The Value of Decentralizing AI - TOKEN2049 Dubai 2024

good morning thank you for being here um I'm going to talk a little bit today about why it's interesting to decentralize AI um our industry talks a lot about decentralization as a shorthand for something that is inherently good it is unclear that this is true in all cases money itself is a social construct and in an industry that is largely concerned with money and finance decentralization is fundamentally a philosophical argument about who controls what and how uh there are very clear positives to financial decentralization such as Financial inclusions stable coins have probably done more to lower the cost of remittances than any other tool in the last 50 years but it also comes with substantial negatives uh the lack of consumer protection is probably the greatest example hence in our industry decentralization is an Article of Faith rather than a fundamental truth but when we get to AI we quickly discover that we move out of the realm of philosophy instead we find a territory where decentralization has substantial concrete benefits decentralization of AI can result in better more accurate AI models and this is good for business and for Society at large and for the advancement of artificial intelligence it's my belief that AI is the most obvious use case for blockchain and in the coming five years it will drive blockchain's mainstream adoption so what are our motivations for decentralizing AI uh why does decentralization present a convincing route for better and more useful AI first we need to start with what is wrong with centralized AI that is AI that is built and operated by singular parties It suffers from fundamentally It suffers from substantial drawbacks that make it far less useful AI models can give poor inferences they can also hallucinate things that do not and have never existed fundamentally this is caused by a combination of bias existing in the data insufficient data and poor model specification what this means for you the user is that the models are far less useful sometimes dangerous for businesses that adopt AI this means they're exposed to reputational financial and legal risk when a chatbot goes Rogue a medical diagnosis is wrong or a self-driving car doesn't see something it should this risk the risk that comes with centralized AI reduces the value of AI to people and to businesses solving for higher quality inferences is the Bedrock of AI research it is the most important thing we have to solve for and at Flair we believe that centralization is one particular route to do this a second aspect that we've seen a lot recently is intellectual property theft the competition to build useful AI has resulted in a disregard for moral and legal risks in the race to get to usable gener gen generative models it's clear that protected intellectual property has been used in the training process this disregard for property rights makes the products of these AIS these AI models less usable because it introduces legal risk can a Hollywood studio legally distribute a film that excessively borrows from other people's work can a farmer company release a drug that was designed using other people's intellectual property again with centralized AI we see that the risk that the risk that comes with it is very substantial and that it reduces the value of using AI to people and to businesses centralized AI that is a technology that competes with the human mind could be the last battle between capital and labor the control of AI results perhaps in the greatest shift in power of any technological innovation in history if both the hardware and the algorithms associated with AI continue to be owned solely by a handful of large companies what power does that bestow on them will you be able to buy groceries if you don't espouse the right viewpoints will your business be able to utilize Services vital to its operation if it operates in a sector that isn't favored or in any in any way competes with an incumbent again the risks that come with centralized AI reduce the value of AI and decentral decentralizing AI looks to alleviate those risks making AI more useful to people to businesses and to society so we've already taken a first step in decentralizing Ai and that's distributed AI uh these are things like Federated learning uh but distributed AI is essentially AI that involves M multiple participants not necessarily within the same organization uh but it also has downsides these systems most often have a central party that collects and Aggregates the results from participants uh that Central party almost always knows the model structure um and this means it's very hard for a participant to protect their proprietary data because the underlying data can be inferred from the model outputs secondly there are no meaningful methods to screen out malicious participants borrowing the use of trust this limits the ability of entities to collaborate rewarding the participant is complex so existing distributed AI models are not optimal for encouraging collaboration and shortly we'll come on to why collaboration can be very important so so what problems uh in AI can decentralization solve if one of the biggest issues of centralized AI is lowquality inferences due in part to bias in data or an insufficient amount of data then having a method to amass data and pay for its value um is is is a useful thing it is however of limited application because data when it's valuable um on the whole most people don't want to sell it or share it and so data marketplaces that are being built in crypto are probably only useful in a limited set of circumstances being able to verify through a group of Arbiters that a data set is free from protected IP reduces the risk of using a model that gets your business in trouble verifying that a particular inference comes from a particular model that's sitting in the cloud Clow so that an application could be sure of the origin of that inference can be useful uh decentralized infrastructure networks such as you know decentralized GPU networks can reduce the Reliance on centralized parties and again the this kind of system can bake in a payment mechanism these are examples of projects that are building broadly in C crypto and AI um but we think you can go a lot further now flare what we're working on is essentially trying to combine and make collaborative AI such that entities can come together to build uh essentially a singular AI model as a combination of other AI models so to introduce consensus learning what is it in short it's a new model of artificial intelligence consisting of a blockchain where the validators train and operate models off chain we call these entities Learners and the models can range from being built from scratch to using large pre-trained models and fine-tuning them to their needs an input on which an inference is required will be submitted on the Chain each learner can then obtain an inference against that input from their own model or various models and provide an initial prediction to the network Learners trans MIT their initial predictions within the network according to a consensus of Gossip protocol during these exchanges the participants can update their inferences over time so if they are less confident about their own inference they can look at the weight of other inferences and they can decide to change their own inference at the end of this phase participants come to an agreement or consensus on the decision seemed optimal given the information available within the network this process is then repeated for any new data outputs so why is this interesting um well if you take one thing with you today it's simply this earlier I mentioned that poor inferences and hallucinations originate from a combination of bias in the data insufficient data and poor model specification I'll try to give some intuition as to what consensus learning can do Tesla Uber Google and Mercedes all have self-driving systems all those systems are trained on their own data sets that are in some way proprietary each of their data sets holds specific information and more importantly lacks specific information unique only to their data set they then have a model or a set of models that they have optimized to their data sets each model has Insight based on that specific specific data set and faults again unique to their data set but if we can take all of these models and all of this data and come up with a combined model that screens out the lack of information of each data set and the specific faults of each model then this results in a higher accuracy model that results in fewer accidents therefore making a self-driving model more useful but there aren't currently good incentives for them to share and to collaborate neither centralized learning nor distributed learning nor the existing projects in Ai and blockchain give these companies a safe method to collaborate to generate a system that benefits them all and facilitates the widespread introduction of autonomous driving this is because their d data would be essentially obtainable by each other their models would be obtainable by each other and they likely don't trust each other not to insert bad inferences as in to be a malicious actor this problem is widespread in Ai and it applies from autonomous vehicles to health care to finance to drug research large language models and the rest the models are trained on data that is incomplete the models have a lack of knowledge and combining those models combining essentially combining that data would create a better product well consensus learning provides a mechanism to correct this but most importantly without those entities sharing their data without sharing their models in the presence of malicious actors and it allows them to get to better inferences to reduce the risk of AI which furthers the utility of artificial intelligence this is why I believe that blockchain goes mainstream with AI this is why I believe that AI presents a far bigger use case for blockchain than Finance thank you