Author: 0xJeff, Crypto KOL
Translated by: Felix, PANews
Competition is the foundation of AI development.
The goals of participants' competition include:
- Training the best model to complete specific tasks
- Jointly training a single model to achieve optimal improvements
- Providing the best insights
- Providing the best trading signals
- Providing the most accurate predictions
- And more competitions
Without competition, innovation would develop at its own pace - often very slowly. Currently, we are witnessing the Bittensor competition in real-time, with many subnet outputs exceeding industry benchmarks in their respective tasks.
Subnet owners can design any incentive mechanism to encourage miners to compete for $TAO rewards, have validators verify miners' tasks, and allow stakers to delegate their $TAO to the best validators (to obtain maximum incentives), making Bittensor a good ecosystem that continuously pushes the boundaries of decentralized AI.
Flock has implemented similar mechanisms in its ecosystem to accelerate initial model development and use federated learning to further fine-tune domain-specific models to suit unique use cases.
What is Federated Learning
Federated Learning: A method where multiple devices (people) train a single model without sharing data. This is particularly useful in privacy-first environments such as healthcare, government, banking, customer data, etc., where privacy/confidentiality is crucial.
Unlike raw data, federated learning shares "gradients" to a central server. The server then aggregates these updates to improve the model and sends it back to the devices used for training. This process is repeated continuously.
Federated learning typically uses edge devices (smartphones, computers, IoT) because they can:
- Generate and store sensitive data locally, complying with regulations
- Be highly scalable due to the variety of smartphones
- Contain personalized data, making them very suitable for training domain-specific models
And since only gradients are shared (not raw data), edge devices with limited CPU and connectivity can run efficiently.
Flock's Products
(This will not use obscure technical terms, focusing on how it works)
Flock's product process is: (i) AI Arena (ii) FL Alliance (iii) Moonbase
- AI Arena is a competition event where AI engineers ("trainers") train their chosen models based on specified tasks (building initial models).
Currently, tasks are manually created by projects/ecosystems, and participants can propose business plans/ideas to Flock through FLock.io and define their desired end-use cases.
Flock will create tasks on the platform based on these needs, which trainers can access and start training. Trainers improve model performance/reduce hallucinations by submitting data and gradients (trainers are similar to miners in the Bittensor ecosystem).
Validators score the model based on the gradients submitted by trainers.
- Trainers and validators need to stake $FLOCK to gmFLOCK to participate (lock-up period optional from 0 to 365 days).
- Trainers and validators with higher gmFLOCK stakes can get more tasks and higher reward multipliers (both have their own $FLOCK incentive criteria, with gmFLOCK staking being one of them).
- If trainers and validators have malicious behavior (training submissions fail validation or validation is inaccurate), their gmFLOCK can be slashed.
- Delegators (stakers) can stake $FLOCK to gmFLOCK and delegate to trainers or validators. Delegators will receive a portion of $FLOCK rewards (annual yield 60%-230%).
After AI Arena completes the initial model training and validation, FL Alliance will adopt these models (the best models) and use federated learning to fine-tune them on edge devices using private datasets.
FL Alliance is a process of further fine-tuning initial models from AI Arena on edge devices using domain-specific datasets through federated learning.
Main Differences between AI Arena and FL Alliance
AI Arena = Competition | Initial model training using traditional machine learning | Public datasets | First step
FL Alliance = Collaboration | Fine-tuning using federated learning | Private datasets on local devices | Advanced fine-tuning for specific domain applications | Second step
Moonbase or AI Model Marketplace
Here, models trained on AI Arena and fine-tuned through FL Alliance can be deployed, used, and monetized.
Moonbase is still in the testing phase, but stages two and three will introduce seamless ways for contributors (trainers, validators, delegators) to own these models/agents. Anyone can pay/subscribe to use models (project owners, researchers, enterprises, etc.), and models can be deployed and integrated on any launch platform.
You can view Flock as a complete, end-to-end agent development platform, starting from trainers competing to build the best initial models, to fine-tuning for domain-specific applications, and then deploying models/agents to solve unique problems.
Recent Initiatives/Ecosystem Partners
- Flock x Qwen: Alibaba Cloud using Flock to train small language models focused on specific domains (such as medicine and finance).
- Flock's FlockOff SN96 on Bittensor: FlockOff is a research-focused subnet dedicated to improvements, incubated by Yuma.
Its goal is to help AI models pick out the most meaningful and representative data points from large datasets, achieving more accurate training without processing all available data.
For example, training a trading model to enhance the trading book - API/SDK scans Binance trading behavior, but with numerous trades, processing all trades would require excessive computation.
SLM picks precise data from Binance that represents the trading behavior on your smartphone, so the FL on your smartphone doesn't need to look at all trades - it might only need to look at 10 out of 10,000 data points that represent the entire dataset.
Flock's Top AI Applications
Before delving into applications built on Flock, it's worth mentioning that models trained on Flock have already surpassed industry-leading models in Web3 tasks.
The model can natively understand complex blockchain logic, interact in real-time with smart contracts and decentralized applications, automate DeFi strategies, manage liquidity pools, and perform multi-chain analysis.
The model was trained and validated through AI Arena and can serve as a base model for more in-depth domain-specific use cases.
1. Flock x Animoca Brands
HeyAni — AI for Venture Capital Research
Flock provided a Web3 model fine-tuned based on Animoca's Investment Committee (IC) 10-year memos. Ultimately, Flock created an experienced Web3 venture capital agent that can parse white papers, GitHub, token contract addresses, X profiles, and provide investment scores and probabilities for venture capital firms.
The agent will also provide summaries of pros and cons and suggestions on how to improve the project.
Animoca uses Ani to help reduce due diligence workload while continuously improving the agent to become a better venture capitalist.
Animoca's @AimonicaBrands also uses Flock models to help refine its trading model.
2. Flock x Eden (Still in Progress)
Eden: SexualFi - Integrating AI technology to mimic OnlyFans performers' behaviors and role-play when they are offline.
The first phase will interact with their personalities, starting with voice.
They are pairing AI agents with sex toys (controlled by the agent), so users can enjoy the toy while having a sexual conversation with the agent.
The ultimate goal is to create an immersive experience through 3D virtual avatars, animations, voice, etc.
Why Be Bullish on FLOCK?
$FLOCK Has Strong Demand
Every participant in the economy needs $FLOCK - task creators, trainers, validators, delegators, etc. all have demand.
Once Moonbase begins actual model usage, delegators/stakers will be able to earn real yield (revenue sharing).
Unlike the tokenization model of tokenizing agent tokens (such as Virtuals), Flock retains all value accumulation generated by the continuous growth of model demand on the platform.
Network participation continues to improve
High staking participation: In its token economics v1 (T+0 to T+20 days staking), staking participation reached over 47%.
In the v2 gmFLOCK model, approximately 25% of circulating supply has been locked for an average of 265 days.
Additionally, the Messari report shows that all indicators for the first quarter are bullish.
Catalysts are abundant in the second half of this year
Moonbase's gates are opening, with AI model access becoming more democratized (similar to Virtuals opening its AI agent tokenization platform). Network effects are beginning to form, and the $FLOCK flywheel effect is starting to spin.
Multiple partnerships and domain-specific collaborations are already in the background, many of which cannot be disclosed yet (but can be speculated based on their past collaborative relationships).
Early investors have long lock-up periods
After investing $150 million to $300 million (with the last round being $300 million), investors have a 12-month cliff period and a 24-month vesting period. There are approximately 6 months left until the cliff period ends. The community's valuation is similar to that of permanently locked venture capital firms.
Due to listings on Upbit and Bithumb, liquidity from the Korean market has significantly increased.
Flock has also staked most of its foundation tokens for a year (just before listing on Upbit/Bithumb)
However, some drawbacks need to be considered.
The incentive design may trigger dynamics similar to Bittensor (i.e., potential daily selling pressure from participants).
By the end of the first year, circulating supply should reach 25%, and by the second year, it should reach 50%. Network growth and actual application speed need to exceed issuance speed. (Otherwise... you know what happens).
Issuance continues for only 5 years and gradually decreases - it's very likely that after the network develops to a certain extent, enterprises and projects will need to pay actual revenue to maintain training on Flock, thereby filling the issuance gap for trainers, validators, and delegators working on the platform.
In other words, enterprises will find it cheaper and more efficient to pay Flock to develop specific domain use cases rather than developing them independently.
Flock also utilizes Bittensor subnet (SN96) to improve FL Alliance R&D, using dTAO subnet issuance instead of $FLOCK issuance. This reduces potential selling pressure on $FLOCK while improving Flock's products.
How does Flock generate profit?
Very simply. When converting gmFLOCK back to FLOCK, Flock charges approximately 5% conversion fee.
Summary
You can view Flock as a combination of Bittensor + Nous Research + Virtuals:
- Bittensor: AI Arena — competing to obtain the best models
- Nous: FL Alliance — collaboratively adjusting single models to suit specific domain use cases
- Virtuals: Moonbase — model marketplace where anyone can deploy, profit from, and subscribe to models/agents
$FLOCK serves as the ecosystem token, a necessity for all operations, integrating the value of demand-side (enterprises/projects) and supply-side (trainers/validators/delegators).
This is the only decentralized AI ecosystem providing an end-to-end model development process for specific domain use cases, while possessing a distribution channel capable of creating real-world economic value.
Meanwhile, the project has gained market attention, and the token's trading price is below the venture capital valuation (with long lock-up and vesting periods).
Related reading: Examining the Next-Generation AI Infrastructure Paradigm from Flock and Alibaba's Computing Alliance
