AI Layer1 Research Report: Finding Fertile Ground for on-chain DeAI

Overview

Background

In recent years, leading tech companies such as OpenAI, Anthropic, Google, and Meta have been continuously driving the rapid development of large language models. These models have demonstrated unprecedented capabilities across various industries, greatly expanding the realm of human imagination, and even showing potential to replace human labor in certain scenarios. However, the core of these technologies is firmly held in the hands of a few centralized tech giants. With strong capital and control over expensive computing resources, these companies have established insurmountable barriers, making it difficult for the vast majority of developers and innovation teams to compete with them.

At the same time, in the early stages of the rapid evolution of AI, public opinion often focuses on the breakthroughs and conveniences brought by technology, while the attention to core issues such as privacy protection, transparency, and security is relatively insufficient. In the long run, these issues will profoundly affect the healthy development of the AI industry and its societal acceptance. If not properly addressed, the debate over whether AI will be "beneficial" or "malicious" will become increasingly prominent, and centralized giants, driven by profit motives, often lack sufficient incentives to proactively tackle these challenges.

Blockchain technology, with its decentralized, transparent, and censorship-resistant characteristics, offers new possibilities for the sustainable development of the AI industry. Currently, numerous "Web3 AI" applications have emerged on some mainstream blockchains. However, a close analysis reveals that these projects still face many issues: on one hand, the degree of decentralization is limited, as key processes and infrastructure still rely on centralized cloud services, making it difficult to support a truly open ecosystem; on the other hand, compared to AI products in the Web2 world, on-chain AI still shows limitations in terms of model capabilities, data utilization, and application scenarios, with the depth and breadth of innovation needing improvement.

To truly realize the vision of decentralized AI, enabling blockchain to securely, efficiently, and democratically support large-scale AI applications and compete with centralized solutions in terms of performance, we need to design a Layer 1 blockchain specifically tailored for AI. This will provide a solid foundation for open innovation in AI, democratic governance, and data security, promoting the prosperous development of a decentralized AI ecosystem.

Core features of AI Layer 1

AI Layer 1, as a blockchain specifically tailored for AI applications, has its underlying architecture and performance design closely aligned with the needs of AI tasks, aiming to efficiently support the sustainable development and prosperity of the on-chain AI ecosystem. Specifically, AI Layer 1 should possess the following core capabilities:

1. Efficient Incentives and Decentralized Consensus Mechanism The core of AI Layer 1 lies in building an open network for sharing resources such as computing power and storage. Unlike traditional blockchain nodes that mainly focus on ledger bookkeeping, the nodes in AI Layer 1 need to undertake more complex tasks, not only providing computing power and completing AI model training and inference but also contributing diversified resources such as storage, data, and bandwidth. This is aimed at breaking the monopoly of centralized giants in AI infrastructure. This places higher demands on the underlying consensus and incentive mechanisms: AI Layer 1 must be able to accurately assess, incentivize, and verify the actual contributions of nodes in tasks such as AI inference and training, achieving network security and efficient resource allocation. Only in this way can the stability and prosperity of the network be ensured, while effectively reducing overall computing power costs.

2. Exceptional high performance and support for heterogeneous task capabilities AI tasks, especially the training and inference of large language models, place extremely high demands on computing performance and parallel processing capabilities. Furthermore, the on-chain AI ecosystem often needs to support diverse and heterogeneous task types, including various model structures, data processing, inference, storage, and other diversified scenarios. AI Layer 1 must deeply optimize its underlying architecture to meet the demands of high throughput, low latency, and elastic parallelism, and provide native support for heterogeneous computing resources, ensuring that various AI tasks can operate efficiently and achieve smooth expansion from "single-type tasks" to "complex diversified ecosystems."

3. Verifiability and Reliable Output Guarantee AI Layer 1 not only needs to prevent security risks such as model malfeasance and data tampering but also must ensure the verifiability and alignment of AI output results from the underlying mechanism. By integrating cutting-edge technologies such as trusted execution environments, zero-knowledge proofs, and secure multi-party computation, the platform allows every model inference, training, and data processing process to be independently verified, ensuring the fairness and transparency of the AI system. At the same time, this verifiability helps users clarify the logic and basis of AI output, achieving "what is obtained is what is desired," thereby enhancing user trust and satisfaction with AI products.

4. Data Privacy Protection AI applications often involve sensitive user data, and in fields such as finance, healthcare, and social interactions, data privacy protection is particularly critical. AI Layer 1 should ensure verifiability while employing encrypted data processing technologies, privacy computing protocols, and data permission management methods to ensure the security of data throughout the entire process of inference, training, and storage, effectively preventing data leakage and misuse, and alleviating user concerns regarding data security.

5. Powerful ecological support and development capabilities As an AI-native Layer 1 infrastructure, the platform not only needs to possess technological leadership but also must provide comprehensive development tools, integrated SDKs, operational support, and incentive mechanisms for ecosystem participants such as developers, node operators, and AI service providers. By continuously optimizing platform usability and developer experience, the platform promotes the landing of diverse AI-native applications, achieving the sustained prosperity of a decentralized AI ecosystem.

Based on the above background and expectations, this article will provide a detailed introduction to six representative AI Layer 1 projects, including Sentient, Sahara AI, Ritual, Gensyn, Bittensor, and 0G. It will systematically outline the latest developments in the field, analyze the current status of the projects, and discuss future trends.

Sentient: Build a loyal open-source decentralized AI model

Project Overview

Sentient is an open-source protocol platform that is building an AI Layer 1 blockchain (, initially as Layer 2 and later migrating to Layer 1). By combining AI Pipeline and blockchain technology, it aims to construct a decentralized artificial intelligence economy. Its core goal is to address model ownership, invocation tracking, and value distribution issues in the centralized large language model market through the "OML" framework (Open, Monetizable, Loyal), enabling AI models to achieve on-chain ownership structure, invocation transparency, and value sharing. Sentient's vision is to empower anyone to build, collaborate, own, and monetize AI products, thereby promoting a fair and open AI Agent network ecosystem.

The Sentient Foundation team brings together top academic experts, blockchain entrepreneurs, and engineers from around the world, dedicated to building a community-driven, open-source, and verifiable AGI platform. Core members include Princeton University professor Pramod Viswanath and Indian Institute of Science professor Himanshu Tyagi, who are responsible for AI safety and privacy protection, while Polygon co-founder Sandeep Nailwal leads the blockchain strategy and ecosystem layout. Team members come from renowned companies such as Meta, Coinbase, and Polygon, as well as top universities like Princeton University and the Indian Institutes of Technology, covering fields such as AI/ML, NLP, and computer vision, working together to drive project implementation.

As a secondary venture of Sandeep Nailwal, co-founder of Polygon, Sentient was born with a halo, possessing rich resources, connections, and market recognition, providing strong backing for the project's development. In mid-2024, Sentient completed a $85 million seed funding round, led by Founders Fund, Pantera, and Framework Ventures, with other investment institutions including dozens of well-known VCs such as Delphi, Hashkey, and Spartan.

Design Architecture and Application Layer

Infrastructure Layer

Core Architecture

The core architecture of Sentient consists of two parts: AI Pipeline and on-chain system.

The AI pipeline is the foundation for developing and training "Loyal AI" artifacts, consisting of two core processes:

• Data Curation: A community-driven data selection process for model alignment.
• Loyalty Training: Ensures that the model maintains a training process in line with community intentions.

The blockchain system provides transparency and decentralized control for protocols, ensuring ownership, usage tracking, revenue distribution, and fair governance of AI artifacts. The specific architecture is divided into four layers:

• Storage layer: stores model weights and fingerprint registration information;
• Distribution Layer: Authorized contract controls the entry point of the model calls;
• Access Layer: Verifies whether the user is authorized through permission proofs;
• Incentive Layer: The yield routing contract will allocate payments to trainers, deployers, and validators for each call.

OML Model Framework

The OML framework (Open, Monetizable, Loyal) is the core concept proposed by Sentient, aiming to provide clear ownership protection and economic incentive mechanisms for open-source AI models. By combining on-chain technology and AI-native cryptography, it has the following characteristics:

• Openness: The model must be open-source, with transparent code and data structures, making it easy for the community to reproduce, audit, and improve.
• Monetization: Each model call triggers a revenue stream, and the on-chain contract distributes the earnings to the trainers, deployers, and validators.
• Loyalty: The model belongs to the contributor community, with upgrade direction and governance determined by the DAO, and usage and modification controlled by encryption mechanisms.

AI-native Cryptography

AI-native encryption utilizes the continuity of AI models, low-dimensional manifold structures, and the differentiable characteristics of models to develop a "verifiable but non-removable" lightweight security mechanism. Its core technology is:

• Fingerprint embedding: During training, a set of covert query-response key-value pairs is inserted to form a unique signature for the model;
• Ownership Verification Protocol: Verifying whether the fingerprint is retained through a third-party detector (Prover) in the form of a query.
• Permission call mechanism: Before calling, you need to obtain a "permission certificate" issued by the model owner, and the system will then authorize the model to decode the input and return the accurate answer.

This method can achieve "behavior-based authorization calls + ownership verification" without the cost of re-encryption.

Model Rights Confirmation and Security Execution Framework

Sentient currently adopts Melange mixed security: combining fingerprint authorization, TEE execution, and on-chain contract profit sharing. The fingerprint method is implemented as OML 1.0, emphasizing the "Optimistic Security" concept, which assumes compliance and allows for detection and punishment after violations.

The fingerprint mechanism is a key implementation of OML, which generates a unique signature during the training phase by embedding specific "question-answer" pairs. Through these signatures, the model owner can verify ownership and prevent unauthorized duplication and commercialization. This mechanism not only protects the rights of model developers but also provides a traceable on-chain record of the model's usage behavior.

In addition, Sentient has launched the Enclave TEE computing framework, which utilizes a trusted execution environment to ensure that the model only responds to authorized requests, preventing unauthorized access and usage. Although TEE relies on hardware and has certain security risks, its high performance and real-time advantages make it a core technology for current model deployment.

In the future, Sentient plans to introduce zero-knowledge.