Written by Linda Bell
In the last round of decentralized AI craze, star projects such as Bittensor, io.net and Olas quickly became industry leaders with their innovative technologies and forward-looking layout. However, as the valuations of these old projects continue to rise, the threshold for ordinary investors to participate is also getting higher and higher. So, in the face of this round of sector rotation, are there still new opportunities for participation?
Flock: Decentralized AI training and validation network
Flock is a decentralized AI model training and application platform that combines federated learning and blockchain technology to provide users with a secure model training and management environment while protecting data privacy and fair community participation. The word Flock first came into the public eye in 2022, and its founding team jointly published an academic paper titled "FLock: Defending malicious behaviors in federated learning with blockchain", proposing the idea of ââintroducing blockchain into federated learning to prevent malicious behavior. The paper explains how to strengthen data security and privacy protection in the model training process through a decentralized mechanism, and also reveals the application potential of this new architecture in distributed computing.
After the initial proof of concept, Flock launched the decentralized multi-agent AI network Flock Research in 2023. In Flock Reseach, each agent is a large language model (LLM) tuned for a specific field, which can provide users with insights in different fields through collaboration. Then in mid-May 2024, Flock officially opened the testnet of the decentralized AI training platform, where users can participate in the training and fine-tuning of the model by using the test token FML and receive rewards. As of September 30, 2024, the number of daily active AI engineers on the Flock platform has exceeded 300, and the cumulative number of submitted models has reached more than 15,000.
As the project continues to develop, Flock has also attracted the attention of the capital market. In March this year, Flock completed a $6 million financing led by Lightspeed Faction and Tagus Capital, with participation from DCG, OKX Ventures, Inception Capital and Volt Capital. It is worth noting that Flock is also the only AI infrastructure project to receive funding in the 2024 Ethereum Foundation Academic Funding Round.
Reshaping the cornerstone of AI production relations: Introducing smart contracts for federated learning
Federated Learning is a machine learning method that allows multiple entities (usually called clients) to jointly train models while ensuring that data is stored locally. Unlike traditional machine learning, federated learning avoids uploading all data to a central server, but protects user privacy through local computing. At present, federated learning has actually been applied in many practical scenarios. For example, Google has introduced federated learning into its Gboard input method since 2017 to optimize input suggestions and text predictions while ensuring that user input data is not uploaded. Tesla also uses similar technology in its autonomous driving system to improve the vehicle's environmental perception capabilities in a local way, reducing the need for massive video data transmission.
However, these applications still have some problems, especially in terms of privacy and security. First, users need to trust centralized third parties. Secondly, in the process of model parameter transmission and aggregation, malicious nodes need to be prevented from uploading false data or malicious parameters, which may cause the overall performance of the model to deviate or even output wrong prediction results. According to research published by the FLock team in the IEEE journal, the accuracy of traditional federated learning models drops to 96.3% when there are 10% malicious nodes, and when the proportion of malicious nodes increases to 30% and 40%, the accuracy drops to 80.1% and 70.9% respectively.
To solve these problems, Flock introduced smart contracts on the blockchain as a "trust engine" in its federated learning architecture. As a trust engine, smart contracts can realize automated parameter collection and verification in a decentralized environment, and publish model results without bias, thereby effectively preventing malicious nodes from tampering with data. Compared with traditional federated learning solutions, even when 40% of the nodes are malicious nodes, FLock's model accuracy can still be maintained above 95.5%.
Positioning the AI ââexecution layer and analyzing the three-layer architecture of FLock
A major pain point in the current AI field is that the resources for AI model training and data usage are still highly concentrated in the hands of a few large companies, making it difficult for ordinary developers and users to use these resources effectively. Therefore, users can only use pre-built standardized models and cannot customize them according to their own needs. This mismatch between supply and demand also means that even if the market has abundant computing power and data reserves, they cannot be transformed into practical models and applications.
To address this problem, Flock hopes to become a scheduling system that effectively coordinates demand, resources, computing power, and data. Drawing on the Web3 technology stack, Flock positions itself as the "execution layer" because, as a core function, it is mainly responsible for allocating users' customized AI needs to various decentralized nodes for training, and scheduling these tasks to run on nodes around the world through smart contracts.
At the same time, in order to ensure the fairness and efficiency of the entire ecosystem, the FLock system is also responsible for "settlement" and "consensus". Settlement refers to motivating and managing the contributions of participants, and rewarding and punishing according to the completion of tasks. Consensus is responsible for evaluating and optimizing the quality of training results to ensure that the final generated model can represent the global optimal solution.
The overall product architecture of FLock consists of three modules: AI Arena, FL Alliance and AI Marketplace. Among them, AI Arena is responsible for decentralized model basic training, FL Alliance is responsible for model fine-tuning under the smart contract mechanism, and AI Marketplace is the final model application market.
AI Arena: Localized Model Training and Validation Incentives
AI Arena is Flock's decentralized AI training platform. Users can participate by staking Flock testnet tokens FML and receive corresponding staking rewards. After the user defines the required model and submits the task, the training node in AI Arena will use the given initial model architecture to train the model locally without uploading the data directly to the centralized server. After each node completes the training, a validator will be responsible for evaluating the work of the training node, checking the quality of the model and scoring it. If you do not want to participate in the verification process, you can also choose to delegate the tokens to the validator to receive rewards.
In AI Arena, the reward mechanism for all roles depends on two core factors: the number of stakes and the quality of the task. The number of stakes represents the "commitment" of the participant, while the quality of the task measures its contribution. For example, the reward for training nodes depends on the number of stakes and the quality ranking of the submitted model, while the reward for validators depends on the consistency of the voting results with the consensus, the number of staked tokens, and the number of times and successes of participating in the verification. The income of the delegator depends on the validator he chooses and the number of stakes.
AI Arena supports traditional machine learning model training modes, and users can choose to use local data or public data for training on their own devices to maximize the performance of the final model. Currently, there are 496 active training nodes, 871 verification nodes, and 72 delegates on the AI ââArena public test network. The current platform staking ratio is 97.74%, the average monthly income of training nodes is 40.57%, and the average monthly income of verification nodes is 24.70%.
FL Alliance: A fine-tuning platform for automated management of smart contracts
The model with the highest score on AI Arena will be selected as the "consensus model" and will be assigned to the FL Alliance for further fine-tuning. Fine-tuning will go through multiple rounds. At the beginning of each round, the system will automatically create a FL smart contract related to the task, which will automatically manage task execution and rewards. Similarly, each participant needs to stake a certain number of FML tokens. Participants are randomly assigned to be proposers or voters, where proposers use their own local datasets to train the model and upload the trained model parameters or weights to other participants. Voters will summarize and vote on the proposer's model update results. All results are then submitted to the smart contract, which will compare each round of scores with the previous round of scores to evaluate the improvement or decline in model performance. If the performance score improves, the system will enter the next stage of training; if the performance score decreases, another round of training, aggregation, and evaluation will be started using the model verified in the previous round.
FL Ailliance combines federated learning and smart contract mechanisms to achieve the goal of multiple participants jointly training a global model while ensuring data sovereignty. Moreover, by integrating different data and aggregating weights, a global model with better performance and stronger capabilities can be built. In addition, participants pledge tokens to show their commitment to participation and receive rewards based on model quality and consensus results, forming a fair and transparent ecological mechanism.
AI Marketplace: How to provide personalized AI services?
The models trained by AI Arena and fine-tuned by FL Alliance will eventually be deployed to AI Marketplace for other applications to use. Unlike traditional "model markets", AI Marketplace not only provides ready-made models, but also allows users to further modify the models and integrate new data sources to cope with different application scenarios. In addition, AI Marketplace also integrates RAG (Retrieval-Augmented Generation) technology to improve the accuracy of models in specific fields. RAG is a method of enhancing large language models by retrieving relevant information from external knowledge bases when generating answers to ensure that the model's response is more accurate and personalized.
At present, many customized GPT models based on different application scenarios have been launched on AI Marketplace, including BTC GPT, Farcaster GPT, Scroll GPT, Ethereum GPT, etc. We can use BTC GPT as an example to understand the difference between customized models and general models.
In December 2023, when the question âWhat is ARC20?â was asked to both BTC GPT and ChatGPT:
BTC GPTâs answer: âThe ARC20 protocol is a protocol built on the Bitcoin network that allows users to create, transfer, and upgrade digital items such as NFTs and interchangeable tokens. Through the ARC20 protocol, non-interchangeable NFTs and interchangeable tokens can be generated on the Bitcoin network. The main token of the ARC20 protocol is ATOM, which is used for various purposes within the protocol.â
ChatGPT's answer: "As of my last knowledge update in January 2022, there is no widely recognized standard or technology called ARC20 in technology, finance, or other common fields. It is possible that a new technology or standard called ARC20 has emerged since then, but I have no information about it."
From the answers of the two, we can see the importance and advantages of customized GPT models. Unlike general language models, customized GPT models can provide more accurate answers by training data for specific fields.
Both support DCG. What are the similarities and differences between Flock and Bittensor?
As the AI ââsector picks up, the token of Bittensor, one of the representatives of decentralized AI projects, has risen by more than 93.7% in the past 30 days. The price was close to its historical high, and the total market value exceeded $4 billion again. It is worth noting that Flock's investment institution DCG is also one of the largest validators and miners in the Bittensor ecosystem. According to people familiar with the matter, DCG holds about $100 million in TAO, and in a 2021 "Business Insider" article, DCG investor Matthew Beck recommended Bittensor as one of the 53 most promising crypto startups.
Although both are DCG-supported projects, FLock and Bittensor have different focuses. In terms of specific positioning, Bittensor's goal is to build a decentralized AI Internet, using "subnets" as the basic unit. Each subnet is equivalent to a decentralized market, and participants can join as "miners" or "validators". Currently, there are 49 subnets in the Bittensor ecosystem, covering multiple fields such as text-to-speech, content generation, and fine-tuning of large language models.
Bittensor has been the focus of market attention since last year. On the one hand, it is due to the rapid rise in the price of its token, which soared from $80 in October 2023 to a high of $730 this year. On the other hand, there are various doubts, including whether its model of relying on token incentives to attract developers is sustainable. In addition, in the Bittensor ecosystem, the top three validators (Opentensor Foundation, Taostats & Corcel, Foundry) have a total stake of nearly 40% in TAO, which has also caused users to worry about its degree of decentralization.
Unlike Bittensor, FLock is committed to providing users with personalized AI services by introducing blockchain into federated learning. Flock positions itself as the "Uber of AI". In this model, Flock acts as a "decentralized scheduling system" that matches AI needs with developers, automatically managing task allocation, result verification, and reward settlement through on-chain smart contracts to ensure that each participant can participate fairly according to their contribution. But similar to Bittensor, in addition to becoming a training node and validator, Flock also provides users with the option of delegated participation.
Specifically:
Training nodes: Participate in the training competition of AI tasks by staking tokens. It is suitable for users with computing power and AI development experience.
Validators: They also need to stake tokens to participate in the network, are responsible for verifying the quality of the miners' models, and influence reward distribution by submitting verification scores.
Delegator: Delegates tokens to miners and validator nodes to increase the node's weight in task allocation and share the rewards of the delegated nodes. In this way, even users who do not have the technical ability to train or verify tasks can participate in the network and earn income.
FLock.io has now officially opened the delegate participation function. Any user can earn income by staking FML tokens, and can choose the best node based on the expected annualized rate of return to maximize their staking income. Flock also stated that the staking and related operations in the testnet phase will affect the potential airdrop rewards after the mainnet is launched in the future.
In the future, Flock also plans to launch a more friendly task initiation mechanism, so that individual users without AI expertise can easily participate in the creation and training of AI models, and realize the vision of "everyone can participate in AI". At the same time, Flock is also actively developing multi-faceted cooperation, such as cooperating with Request Finance to develop an on-chain credit scoring model, and cooperating with Morpheus and Ritual to build a trading robot model. Providing a one-click deployment training node template allows developers to easily start and run model training on Akash. In addition, Flock has also trained the Move language programming assistant for Aptos to serve developers.
In general, although Bittensor and Flock have different market positioning, both are trying to redefine the production relations in the AI ââecosystem through different decentralized technology architectures. Their common goal is to break the monopoly of centralized giants on AI resources and build a more open and fair AI ecosystem, which is exactly what the market urgently needs.
