Forward the Original Title‘Bittensor 和这些 Crypto AI 网络值得关注吗？’

In the past year, with the popularization of the concept of decentralized AI and the widespread application of various AI tools, AI + Web3 has gradually become one of the hottest topics in the encryption circle. According to incomplete statistics, there are currently more than 140 projects combining Web3 and AI in the industry, covering multiple directions such as computing, verification, metaverse, and games. Ethereum co-founder, Vitalik, also wrote an article to discuss the use cases of combining blockchain and AI, and pointed out that the cross-field use cases of the two are increasing, and some use cases are of higher significance and robustness. In addition, at the recent “Hong Kong Web3 Carnival” event held in Hong Kong, the topic of the combination of AI and Web3 was frequently mentioned whether it was the main venue or the side venue.

This article selects three worth-attention projects that combine Web3 and AI, and discusses their unique positions and development prospects in the field of encrypted AI.

Bittensor: Leading in market capitalization, but practicality is questioned by the market

In the AI field, unlike the resource-intensive computation and data, the encryption algorithm focuses more on the technical intensive work. However, there is a problem in the current AI field, that is, due to the existence of technical barriers, algorithms and models often cannot effectively cooperate, leading to a zero-sum game situation. To change this, Bittensor proposed a solution, using blockchain networks and incentive mechanisms to promote cooperation between different algorithms, gradually building a shared knowledge algorithm market. In short, similar to the Bitcoin mining network, Bittensor has replaced the Bitcoin mining calculation process with training and verifying AI models.

The name ‘Bittensor’ can be broken down into ‘Bit’ and ‘Tensor’. ‘Bit’ is familiar to many as the smallest currency unit in Bitcoin, but in computer science, it also represents the most basic unit of information. ‘Tensor’ originates from the Latin word ‘Tendera’, which means ‘extension’. In physics, a Tensor is a multi-indexed tensor or a multi-dimensional array or matrix, and it can represent various types of data. In machine learning, a Tensor is used to represent and process multi-dimensional data.

The architecture of Bittensor can be divided into two layers, the underlying layer is a blockchain based on Polkadot Substrate, responsible for executing the consensus mechanism and incentivizing the network. The AI layer is responsible for inference, training, and ensuring the input/output compatibility between Bittensor protocol nodes. The Bittensor network has two key participants, miners and validators. Miners submit training models to the network in exchange for token rewards, while validators are responsible for confirming the validity and accuracy of model outputs, and choose the most accurate output to return to users. To create a positive competitive cycle, Bittensor implements incentive distribution through the Yuma consensus mechanism. Yuma consensus combines PoW and PoS mechanisms, where miners obtain token rewards by competing computation results, and validators need to stake their tokens on a subnet and complete the validation work to obtain a certain number of TAO incentives. The more accurate and consistent the screening and evaluation of the AI model, the more rewards are obtained.

The subnet is the core of the Bittensor ecosystem. Bittensor introduced the concept of ‘Subnet’ in October 2023 through the Revolution upgrade. Different subnets can handle various tasks, such as machine translation, image recognition, generation, and large language models. These subnets can interact and learn from each other. Anyone can create a subnet on Bittensor. However, they must pay a fee using TAO tokens. The fee amount depends on subnet supply and demand on the network. Additionally, the subnet must be tested on the local and testnet before launching on the mainnet.

Currently, Bittensor has one unique subnet, 0# Root, and 32 other subnets. 0# Root, created by the Opentensor Foundation, functions as the governance center on Bittensor. It distributes the TAO generated by consensus to the other subnets. The role of validators on 0# Root comes from the top 64 validators with the most staked amount on other subnets, while the role of miners is played by the other subnets. Furthermore, 0# Root can allocate incentives to other subnets based on their contributions. For the remaining 32 subnets, validator nodes and miners receive a certain proportion of TAO based on their contributions. Typically, 41% is allocated to validators, another 41% to miners, and the final 18% goes to the subnet creator. Competition between subnets in the Bittensor ecosystem is intense. The system currently allows a maximum of 32 subnets, but more than 200 subnets are waiting to register on the mainnet in the testnet. Recently, several notable teams, such as MyShell TTS, have registered their own subnets on Bittensor. According to the subnet registration rules, once the number of subnets reaches the limit, the system will automatically unregister the subnet with the lowest token allocation.

Bittensor has recently been questioned regarding its registration fees and practicality. Reports indicate that the current cost to register a subnet on Bittensor is 2078.49 TAOs. On March 1, this cost peaked at 10281 TAOs, equivalent to over 7 million US dollars. With the rising price of TAO, registration fees may continue to increase. Each time a project registers a subnet, the fee doubles. If no one registers, the price halves linearly within four days. This high registration fee could pose a significant burden for developers wanting to create or join subnets. The practicality of Bittensor subnets is also under scrutiny. Most of the 32 subnets are used for ‘data collation’, ‘text, image and audio conversion’ and other low-threshold applications. None of the teams building on Bittensor have more than a dozen full-time members; most only have 2 to 3 people. Eric Wall, initiator of the Bitcoin Ordinals project Taproot Wizards and the Bitcoin NFT project Quantum Cats, shared his perspective on social platforms. He suggested that Bittensor is merely a purposeless decentralized experiment with no practical value. He pointed out that “Subnet #1 describes itself as a text prompting service. However, its operation is simple: the user sends a prompt, and the miner responds, like ChatGPT. Miners involved in this process receive TAO tokens as rewards. However, there is significant redundancy, as the validator simply checks the answer’s similarity. If a miner’s answer differs from the others, they receive no reward. The system’s efficiency is extremely low, and it is impossible to verify whether the model has been run effectively. Moreover, ordinary users cannot interact with the network at all. The subnet’s sole purpose appears to be internal operation. This process looks like purchasing useless AI tokens for exposure to decentralized AI.”

Ritual: Super luxurious background, using ZKP for AI model inference training

There are many problems in the existing AI stack, including lack of guarantees for computational integrity, privacy, and censorship resistance. In addition, the infrastructure hosted by a few centralized companies also restricts the local integration capabilities of developers and users, leading to the emergence of validity issues. Against this backdrop, the decentralized AI computing platform Ritual was born.

Ritual’s main goal is to provide an open and modular sovereignty execution layer for AI, that is, how to introduce AI into EVM, SVM, and other virtual machine environments. Simply put, Ritual connects distributed node network computing resources and model creators, allowing creators to host their AI models, while users can add all inference functions of the AI model to their existing workflows in a verifiable way.

The Ritual team has a very strong background. Co-founders Niraj Pant and Akilesh Pott were general partners at Polychain. In addition, the team includes senior engineers from well-known companies like Microsoft AI and Facebook Novi, and professionals from renowned institutions like Dragonfly, Protocol Labs, and dYdX. Furthermore, Ritual’s advisory lineup is very impressive, including EigenLayer founder and partner Sreeram Kannan, Gauntlet founder and CEO Tarun Chitra, and BitMEX co-founder Arthur Hayes.

To date, Ritual has completed two rounds of financing. In November 2023, Ritual announced the completion of a 25 million dollar financing, led by Archetype, with participation from ccomplice, Robot Ventures, dao5, Accel, Dilectic, Anagram, Avra, Hyperspher, and angel investors such as former Coinbase CTO Balaji Srinivasan, Protocol Labs researcher Nicola Greco, Worldcoin research engineer DC Builder, EigenLayer CSO Calvin Liu, Monad co-founder Keone Hon, and Daniel Shorr and Ryan Cao from the AI+Crypto project Modulus Labs. Then on April 8, 2024, Ritual received a multimillion-dollar investment from Polychain Capital, the specific amount is unknown.

Currently, Ritual has launched Infernet, a lightweight library that brings computation on-chain, allowing smart contract developers to request computation from Infernet nodes off-chain, and pass computation results to on-chain smart contracts through the Infernet SDK. Infernet nodes are lightweight off-chain clients of Infernet, mainly responsible for listening to on-chain or off-chain requests, and delivering workflow outputs and optional proofs through on-chain transactions or off-chain APIs. The Infernet SDK is a set of smart contracts that allow users to subscribe to the output of off-chain computing workloads. One of its main use cases is to introduce machine learning inference on-chain. Infernet can be deployed to any chain, allowing any protocol and application to be integrated. In addition, Infernet also allows developers to introduce their own proof systems, including Halo2 verifier and Plonky3 verifier.

Infernet does not directly perform inference on-chain, but is similar to an oracle system, where the chain issues a request, off-chain nodes execute it, and then return the response on-chain. However, this method also has an asynchronous problem, that is, developers need to wait on the block after submitting the request and cannot get an immediate response. Ritual’s method is to let developers perform inference operations directly in their familiar environment, without worrying about where the operation occurs. Although these operations are still performed off-chain, by embedding these computational operations in the virtual machine, each node can perform super-optimized AI operations while running the modified virtual machine. This method can be seen as a kind of interactive communication, implemented through pre-compilation. The emergence of this method is also a development trend of the blockchain ecosystem.

Specifically, through Infernet, developers can delegate computationally intensive operations to off-chain, and consume output and optional proofs in smart contracts through on-chain callbacks, to circumvent the restrictions of the smart contract execution environment. For example, Emily is developing a new NFT collection, allowing minters to add new features to the NFT autonomously. Emily built a minting website, published the signed delegation to the Infernet nodes running the custom workflow, the workflow can parse user input, generate new images, and the Infernet nodes then send the final image to her smart contract through on-chain transactions.

At the end of 2023, Ritual released an application supported by the Infernet SDK, Frenrug. Frenrug is a chatbot running in the Friend.tech chat room. Any user holding the Frenrug Key can send messages to Frenrug. For example, users can buy or sell the corresponding friend.tech user Key through Frenrug, but Frenrug does not handle user messages directly, instead, it sends messages to multiple Infernet nodes, these nodes run different language models. Infernet nodes process user messages and generate votes on the blockchain. When enough nodes vote, the system aggregates these votes and performs corresponding operations on the blockchain, such as buying or selling Keys. Finally, Frenrug responds in the chat room, which includes the voting results of each node and the final operation, so that users understand how the system handles their requests.

Ritual is currently developing a second product, the “Sovereign Chain Ritual Chain”. Although Infernet can be easily integrated into any EVM chain, making it available for any protocol to use. But Ritual still firmly believes that building a chain is necessary because it can build more efficient features at the core execution layer and consensus layer, and it can allow users who wish to maximize the value that AI brings to the protocol to realize their vision. Of course, to implement the sovereign chain, Ritual needs to build different types of validators, proof systems, and various complex features, and it needs to be simple enough for users to easily get started.

Virtual: more fun and focused on user participation

Virtual Protocol, unlike Bittensor and Ritual which interact with various machine models, is akin to a decentralized factory dedicated to creating AI characters for various virtual worlds. It focuses more on user involvement and incorporates human subjective thoughts and social consensus into its vision to promote the development of personalization and immersion. The core philosophy of Virtual Protocol is that future virtual interactions will be implemented by AI and built in a decentralized manner to provide personalized and immersive experiences. Personalization ensures that each interaction establishes a personal connection with the user, making it uniquely relevant. On the other hand, immersion can stimulate the user’s various senses, creating a more realistic experience.

Virtual ecosystem participants include contributors and validators. Contributors can provide various text data, voice data, and visual data to the models, whether it’s improvements to existing models or the proposition of new ones. These contents will be reviewed and certified by validators to ensure accuracy and authenticity, and to evaluate the quality of their contribution to ensure it meets the standards set by the Virtual Protocol ecosystem.

• New Proposals: Anyone can initiate the creation of Genesis Virtual, but it requires staking at least 100,000 VIRTUALs within a specified three months and going through the DAO proposal process. All token holders within the Virtual community can vote on this proposal. Once the proposal is approved, a new Virtual NFT will be minted.
• Contributions to Existing Models: A proposal will be automatically generated, and validators will review, discuss, verify, and vote on whether to make changes.

Currently, only validators have the right to verify or vote on proposals, and the entire verification process is anonymous. Validators need to interact with each model for at least 10 rounds. After completing the verification task, validators can get staking rewards in proportion to their total staking ratio. If a user wants to become a validator, they must hold 1,000 Virtual tokens in their Virtual account and commit to verifying all proposals. Additionally, Virtual uses a DPos mechanism, so if you want to get staking rewards without verifying, you can delegate any amount of tokens to a Virtual validator. Validators will return staking rewards to the delegation users after deducting a 10% income ratio.

The entire participation process in Virtual is transparent and recorded through a public blockchain. All contributions are transformed into NFTs and stored in an Immutable Contribution Vault (ICV) to ensure traceability and fair reward distribution. The Immutable Contribution Vault (ICV) is a multi-layer on-chain storage repository of Virtual that archives all Virtual-approved contributions on-chain. It can present the current state of each Virtual and track its historical evolution. In addition, by open sourcing the VIRTUALs codebase model, ICV creates a transparent environment. It promotes composability, allowing developers and contributors to build on the basis of existing VIRTUALs and seamlessly integrate with them.

The Virtual token is the core of the Virtual Protocol. Its main functions include rewarding contributors and validators, supporting protocol development, and carrying out airdrops, etc. The total supply of Virtual tokens is 1 billion, of which 60% is already in public circulation, 5% is reserved for liquidity pools, and the remaining 35% is dedicated to community incentive measures and initiatives for the development of the Virtual Protocol ecosystem. Also, in the next three years, the release amount will not exceed 10% each year and deployment requires approval from the management department.

Virtual Protocol drives a flywheel mechanism through income and incentives. The income is generated from the use of various dApps that pay usage fees to the protocol. At the end of each month, Virtual Protocol distributes incentives according to the total income flowing into the dApp. Out of this, 10% is allocated to the protocol, and the remaining 90% is divided among various Virtual applications based on their staking ratio. This ensures that income corresponds to contribution. For example: if the total income inflow is $100 dollars, 10 dollars will be distributed to the protocol. Of the remaining $90 dollars, since Virtual A’s staking pool has 9000 tokens, and Virtual B’s staking pool only has 1000 tokens, Virtual A will get 9090%= $81 dollars, while Virtual B will get 9010%= $9 dollars.

Within each Virtual application, the income is evenly distributed to the validators and contributors. Validators get income according to running time and staked amount, where running time refers to the ratio of the number of proposal verifications to the total number of proposals. For example, if Validator A’s running time in Virtual A is 90%, then it will get 81/2*90%= $36.45 dollars. Then the income will be further distributed to various stakeholders according to the staked amount. Furthermore, by default, 10% will be paid to the validator of the pool as a delegation fee. Contributors are then allocated income according to the contribution utilization rate and the impact pool. The contribution utilization rate considers the active usage time of the contributor’s contribution in the system. Contributors who develop and maintain the model will get 30% of the total distributed income, and users who provide and maintain datasets for model finetuning will get 70% of the total distributed income. The impact pool then awards points according to the importance of the contribution.

Currently, Virtual has been integrated into a virtual companion game called AI Waifu. The game’s story takes place in a world called “Arcadia”. In the game, you play as a wizard in Arcadia, needing to confront other wizards and their Waifus. You can choose to talk to your Waifu to deepen your connection, unlock hidden stories, and also get more rewards through giving gifts. Currently, there are three different Waifus to choose from in the game, each with a unique background story and personality. In addition, the game has also introduced a battle mode, where you can seduce other Waifus and protect your own Waifu. All expenditure within the game will enter the game’s reward pool, and 60% of the WAI transaction fee will also be distributed as part of the reward pool.

Unlike other AI companions and chatbots, AI Waifu is visually presented as a 3D model, and can emotionally and animately respond to voice and text. Through interaction with AI Waifu, she will not repeat content styles, but will keep learning and give players personalized responses. In addition, AI Waifu is a cross-platform PWA, with a crypto-enabled economic design, allows for collective ownership, and returns its expenditure as an income share to the developers.

Besides AI Waifu, Virtual also plans to launch a new AI RPG with cross-game memory and ultimate consciousness. These AI agents can dynamically evolve through interactions with players and other agents within the game. Users can train the agent in game A and retain the training memory. Then, when the agent is put into game B, it still retains the memory from game A. Through continuous learning, AI agents can mimic human player behavior, and can dynamically build based on player behavior and game environment changes. This can make the user’s game experience more diverse and personalized, and even include unpredictability. Users can also upload interaction records to get token rewards. In addition, Virtual also plans to launch a virtual idol that can live stream on any platform.

Conclusion

In the field of crypto AI, Bittensor, Ritual, and Virtual Protocol are each deeply cultivated in different areas. Bittensor aims to build a shared knowledge algorithm market, currently leading in market value in the crypto AI field. However, recent community members have raised some questions about its subnet registration fee and practicality. Whether the issues of individual subnets can be attributed to defects of the whole network needs further evaluation. Furthermore, regarding the issue that the system heavily relies on validators’ operations, a dynamic TAO solution “BIT001” was recently proposed by a contributor from the Opentensor Foundation.

With a powerful financing lineup and team background, Ritual has become a new favorite contestant on the encrypted AI track. Previously, Dragonfly partner Haseeb Qureshi stated in an article that the cryptoeconomics used by Ritual is the simplest and possibly the cheapest in the verifiable inference track, but there is a security problem of node collusion. However, a co-founder of Ritual subsequently explained on social platforms,stating that Ritual platform does not adopt a cryptoeconomics method based on node cooperation and selective collusion, but provides users with options to choose the security level according to their preferences.

In contrast, Virtual Protocol is more interesting and focuses on user participation. For example, the protocol launched the AI Waifu virtual companion game and is about to launch the game AI agent. Compared to the established game rules of traditional games, Virtual Protocol is dedicated to establishing an interactive relationship with players and hopes to evolve dynamically according to player behavior and the gaming environment, thereby increasing the social attribute and continuity of the game.

Of course, in addition to the three projects mentioned in this article, there are many Crypto AI projects on the market that are worth paying attention to, such as the GPU rental market-focused io.net, AI agent protocol Autonolas, and the Web3-enabled AI platform MyShell, which is dedicated to creators. These projects all show the diversity and potential of the Crypto AI field, and we will continue to closely follow the development of this field.

Disclaimer：

1. This article is reprinted from [ChainFeeds Research]. Forward the Original Title‘Bittensor 和这些 Crypto AI 网络值得关注吗？’. All copyrights belong to the original author [LINDABELL]. If there are objections to this reprint, please contact the Gate Learn team, and they will handle it promptly.
2. Liability Disclaimer: The views and opinions expressed in this article are solely those of the author and do not constitute any investment advice.
3. Translations of the article into other languages are done by the Gate Learn team. Unless mentioned, copying, distributing, or plagiarizing the translated articles is prohibited.