#1 Coordinating Intelligence

The intersection of Artificial Intelligence (AI) and cryptocurrency/blockchain (crypto) has become a hyped narrative, but its importance extends far beyond the buzz. This series will explore how both blockchain technology and philosophy can address many of the problems introduced by AI regarding centralization and access, aiming to create a more positive-sum outcome for society.

In this first essay, we explore the fundamental why of AI x Crypto. We'll examine the current landscape of technology coordination, explore how crypto enables bottom-up coordination, and analyze how this model could be applied across the AI value chain – from data collection to inference. 

If you finish this piece with more questions and a heightened interest in exploring how crypto might solve some of AI's problems, we've achieved our goal! Future essays will focus on the what and how, deep diving into different parts of the AI stack (data, training, inference, etc.), and exploring specific use cases.

Coordination is about getting people or groups to work together smoothly and efficiently. It's the glue that holds societies, companies, and communities together. When we talk about coordination, we're really exploring how people can align their efforts and decisions to achieve shared goals.

There are two main flavors of coordination: top-down and bottom-up.

Top-down coordination refers to a hierarchical structure where those at the top make decisions that impact the rest of the group. Top-down coordination resembles a pyramid. Think of a traditional company structure: decisions are made by executives at the top, cascade down through managers, and eventually reach the broader workforce. In this model, as long as the folks at the top agree, that's what happens. Power is concentrated in the hands of a few.

Bottom-up coordination refers to a network-like structure where individuals or groups collaboratively contribute to decision-making processes. Bottom-up coordination looks more like a web. Imagine a community where everyone has a say: decisions emerge from the collective input of all participants. Power is distributed to its participants.

Both approaches have their strengths and weaknesses. Top-down can be quick and efficient, especially in crisis situations. Bottom-up can be slower, but often leads to more buy-in and creative solutions. These coordination models serve as useful models to help us understand the challenges and opportunities that arise when dealing with powerful technologies, like AI, that can impact billions of people. 

The predominant coordination model in technology today is top-down: where a few companies and individuals within these companies decide the allocation of resources and set policies for technologies that impact billions of users. This model has persisted due to its perceived efficiency in coordinating teams toward specific goals and generating significant financial returns. While this aligns with some aspects of the theory of the firm—such as reduced transaction costs and improved coordination—the extreme concentration of power in a few tech companies goes beyond what the original theory predicted or justified.

These advantages have allowed infrastructure and Web2 companies to become some of the largest and most influential companies in the world. However, this concentration of power comes at a cost to users. For example, most of the social media landscape is dominated by a few companies setting platform policies that impact and dictate the experience and rewards of billions of users. You can post whatever you think on social media as long as "whatever you think" falls within the confines of the social media platform's policies. Falling outside these confines can result in your post getting deleted, and your account can get banned. Similarly, you can also use the large tech platforms to make a living, as long as they can take 30% (or more) of what you earn. In other cases of social media platforms, the take rate is "100%" since no revenues pass back to the creators on these platforms.

The top-down coordination model, while efficient for rapid decision-making and resource allocation, leads to zero-sum games and occasionally negative sum games. This means companies must constantly try to outcompete each other to maintain their dominance. For example, if one social media platform doesn't leverage its users' data to serve them ads, another competitor will. This pressure is exacerbated by the expectations of shareholders and the need to justify and generate returns on investment of capital. 

Centralized power structures that concentrate both control and rewards at the top, often do so at a cost to users. The internet's original vision of an open, egalitarian, and peer-to-peer system that was meant to eliminate traditional gatekeepers, was replaced by a new set of scaled internet-native gatekeepers that came to power through a predictable "attract then extract" cycle

In AI, many of the common ingredients are in place for this cycle to repeat itself:

1. A few centralized entities controlled by a handful of people behind closed doors deciding the policy for a technology that has the potential to impact billions of users.

2. Fierce competition to build the biggest and best model to attract as many users as possible.

3. Massive amounts of capital spent to win the race, with a looming pressure to recoup costs and generate profits, which might deprioritize user interests and the company's initial ideals.

We are starting to see this play out in AI, albeit AI is in the early innings (and evolving rapidly). Platforms have taken user created data and licensed it for profit, for example Reddit struck a deal with Google to license Reddit data for $60M per year. Other companies, like Adobe and Slack received user backlash for stating in their policies that they can use user data to train AI models. This creates an extractive relationship between users and platforms since the valuable data on these platforms – messages, interactions, and posts – are created by users, but they don't participate in any of the value creation that comes from this data. 

The user can then use the AI models which these centralized giants create (based on the users' data), but implicitly must agree to the social (and potentially political) preferences of the company, which might influence the outputs of the model (See Google Gemini launch). Relatedly, we’ve seen OpenAI go from a non-profit committed to open-source AI development for humanity's benefit to a closed and corporate for-profit structure with the introduction of outside capital. Once again, creating a situation where there are incentives to prioritize shareholders over users.

The risk that AI could trend towards this extractive model is enough to ask whether an alternative coordination approach – one that is bottom-up like the one enabled by Crypto can offer a better solution, even if it might not fix everything.

Bitcoin introduced the concept of censorship resistant money not controlled by centralized entities but rather, secured by cryptography, software code, and economic incentives, all built on a decentralized network. Blockchain technology builds on the idea of open source and offers a new model of coordination that is bottom-up. The role of coordination shifts from centralized parties to the mechanism and incentive design of these distributed networks, all powered by open source software code, and governed by individuals and groups in a peer-to-peer system without centralized trust.

Crypto is the combination of three major areas of academia: Cryptography, Networking & Distributed Systems, and Game Theory. Cryptography is a discipline of mathematics that allows for the security of information across networks through the encryption and decryption of messages. Networking and Distributed Systems is a branch  of computer science, which is often credited as the architecture underpinning peer-to-peer infrastructure. Game Theory—instrumental to the field of mechanism design—comes from Economics, and creates the incentive structure that allows for the coordination of different parties to achieve a goal, such as the continuity and upkeep of a system.

The combined innovations across these fields have enabled many of crypto's core primitives that allow for bottom-up coordination:

• Immutable Ledgers: One of the most important characteristics of Blockchains is that they are immutable. They are unchangeable sources of truth that are nearly impossible to delete or alter. In the age of AI-powered creation, blockchains provide the only trustless method for digital provenance, ownership, and attribution. 

• Ownership at the Individual Level: Crypto enables a user to own assets on the Internet without the need for a trusted platform as an intermediary to verify that ownership: enabling trustless contractual agreements between parties in a peer-to-peer fashion. These assets take the form of tokens, which are programmable and can represent any arbitrary piece of data or object. With crypto based asset ownership, users have full control of not just their tokens that represent financial goods, but tokens that represent their identity, data, and preferences.

• Decentralization: Blockchain networks are built atop a distributed network of nodes in a peer-to-peer fashion. Decentralization in this context is the idea that no single entity or jurisdiction has the ability to change, corrupt or have control of the network without following the protocol. In a decentralized system, policies are set and decisions are made collectively by the participants and network users which includes a variety of different stakeholders such as users, validators, investors and the broader community members, through well-defined consensus mechanisms outlined in open source code. 

Crypto creates a unique design space that allows for the coordination of individuals or groups, in a grassroots manner that is facilitated and enforced by code. Crypto doesn't propose one form of coordination, but rather philosophically allows users to opt-in and opt-out based on their preferences. If a user or a set of users disagree with any part of a network/protocol, they can go join another network and take their assets and data with them, and in some cases even start a new network altogether.

These primitives enable the formation of trustless, transparent, and decentralized ecosystems. AI faces centralization risks as described in the top-down model that was previously highlighted. The big question on our minds is can AI benefit from a bottom-up coordination model enabled by crypto?

To understand how crypto's bottom-up coordination model could benefit AI, let's examine the AI value chain – the different stages of creation and deployment – and compare the top-down and bottom-up approaches at each stage.

Everything on the Internet is data in some shape or form. Any kind of content, intellectual property, software code, image, content, music, how quickly you swipe or click, etc. is a form of data. Data is one of the core ingredients for training AI models. The most popular datasets for training foundational models today include Common Crawl, a data set that has 100 Trillion tokens (1 Token roughly represents 1 word) and Image Net, a data set of 14M annotated images.

A key coordination question around data is data sourcing, which at its core includes data ownership. 

In the top-down coordination model, AI companies assume they can use data they scrape online and claim fair use without compensating users who create that data. However, creators, users, and data owners are pushing back on the legality and ethics of AI companies training on their data without their explicit permission. There are numerous lawsuits (NYtimes vs OpenAI, Newspapers vs OpenAI, Music Industry vs AI Startups) claiming that AI companies cannot just use scraped data. Even centralized platforms like X are turning off their APIs and using rate limits to prevent bots from scraping their websites.

In a bottom-up coordination model facilitated by cryptoeconomic mechanisms, users might register and tokenize their data, content and intellectual property on a blockchain. Once onchain, users can set the rights to how others can legitimately use their data, including deciding if they want to be compensated for providing data. This creates a system where users can benefit when (AI) companies use this data, and also allows (AI) companies to legitimately use this data without operating in a gray area as they do today.

Training refers to the process of teaching a machine learning model to recognize patterns and relationships in its training data. The goal of this step is for the model to develop parameters or weights describing the relationship between different data points within its training data. 

Training the next generation of AI models requires significant computational resources and time. Models with more parameters require more data and compute power to train (i.e. larger GPU clusters). For instance, training GPT-4 cost OpenAI ~$100M and training Gemini Ultra cost Google ~$200M.

A key coordination question around the training step is how to fund the development of AI models, given the high underlying costs.

In the top-down coordination model, AI is funded exclusively through big technology companies, private funds and institutional investors. There is no concept of crowdfunding at the scale required to develop the next generation of AI models. This model of funding optimizes for a small group of participants and is likely at the cost of users. AI companies have to prioritize value accrual for shareholders, despite users and other stakeholders (e.g. 3rd party developers) having disproportionately large contributions to the eventual success of the AI model.

In a bottom-up coordination model enabled by crypto, anyone with useful resources can contribute to the development of AI models and capture value that is created from those AI models. Users can help fund the creation of an AI model and own a part of that model through the use of crypto tokens. Beyond contributing capital, crypto-based systems can also be used to incentivize users to contribute other resources, such as their data and/or excess compute capacity, which is a large part of what centralized companies raise money to access in the first place.

The astute skeptic might call out that there is a healthy open source movement in AI that is already coordinating effectively without crypto. But without incentives to contribute, open source lacks the robust coordination system that it needs to achieve its long term promise of being a true alternative to centralized closed solutions. Said another way, open source AI today lacks an incentive layer that allows it to coordinate, at scale, contributors and collaborators from around the world and allows them to capture the value they create.

Once a model is trained, a user can input a prompt and ask the model to return some output based on the prompt. The model will do so based on the parameters developed in the training phase. This process of the model producing an output is called inference.

In the top-down coordination model, private models are opaque systems that don't give the user or community insight into the parameters of the model, serving as black-boxes. Over the past few years, we've seen biases emerge in models, particularly based on the underlying preferences of the organization that created that model. For example, when Google launched Gemini, the model failed to generate accurate representations of historical figures based on their race and gender. Many other centralized models prevent the user from asking certain types of questions based on the policies their management teams set.

In addition, verifying that inferences were done correctly by the right model is a matter of simply trusting the platform. This trust-based system opens the door to potential manipulation, especially as AI companies face pressure to monetize their investments.

In a bottom-up coordination model enabled by crypto, verifying that inferences were done correctly happens through onchain verification and provable cryptographic based techniques, leading to trustlessness. This approach could mitigate bias and increase transparency in AI model outputs.

As AI becomes more integrated into our daily lives and enables us to do more things at scale, it raises deeper coordination questions that go beyond the technical aspects of data, training, and inference. 

These questions revolve around two main themes: ownership and attribution, and legitimacy and scarcity.

• Who owns the image that an AI model creates? 

• What happens when that image is remixed (cropped, edited, turned into a video, etc), and a derivative is created?

• If the outputs of these models are used for commercial purposes, who should be compensated and for what?

• Should the creators/owners of the data used to train AI models be compensated for any output the models create? 

• What if an AI Agent is acting on your behalf and uses an AI model that was trained on biased data to create something you monetize, who bears responsibility if this agent makes a mistake or causes harm?

• As AI enables infinite abundance and AI-generated content floods the Internet, how do we distinguish between "real" and "fake" content?

• If people develop emotional attachments to AI companions, who owns these companions? Who controls them? Can the creator/owner benefit from influencing the person's decisions or behaviors?

• As AI companions become more personalized, should the interactions with them serve as training data for the platform to use freely?

These questions highlight the complexity of applying traditional concepts of ownership, attribution, and scarcity to AI-generated content and interactions. The bottom-up coordination model enabled by crypto could provide new frameworks for addressing these issues, potentially allowing for more nuanced and fair systems of ownership and compensation.

With a technology as powerful as AI, one would think we should be living in a digital renaissance. Instead, we are witnessing an arms race for the most intelligent AI, where the largest stakeholder and contributor of data, the user, is often not prioritized.

Crypto enables a new design space for bottom-up coordination, which is particularly relevant to AI due to the potential for extreme concentration of power. Unlike previous technologies, the risks associated with AI concentration could lead to negative sum outcomes, making the need for alternative coordination models more pressing (and interesting).

E.O. Wilson said, “The real problem of humanity is the following: We have Palaeolithic emotions, medieval institutions and godlike technology.” We find ourselves wielding godlike technology in AI, yet our methods of coordination and governance remain rooted in outdated, top-down models that resemble medieval institutions. This mismatch might end up being society's most critical coordination failure.

In future posts, we will deep dive into specific areas and questions that we presented in this intro post. We'll examine different parts of the AI stack and explore which AI use cases might benefit from crypto. We'll also investigate which specific aspects of crypto – from its technology to its incentive mechanisms and governance models – could be most relevant to addressing AI's challenges.

The problem with studying general-purpose technologies is that there are an infinite number of ways to talk about the elephant in the room. Rather than making broad swathe statements on how crypto will solve every problem in AI (we don't think it will), we'll approach it with a more fine tooth comb and dive into specific examples. 

What makes this stimulating to study is that both technologies are relatively early in their development, but are rapidly advancing through experimentation; the final rules have not been written yet.

To learn more, subscribe to follow along! DMs are open.

We are grateful to the contributors of this post who helped shape our ideas through conversations and who reviewed and provided comments on drafts of this post.

Contributors: Jason Zhao, Scott Kominers, Luke Pearson, Cian Lalor, Leo Chen, Mike Hanono, Ben Ebner 

Publication: https://paragraph.xyz/@aixcrypto

LJW X/Twitter: https://x.com/WhatTheLJW

Sven X/Twitter: https://x.com/svenwelly

Story Protocol X/Twitter: https://x.com/StoryProtocol

Polychain X/Twitter: https://x.com/polychain

This post is provided for educational and informational purposes only. Nothing written in this post should be taken as financial advice or advice of any kind. The content of this post are the opinions of the authors and not representative of other parties.