content between information uh put together so it's like any type of innovation when you uh innovate you take existing stuff and you mix it together differently and you have something new it's not like you start completely from scratch because that would mean you have to come up with new particles molecules and they built on top of that that's I think kind of the same with the content that we put out there so a lot of unstructured data can be put into new combinations through synthetic data and come on we can come up with new things and um not sure if this true but it makes sense so then everything what I said before is it is not is it not an actual problem anymore because we have this um no because this is largely for unstructured data we still have the same problem for all the business data all the other data reference data that's structured for which you cannot generate uh synthetic data well you can but it doesn't really make sense because then you're actually going to make factual data not factual anymore because you're GNA extrapolate the data with a bunch of it's not true anymore so what we need to move forward and this is what we are building at uh at nucle uh I'm founder and CEO of nucle nucle is building uh layer one infrastructure to exactly accommodate all of this what I've just mentioned so easy onboarding of data so the data is immediately interoperable the data is uh in the same format data from completely different Industries different countries different formats can be all accessed in one query without building the different connectors for every data set have access to reference data where the data can be easily contextualized so that this data can be understood by AI by humans where data can be infinitely combined to create anthologies and uh where it can be leveraged as like one big API because now having access to a lot of different data sources you have a new API you need to connect a database it's so it's really time consuming it's really expensive so having infrastructure that allows Builders to build things on top of it leveraging all of the data that's in it going to save a lot of time and the important thing is that that ecosystem needs to be collaborative it needs to be a place where people can contribute freely can be rewarded for the contributions and um can develop Solutions can build their own startups and and this is why blockchain matters and how it helps to build the next generation of AI because look having an EOS system with all the data it cannot be owned by a single entity it cannot be owned by the next Google or meta or whatever it has to be decentralized because if we put all the world's data there there cannot be one owner we have to all collectively own it and maintain it so that's why we believe our contribution is building in that decentralized infrastructure to accommodate all of this that's it thank you any questions uh so I have a question can you like value our personal data of an average person like uh how much uh my data on social media my um where do I go with my car like can you value it as a one person I um no well I can't for sure um and I also don't think that we should look at it like this individually and there is a lot to say about personal data social media data we don't really focus I don't really want to focus on personal data it's more about the the the data that that we generate around us like we drive a car it can be about what kind of inputs that we give into the car without refilling our name it's unnecessary right but um if we collect personal data or any data and we put that together in a single data set uh with all the contributions of all the different people that data set is worth more than just your data so we'll make it more um and also more useful can leverage for more use cases so that bunch of data with bunch of records from a lot of people um can be very valuable to a business and it will pay for that entire data set and then everyone who did a contribution will get a uh a percentage of the revenue generated the fair percentage of uh uh concerning of the contribution that you made the size of your contribution right make sense yes yes yes and so we can't say like uh when we have a group of people we can't say What is the most value data from day because it's it depends on what our language model do yes sorry can you elaborate a little bit it's pretty hard uh like is there any data that is the most important uh for like the big group of people for companies or no depends what they want yeah of course it's use case depending like I I speak about large language models because it's very on top of our mind right uh and everyone I'm sure here probably everyone has been using one or testing with it or use it on a daily basis for development or whatever and um there are a lot of use cases around this for uh code generation but also for a business for their uh C customer assistant uh llm and stuff like that there's also a lot of really boring AI cases to um to improve supply chain um which also requires an AI model which is incredibly boring but it also need Solutions like this so but all Next Generation AI or any type of innovation um will need data and we need more of that um so all the problems are are are the same and there's no like I think a lot of personal data is very valuable because of the fact that um it can be used for marketing purposes uh there's supply chain uh improvements require a lot of data around uh truck routs and uh Geographic stuff and and uh stock uh values um yeah it's most valuable to them is that kind of data for example I have a little bit of a random question ni so uh there's this concept of Internet of things and it feels to me that this product kind of fits into that concept however there was a lot of talking about it I feel several years ago but then it quieted down so what what is happening is internet of things coming or not Internet of Things yeah like this concept where you have one network where uh basically all the devices that can connect to the internet can easily connect and integrate with it yeah yeah good question because that way that it would make it very easy to collect the data for the for your exactly of things is huge um it's it's really big at the moment there's a lot of um devices out there on the street in houses run on a SIM card that phone home with data Internet of Things rely on data in order to because they have no interface usually they rely on data exchange very important and in uh the world of web 3 Internet of Things is deeping and um actually deep in and everything that we stand for they go hugely hand inand uh with for example we uh have a partnership with Peak Network um especially to work on all things deep in around data exchange TR trustful data exchanges verifiable data exchanges and stuff like that all right so still there's a lot of Technologies is not in place but it is moving forward rapidly you would say absolutely and there is also already a lot in place we just contribute there's one uh there for a while uh thank you for your presentation Alexander kov uh ftech Association and uh my question uh I read that uh for launch validator of your network uh I need uh one and half million million uh tokens yes and it's 4,000 yeah right now could be yeah yeah yeah uh and uh I I need uh uh stake this amount uh but can I get it from maybe you have Dow and I can uh apply it like uh validator with uh a delicated state yes yes it's possible I mean it's possible once we have main net we are still on test net right now right now test net yes exactly yeah and I and I uh can uh be participation in in uh test in the next two months when we're going to be uh um uh accepting outside validators right now we run all of them ourselves it's more of death net than a test net but test net uh where you uh where we will invite people to run their own validator is coming in like the next two months something like okay well you you just know iting in okay well let's connect yeah thank you the last one thank hello hello hello perfect um I I strongly agree that like blockchain data is highly structured and they'll be very valuable uh you mentioned contextualizing it makes it more important for large language models that are able them to like get the semantic meaning of this data out um I would kind of have the feeling that like the contextualization is the fastest thing to degrade with blockchain data because it's very expensive so you're very selective about what you put on chain and the surrounding context is like almost too expensive to put there so how are you guys thinking of tackling that in the long term and like preserving this uh this contextualization yeah easy we don't put the data on chain so we are not decentralizing data itself if you decentralize all the world's data is very highly inefficient like data storage is not so expensive anymore but it's going to become very expensive um so it's more about um like many businesses will not also allow to put their data on chain it's impossible like uh they want to keep data governance themselves so we need to allow them to give them decentralized controls mechanisms um to allow someone to trustless purchase access to this data and um to have access to like distributed GPU networks and stuff like that and have the this the centralized reward mechanisms in place so that all the contributors will always be fairly rewarded when uh V flows through the E system thank you one of audience