FLock’s Business Notes

Overview

Federated Learning on Blockchain (FLock), which anchors our business goals and inspires our project name, serves as an antidote to the over-dependence on centralised AI systems. FLock enables AI engineers to train models in FLock while data stays local. As an industry leader in the decentralised artificial intelligence (deAI) training sector, we leverage our team's deep expertise in Web 3.0 and machine learning.

We launched our testnet in mid-May 2024, and as of [September 30], 2024, our platform is utilised by [1,400] AI engineers, with activity measured by the number of GPUs. Additionally, we have over [300] daily active AI engineers, based on the count of active GPUs. As of [September 30], 2024, our platform has seen over [15,000] model submissions. Following [1.6] million validations, [37] AI models for applications such as AI assistants, transaction agents and health monitoring have been selected as canonical models, serving [611,600] model end-users. In terms of growth, we recorded daily increases of over [300] model submissions and [3,000] validations in [September] 2024. Furthermore, the Ethereum Foundation awarded us an academic grant, recognizing us as the sole AI infrastructure recipient of their 2024 Academic Grants Round. We were also honoured to present our groundbreaking research at the Royal Society in October 2024, marking the first introduction of a Web 3.0 perspective to this prestigious scholarly community. Additionally, one of our research papers on improvements to traditional federated learning was selected as the best paper at NeurIPS in 2022.

Democratising AI Development and Creation

FLock is designed to democratise AI by combining the efforts of individual data owners (whose personal data is the most valuable), compute owners and AI developers into an “AI training federation.” This federation uses crypto-economic incentives to address the increasing centralization of AI training, which stems from high requirements for data, compute and AI expertise. FLock counters this trend by enabling everyone to contribute, with the entire process governed by smart contracts. On the FLock platform, community members can contribute to the creation or development of AI models, thereby reducing reliance on the governance and provisions of tech giants. Community members can propose the models they need, support AI developers, while developers themselves can contribute compute, data or AI algorithms in a composable manner. To orchestrate decentralised AI development, ensuring data security and quality is critical—this is where FLock’s award-winning design plays a crucial role. Thanks to federated learning, data can remain decentralised, allowing models to be trained without the need to collect the data centrally.

An Ultimate Version of Federated Learning

To address the drawbacks of traditional centralised federated learning (ceFL), we utilise blockchain technology to further decentralise the supervising node, enhancing data privacy and promoting the mass adoption of federated learning. In ceFL, a supervising node, often called a central server or coordinator, plays a critical role in managing the learning process across multiple decentralised devices or nodes. This setup is designed to handle scenarios where data cannot be centrally collected due to privacy concerns, bandwidth limitations or other practical considerations. However, the supervising node operator can potentially access the primary data being trained within the local training nodes. For instance, Google and Apple use ceFL for their native keyboard applications, claiming they are free of privacy risks and do not inspect the raw typing data. Technically, however, it is feasible for them to peek at the data behind the scenes. FLock decentralises the supervisory node on the platform, ensuring complete privacy and risk mitigation.

CeFL, on the other hand, cannot scale to a large number of training nodes in most settings because any malicious node can compromise the entire model. According to a paper published by the FLock team in an IEEE journal, model accuracy decreases to 96.3% with 10% of the nodes acting maliciously, and further declines to 80.1% and 70.9% when 30% and 40% of nodes, respectively, are malicious. (See https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10471193) This vulnerability is why companies like Google and Apple only use FL techniques under conditions where they control all nodes, thus ensuring there are no malicious actors. Leveraging FLock’s technology, model accuracy remains as high as 95.5% even with 40% malicious nodes. FLock allows federated learning to scale to everyone, empowering communities to benefit from the privacy of FL-trained models.

Our Products

Our product suite covers the complete lifecycle of decentralised AI development through on-chain federated learning, which includes (1) AI Arena, a competitive venue for base model training using decentralised compute; (2) FL Alliance, a collaborative and privacy-preserving platform where developers form an alliance to augment base models; and (3) AI Marketplace, where models trained on FLock are hosted and accrue revenue to contributors of data, compute, and AI algorithms, while keeping the training and fine-tuning process evergreen.

AI Arena

AI Arena is FLock’s platform for base model selection and training. It supports a conventional machine learning model training paradigm, allowing models to be optimised directly on trainers' devices using their own or public data. For any specific model task requests from the community, trainers must select the best available base model for further training and competition. The winning optimised base model is then advanced to FLock's deFL training client for further enhancement.

As of September 30, 2024, there are [1,400] AI developers actively contributing models on AI Arena, with more than [300] daily active developers. These developers have created over [15,000] AI models to compete as the best base models on FLock. For comparison, HuggingFace, the largest repository for AI models, has just reached [1] million models as of the date of this document. Since the platform's launch in mid-May 2024, compute providers, who are required to have at least a 3090 or 4090 GPU, have continuously run model validations, accumulating over [1.6] million validations. These training and validation efforts are verifiable on-chain. Notably, our AI developer base is of high quality, with each developer required to pass an active developer test based on GitHub verifications. Among them, there have been at least [4,000] active contributions to open source software development. These users have demonstrated strong platform loyalty, with [98]% of all issued tokens staked in the training contracts.

Compared to our industry competitors, the tokenomics underlying AI Arena are more organic and reasonable. A training task is typically set for a specific period to determine the winner of a designated base model. Further competition beyond this period is considered uneconomical and redundant, with no protocol emissions. In contrast, our industry peers continue to emit tokens for trainers and validators long after the training subnets have been operational, promoting continued exploitation of their protocols without providing any significant marginal utilities.

FL Alliance

After AI Arena has optimised the model parameters, base models are further refined in FL Alliance with proprietary data, leveraging our deFL capabilities. FL Alliance enables thousands of participants to collaboratively train a global model, where data sovereignty is preserved by ensuring that no local data are transmitted at any stage of the training process. The model aggregation component allows participants to upload weights from models trained on their unique local data. These weights are then aggregated to build an optimal global model, enhancing its generalisation capabilities and performance. The integration of training task automation and deployment orchestration components simplifies the process for users to join tasks and contribute valuable knowledge extracted from their data.

Further leveraging the blockchain technology by decentralising the supervising node operators, we effectively get rid of the “peeking” problem and promote the mass adoption of federated learning to the general community. See “Overview – FLock, An Ultimate Version of Federated Learning” for details.

AI Marketplace

Once models are trained and fine-tuned through AI Arena and the FL Alliance, they can be hosted on our platform, which provides a space for people to build, improve, and use models. Through the AI Marketplace, model hosts can recycle data contributions from end users to continuously fine-tune their models on FLock’s platform, keeping the training and fine-tuning process evergreen. We will ensure all data contributions follow a tokenomics similar to that of an AI data layer provider, rewarding all data contributors while they are contributing to the overall refinement of underlying models.

Our Ecosystem

We envision ourselves as not only a solid deAI training platform but also an irreplaceable hub across the entire deAI sector. Our community features protocols and companies seeking AI transformation, AI engineers, model users and investors. Furthermore, by leveraging our product offerings, we enable our deAI fellows to experience enhanced utilities, better tokenomics and broader possibilities.

Within FLock

Model demands. FLock’s training platform is basically Uber for AI. After our future mainnet launch, anyone can propose a model training task by staking a certain amount of our tokens, referred to as $FML. Model trainers and validators will conduct the training using the FLock product lineup. For time-sensitive tasks or those that aim to attract top-tier trainers—typically for protocol or business customers—task creators can offer additional bounties on top of their staked tokens as bonuses for the trainers. In the future, FLock will also support individuals with no prior AI knowledge in creating tasks, truly making it the Uber for AI.

AI engineers. By participating as trainers, validators and deFL nodes, AI engineers can earn $FML tokens as a reward for their contributions, featuring a fairer and more equitable incentive mechanism proposed and designed by the FLock team.

Model users. See “Our Products—AI Marketplace” for details.

Investors. Delegation as investments enhance the FLock system by supporting the staking process of other participants, thereby boosting the network's validation capacity without being directly involved in task training or validation. They provide their stake tokens to others, increasing the delegatees’ chances of being selected for task assignments and affecting the overall reward distribution mechanism. Delegators earn a share of the rewards garnered by their delegatees, with distributions based on predefined algorithms that consider their staked contributions. This role enables individuals to engage in the network’s economic and validation activities by using their tokens to support delegatees, even if they lack the technical skills to train or validate tasks themselves. The delegation feature may not be accessible in certain regions due to local laws and regulations.

Outside FLock

Decompute providers. As a critical part of deAI, we connect compute power, the data layer and the end application layers. For decompute providers, we serve not only as their demand side but also as an auditor of their decompute power. As we rolled out the PoAI with IO.NET and other deAI compute providers, we act as their examiner to validate the integrity of local GPU providers. Without our involvement, decompute providers would struggle with counterfeit GPUs, which could ultimately jeopardise their token emissions.

Data layer. Furthermore, the deAI data layer's tokenomics, which rewards individuals for providing data, also relies on us to calculate the weight of their rewards. Since there can be redundant or irrelevant data providers within these protocols, and because we have access to the final model host through the training process, we can identify what data is useless and repetitive. Thus, we essentially act as a proof of contributions for the deAI data layer.

Agents and dApps. Our role is crucial to the application layer. We assist in fine-tuning the models and evolving their end products, and assist their tokenomics design and rewarding system.

Our Revenue Model

As a key orchestrator for decentralised training, FLock generates revenue from both the demand and supply sides of model training.

On the demand side, FLock offers AI models to communities, protocols and companies seeking custom solutions. We charge an ongoing fee for model use and continuous fine-tuning. As of the date of this document, FLock has delivered trained models to a variety of Web 2.0 and Web 3.0 clients. Notably, we have supplied glucose prediction models based on federated learning to over 100,000 patients across three hospitals globally. In the Web 3.0 space, we have provided a series of AI assistants to Morpheus Network, accruing monthly revenue equivalent to over $700,000 in staked stETH. FLock also contributes network consensus to DePIN projects using our proprietary consensus algorithms. For example, our PoAI collaboration with IO.net has generated over $350,000 in protocol grants. Additionally, FLock offers advisory services to our customers on implementation, continuous refinement, inference and application deployment.

On the supply side, FLock charges a platform fee from AI contributors (data, compute, AI algorithms) and requires participants to stake $FML tokens as initiation entry tickets. We accrue transaction fees and take a share of staking and bounty rewards for training nodes and validators as protocol revenue, which is partially burned to create value for token holders.

Appendix:

Web 2.0 Use Case – Healthcare Privacy-Preserving Personalised AI Models

Challenge: In the healthcare sector, strict data protection regulations and security concerns prevent patient data from being shared across institutions. However, individual hospitals often struggle to gather enough local data to effectively train AI models, especially for sensitive tasks like monitoring glucose levels in diabetic patients.

Solution: FLock addressed this issue by connecting patient data across multiple hospitals using a decentralised, federated learning framework. This allowed the training of a glucose monitoring algorithm without any patient data leaving local servers, thereby preserving privacy and ensuring compliance with data protection regulations.

Results: The glucose monitoring algorithm is now live and being used by 100,000 patients across multiple hospitals. FLock’s privacy-preserving, decentralised solution has allowed hospitals to harness the power of AI while adhering to data security requirements.

Web 3.0 Use Case – Specialised Web 3 Agents for Search, Transactions, and Coding

Challenge: In the Web3 ecosystem, there is a growing need for specialised AI models that can perform complex tasks such as on-chain data analysis, transaction facilitation, and coding for blockchain-specific languages. Users in this space often find general-purpose AI solutions, such as OpenAI’s API, insufficient for their precise needs, especially when dealing with sensitive or proprietary information, which is commercially impossible for centralised AI models to collect and aggregate for training. That is why decentralised training is needed, to preserve the decentralised community value and to empower end users in AI model co-creation, evaluation and contribution.

Solution: FLock developed several specialised AI models to address these demands:

• A query parser model for on-chain data analysis, now used by 0xScope, a leading on-chain platform.

• Text-to-SQL models for Chainbase (MOU signed) to streamline complex data queries.

• Transaction models for Morpheus, earning an ongoing emission of $700k worth of stETH.

• A coding agent for Aptos Move, which has also won a grant.

• AI assistants deployed across several Layer 1 and Layer 2 blockchains like Scroll, Berachain, and BNB, guiding users through documentation and answering quiz questions.

Results: These initiatives have attracted 600,000 users and facilitated 1.4 million model interactions. FLock’s decentralised approach ensures that users retain control over their models and data, all while enabling high-precision AI services. Partnerships with organisations like Ritual and Chainbase are expanding, while implementation with Animoca Brands around crypto-focused AI decks are ongoing.

Additional Use Cases – Business Customers with Heightened Fiduciary Duties

FLock's deAI approach is gaining traction in sectors with high fiduciary duties, including finance, consulting, legal, and healthcare.

In venture capital firms, for example, deFL protects sensitive internal data while facilitating inter-departmental collaboration without breaching confidentiality. Typically, a VC firm may operate separate primary market investment and secondary market trading teams. Regulatory and compliance mandates necessitate a stringent separation—often referred to as a “Great Chinese Wall”—to prevent conflicts of interest and insider trading. Yet, both departments possess unique insights that could benefit the entire organisation if shared securely. By employing deAI and deFL techniques, FLock is addressing these privacy concerns and enhancing capabilities. This is exemplified by companies like Animoca, which collaborate with FLock to develop AI models for their investment and trading desks, ensuring secure, compliant data sharing and improved operational efficiency.

On the other hand, for law firms, it is a firm policy that ChatGPT and Gemini cannot be used for client matters. For example, at Wilson Sonsini P.C., they are tasked with drafting the annual report for Nvidia. Utilising ChatGPT for drafting or proofreading such documents could pose a risk, as unauthorised access to the data through centralised AI training might lead to market front-running. By implementing decentralised federated learning, law firms can train their specific models using local data without concerns about privacy breaches.

Through its offerings, FLock enables organisations across various industries to develop custom AI models without compromising privacy or security.

FLock Model Pipelines: