Synergies and conflicts between blockchain and artificial intelligence

In recent years, the blockchain field has been full of gongs and drums, and the artificial intelligence (AI) field has been full of firecrackers, but these two transformative technologies seem to have little intersection. However, from a conceptual point of view, blockchain and AI have many complementary aspects. For example, the inherent decentralized nature of blockchain technology may help solve the centralization problem of AI, and the transparent and verifiable nature of blockchain may help solve the opacity problem of AI models.

Some time ago, the concept of "blockchain x AI" hype caused a large increase in the market value of many related cryptocurrencies, and the market value of the entire sector once exceeded 20 billion US dollars, as shown in Figure 1. This shows that the market is quite optimistic about this combination and investors are quite confident.

Figure 1: Market value of cryptocurrency in different segments, data as of April 18, 2024

However, the integration practice of blockchain and AI has also exposed some conflicts between the two. For example, AI requires intensive computing and large amounts of storage, while the distributed ledger architecture of blockchain emphasizes redundancy - each node stores and calculates the same information.

Recently, a research team from Tsinghua University and Fraunhofer HHI and other institutions published a paper "Blockchain and Artificial Intelligence: Synergies and Conflicts", analyzing the technical synergies and conflicts between blockchain and AI. It is worth noting that the team did not bury their heads in analyzing theories, but instead focused on the cryptocurrency market, analyzing "blockchain x AI" projects with a market value of more than $10 million and some specific use cases.

Let's take a look at what this paper says and what interesting or useful insights we get.

Blockchain x AI: Synergies and Conflicts

Figure 2 shows the complementarity and opposition between blockchain and AI.

The synergy between blockchain and AI

Decentralization and centralization. The training and maintenance of the current best large language models (LLMs) such as GPT require a lot of computing power, electricity, and data resources. For example, the computing cost of the training process of GPT-3 released in 2020 is about 4.6 million US dollars. Such high costs have made AI large models a battleground for a few large technology companies, which have basically become monopolies in the AI ​​market. Such monopolies may hinder competition, which is also a concern often expressed by policymakers in regions such as the United States and Europe, which are very active in enforcing antitrust laws to maintain market balance and prevent the market from being dominated by a single entity. In contrast, blockchain technology is decentralized; this property may be used to address the centralization problem of AI systems. If implemented properly, the decentralized nature of blockchain prevents any one party from controlling the entire network. This property can help implement some kind of regulatory mechanism within AI systems, achieve a more balanced distribution of power, and promote collaboration among all parties. Therefore, the integration of blockchain technology is expected to resolve the debate about regulation and monopoly in the field of AI, making AI governance more inclusive and fair.

Transparency and Black Box Nature. Another major feature of blockchain technology is transparency, and transactions and records on it are verifiable and cannot be tampered with. On the other hand, AI is like a black box - it is difficult for people to understand the reasoning behind its decisions. Perhaps we can use blockchain ledgers to record the decision-making process of AI and achieve a transparent audit trail, thereby improving the credibility of AI applications. In addition, blockchain can also integrate advanced cryptographic techniques (such as zero-knowledge proofs such as zk-SNARK) or use secure hardware (such as Trusted Execution Environments/TEE). These technologies can help verify that specific computational steps are executed faithfully and accurately.

Data management and dependencies. Blockchains can regulate data and data access through protocols such as smart contracts and the InterPlanetary File System (IPFS).

Open source and closed source. Blockchains can address the limitations of proprietary AI models by enabling shared ownership through cryptographic protocols, which in turn enables fine-grained privacy configuration. Transparency in AI development would be greatly improved if shared AI systems (jointly trained and controlled by participants) could perform at the level of commercial models. This would also promote the creation of more fair and comprehensive AI solutions.

Conflicts between blockchain and AI

Despite the above synergies between blockchain and AI, there are significant conflicts in their operational requirements that hinder the integration of the two.

Computational cost and load. For large language models (LLMs) such as GPT-4 and Llama 3, training and inference require a lot of computing resources. Blockchain's consensus mechanism, cryptographic operations, and unfavorable data structures all increase the computational burden, which affects scalability.

Storage limitations and data density. The decentralized nature of blockchains, while ensuring security and redundancy, also results in significant storage requirements, which is undoubtedly costly and inefficient for data-driven AI systems. In general-purpose blockchain systems (GBPS) such as Ethereum, every node must store all information, because redundancy ensures the security and resilience of the blockchain network, but is not conducive to scalability. Since new data on the Ethereum Virtual Machine (EVM) is stored in a transaction format, common data on the EVM structure may hinder the speed of retrieval. On the other hand, AI applications generate and process large amounts of data, which requires efficient and scalable storage solutions.

Pseudo-anonymity and security challenges. Blockchains allow permissionless, pseudo-anonymous access through asymmetric encryption; and the network can be protected by setting up computational or financial barriers against possible Sybil attacks. In addition, some use cases use blockchain as a platform for enhanced privacy protection and distributed AI training, using technologies such as federated learning; if these use cases support pseudo-anonymous participation in the training process, risks may arise.

These methods are vulnerable to adversarial federated learning attacks, and it is very difficult to identify malicious attackers because the contributions submitted to the overall AI model are private and difficult to measure by design.

Operation mismatch. Most blockchain virtual machines use fixed ledger operations to ensure that the results are deterministic - this is important, after all, financial transactions involve money. Floating-point operations may cause precision loss in calculations, especially when calculating multiple values ​​with large orders of magnitude. However, a common practice in AI training is to normalize floating-point parameters to between 0 and 1, because this helps to achieve stable and efficient gradient flow and provides implicit regularization, thereby improving the overall training effect.

Blockchain X AI: Use Case Study

Based on the above synergies and conflicts between blockchain and AI, we can look at use cases. The team investigated some of the projects that have done the best job in integrating blockchain and AI. They focused on projects that have existing products, issued tokens, and have a market value of more than $10 million. In addition, there are some projects with a market value of less than 10 million US dollars but novel use cases. They classified these projects based on three research questions:

1. To what extent are blockchain and AI technologies integrated in the project?

2. What is the role of blockchain in the project?

3. What is the role of AI in the project?

The cluster analysis results are shown in Figure 3, which contains 4 main clusters: AI is a peripheral technology of blockchain, AI participates in blockchain, blockchain manages AI processes, and blockchain is the core infrastructure of AI.

AI is a peripheral technology of blockchain

AI can help improve the user experience of interacting with blockchain, realize intelligent analysis, simplify the development process of blockchain applications, etc.

AI participates in blockchain

AI can actively participate in the blockchain ecosystem and governance structure. The team gave two exploration directions in the paper: one is to let AI agents join the distributed network as participants or stakeholders, such as letting AI speculate on Dex by itself; the other is to let AI participate in the governance of DAO (decentralized autonomous organization), but it is still difficult at present.

Blockchain manages AI processes

Nowadays, people are increasingly using blockchain technology to manage AI processes, creating a decentralized framework for resource sharing, data management, and application deployment.

Blockchain is the core infrastructure of AI

General purpose blockchain systems (GPBS) such as Ethereum face this triangular trade-off of scalability, security, and decentralization.

Take Ethereum as an example, its security is guaranteed by thousands of nodes scattered around the world, and the number of validators has exceeded 1 million. In order to reach consensus and finality, each new block of information must reach and be verified by every node in this global network, and each block is in the kilobyte range and is created every 12 seconds, which results in high storage and computing costs.

Therefore, it is unrealistic to execute or store computationally intensive AI operations directly on the chain; but now Layer2 rollup is becoming a more popular paradigm. Simply put, Layer2 rollup means processing transactions off-chain, and then aggregating the results before recording them on-chain; this solution can both increase throughput and reduce costs, making it cost-effective.

Similarly, blockchains developed specifically for AI use cases must (1) overcome challenges related to high computing and storage costs, public access, and limitations of the underlying virtual machine and (2) use blockchains as a pure management, governance, and security layer. Table 1 shows new systems that use blockchains as core infrastructure.

Table 1: Using blockchain as infrastructure for AI, where DAI = Distributed Artificial Intelligence, BC = Blockchain, DT/FL = Distributed Training/Federated Learning, C-Layer = Computation Layer, TA = Technical Analysis, DM = Distributed Management, PoS = Proof of Stake, DPoS = Delegated Proof of Stake, dBFT = Delegated Byzantine Fault Tolerance, FL = Federated Learning, DID = Decentralized Identity, ZK = Zero Knowledge, DePIN = Decentralized Physical Infrastructure Network, DC = Distributed Computing, DD = Distributed Data, ASBS = Application Specific Blockchain System, IPFS = Interplanetary File System, * = Low maturity / no public code, ? = Information not available.