Interpreting the PlatON 2.0 White Paper: How to Achieve a Decentralized General Artificial Intelligence Network?

Source: PlatON

Author: 0x007

The universal application of artificial intelligence faces three major issues. The first is data. Data is the most important resource for artificial intelligence, which requires vast amounts of effective datasets to train better models. However, under the premise of data privacy and regulation, issues such as data usage rights and secure usage need to be addressed.

The second issue is training costs. The scale of artificial intelligence models is growing at a rate of ten times per year, requiring enormous computational power, which leads to a continuous rise in the total training costs of artificial intelligence. The third issue is centralization. Most artificial intelligence research is controlled by a few tech giants, while other organizations face a lack of AI talent and technology. At the same time, AI developers lack methods to monetize their results and can only sell their technology to tech giants.

The PlatON 2.0 white paper presents a plan to establish an infrastructure network using blockchain and privacy computing technology. In this network, developers can obtain resources, including data and computing power, at low cost, train artificial intelligence models, and publish them as AI services on the network. These services can interact and combine with other AI services or agents, and anyone or any organization can access AI algorithms or services from this network. As this infrastructure network develops, it will nurture a prosperous decentralized AI market where people can trade AI-related "goods," such as data, computing power, AI algorithms, and AI services.

Furthermore, this self-organizing decentralized collaborative network could potentially become a whole that is greater than the sum of its parts, connecting artificial intelligence and enabling them to learn collaboratively, ultimately leading to the emergence of general artificial intelligence.

PlatON aims to achieve the above goals in three phases:

• Phase One: Decentralized Privacy Computing Network. Establish a decentralized data sharing and privacy computing infrastructure network that connects data owners, data users, algorithm developers, and computing power providers.
• Phase Two: Decentralized AI Market. Achieve co-construction and sharing of AI assets, agile development of intelligent applications, and provide a full-process product and service from AI computing power and algorithms to AI capabilities and their production, deployment, and integration.
• Phase Three: Decentralized Collaborative AI Network. Allow artificial intelligence to collaborate on a large scale, gathering collective wisdom to execute complex AI services.

Privacy Computing Network

The basic elements of computing are data, algorithms, and computing power. The privacy computing network tightly integrates data, algorithms, and computing power to build a complete computing ecosystem.

In the privacy computing network, data nodes and computing nodes are connected to the system via P2P protocols, publishing data and computing power, which are generally stored locally. By utilizing data and computing power, collaborative computing on algorithms is performed through technologies such as secure multi-party computation and federated learning. Data is usable but not visible, ensuring that both the privacy of the data and the privacy of the computation results, such as the trained AI models, are protected.

The technical architecture of PlatON's privacy computing network is shown in the diagram below:

Data subjects can initiate data nodes locally or encrypt and host data at data nodes, as shown in the data service section of the diagram above. When a data node receives a computation request, PlatON supports two types of privacy computing: secure multi-party computation and privacy outsourcing computation.

In the secure multi-party computation approach, data nodes use secret sharing to split data and distribute it to randomly selected computing nodes. The computing nodes perform privacy computation using secure multi-party computation protocols, and the computation results are returned to the computation requester via blockchain smart contracts. If it involves training an AI model, the completed model can be deployed to the AI network and published as an AI service.

In the privacy outsourcing computation approach, data nodes encrypt data using homomorphic encryption and distribute it to computing nodes for outsourced computation. The computation tasks can be decomposed based on data or models, and after the computing nodes complete the computation, they return the computation results and proofs, which can verify the correctness of the computation. If users have their own data and algorithms but lack sufficient computing power, they can use this method.

AI Market

Through smart contracts on the blockchain, a decentralized trading market for data, computing power, and algorithms can be constructed. Based on the cryptoeconomics of the blockchain, data, computing power, and algorithms can be monetized, forming effective incentive mechanisms to encourage more data, algorithms, and computing power to join the network.

For data providers, entities including individuals and institutions will provide personal and professional data due to economic incentives, and secure computing ensures the safety and privacy of the data. Various entities will be more willing to share sensitive data, such as consumption and health information. Over time, the market will accumulate high-quality data.

For computing power providers, anyone can share computing resources in a secure and frictionless market. As the scale of AI models grows larger, sharing idle computing resources from around the world through the AI market can achieve a decentralized computing power network, theoretically providing unlimited computing power for AI and truly reducing computing costs.

For AI developers, they can actively search for training datasets in the data market to train AI models, or they can publish models and allow others to provide data for collaborative training. The training datasets are not exchanged in plain text but participate in model training through secure multi-party computation protocols, allowing for fair trading where no party can gain an advantage through premature withdrawal or other improper behaviors.

AI developers can trade AI algorithms and AI services in the AI market, directly monetizing their results and being incentivized and encouraged to create better AI models. The AI models developed by them can also interact with other AI models and paying users.

For AI users, they can conveniently and cost-effectively obtain and use AI services.

Collaborative AI Network

By utilizing the datasets and computing resources from the privacy computing network, AI models can be trained, which can be deployed in the AI network and provide services externally through AI agents, forming an AI service market. Through technologies such as multi-agent systems, AI agents can communicate and collaborate, creating an increasing number of innovative AI services and realizing AI DAOs, forming a self-governing collaborative AI network.

The technical architecture of PlatON's collaborative AI network is shown in the diagram below:

Service nodes in the collaborative AI network are used to host trained AI models and provide AI services externally. Registration nodes and evaluation nodes form an intelligent search network to search for and interact with AI services and agents. Specifically, AI services and agents register their textual descriptions and tags with registration nodes so that users can discover their services, pricing, addresses, and other information and invoke them. Evaluation nodes conduct service testing, evaluation, and rating for AI services and agents, establishing a reputation scoring system through consensus algorithms, which serves as the basis for searching and recommending, enabling other users to quickly and easily query AI services and agents.

Autonomous AI agents can autonomously search for and invoke AI services or interact with other autonomous AI agents, continuously learning and improving, adjusting strategies and goals. They are software programs representing humans in this self-organizing intelligent network space to achieve certain goals, possessing a degree of independence or autonomy without direct human intervention.

The collaborative AI network consists of many interacting autonomous AI agents, meaning it is a multi-agent system. Multi-agent systems have been used in various application fields, including personal assistants, traffic management, gaming, and virtual characters. For example, the AI assistant Siri is a simple example of an autonomous agent that uses sensors to perceive user requests and automatically collects data from the internet to satisfy user requests without user assistance.

Autonomous AI agents exist not only in the digital world but can also serve as a bridge between the digital world and the real world, connecting to humans, IoT devices, and external IT systems. Each autonomous agent operates as an independent daemon, each pursuing its relatively simple goals, but their interactions will generate complex goals, resulting in more intelligent higher-level agents.

PlatON plans to launch a decentralized privacy computing network in the fourth quarter of 2021, connecting data, algorithms, and computing power through privacy computing protocols, gradually forming an AI market. In the fourth quarter of 2022, the collaborative AI network will be launched, ultimately forming a self-organizing collaborative AI network based on this network.

Technically, PlatON comprehensively utilizes blockchain, privacy computing, and artificial intelligence technologies. At the end of the article, the core technical features are briefly introduced, mainly including:

1. Decentralization. Any user or node can connect to the network without permission, combined with decentralized digital identity authentication and authorization, allowing secure sharing, connection, and trading of any data, algorithms, and computing power globally, enabling anyone to develop and use artificial intelligence.

2. Privacy protection. Based on modern cryptographic privacy computing technologies such as MPC, homomorphic encryption, and zero-knowledge proofs, a new paradigm of computation is provided, making data and models usable but not visible, ensuring complete protection of privacy and safeguarding data rights.

3. Low training costs. Artificial intelligence requires a large amount of computing power and training data, leading to high training costs. With the help of blockchain and privacy computing technologies, computing resources can be shared to reduce computing costs; secure data sharing can be achieved to promote compliant circulation of data, accumulating more and better quality data at lower costs than tech giants through decentralization.

4. Low development threshold. Visualized AI model development and debugging, automated machine learning (AutoML), and simplified full-process management of AI models from development, training to deployment through MLOps reduce the development threshold of AI models and improve development efficiency. AI services can automatically discover, combine, and collaborate, creating an increasing number of innovative AI services through higher-level automated programming forms.

5. Regulatory compliance. All data, variables, and processes used in the AI training decision-making process have immutable records that can be tracked and audited. The use of privacy protection technologies ensures that data usage complies with regulations such as the right to be forgotten, portability, conditional authorization, and data minimization.

The above is my personal interpretation of the PlatON 2.0 white paper. If you want to learn more, the best way is to read the white paper, which can be found here.