Disassembling the training platform Flock: The "new Bittensor" for AI applications

Author: 0xJeff , Crypto KOL

Compiled by: Felix, PANews

Competition is the foundation of AI development.

Participants compete for the following goals:

• Train the best model for a specific task
• Co-training a single model to achieve the best improvement
• Provide the best insights
• Providing the best trading signals
• Provide the most accurate forecast
• And many more competitions

Without competition, innovation moves at its own pace — often slowly. The Bittensor competition is currently being witnessed in real time, with many subnetworks outputting results that surpass industry benchmarks on their respective tasks.

Subnet owners can design any incentive mechanism to let miners compete for $TAO rewards, let validators verify miners' tasks, and let stakers delegate their $TAO to the validators who are best at verification (to get the maximum incentive), which makes Bittensor a great ecosystem to continue to push the boundaries of decentralized AI.

Flock has implemented similar mechanisms as Bittensor in its ecosystem to accelerate the process of initial model development and leverage federated learning to further fine-tune domain-specific models to fit unique use cases.

What is Federated Learning

Federated Learning: A way for multiple devices (people) to train a single model without sharing data. This is particularly useful in privacy-critical environments such as healthcare, government, banking, customer data, etc., where privacy/confidentiality is critical.

Instead of raw data, federated learning shares “gradients” to a central server. The server then aggregates these updates to improve the model and sends them back to the device used to train the model. This process repeats.

Federated learning often uses edge devices (smartphones, computers, IoT) because they can:

• Generate and store sensitive data locally to comply with regulations
• Highly scalable due to the wide variety of smartphones
• Contains personalized data, which is ideal for training domain-specific models

And because only gradients (not raw data) are shared, it can run efficiently even on edge devices with limited CPU and connectivity.

Flock Products

(I won’t use obscure technical terms here, I’ll just focus on how it works)

Flock's product pipeline is: (i) AI Arena (ii) FL Alliance (iii) Moonbase

• AI Arena is a competition where AI engineers (“trainers”) train a model of their choice based on a specified task (building an initial model).

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.

Based on these requirements, Flock creates tasks on the platform that trainers can access and start training. Trainers improve the model by submitting data and gradients, thereby improving model performance/reducing hallucinations (trainers are similar to miners in the Bittensor ecosystem).

The validator scores the model based on the gradients submitted by the trainer.

• Trainers and validators need to stake $FLOCK to gmFLOCK to participate (lock-up period can be selected from 0 to 365 days).
• Trainers and validators with higher gmFLOCK staking can get more tasks and higher reward multipliers (both have their own $FLOCK incentive standards, gmFLOCK staking is one of them).
• If trainers and validators behave maliciously (training submissions fail validation or validation is inaccurate), their gmFLOCK may be slashed.
• Delegators (stakers) can stake $FLOCK into gmFLOCK and delegate it to trainers or validators. Delegators will receive a portion of the $FLOCK rewards (annualized rate of return 60%-230%).

Once AI Arena completes training and validation of the initial models, FL Alliance will take those models (the best ones) and fine-tune them using private datasets on edge devices using federated learning.

• FL Alliance is a process where the initial model of AI Arena is further fine-tuned on edge devices using domain-specific datasets (via federated learning).

The main differences between AI Arena and FL Alliance

1. AI Arena = Competition | Initial model training using traditional machine learning | Public datasets | First steps

2. FL Alliance = Collaboration | Fine-tuning using federated learning | Private datasets on local devices | Advanced fine-tuning for domain-specific applications | Step 2

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 beta, but phases 2 and 3 will introduce a seamless way for contributors (trainers, validators, delegators) to own these models/proxies. Anyone can pay/subscribe to use the models (project owners, researchers, enterprises, etc.), and the models can be deployed and integrated on any startup platform.

You can think of Flock as a full-loop, end-to-end agent development platform, starting with trainers competing to build the best initial model, to fine-tuning it for domain-specific applications, to deploying the model/agent to solve unique problems.

Recent Initiatives/Ecosystem Partners

• Flock x Qwen: Alibaba Cloud uses Flock to train small language models focused on specific fields such as medicine and finance.
• FlockOff SN96 by Flock on Bittensor: FlockOff is a research subnet focused on improvement, incubated by Yuma.

The goal is to help AI models pick out the most meaningful and representative data points from large data sets, allowing for more accurate training without having to process all available data.

For example, training a trading model to enhance the trading book - the API/SDK scans Binance's trading behavior, but the number of transactions is so large that the amount of computation required to process all transactions is too large.

SLM selects precise data from Binance that represents the trading behavior on your smartphone, so that FL on your smartphone does not have to see all the transactions - maybe only 10 data points out of 10,000 that represent the entire data set.

Flock's Top AI Applications

Before we dive into applications built on Flock, it’s worth mentioning that models trained on Flock have outperformed industry-leading models on Web3 tasks.

The model can natively understand complex blockchain logic, interact with smart contracts and decentralized applications in real time, automate DeFi strategies, manage liquidity pools, and perform multi-chain analysis.

The model is trained and validated through AI Arena and can serve as a base model for deeper domain-specific use cases.

1. Flock x Animoca Brands

HeyAni — AI for venture capital research

Flock provides a Web3 model that is fine-tuned based on 10 years of memos from the Animoca Investment Committee (IC). As a result, Flock has created an experienced Web3 VC agent that can parse white papers, GitHub, token contract addresses, X configuration files, and give a score and probability of a VC investing in your project.

The agent will also provide a summary of strengths and weaknesses and suggestions on how to improve the project.

Animoca uses Ani to help reduce the burden of due diligence work while continuously improving its agents to become better venture capitalists.

Animoca's @AimonicaBrands also uses the Flock model to help refine its trading model.

2. Flock x Eden (still in progress)

Eden: SexualFi - Integrates AI technology to imitate the behavior of OnlyFans actresses and perform role-playing when they are offline.

The first stage will be interacting with you with their personalities, starting with voice.

They are pairing an AI agent with a sex toy (controlled by the agent) so that one can enjoy the toy while having a sex call with the agent.

The ultimate goal is to create an immersive experience through 3D virtual images, animations, voice, etc.

Why are you optimistic about FLOCK?

$FLOCK has strong demand

Every participant in the economy needs $FLOCK - task creators, trainers, validators, delegators, etc.

Once Moonbase starts to actually use the model, delegators/stakers will be able to earn real returns (revenue sharing).

Unlike tokenization models that tokenize agents (such as Virtuals), Flock retains all value accrued from growing demand for models on the platform.

Network participation continues to increase

• High staking participation: In its token economics v1 (T+0 to T+20 days staking), the staking participation rate reached over 47%.

• In the v2 gmFLOCK model, ~25% of the circulating supply has been locked for an average of 265 days.

Additionally, Messari reported that all indicators were bullish in the first quarter.

Catalysts are emerging in the second half of this year

The floodgates of Moonbase are opening, and access to AI models will be more democratized (similar to Virtuals opening up its AI agent tokenization platform). Network effects are starting to form, and the flywheel effect of $FLOCK is starting to kick in.

There are multiple partnerships and area-specific collaborations going on behind the scenes, many of which cannot be announced just yet (but one can guess at the level of these based on their past collaborations).

Early investors have a long lock-in period

Investors have a 12 month cliff and 24 month vesting period after investing $150M to $300M ($300M in the last round). There are about 6 months left before the cliff ends. The community’s valuation is similar to those of VCs that are locked in forever.

Liquidity from the Korean market has increased significantly due to listings on Upbit and Bithumb.

Flock also staked most of the foundation tokens for a year (just before Upbit/Bithumb listing)

But there are some disadvantages to consider.

The incentive design could potentially induce similar dynamics as in Bittensor (i.e., the selling pressure that participants could generate on a daily basis).

​​By the end of the first year, the circulating supply should reach 25%, and by the second year, 50%. The rate of network growth and real-world adoption needs to outpace the rate of issuance. (Otherwise… you know what will happen).

Issuance only lasts for 5 years and gradually decreases each year - it is likely that once the network develops to a certain extent, businesses and projects will need to pay real income to maintain training on Flock, thereby filling the issuance gap for trainers, validators and delegators to work on the platform.

That is, companies will find that paying Flock to develop domain-specific use cases is cheaper and more efficient than developing them themselves.

Flock also uses the Bittensor subnet (SN96) to improve FL Alliance's research and development, using dTAO subnet issuance instead of $FLOCK issuance. This reduces the potential selling pressure on $FLOCK while improving Flock's product.

How does Flock make money?

It's very simple. When exchanging gmFLOCK back to FLOCK, Flock will charge a conversion fee of about 5%.

Summarize

You can think of Flock as a combination of Bittensor + Nous Research + Virtuals:

• Bittensor: AI Arena — Compete for the best model
• Nous: FL Alliance — Collaboratively adapting a single model to fit domain-specific use cases
• Virtuals: Moonbase — Model marketplace where anyone can deploy, monetize, and subscribe to models/agents

$FLOCK, as an ecosystem token, is essential for all operations and integrates the value of the demand side (enterprises/projects) and the supply side (trainers/validators/delegators).

It is the only decentralized AI ecosystem that provides an end-to-end model development process for domain-specific use cases, while having a distribution channel that can create real-world economic value.

Meanwhile, the project has gained traction and tokens are trading below venture capital valuations (with long lock-up and vesting periods).

Related reading: The next generation of AI infrastructure paradigm from the computing alliance between Flock and Alibaba