AI Governance: Wulf Kaal, AI Learning Ecosystem

Transcript

okay hello and welcome my name is Wolf Carl thanks for being here before I get started um I want to thank Alexander and sasia for hosting and inviting me again I spoke here about a month ago at a different event and I see so many friendly faces thanks for being here before I get started can I get a show of hands how many people are using chat GPT Gemini Claud everybody okay how many people are using it monthly weekly daily hourly okay okay excellent how many

people have problems with hallucinations not themselves with h the the AI oh quite a few okay excellent okay so let let me just say it's getting better so we're all using presumably four three or four right now and if you think about the data that was used to train them the data cut off for CH gbt 4 I believe was August 23 um with chat GPT 5 just in terms of the amount of data that is that has been used to train the model and the model quality that comes out of this will be

significantly above what we're currently using right so if if we're worried about the hallucinations we may still be worried about it with ch gbt 5 but maybe less so yeah over time and what what do we take from this well it's exponential growth yeah um I used to teach charts with exponential growth where you show the lines but I felt that this model uh or this picture in in this overview is is even more powerful uh so the biggest uh the biggest animal on the planet Earth uh equivalent to Chad GPT 5 I'm not sure what comes afterwards but it's quite significant so AI learning AI uses deep learning most or many of you know about this deep

learning en enables AI to recognize complex patterns in data sets that we all experience and AI can generate accordingly insights that are unparalleled I mean I've been using it for over a year in my own research and I've been writing in web 3 and web 3 governance for over a decade and I swear it's better than I am already yeah so when I use it it's just it's just amazing what it what it can see um so in terms of ter terminology so of of course on the outer roomm it's computer science cognitive Science Biology philosophy math um then we talk about artificial intelligence as an umbrella term machine learning is uh the most common type of

that and then deep learning of course um at the core uh to to experience what we see as users today um but it all starts with data and let me say up front um at Al a um AI learning ecosystem we we are focused exclusively if you want at the front end at the data input um um parameters for this uh for the for these model optimizations that we we're seeing so there's the output on the experts uh uh system in the performed task environment that we all see um there's the learn patterns the made the predictions and then the the

per performed task everything starts with data yeah there is no such thing as AI without data and that's where AI learning starts um um and the the disruption that comes out of this as many many know um we've seen the New York Times scul and sacks all predicting that about 44% of or so of all tasks across Industries will be disrupted and taken over by these systems there of course is still uh public debate about this but to me this is this is inevitable right so Predictive Analytics power powered by AI um the Competitive

Edge that comes out of this is unparalleled um that the costs are coming down it's all happening uh in parallel if you want the automation that comes out of this so significant and the disruptions and the scaling that that come out of this is is unparalleled in in business as far as I can tell um where do we see this the most let me say something provocative don't die in the next 10 years we don't want to die either way but don't die in the next 10 years why because these models can create new solutions for health care in our personal health be cancer longevity um that are that we couldn't

create before yeah the learning that comes out of this is so significant that we we have the ability in my humble opinion uh never mind the you know the long pharmaceutical trial uh period to develop new technology driven solutions for health care that will Propel us to longevity objectives that we couldn't even talk about before yeah so in Silicon Valley some of the folks they're talking about realistically people making it to 100 as a baseline yeah so um and Beyond of course now of course there this is across different industry retail manufacturing Finance um all all affect affected in this in this logic now what

what's the challenge for this the challenge is data right so it doesn't matter which model you're looking at whether it's a l large neural network medium small or traditional machine learning they all depend on data and the correlation is the performance of the model is correlated with the data and the data quality that goes into the model yeah and if we look at only on the on the right side here right what is what it's taking um to get to to get to this model so only in blue is only the the actual model at the the actual model tuning the uh model algorith uh algorithm development and the uh model training all in blue everything else that's associated with the metal model

development at about 80% of time spent on the machine learning projects themselves is all related to data yeah so the data is is the is the the Lynch pin the the starting point right so it's data cleaning labeling augmentation aggregation identification and so forth right it's all data yeah and so the pain point is data quality right so you many of you will know that a lot of these models they're in silos right now and so the industry comes in they buy a model they develop their own model they train it on their proprietary data sets and it's siloed in in in that data set that that that is available right so

the healthare industry the data cleaning alone will probably take another three years or so right to be able to get to the next iteration of how to use these models and because it's siloed they the the industry in across Industries requires proprietary data developed that that is used above and beyond the proprietary data that they already have yeah so where is all this data coming from high quality diverse data real paino for AI development yeah and um so and this goes hand inand with the scarcity and the need for humans to actually put uh put together and create

these data sets and that's where I learning ecosystem comes in I'll explain this in more detail in a second um so some of you may have heard about other other forms of synthetic data um and this the rise of synthetic data right so Alex knows about this um so there's a lot of data on people creating their own data madeup data quote unquote synthetic data to train AI models that is true that the growth is is quite significant but all driven by human data that generated by humans yeah so the synthetic data and the the human generated data they go hand in hand right um and there's quite significant

inefficiencies right now in this market and we haven't used the 1.4 billion people who are unbanked right now to help help generate these these data sets that are inevitably needed for the evolution of AI systems and the market is quite significant and this is we we looked at different data sets for this um and this is the most conservative growth model for the for the market with a compound annual growth rate here of over 28% and you see the breakdown it's very much predominated by uh North America Europe Asia Pacific yeah um and this is a huge opportunity this is huge business

opportunity given this growth data uh for data providers and they are both data providers in web 2 or traditional systems as well as in in our humble opinion in web 3 systems and what I will explain here is that web 3 systems in many ways have a Competitive Edge in this as we're building out these the these Solutions yeah um now what is microtask work many of you will know that about 14 years ago Amazon decided to launch mechanical t T and mechanical turque is this idea of allowing humans to annotate data validate data provide data cleaning

services and data augmentation yeah because 14 already 14 years ago it became clear that there's a real Market a need for this and Mechanical Turk and crowd flow to the traditionals provide providers um were mandating and and setting this up at the time already um now here here in this context the growth of this Market is quite staggering in the context so we've seen the T for the overall uh data uh generation market for AI and other systems and it feeding this Market if you want is the microtask market with 35% compound annual growth rates yeah um so these are billions of

microtasks that are completed yearly with demand demand is consistently Rising across the board um and these microtasks involve breaking down complex tasking to smaller manageable ones complex tasks are divided into microtask that can be completed by independently by human workers um and so forth now there are significant shortcomings in this existing model that we see with cflow and Mechanical Turk look I mean they they have launched this market and we we need to give them credit for if you want leveling the playing field and starting this but they're subject to significant um shortcomings and those include overpricing right and the way

these centralized platforms um struct structure their business and how they inflate prices for microtask work for their clients this is often driven by the need to duplicate the work right so if you have a data set um to train a machine um algorithm on uh what are cats and you have this typical microtask worth you have people uh individuals clicking on pictures of cats without ears to teach the machine that cats without ears are still cats yeah um in order to maintain quality in these data sets the existing systems duplicate the work 15 times so 15 different people have to create the same data set because

of the human fallibility in order to generate high quality data sets that can be used for machine learning and other systems yeah um and at at Al we believe that we can bring those numbers significantly down let me say this up front quickly um reputation systems in web 3 I've worked in web 3 governance for over a decade um in this in in in in in the AL context we can use reputation systems to significantly bring down that duplication cost um I personally believe the number is below half of what what the current systems are using but we we have to show the the data for that yeah so we believe we can bring it down by by

at least half um in terms of duplication simply by ranking people uh people output and giving them reputation scores yeah um lack of access is a huge problem in the existing Market uh many of you will know that in the last Bull Run there were several startups that um tried to address this with web 3 systems um and we're not allowing the 1.4 Milli billion unbanked to actually help uh participate in this growth the growth of this Market um and so access is a real problem that that can be addressed with web 3 systems um let me go through this I'm running out of time um how does Al work so the way we approach this is by simply

starting with with one um having a client upload the request to create a data set and the the respect data set request then goes to the workers on the platform and all the eligible workers can graduate into the system through reputation metrics right so you start with bot or not swipe left or right to create a data set on what is a machine once you have a reputation score in that Community um you can then graduate to more complex task with more complex assignments if you want everything is consensus driven yeah so the the reputation algorithm that we use always is aim aiming for consensus yeah

so a data set uh goes into the system workers perform tasks the consensus of the uh the underlying consensus requirements in the system ensure that we can bring cost down yeah because there a community consensus on the work performed which allows us to create uh reputation scores that are driven by consensus which brings down the duplication cost yeah and then um at the end of a given work performed the reputation and the the respective Al payouts are now being distributed so for companies we have we provide a global Workforce High high quality outputs as I said economies of scale quite

significant because of the reduction of the duplication cost that are otherwise associated with u traditional um providers it's a decentralized uh diverse Global Workforce it's reputation driven skill-based communities for specialized task and gamification is quite important here so how many people know XY Infinity okay so huge proof point right you have millions of people in Southeast Asia quitting their day jobs to start playing an ethereum based game to make a living to me that is the ultimate uh proof point that people are

looking for alternative means of making income and we believe at Al that we can combine the gamification that comes out of x infinity where it was a pure game with actual underlying AI learning uh task creation and that's that's a key objective for us whether that's a fullon game that's played in this context or if it's a gamified engagement to to develop microtask works for AI out data outputs um we think it's a huge economic opportunity for um microtask workers and uh there is Merit based compensation we actually believe we can pay our microtask workers more because of the efficiencies that we create through web

3 systems and the consensus driven reputation logic um that to us is a form of liberation of the workforce especially in the 1.4 billion unbanked and that scales Equity I already spoke about the reputation to remove cost and I see I'm running out of time um last point I want want to make is there's significant transaction cost minimization alone so if we don't even think about the reputation logic and bringing bringing down duplication cost through uh web 3 systems and governance and consensus driven algorithms um the just the use alone of web 3 systems brings down transaction costs uh dramatically which helps the systems to scale and with that I see I'm running

out of time happy to take any questions thank you for listening [Applause] Alex I know you want to ask a question I I believe you're right there will be an initial platform driven pricing that is subject to Market factors though right

so this this environment is starting with the pricing Basics that come out of the traditional systems where web 3 will uh will underprice that I think that's that's pretty pretty clear and then in the web 3 environment there will be Market factors that Drive pricing because we need to attract the workers to our platform so if the pricing is better than another platform we will have to adjust our pricing but as I said it's the fact that we can use reputation systems allows us to to underprice consistently as the system scales specifically when it comes to governance I assume this model provides an incentive for participants to vote

and engaging governance through the yeah so there different governance metrics involved I wrote a book about uh this this core logic of how web 3 systems governance is has to be um attack resistant um in the initial communities we literally have a bot or not environment where a collective of people uh left swipe right swipe will identify with a right swipe whether something is a bot or not and the consensus is generated by those people who who swiped right identifying the bot as a bot or not yeah now that is a very simplified form of governance and web 3

systems are especially prone to attacks as you may know so as the tasks get more complex we need to use more complex governance metrics um that's a much bigger question I'm happy to answer that later if you if you're interested happy to to go through it but as we to to finish this as we build these systems the more complex the task the more we have to put governance metrics out to ensure the quality under of the underlying work performed yeah we do thank you everybody appreciate you very much