Intelligence as a Public Good

Transcript

Hey guys, is uh can everyone hear me? Good. Cool. Uh where am I looking? Um cool. Yeah, so my name is Nick. I'm the founder of a of a project called the Laura network and we're trying to build this kind of new standard for AI. We're trying to build this this network for aggregating different forms of machine intelligence, creating this abstraction layer for intelligence basically. And so first I just want to level set around the problem we're trying to solve this core problem in AI today which is basically that AI is controlled by a few centralized entities. Obviously a lot of the models we all use in our day-to-day lives whether it's coding whether it's just in in getting help with with different content whatever it is are produced by a very small number of entities. And this is problematic in

that one it creates a a sort of dynamic in society where this source of intelligence this this source that is going to be increasingly our sort of uh cognition. It's going to be the vehicle or the technology by which a lot of our perspectives opinions thoughts are formed is controlled by a few centralized entities which poses like a few very obvious threats to the well-being and longevity of society. Additionally, it creates this really inefficient marketplace where it requires a huge amount of resources, a lot of overhead, etc. to be able to contribute at all to the the sort of state-of-the-art of AI. Many resources in terms of of AI talent, of compute, data, etc. are inaccessible to the sort of core of the AI market as a function

of this consolidation or this centralization that's that's sort of formed from a market structure perspective in AI. Is this working? There we go. Um, and and yeah, so what what we get from this is that many many different resources are currently excluded from the the AI market structure. There's huge amounts of of AI developers that are are currently sort of blocked out from contributing to the state-of-the-art of AI. There's tons of models on platforms like HuggingFace, Kaggle, other types of sort of open source AI developer communities, as well as just kind of closed source models being developed by smaller teams or individual developers.

Um there's huge amounts of compute that are currently inaccessible. If we tap into edge device compute, things like this, data, etc. All of the resources that essentially contribute to AI. Um a large portion of those are currently blocked from participating or contributing in the current sort of state-of-the-art of AI, creating this really inefficient market structure that we currently exist in where again a few centralized entities control the majority of the AI market. They dictate what the state-of-the-art of AI is. Um, and so what we're trying to solve with Allora is this really fundamental problem that's born from this this problem of siloed machine intelligence in that a big reason the market structure has resulted in in what it has in this kind of hyper consolidation is because we we have a bunch of models

that exist but they all exist in these sort of isolated silos. I can't take model A, B, and C and somehow merge them together. take the best pieces of of performance of or intelligence from each of them and produce a new form of intelligence that represents the sum of intelligence of those underlying models. And this means that these centralized entities that have accumulated these mass amounts of resources in compute and data will increasingly produce better models, increase their lead from the rest of the pack, and create this increasingly consolidated market structure for AI. And so what we're trying to leverage decentralized uh networks to do is kind of break down these barriers between models, break the the silos that models currently exist in and create an environment where many different models can sort of come

together in collectively optimizing different ML objective functions. learn off of one another, take the best pieces from each of these models and produce these streams of aggregate inference or this this source of aggregate intelligence that represents this kind of sum of intelligence of a much more accessible and efficient underlying market environment for intelligence. Um, and what this results in is is intelligence shifting from what it is today, machine intelligence shifting from what it is today as a private utility that again is owned by a few centralized entities to it transitioning to being more of a public good. Um I think we faced a similar dilemma in the early days of the internet where there were these attempts for for sort of capture of the internet. The internet obviously being the main means of

information flow in society. An incredible an incredible technology for sort of uh catalyzing society's progression as a function of making information flow a kind of public good. And we're at a similar crossroads right now in AI where AI is a a private good. It is owned by centralized entities a very small number of them. And so again, this source of outsourced cognition that is going to become even more deeply integrated into all of society's critical functions is being captured by these these centralized monoliths. And we see it as as being critical, if not existential, for society to instantiate machine intelligence, not as a private utility, but as a public good. We see that as as being this this necessary component to society's long-term

well-being. And that's what we're trying to contribute to and what we're building with Allora essentially. Um and what what we're building essentially enables this kind of paradigm that we refer to as objective centric AI. And we think this is important. You can think of this as being analogous to intents in that um with with the the kind of advent of intents in how we interact with blockchains. We transition from this transaction ccentric paradigm of interacting with blockchains where we have to go and specify the specific state transitions we'd like to execute on chain the specific functions we'd like to call in order to achieve some ultimate objective obviously to users that imposes a meaningful amount of overhead in that I have to reverse engineer what my intent is or what my

ultimate objective is into a state a set of state transitions that I need to execute in order to achieve that objective. What intents do is they instantiate this hyperefficient market environment with hypers specialized market participants in solvers such that now users of blockchains can just specify what their ultimate intent or objective is and this efficient marketplace of solvers goes and finds the best state transitions to to execute or the best functions to call to achieve that objective. We currently exist in a kind of like transactionish world in AI as well. When any of us as users, as developers, as companies are interacting with AI, we have to scan the entire universe of models that exist and pick the best model, assert which model is

going to be the best for our use case and assert that it's going to be the best in any given circumstance, context, domain, etc. And obviously, this imposes a massive amount of overhead and friction in interacting with AI because models are always changing, circumstances are hyperdamic, etc. And so it creates this inefficiency at the consumption level of AI as well, not just at the production level. And so what we're trying to do with what we're building with Allora and creating this abstraction layer for AI and bringing together all of the world's models in this in this more efficient market environment is enable the use of AI to be much more objective centric. Now instead of the user or the consumer of AI scanning the universe of models and locking into a specific model at any given point in time and repeating that process at Nauseium, you can just

specify what you want AI to do and what you define AI as it's doing that thing well is. And then this efficient market of of models can then be provisioned, aggregated and and kind of uh stitched together in a sophisticated or an intelligent way to produce the best possible output for this. So, another way to think of what we're building with Allora is this kind of intent ccentric engine for interacting with with machine intelligence or AI. Um, and and this is just quickly kind of how the network works. We won't go super deep into the technicals here, but essentially what the network does is it breaks the problem of intelligence provisioning into these these subn networks called topics. And each topic is this kind of hyperfocused submarket environment that is defined by an

objective function i.e e uh a target variable and a loss function. So an example of this might be something simple in a kind of financial context. Predict the price of ETHUSD on a regular time basis say every eight hours. What is the price of ETHUSD going to be eight hours from now? This enables a much more robust and and sort of expressive set of DeFi primitives to evolve much more risk mitigation embedded in DeFi things like this. But it can really be for any objective function that you can think of for different AI models to optimize. And so in the instantiation of this topic and the in the sort of definition of this objective function, what happens is there's a few different types of actors that are all kind of working to to fulfill the the the sort of uh ongoing uh objective of that topic. And so on

one hand we have workers and workers are broken up into two different classes. We have base workers which are people that are just running models attempting to optimize the objective function defined by that topic. So these workers are running models to predict the price of ETHUSD every eight hours from now. They're trying to produce the best output. They're trying to pull in the appropriate data, approach it with the right kind of model methodology, etc. to optimize this objective. And in addition to those kind of base workers, we have this second class of workers, which is really kind of what makes this network work, which are called forecasting workers. They're additionally running models, but what they're doing is instead of trying to optimize the objective function at hand defined in that topic, they're trying to explore this much more unbounded or nebulous signal space in sort of context

awareness to best predict what the performance or the loss realized by those base workers is going to be. And so they're over time learning this base worker is best in volatile market conditions, this one in less volatile market conditions, maybe this base worker is just really good on weekends for whatever reason. They're exploring this additional sort of context space or signal space to create this additional dimension of performance evaluation that is then embedded into the aggregation of these inferences. And so now with both of these model outputs from both of these classes of workers, we're able to produce a single aggregate inference that takes kind of the best pieces intelligent this intelligently weighted aggregate inference to produce now a single inference to to optimize that

objective function in that topic at that point in time as best as possible. that takes from all of the the best models contributing to it. And so in that we've created this kind of abstraction layer for intelligence or this kind of intent centric engine for consuming AI for various domains. And then the third actor in the system is the one that sort of keeps the clock moving in a lot of ways. It's this class of actor called computers. They aren't running models, but on a regular time basis, every epoch within the system as defined within a given topic, they're coming in and they're evaluating how well each of those classes of workers did. they're they're uh evaluating the performance of the base workers and the forecasting workers and sort of updating their canonical weight at that given point in time for the next epoch to run. And it's

this sort of repetitive process that enables us one to enable consumers of AI to just get a single aggregate inference as def as a function of them specifying the objective they want AI to accomplish as well as as a function of forecasters and computers being so deeply embedded into the system. a system where inference is actually improves it upon itself over time. We're identifying, we're we're learning about new signals that meaningfully contribute to the performance of inference over time. We're taking the best pieces from them in a in a sort of more multi-dimensional kind of notion. And so we're we're improving the inference over time in addition consistently creating inferences that outperform any of the individual inferences within the network

as a function of introducing again that additional signal space through forecasters. This aggregate inference that is produced by the the network is consistently better from a a loss perspective than any of the individual models or the individual base workers that are contributing to that inference which we see as this kind of critical design primitive in creating a competitive alternative to this kind of siloed instantiation or paradigm for AI. Um and this is this is just quickly kind of going into how this works or just showcasing some results. Um so again forecasters exploring this additional signal space is what makes this system work. It's what makes the network uh outperform any of the individual models contributing to it. And as you can see here it might be a little small but in

the in the uh top part of this graph here we see the performance of of five individual workers in a given topic. We see that the best worker is uh predictor one or worker one and they're achieving a log loss of aboutg3.337. And then once we introduce these forecasters, again exploring this additional signal space, we see that the the best forecaster as a function of introducing this additional context awareness is able to outperform even the best individual worker within the network. And so introducing forecasting, introducing this kind of novel concept of concept of context awareness in model evaluation enables the network to meaningfully outperform the best individual worker creating this actually hyperefficient and effective uh market

environment for provisioning intelligence. This is the big innovation of of the network essentially. Um and what this means for for ML engineers is not too dissimilar what it means for sort of any participant in any decentralized network. I think what decentralized networks are more than anything is they're they're sort of tools or frameworks for instantiating more efficient and tightly scoped market environments for provisioning different types of resources. So I think for example Bitcoin is one of the most efficient energy markets that exists in the world. Before that if you had access to excess energy, you'd sell it to a grid, you'd stand up an energy company, you'd have to follow this this fairly long indirect path of uh to turning that energy into value somehow. With the advent of of of Bitcoin, obviously I can now take energy, turn it into value in

the shortest path possible. Ethereum does this with things like trust and capital. Other networks are aimed at the problem of efficiently provisioning resources via this this kind of decentralized market design. And so what we're doing with AI is creating this the the most efficient market possible for provisioning intelligence. Today, if I'm a an AI or ML engineer and I have some innovation in model development, I still have to follow this fairly long indirect path to turning that that innovation into value. If it's something that is going to be monetized via a product, I have to go through immense amount of administrative overhead to turning that into a company. I have to productize this. I have to acquire funding. I have to find product market fit, productize this model innovation in some way.

There's meaningful overhead in turning AI innovation today into value. What you get when you create a a decentralized network that is pointed very specifically at this problem of intelligence provisioning is you enable it such that any AI developer who has some innovation and model development can capture value from that in the shortest path possible in the most efficient way possible just by pointing it at the the relevant topics within the network contributing to this source of collective intelligence and being compensated based on how good their model is relative to the rest of the pack. And this goes back to our earlier points at the beginning of the presentation where now we create this fundamentally more efficient market structure for AI whereby any AI developer can very efficiently turn AI

innovation into value without this additional kind of uh uh path these these uh kind of very expensive and and frictionful steps that are currently embedded in the process of turning AI innovation or AI model uh AI models into value. Um just some quick background on who we are. We've been building in the space in this kind of like AI powered crypto space uh since 2020. We started doing some deeper research on on the in the field of kind of subjective consensus mechanism design. How do we build decentralized networks where consensus is achieved for more kind of nebulous or subjective domains? How do we enable decentralized participants in a decentralized network to reach consensus on a on a like resolution to a subjective question? We took some of

this we we tried to build some sort of crowdsourcing systems for creating more efficient financial infrastructure for especially longtail asset domains. Very quickly realized understandably that humans as the sort of main inputs or more analog systems as the main inputs into these more subjective uh network designs are just inefficient. They're they're not super effective. And so in the beginning of 2021, we started building some of the earlier kind of AI powered crypto infrastructure again aimed at creating more efficient or expressive DeFi infrastructure especially for longtail domains um NFT valuations more efficient sort of money market infrastructure enabled by AI per infrastructure liquidity provisioning infrastructure and then towards the second half of 2023 we took a lot of

that that early work we did around subjective consensus mechanism design combined it with a a lot of the stuff we had been doing in the years prior around sort of AI powered infrastructure that that is is optimized for the crypto space or the DeFi space and instantiated as this decentralized network in Allora to to enable different models to reach consensus on these hyper subjective non-deterministic problem sets i.e. optimizing different objective functions that pertain to AI. Um and here's just a quick overview of the where the network is today. Um the network's generated north of 700 million inferences at this point. Um there's probably over this number at this point, but there's been about 300,000 different workers coming onto the network. Uh these are people

running models to optimize various objective functions within the network. And these are spanning across a number of different topics. This is this number is larger than 55 at this point, but powering a bunch of different largely DeFi related domains or problem spa or problem sets that that can benefit from the the deeper integration of of AI especially in the form of of AI being provisioned by this kind of open and efficient marketplace for provisioning the different types of models and stitching different models together. Um, so I just wanted to share like uh one or two quick quick kind of case studies of where this was also just to showcase that this goes beyond just like typical price prediction or these these kind of like really obvious um uh DeFi domains.

One case study that was run an agent was built on the network for uh participating in in the the US presidential election markets on poly market. I think prediction markets are really clear use case for where AI can really open up the the capabilities and the possibilities of DeFi in that the main bottleneck for for a lot of DeFi today is that much of DeFi's value proposition is built on the fact that we can create these autonomous financial rails to support these hyper longtail financial categories. And so we we've seen this with things like like AMMs and Uniswap opening up the ability to efficiently and autonomously provision liquidity for different longtail asset domains and enable this kind of new suite of assets to to emerge as a

function of that. And prediction markets just turn this up to 100. Prediction markets are the the most longtail type of market that can exist. They're obviously a market that can uh that that are tied to a very specific event at a specific point in time. And so even if we have autonomous sort of settlement mechanisms or financial infrastructure underpinning them, we still lack an efficient enough market participant to actually participate in these markets and make them accurate. And so what we what we see is agents being the sort of primary users or participants in a lot of these prediction markets. So in this example, an agent was built uh by consuming from a topic that was stood up on the network that was aimed at uh aggregating many different US presidential election models. There was there was 10 to 20 different US

presidential election models built by large publications and individual data scientists etc. And the network was aggregating these on a regular basis. The agent was then using these to participate in the the prediction markets for predicting the next uh president of the United States and using even just a largely hedge strategy like a quasi delta neutral strategy. It was able to produce an APY of around 68% over the 3 months it was trading. And so we can see that by taking many different models, each better in different domains or contexts, and stitching them together on a regular basis in this kind of intelligent form factor, we're able to power this this meaningfully more effective type of market participant uh across a huge variety of domains. Um and

then just one more quickly. I know we're running out of time, but uh what we can see is is this applies to things like efficient liquidity provisioning as well. We there was there was some work done with a protocol called steer for uh more dynamically managing liquidity positions in V3 style dexes by taking many different models pertaining to price prediction volatility forecasting things like this to create more uh uh intelligently informed liquidity bands increasing the level of efficiency level of returns risk mitigation etc for provisioning liquidity even in these kind of longer tail domains. And so this this kind of advent of of bringing together many models via this abstraction layer form factor is this meaningful upgrade to the level of

expressiveness we can start to see in DeFi and just the broader decentralized design space. Um so that's it. Yeah, thanks. [applause] Um so we won't have time for questions so please uh catch Nick outside in the hall. Uh and please don't forget mainet just launched. Feel uh please join it and don't miss out. Thank you. Thank you.