1. Introduction
1.1 Preface
In the previous report, we delved into the development history and strategic vision of ICP. At a recent technical forum held in Argentina, Dominic Williams, the founder of ICP and President of Dfinity, delivered a keynote speech, sharing his vision for the future of decentralized computing and sparking widespread discussion within the industry. In his speech, he emphasized how ICP breaks boundaries by combining the power of blockchain with artificial intelligence, allowing people to see how these technologies can change the way we build and execute applications in the future. Additionally, according to the latest announcement from ICP, the DFINITY Foundation of ICP will establish a new office in San Francisco, centered in Silicon Valley, to develop AI, distributed computing power, and other technologies, and carry out strategic planning.
Based on this background, we will delve into the technical logic and ecological layout of how ICP empowers AI in this article, analyzing its key breakthroughs in technical architecture, practical applications, and future development directions.
Today, many people are already using ChatGPT to explore creativity, access information, and analyze and create content. ChatGPT is a large language model (LLM) with a massive number of parameters, trained through a large amount of data. These models have propelled AI applications to new heights. In fact, before applications like ChatGPT emerged, there were already Web2 products related to AI algorithms. One of the earliest large-scale manifestations was services like TikTok and Instagram Reels. They are not just traditional social media services, but are actually driven by powerful AI engines. AI accurately recommends content by analyzing video content and user interaction behavior (such as user viewing duration), thereby enhancing user experience. This personalized service makes them highly attractive. Moreover, this trend is expanding in broader directions.
However, as AI rapidly develops across multiple industries, problems with AI applications in traditional Web2 architecture are starting to emerge. For example, these AI models are mostly centralized, which means they rely on single-point control or a limited number of nodes to execute, thus facing many limitations such as data privacy, centralized computing resources, and lack of transparency. Additionally, traditional IT infrastructure (such as AWS) involves tedious configuration and maintenance processes, from cloud account registration, server configuration, to database installation, security updates, etc., which are not only time-consuming and labor-intensive but also prone to errors and insecure by default. Moreover, the upgrade cycle of traditional IT is long and complex, making it difficult to support the real-time requirements of new modes.
The emergence of Web3 brings new opportunities for the development of AI. With features such as decentralization, transparency, and autonomy, Web3 can achieve smart contract automation, promote the democratization of AI, enhance interoperability, promote fair governance, and improve network security. Some platforms have already started to address these issues. For example, Vercel (a provider of cloud platform as a service) provides AI application services by customizing infrastructure. These platforms can alleviate some problems but are still far from ideal. More importantly, software generated by AI will be bound to these dedicated platforms, and relevant data may also be controlled by these platforms, resulting in users being locked into their ecosystems (customer lock-in). In other words, applications and services in this model cannot achieve decentralization.
The world is beginning to realize that to truly unleash the potential of combining AI with Web3, a powerful infrastructure is needed. In this context, ICP is searching for new solutions. In the traditional AI environment, training models are like conductors controlling an orchestra to create beautiful music. Here, the conductor is the central server, which needs to process a large amount of data and possess powerful computing capabilities. This is similar to how companies like OpenAI train large models on large central servers.
However, in ICP, the training method for AI models is different. It is not a single central conductor controlling everything, but rather, each participant is both a conductor and a musician, collaborating with each other to complete tasks. This means that every node or device in the network can contribute to the training, decision-making, and execution of artificial intelligence models.
Decentralized AI models on ICP have the following advantages compared to centralized AI models:
Trust and transparency: Decentralized Artificial Intelligence (DeAI) models on ICP are fully executed on the chain, possessing immutability and openness. Users do not need to blindly trust centralized servers but can verify the training and inference processes of the models. This addresses a key issue in centralized AI, where users often cannot understand how data is used and how models behave.
Data security and control: ICP enables secure access to data from different blockchains through Chain Fusion technology (details to be explained later), providing advantages in data security and control. Users can retain ownership of their data while allowing DeAI models to access and learn from the data, which is different from the centralized AI system where data is usually centrally managed.
Censorship resistance: AI models executed on ICP can resist censorship. Centralized AI models may be subject to control and manipulation by operators. DeAI on ICP provides a more open and censorship-resistant platform, facilitating fair development and deployment of artificial intelligence.
Scalable packageability: Compared to traditional AI systems, DeAI on ICP has stronger scalability. Through connected node networks, DeAI can flexibly scale packages and perform parallel processing tasks, thereby enhancing overall capacity while maintaining high levels of security and performance.
Inclusiveness: ICP provides permissionless and composable access, promoting inclusiveness and fairness. Individuals and small companies can also participate in the development and decision-making of artificial intelligence, encouraging innovation and collaboration.
ICP transforms the internet into a massive decentralized world computer, where computation is conducted securely through node machine networks. By combining decentralized node hardware, ICP provides developers with the ability to host and build software applications without the limitations of traditional cloud services. With decentralized infrastructure, AI hosted on ICP is fundamentally immune to network attacks, meaning sensitive data is always protected from intrusion and tampering.
In fact, ICP has already demonstrated its capabilities by executing neural networks for tasks such as image classification and facial recognition. Furthermore, the capabilities of larger AI models such as Llama 3 executed on ICP open the doors to more advanced AI-driven applications. AI executed on the public network is immune to network attacks, runs uninterrupted, and remains accessible at all times. This concept represents a fundamental shift in how AI is deployed in critical areas such as healthcare, finance, and government within human society.
1.2 Underlying Technologies Supporting AI on ICP