The AI track has recently been making a splash in the industry, and the coins dominated by Ai Agent have seen significant increases.
Let’s adjust our perspective and focus on the traditional “blockchain + AI” model training field, and learn about today’s project - FLock.io @flock_io
1. Why does AI need to be decentralized?
First, the FLock.io team believes that current AI models such as ChatGPT, Gemini, etc. are too centralized, resulting in the AI trends being led by these large companies. The public can only follow the giants and cannot participate on their own. Secondly, the amount of data required for AI model training is huge, and basically comes from users. Once this data is leaked, it will have a great impact. Finally, traditional AI models are basically single-point models, which can easily be paralyzed after a network attack.
2. How does the FLock.io project implement "blockchain + AI" and ensure effective and secure model training?
The entire project is built on a Layer 2 of Ethereum, the Base chain, and builds an end-to-end AI training link of "AI Arena - FL Alliance - AI Marketplace". At the same time, the Base chain is introduced to realize decentralized computing and decentralized data storage. All participants in the project must stake FLOCK tokens (even the system operation and maintenance personnel of the project party), which means that as long as one party does evil or is lazy, it will face punishment, which effectively ensures the security of the entire AI training process.
3. Let’s take a look at the specific process.
(1) Preliminary training: AI Arena creates different model computing tasks (currently created by the project party, and may be created by community personnel in the future), and then dispatches the tasks to different training nodes. The training nodes pass the completed models to different validators, who then score them. Finally, the scores and results are summarized and uploaded to the Base chain. This model is called the "consensus model". At the same time, each role receives token rewards based on the amount of stake, training quality, and verification quality.
The project will later add the role of data provider to provide more data to the training nodes to make the model more accurate.
(2) Advanced training: The FL Alliance client pulls the "consensus model" and jointly decides whether to fine-tune and update it. The data used is its own local data. The fine-tuned model is called the "global model". A task can be created by an initiator. The proposer uses his own data to perform local training and proposes an update to the "global model". Voters can be responsible for counting local model updates, evaluating the "global model", and voting for or against the proposed update.
(3) Hosted deployment: The FL Alliance client uploads the fine-tuned "global model" to the chain. Both the "consensus model" and the "global model" will be hosted on the AI market for users or developers to pay for, use, build, or improve. This AI market can be taken over by anyone as long as you meet a certain amount of FLOCK token staking and hardware requirements. The AI market is currently a complex chatbot driven by the most advanced LLM (probably the API of the Web2 model).
(4) Model use and development: In the AI market, project owners are encouraged to propose models, contribute data and idle resources, evaluate models, optimize data, and use these models in various applications. This helps further development and use of the model. Users or developers need to spend money to use the model. The document states that the test currency used is FML, and the token may be switched to FLOCK tokens when it goes online.
(5) Special role delegator: In addition to the above roles, the project has established a delegator role. The delegator can delegate his tokens to any qualified participant and share the benefits of his contribution to the participation model.
4. What is the current status of the project?
Currently, the project has applications. One is an AI programming assistant, which mainly assists in the development of Move language smart contracts on the Aptos chain. According to the project party, the accuracy is better than Gpt 4o. The other is a Web3 language assistant. One is the social platform Farcaster GPT, which only bears the name of Farcaster, and the other is Scroll GPT, which still bears the name of Scroll. In essence, they are both AI language assistants developed by themselves.
Because the project is in its early stages, the data is not particularly impressive. There are 700,000 model users (most of whom I personally think are taking advantage of the airdrop), 42 models have been created, 37 validators, 600 delegators, but only 16 developers.
To sum up
Through research, we can find that FLock.io's ambition is actually very big, and the pledged funds and architecture of each participant ensure decentralization and security. Unlike Bittensor (TAO), FLock.io introduces two different models - "consensus model" and "global model", and any user in the market can get rewards by participating in the model contribution.
However, the project is still in its early stages, and the training efficiency seems a bit too slow, which may be a common problem of decentralized AI platforms. Currently, there is only one programming assistant that can be used, and this programming assistant is also very niche, and the amount of data involved in training is also relatively small. Finally, I hope that the project party can keep the original intention and continue to build, and carry forward AI + blockchain.
