Digital Synergy: How the Merger of AI and Blockchain Will Play Out

In November 2022, OpenAI introduced the chatbot ChatGPT. Two months after its launch, the number of active users of the service reached 100 million. For comparison, it took TikTok about nine months to reach this mark and Instagram more than two years.

By then, generative AI was already quite popular, and the new product further fueled interest in the segment. Not surprisingly, the wave of hype around artificial intelligence (AI) did not bypass the cryptocurrency industry.

Since the end of 2022, it has periodically been possible to observe local rallies in the tokens of projects that, according to developers' assurances, used the technology. However, some are skeptical about such assets, believing, not without reason, that the degree of integration of any algorithms is at a low level.

However, while there are different and rather polarized views on existing products, there is some consensus on the potential synergy between blockchain and artificial intelligence.

Many players, including crypto exchanges and Web3 accelerators, believe that merging AI and blockchain would benefit both industries by allowing each to solve existing problems. Some venture capitalists hold a similar view. For example, in May 2023, it was reported that Paradigm would expand its interests in AI.

Why do AI algorithms need decentralization?

The narrative about the synergy between artificial intelligence and blockchain is nothing new. However, interest in the field has grown significantly over the past few years, as relevant research data indicates.

Integrating distributed networks with AI services has many long-term benefits for AI developers. Blockchain has the potential to eliminate or at least alleviate a number of critical barriers, such as those related to a lack of computing power.

This synergy also opens up access to innovative interoperability options. For example, DLT technology can enable fine-tuning neural models and collecting more representative datasets for training algorithms.

Integrating systems with artificial intelligence—especially on-chain and smart contracts—will also benefit the blockchain industry. AI can potentially improve the performance of distributed networks and become a major growth driver for the decentralized finance (DeFi) sector.

Everyone has problems

The history of artificial intelligence as a scientific field is almost 70 years old. However, the industry has never managed to remove some of the barriers that prevent its widespread adoption. Moreover, as the industry has evolved, new challenges have emerged.

Below, I detail some potential scenarios where distributed networks can offset certain limitations.

Power shortage

Graphics processing units (GPUs) are important in training algorithms and inferring user queries. This is particularly evident in Nvidia's report for Q1 FY 2024.

Against the backdrop of the growth of the AI sector, there was a sharp increase in demand for GPUs - leading to a significant shortage of microcircuits. The situation was so serious that major cloud service providers like Google and Amazon even started imposing restrictions on their customers.

Many companies involved in AI have turned to alternative suppliers (e.g., Lambda), but they were also close to their capacity limits.

Decentralized computing networks can fix the problem. They are de facto intermediaries connecting organizations that need computing power with system owners who have the necessary resources.

Such solutions offer lower prices compared to centralized service providers. This is mainly due to the absence of additional costs for the providers connected to the system.

There are two main types of such computing networks:

• General purpose (GP). These include decentralized cloud services that provide resources to various applications. Their operating model resembles a marketplace—customers rent server space from providers who set their own prices.
• Special purpose (SP). These, on the contrary, are designed for specific usage scenarios. Usually, their architecture combines computing resources into a single pool—a kind of decentralized computer. In this case, the cost of services is determined by market dynamics or parameters controlled by the community. An example here is Gensyn, which is designed to train ML models.

Decentralized networks democratize access to computing power. This reduces the cost of training, fine-tuning algorithms, and handling user queries, which are even more computationally intensive tasks.

However, the community is concerned about the speed of training ML models on a distributed resource. According to Mohamed Fouda, an Alliance member and partner at Volt Capital, it may be one or even two orders of magnitude slower than centralized methods.

The teams are already working on optimizing the decentralized learning process. The Together developers were creating a solution theoretically eliminating the bottleneck, while Gensyn was trying to alleviate issues arising from connecting different hardware to the network.

However, the community will likely have to compromise on slow learning to save money.

Machine learning with zero disclosure

I want to single out the projects focused on zero-knowledge machine learning (ZKML) separately.

Various mechanisms such as trusted execution environments (TEE) and reputation models are used to ensure correct operation in computing networks. But each approach has its own limitations and drawbacks. For example, a TEE may have a potential hardware attack vector.

Therefore, a new wave of projects (Gensyn, Modulus Labs, and Giza) have started experimenting with applying zero-knowledge proof (ZKP) to verify computational integrity for ML.

ZKP is a cryptographic protocol that allows one party (the proving party) to confirm the truth of an assertion to another party (the verifying party) without revealing any additional information. The protocol is quite popular in the blockchain industry as it allows developers to build scalable and secure applications.

When applied to machine learning, ZKP hides part of the input data or the model itself, if necessary. This is especially relevant when algorithms are working in highly regulated industries such as healthcare and finance.

ZKML has other advantages as well. The method, for example, allows proving that a particular algorithm has been trained on a strictly defined data set. In the case of proprietary AI, it also makes it possible to verify that the same model is available to all users.

A disadvantage of the approach is the process of generating the evidence itself—it is a resource-intensive task that can cost more to perform than the original operations. This means that, in some cases, it is impractical to compute it.

Nevertheless, ZKML is a vector of decentralization for the AI industry. It is important in a situation where the concentration of technology in the hands of a narrow pool of players is a cause for concern.

Content authentication

The development and proliferation of generative AI have led to the emergence of realistic deep fakes. Examples include fabricated images of Pope Francis in a Balenciaga down jacket and a video of a bombing scene near the Pentagon.

Cryptographic signatures can be used to combat such deep fakes—the identity of the content creator is verified by matching the private-public key pair. One example of implementation is decentralized social networks. Lens Protocol-based projects link user accounts to addresses in a public blockchain, simplifying identification.

The Bundlr and Arweave teams are also working on industry-wide standards. Arweave envisages introducing specifications requiring the integration of immutable cryptographic signatures and timestamps into digital content recorded in a distributed registry.

AI and blockchain synergy toward efficiency

In the long term, blockchain will improve the efficiency of neural model training and may change how the industry conducts research.

While most research on blockchain was conducted in academia in its early days, it is now dominated by large technology companies. This situation limits the participation of local laboratories and individuals due to the lack of incentives and opportunities for collaboration.

Decentralized platforms like Bittensor can fix things. These are marketplaces where participants are rewarded for their contribution to development and can share data to train models. Such platforms are especially attractive when creating open-source AI.

Blockchain also facilitates the application of Reinforcement Learning from Human Feedback (RLHF). This is a method that incorporates human feedback into the process to fine-tune the neural model.

RLHF allows you to “polish” the model, reducing the number of inaccurate or biased results. For example, OpenAI used it to debug GPT-3 and to develop ChatGPT.

Fine-tuning increases algorithms' performance and enables them to gain domain-specific expertise. As the demand for such highly specialized models grows, so does the need for experts to provide feedback.

Multicoin proposes a way to scale the RLHF through incentive payments in the form of tokens. However, this approach has at least two problems:

• Experts must agree to accept tokens as compensation, which limits the range of individuals involved in the learning process.

• Such a system needs to be protected from manipulative attacks to maintain the accuracy of the feedback.

  
  

However, projects like Hivemapper have already put the method into practice.

How intelligent can a smart contract be?

There are many areas where blockchain platforms can utilize artificial intelligence at a variety of levels, from infrastructure to application.

However, the scenarios of greatest interest to the cryptocurrency industry are those in which AI operates directly in a distributed ledger. In a general sense, there are two ways to move the activity of algorithms to the blockchain:

• Allow autonomous agents to use the Web3 infrastructure to make payments and access digital resources.
• Allow neural models to interact with smart contracts via ZKML.

Interesting, isn't it?

Payments and resources for autonomous agents

Autonomous economic agents (AEAs) are autonomous systems based on machine learning algorithms that perform specific tasks on behalf of their owners without their direct intervention in the process.

Experts expect that as technology advances, AEAs will become more highly specialized, leading to the proliferation of “multi-agent systems”.

This, in turn, will entail the emergence of a market in which some agents can “hire” others and pay them remuneration for performing certain tasks. In this context, cryptocurrency payments will likely be preferable to fiat payments for several reasons:

• Existing laws do not allow AEAs to open bank accounts or use account holders' accounts to conduct cross-border transactions.
• Cryptocurrencies eliminate the need for trusted intermediaries.
• Blockchain infrastructure enables faster and cheaper settlements.

AEAs will be able to interact with payment and decentralized physical infrastructure networks (DePIN). DePINs integrate hardware devices—the computing systems discussed above can also be attributed to this segment.

DePINs will give AI access to digital resources like disk space and computing power. For example, if an algorithm needs to create a 3D model, it can use the Render Network for rendering and Arweave for data storage instead of relying on centralized solutions.

Integrating AI into smart contracts

The application of AI models in smart contracts significantly expands their capabilities. Neural networks will not only open access to innovative use cases but also increase the efficiency of existing tools.

Much of this integration is hindered by the high computational costs associated with deploying algorithms in the blockchain. However, using ZKP to validate the exact execution of off-chain models could solve this problem, as only relevant evidence can be placed in a distributed registry.

Such an approach will allow smart contracts to make decisions based on dynamic data without being limited to a set of hard-coded rules. In this way, they will become more autonomous, flexible, and sophisticated.

ZKML can be used in multiple industry sectors, including DeFi, GameFi, DeSo (Decentralized Social), and DePIN.

For example, in decentralized financial applications, AI can adjust protocol parameters based on current network parameters. One possible use case is a lending protocol that uses an ML model to adjust the collateral factor in real time.

Other scenarios include automated treasury management, credit on-chain scoring, and AMM liquidity management.

Conclusion

Currently, there is a contradiction between the AI and Web3 industries at the basic logic level: the former is highly centralized. At the same time, the latter is built on the principles of widespread decentralization. At times, this situation makes it difficult to integrate applications.

However, the same contradiction allows products from these two sectors to effectively complement each other and promote mutual development.

There is no guarantee that blockchain will be the foundation for future neural models or that algorithms will run at the core of decentralized platforms.

But it is safe to say that the combination of the two technologies will produce many new narratives, some of which will prove quite viable.