The Case for User-Owned AI - TOKEN2049 Dubai 2024

thank you thank you and uh welcome to everyone who survived all all of this and uh staying strong for the last session uh and today I want to make a case for user owned AI so for those who are not familiar I kind of before crypto journey I was actually doing machine learning research a research I worked at Google uh across uh natural language understanding question answering if you ask questions on google.com uh we actually build models that were facilitating those answers and I worked on the technology uh which is called Transformers which is actually transformed the world and pretty much is the te in GPT it's what powering the chat GPT mid journey and all of the current Innovations in AI now I left Google to start a near AI startup it was an AI startup that was teaching machines to code and it was pretty much trying to innovate on how we interact with Computing now this was 2017 it was a few years too early and about100 million uh lack of capital to build what uh open ad did but we're trying to be smart we tried to build um a crowdsourcing platform where we had students around the world actually writing code for us and describing code for us and we had this challenge of paying them because people in China in Eastern Europe have monetary control some of them don't have bank accounts and so we faced that you know normal challenge of how do you do Global Payments uh at scale and especially micro payments 2018 even back then fees were too high on bitcoin and ethereum and that's how we got to near protocol which is now a scalable easy to use and uh uh kind of massive user based block chain oh so for those who are not familiar the language models is not a new technology this has existed since 50s it's really a statistical tool that was applicable in various ways now there's something interesting happened in 2013 where it went from a kind of a pure symbolic approach where you know what's the probability of where uh and R after where to embedding words converting words into vectors into multi-dimensional space and pretty much unlocking to do a lot of math um on top of the words and on top of the language so that was word embeddings and it kind of unlocked the Deep learning over natural language so this is where actually I joined Google after seeing this because my belief always that language is how we advance intelligence language is something that has all the World Knowledge and allows us to have reasoning and kind of capabilities and we able to test machines if they actually understand something now we were using recurr networks which are not um also that new technology but importantly they mimic how humans read and understand language they read one word at a time the problem is if you're asking Google you don't want it to like go and read an article for five minutes to give you an answer you want it to happen instantly and so for us the idea of Transformers came from this real need to actually scale ability to process information by machines in such a way that they can actually read everything at once leverage this massive parallelism capability that we have in machines and gpus and other accelerators and really then able to reason and give you an answer and so Transformers are this massively parallel kind of machine and as we saw they are very well scalable and you know if you feed them tens of trillions of tokens they become actually extremely uh kind of powerful in reasoning and understanding obviously still a lot of research but kind of the state-ofthe-art right now is chat GPT we just saw llama 3 showing extremely powerful results as well we have ex kind of a new capability that we didn't have before to understand language to make uh predictions and reasoning now what happens with this is we have this AI Innovation that's happening where before you needed a data scientist machine learning engineer who was interpreting results right somebody was looking at screens you know looking at tables and giving you kind of business decisions or integrating it into some product now we have something that can talk directly to people who are not technical it also able to interact and take inputs that are uh natural language English and so we have a kind of a transition of how we interact with computing that's happening in front of us now just to give you some context these models are not built on you know your gamer GPU right when people say oh we need more gpus there's lack of gpus these are a superc computer pretty much in disguise right so uh each node has eight gpus each GPU cost $30,000 and you usually have eight of them on one computer and you have uh eight to 10 computers on a rack and they all interconnected with a uh special cable connectivity to ensure that uh you can move pretty much information between gpus directly without even hitting your traditional CPU your kind of traditional processors so models of GPT 4 size are trained on 10,000 h100s uh for about 3 months and it costs around 64 million so this is just to give you an idea what it takes to build something like this now importantly when I mention these things are interconnected they're not connected with your traditional uh kind of networking this is a specialized EnV link connectivity and the current generation is 900 GB per second Compares this to your CPU to Ram connecting at 200 GB per second it's faster to move data between one GPU to another GPU in the same data center than to move information from your CPU to ram on one computer now the New Black Wall uh is actually at 1 18.8 terab bytes per second throughput it's insane connectivity and it's leads to a very weird situation where only big companies are able to train this models because they need to have access to this highly interconnected very expensive Hardware now a big company has a very uh specific challenges right when they training this models right it starts with data and in in in case of data set right somebody somewhere decides what goes into the data set right if you just scrape the internet there's a lot of weird stuff going on in Internet and somebody decides that we're going to not include this information do we include the qon website do we include Alex John's website do we include through social right all of these are decisions made by some team now there's a legal team that also decides oh this is copyrighted we know want to include it this is you know illegal there's ethics team we don't believe this is ethical we don't the line with this and so all of this affects how your model works and you have no understanding how that happened you have no as a user or developer you have no visibility what went into the models these models are statistical at the end and they pick up the bias that you have in the training data now even worse when after the model already trained there's a bunch of this same legal and ethics teams that are actually affecting the outcome right they're actually biasing the model's output depending on what is the current kind of decision making made and now on top of this these models are expensive to run and so right now you're assuming the output is coming from a model but actually it can be coming from an ad right they can actually run an ad on you and figure out what would not be the best answer Based on data right you're like well hey give me a review of all the cars on the market right now it can be hey it's from training data here's the different options or it can be like that you know some manufacturer paid this company to prioritize their output and so this is extremely dangerous and kind of creates a lot of risks and we see this happening right now that a lot of kind of the things are getting battled out in court right because we have a lot of the kind of uh produ like content producers you have a lot of uh content creators copyright owners pretty much suing open AI Google for variety of different uh topics because there's so much kind of uh obscurity that's happening during this process now on top of this right we have Regulators who are kind of very uh scared of the repercussions of this models but at the same time they don't fully understand how to regulate this and in and instead of like creating a more competitive landscape requiring to have the content and kind of models to be open and accessible they require to have data teams and inspection teams and ethics teams that you know just add more complexity and make sure that startups are not able to do anything we have users who are not able to kind of distinguish what's misleading content what is you know comes out of an ad or what is a manipulative content that trying to extract more value from them you have this challenge with economics of big tech companies where they need to continue growing revenue and in turn pretty much manipulating people to like click on an ad and pretty much uh kind of generate more renue through that and you have users in communities that have no idea what is going on in this model right what content went in right we are here in UAE which is very different standards from us which has very different standards from you know Netherlands or China and so all of the different countries right they need a different data set curation to go into the model that they use and so this is a case for user owned AI this is really really about how do we have models that are not trying to profit a company a corporation but actually are on your side and actually able to deliver you the value be aligned with your you know personal and kind of Financial Health mindset reasons and be able to be kind of serving you and defending you from kind of the attack of the rest of the world so there's a whole stack of technologies that need to happen for this to really be realized and it kind of spans from data to infrastructure to Applications there's a lot of applications uh but I'll cover a few very specific things uh just to give you an idea how this works so first of all it starts with this content right a lot of data right now uh being scraped from internet but the thing is like now because people know uh that this is happening you actually can start poisoning training data for the future models by putting incorrect cont or biased content into the future training data right right now if you just start posting stuff on Reddit the Google and other companies who buys this data from Reddit will now be affected and be you know the model will respond differently based on what's Happening and so what you want is you want an ability to have a cryptographic reputation of the data you want to have this information available you want the community to be able to actively uh kind of comment and provide uh context for for any of the information the best kind of closest example for this is community noes on Ax where you're able to see if you know different opinions about somebody's like popular tweet and so you can have similar idea using blockchain using cryptography to really link the content and again importantly this is not like pure AI problem the misinformation been existing for forever right you can literally Mis like create a misinformation by just editing a page you know in inspect in your browser posting the photo like photography of that on Twitter and people now confused and concerned right so this is kind of an important tool in general for preventing misinformation and misleading now next step is the kind of Creator economy that's happening right now that is based on the attention the advertisement model right the Tik toks and you know New York Times and other platforms they all based on you know somebody posts a content and then a lot of people look at it and there's ads intermixed into this then creators are rewarded is going to get disrupted by AI right AI models will be pretty much consuming all this content and then reinterpreting it pretty much with AI model to give you personalized output right your news source will not be New York Times anymore it will be your personal model that read New York Times that read uh rors that read Bloomberg and so how do we reward creators in this new economy right and right now we have all those lawsuits actually trying to battle this out but they're missing the core component which is a way a way to distribute reward to this people and so the idea here is we have creators be the author the photog like graphic designer photographer musician who contribute data to a broader Data Network all of this data then is used to train models and the models are paying kind of royalties and users of that models pay royalties back and now we distribute those royalties based on how Noel the data is and this is important because it's different from traditional method where you distribute based on how much views that got or you know fixed fee you're now Distributing based on how novel the information is for the model so that's a kind of you know data contribution network data cation that needs to happen as well in pair with cryptographic reputation to really a kind of affect and clean the data going forward there is another piece which is provable inference right right now when you're go to chat GPT you're entering your request you have no idea what output you're getting is this output from a model is there like an if statement somewhere that says if if your email you know output this if for example you you know trying to analyze some trading or some Financial uh information maybe somebody's trying to affect your decision on the other side you have no idea and so provable inference is this idea that you want to have a proof an ability to verify that the output actually came from a model that you asked it for now the challenge with this is that it's extremely complex already to get an inference from a model it requires a lot of hardware and it's also non-deterministic because of different Hardware um has different architecture of implementing floating Point operations and so it's really challenging to achieve this but it and so it's only valuable for operations where you you are making financial decisions or maybe Health decisions and probably not as important for cases where you are you know trying to post on Twitter so extremely important use case there's a lot of companies working on it uh but also extremely complicated and and hard to achieve now the other interesting idea is right now we developed programmable assets right we have you know defi kind of exploding with all kinds of different options but with now this intelligent tools with llms you're able to have intelligent assets assets that react and act based on different you know outputs in the real world or something that's happening and you can have everything from natural language oracles to yield optimization to even intelligent walls that rebalance kind of the assets they have based on you know different trading strategies uh now the big caveat with this is that uh the CMS right now are not robust for adversarial usage right they're trained to take everything that you give them on kind of as truth they're generally not very good with ambiguity and kind of of ability to determine when they there's s not confident or when they fed misinformation and so right now this is extremely dangerous and like a lot of people who are trying to do this are kind of potentially have will have a lot of issues now on near we actually have a really cool mini app to Showcase this but also collect training data it's called Shi GPT and it allows to users to pretty much convince the model to rebalance the The Volt of assets it has to whatever the asset they like right so you pretty much open it up and use the fact that the models are convincible and you having this PVP game between different people who are trying to convince it to rebalance in different ways the benefit is you collecting training data to see the actual adversarial examples of how people are trying to convince models so I'll just finish on this note right we have an amazing advancements in technology right I mean I've worked on AI like I've been excited about this idea from I don't know age 12 and the kind of challenge we're facing in front of us is that for-profit companies have this extremely um negative incentive because when they makeing decision they will always make decision how to generate more Revenue out of the user how to make more profit and this leads to manipulative Behavior this leads to you know trying to use more of your private data and it's you know kind of negative like Black Mirror type uh outcome and we have now web 3 which is a tool to disperse ownership to have coordination of people to have something where we don't need to continue post new gross revenue we can have people owning and deciding how they uh the prodal should evolve and so we are this kind of fork in the road and I'm just encouraging everyone to continue seeing how we can contribute to going into the more user owned uh Direction and really benefit from kind of building this out even though it will take some time and it will require to build new infrastructure to really build kind of truly user owned open source AI thank you follow me on Twitter and uh hope you had a great conference