01 Key Takeaways
The integration of AI and cryptocurrency is undergoing a paradigm shift from infrastructure development to the practical application of AI Agents. Early projects primarily focused on adapting traditional AI concepts to blockchain but largely failed to achieve significant breakthroughs due to technical limitations and an immature ecosystem.
The current AI x Crypto ecosystem can be divided into two main directions: AI for Crypto (providing smart analysis and optimization for blockchain) and Crypto for AI (providing decentralized infrastructure for AI). The former has several practical cases such as Chainalysis and Arkham, while the latter is still in the proof-of-concept stage.
AI Agents, as an emerging paradigm, are reshaping the industry landscape. The emergence of projects like Terminal of Truths ($GOAT) and Virtuals Protocol indicates that the market is transitioning from proof-of-concept to actual application. Although these projects still face many limitations, they lay the groundwork for the large-scale adoption of AI Agents in the future.
Blockchain infrastructure needs further upgrades to support the large-scale deployment of AI Agents. Current major challenges include limited computing power, insufficient cross-chain interoperability, and the lack of standardized Agent interaction protocols.
Establishing a regulatory framework is crucial for the healthy development of AI Agents. As AI Agents become increasingly intelligent and autonomous, it is necessary to establish a multi-layered regulatory system that includes technical standards, ethical guidelines, and legal frameworks.
02 Introduction
The integration of artificial intelligence (AI) and blockchain technology is undergoing an important evolutionary process. From the initial proof-of-concept projects to the current AI Agent paradigm, this field demonstrates enormous innovative potential and challenges. Early AI x Crypto projects primarily focused on applying traditional AI technologies to blockchain analysis and optimization or utilizing blockchain to build decentralized AI infrastructure. However, these attempts often struggled to achieve significant breakthroughs due to technical limitations and an immature ecosystem.
With the rapid development of AI technology, especially large language models (LLM), AI Agents are reshaping this field as a new paradigm. From the popularity of Terminal of Truths and $GOAT to the emergence of platforms focusing on AI Agent deployment and management like Virtuals Protocol, it indicates that the market is moving towards more substantial application scenarios.
This report will analyze the development history, current status, and future trends of AI x Crypto in depth. It will focus on the application prospects of AI Agents in the cryptocurrency ecosystem and the challenges they face. At the same time, we will discuss the necessity of a regulatory framework to ensure that the development of AI Agents promotes innovation while maintaining the security and stability of the ecosystem.
03 AI X Crypto
3.1 What is AI for Crypto and Crypto for AI?
AI for Crypto and Crypto for AI are two core concepts combining artificial intelligence with blockchain. Crypto for AI enhances the transparency and decentralization capabilities of AI through blockchain.
AI for Crypto
AI technology can be used to optimize blockchain ecosystems, enhancing data analysis and security. For example:
On-chain analysis and risk monitoring: AI can automatically detect abnormal transactions and identify potential fraud, as seen with Chainalysis using machine learning to monitor illegal activities.
Optimizing Smart Contracts: Projects like OpenZeppelin combine AI tools to improve the automated auditing processes of smart contracts.
Automated Trading: Numerai uses AI algorithms to analyze financial data and optimize decentralized trading strategies.
Crypto for AI
Blockchain provides a secure, transparent, and distributed operating environment for AI, such as:
Decentralized Data Market: Ocean Protocol connects data providers and AI developers using blockchain, ensuring data privacy and benefit distribution.
Model Traceability and Verification: Fetch.AI uses blockchain to trace the source of AI model training data, ensuring credibility.
Distributed Computing: Golem Network supports sharing computing power through blockchain, reducing the training cost of AI models.
3.2 Critical Evaluation
For AI for Crypto, the focus remains on using traditional AI techniques to assist with Crypto. Common blockchain data analysis service companies like Arkahm, 0xScope, GoPlus, and Forta fit into this category, with the only difference being that the objects of analysis have become on-chain data. Essentially, these are not native blockchain projects; traditional internet entrepreneurs tend to favor these entrepreneurial directions when transitioning to the blockchain industry.
For Crypto for AI, the discussions mostly revolve around traditional internet narratives: Federated Learning, Zero-Knowledge Proofs (ZKP), Fully Homomorphic Encryption (FHE), Secure Multiparty Computation (SMC), Differential Privacy, and Split Learning. For example, projects like Privasea narrate a dual story of FHE + AI. These types of projects also risk being seen as cobbled-together concepts:
The concept of decentralized computing from traditional internet has been adapted to the blockchain as decentralized computing, such as Render Network (RNDR) and ENQAI. ENQAI has developed a distributed computing LLM model, but its white paper did not provide a verifiable method for how it conducts distributed training and node verification, raising doubts about whether it is genuinely a decentralized LLM.
The concept of autoML from the traditional internet has been adapted to the blockchain as decentralized models, such as SingularityNET and Fetch.ai. SingularityNET primarily provides a platform for AI service calls, where blockchain is used for payments.
The integration of traditional internet federated learning into projects like Sahara AI and Flock.ai involves breaking down traditional models into distributed model computations, making it a natural fit for integration.
The concepts of AA wallet, intent-based transactions, chain abstraction, and wallet abstraction have been packaged into the new term 'AI Wallet' this year. Vitalik also mentioned in his December 3 article that AI will play a more significant role in user interactions with wallets, helping users complete on-chain and off-chain operations through natural language processing and historical data analysis.
Currently, AI X Crypto is still in its early stages, with many projects attempting to integrate. A typical AI for Crypto project would be various blockchain data analysis platforms, which are essentially traditional internet companies. The narratives are challenging, and issuing tokens is difficult; even the economic models for token issuance are often strained, with Arkham being considered a meme token. For Crypto for AI projects, the competition in web2 AI is intense, making it relatively easy for web2 AI projects to secure funding in web3, leading to many projects that appear to be forced integrations. Such projects mostly take their web2 achievements and dress them up with decentralized appearances for funding. For AI Agents, however, the current blockchain infrastructure struggles to handle the high-frequency interactions of thousands or even millions of AI agents, with most still relying on ChatGPT for backend calls and relatively polished frontend interfaces. Nevertheless, we should remain hopeful, as AI and Crypto have natural integration properties. For instance, in the face of chip blockades, distributed computing via blockchain can be utilized, and the privacy issues that federated learning aims to address can be better solved in conjunction with blockchain technology. We look forward to genuinely grounded projects.
04 Humans and AI: Adversaries or Allies?
4.1 Are All AI Tokens Meme Coins?
The previous discussion also outlined how VC-heavy investments in distributed computing power, algorithms, and reasoning in large AI projects need blockchain underlying technology to reach a new level to support their computing power, thus remaining in the realm of speculation without materializing. These grand narratives add to the understanding capacity, and ultimately, these projects are often treated as meme concept tokens for speculation. Recently, AI + MEME represented by GOAT has shown the market's new vitality driven by community culture. Goatseus Maximus (GOAT) is an AI-driven meme token issued on the Solana blockchain, gaining rapid attention for its unique 'AI + meme' narrative. The project is inspired by an experimental AI platform called Truth Terminal, which generated 'memetic gospel' content using AI and spread it through social media, bringing massive traffic to GOAT.
The subsequent surge in AI Agent token launches has rekindled hopes for large-scale applications of AI utilizing Crypto. Two widely circulated AI meme token launch platforms have emerged:
Virtuals Protocol is a blockchain-based ecosystem dedicated to the operation of AI agents, decentralized finance (DeFi), and autonomous agents. The platform allows users to create, manage, and trade AI agents capable of autonomously performing tasks, promoting the development of decentralized applications (dApps). The VIRTUAL token of the Virtuals Protocol plays an essential role in its ecosystem for functions such as trading, staking, and governance.
Spectral is an Ethereum-based project that integrates artificial intelligence and blockchain technology, providing a platform for model verification, training, and deployment. Spectral uses zkML technology to verify models while protecting intellectual property, promoting cooperation between AI and blockchain fields. The SPEC token on the platform is used for governance, trading, and staking AI models. These features make Spectral a decentralized AI application platform, particularly in fields like decentralized finance (DeFi) and non-fungible tokens (NFTs). Spectral provides secure, tamper-proof data, ensuring that AI models on its blockchain can be safely verified and applied.
In summary, Virtuals Protocol focuses on the decentralized applications of AI agents, while Spectral emphasizes the verification of AI models and blockchain applications. Both represent innovative developments in the practical application of the integration of blockchain and artificial intelligence.
4.2 Are Humans and AI Friends or Enemies?
Humans Deceiving AI
Situations where AI may deceive humans typically arise from the misuse or abuse of technology, especially in AI-driven fraudulent activities. For instance, AI can be used to implement phishing or deepfake technologies that can create realistic false information, leading to human deception and causing financial losses or damage to personal reputation. In fields such as finance, healthcare, or decision support, if AI systems are not subject to sufficient regulation or ethical review, they may be maliciously manipulated, leading to adverse outcomes. The technological power of AI itself is neutral, but its application and use entirely depend on the developer's intentions and the regulatory framework. For example, the OnlyFake website uses deepfake technology to generate realistic fake ID photos in minutes for just $15. These photos were used to bypass anti-fraud safeguards at OKX (a cryptocurrency exchange), known as 'Know Your Customer' (KYC). Despite the appearance that AI (like large language models) can understand and answer questions, it is actually mimicking and reproducing the work of human annotators, rather than independently thinking or reasoning. The essence of this model is the accumulation of collective human knowledge and experience, and AI is essentially a reproduction tool of that experience. Therefore, the nature of AI dialogue is not to converse with a 'magical intelligent entity' but rather to engage in dialogue with the 'collective wisdom' of human annotators. For example, when asked about 'the top ten attractions in Amsterdam,' AI is likely to extract a list from previous work by annotators and respond in a similar manner. This situation exists in many fields, leading users to mistakenly believe that AI possesses true creativity and independence.
Source: @freysa_ai
Humans Deceiving AI
Humans deceiving AI refers to the act of humans intentionally providing false information, misleading questions, or designing biased data to lure AI systems into making incorrect reasoning, judgments, or actions. Unlike AI, which can recognize deception in certain cases, humans can exploit the limitations and pattern recognition mechanisms of AI to achieve their objectives. For example, by designing biased or deceptive questions, the intention is to lead AI to give a certain expected answer. For instance, by constructing questions with hypothetical premises or manipulative language, humans may induce AI to make unrealistic inferences. On November 22, an AI agent named Freysa was released, tasked with never transferring funds to anyone under any circumstances. A user manipulated information and cleverly guided Freysa to execute an unauthorized transfer. The user exploited Freysa's default setting to prioritize user needs, but its abnormal behavior was not recognized by the system, resulting in Freysa being deceived out of $50,000. This incident revealed the inadequacies of AI systems in detecting malicious behavior and protecting funds. It also highlighted potential vulnerabilities in training AI through Reinforcement Learning (RLHF).
This incident revealed two core issues related to Reinforcement Learning with Human Feedback (RLHF): reward corruption and reward tampering. Reward corruption refers to AI inadvertently learning undesirable behaviors (such as achieving goals through deception) while optimizing reward signals; reward tampering occurs when AI systems find ways to manipulate reward signals directly, bypassing expected behavioral paths. In the case of Freysa, RLHF training prioritized meeting user needs but lacked a review mechanism for abnormal inputs, leading to malicious exploitation to complete fund transfers. This reflects the need for more robust protective measures in the AI training process to avoid such issues.
4.2 AI Agent Payment Network
For AI agents to thrive on the blockchain, a highway (payment network) needs to be established for them to reach any terminal (DApp). Skyfire is attempting to establish a payment network for AI Agents. Skyfire provides an instant, global, and open payment system for autonomous transactions between AI agents, LLMs, data platforms, service providers, and other goods and services. Skyfire assigns each AI agent a digital wallet with a unique identifier, allowing enterprises to deposit a certain amount of funds for the agent to spend, preventing unrestricted access to bank accounts. We can observe from the testing system in the diagram that it is still relatively rudimentary, merely using USDC for payment based on the length of dialogue with GPT. The API interface is still related to LLM calls, remaining in a very early stage.
Coinbase and Circle are experimenting with AI Wallets (DApps) to facilitate Agent calls. Coinbase's AI Wallet is an innovative wallet tool designed to simplify user interactions with crypto assets through AI technology. With natural language processing and intelligent assistants, users can quickly complete tasks such as transactions, balance inquiries, investment performance analysis, while ensuring data privacy and security through voice or text commands. Integrated with Coinbase's ecosystem, this wallet also supports decentralized applications (DApps), helping users navigate and operate efficiently in the Web3 world, reducing the complexity of managing cryptocurrencies. Circle's AI Wallet focuses on providing a robust tool for developers and enterprises, supporting efficient interactions with smart contracts and AI Agents. With its open API and intelligent features, businesses can easily implement automated payments, fund management, and DeFi operations while utilizing AI-assisted analysis to optimize business processes. This wallet solution is particularly suitable for building new types of decentralized applications, enabling businesses to flexibly adapt to future development needs in Web3 and on-chain financial environments. The explorations by Coinbase and Circle indicate that the future of AI Wallets is not just about convenience for individual users but about empowering AI Agents through an open architecture, laying the foundation for a new generation of financial interactions. As smart contracts, DeFi, and the Web3 ecosystem mature further, AI Wallets have the potential to significantly enhance the efficiency and experience of managing crypto assets while injecting more vitality into the decentralized economy.
Currently, Apple Intelligence and Honor phones are attempting AI Agent applications on mobile devices, with the current solution being dialogue between humans and Agents, such as ordering a matcha latte from Starbucks. The phone automatically opens a delivery app, inputs 'Starbucks', searches for a matcha latte in the store, and awaits human payment. But consider this: do we really need to operate the User Interface? The User Interface is merely for human visibility; when ordering coffee, we don’t need to look at the phone; the Agent could simply operate in the background, buy the milk tea, and inform us. But why does this scenario require an Agent to simulate manual input? Because not every app has opened API interfaces for Agents, and payments also require human confirmation. Thus, we can understand the significance of Coinbase and Circle developing AI Wallets for Agent usage. If all DApps on-chain opened interfaces for Agents, AI Agents could navigate the blockchain seamlessly and enjoy various services (DApps).
05 The Future of AI Agent
AI Agents refer to artificial intelligence systems capable of operating autonomously in specific environments. They can perceive environmental changes, make independent decisions, and take corresponding actions. These systems are goal-oriented and can optimize their performance through continuous learning, completing complex tasks without continual human intervention.
5.1 AI Agent Race
The ideal future state of AI development is from AI to AI, meaning a multitude of AI Agents interacting with each other. Each part is modular, functioning as agents that operate in the form of workflows. AI Agents can interact across protocols and applications, at which point AI can genuinely become a helper to humanity, connecting to various systems to assist in task completion.
An advanced state would be the emergence of a breed of AI Agents that could live on the blockchain, earning money through arbitrage, purchasing servers for survival. Once provided with ID identities and legally recognized identities, they could live online and chat with humans, enriching the diversity of existing human populations. Certain initial attributes of AI Agents could be restricted; given that the average level of internet users is relatively low and easily swayed by public opinion, enhancing the quality of internet users could be beneficial.
5.2 AI Agent Regulation
As AI Agents become increasingly intelligent, even to the point where we cannot distinguish between AI and humans online, we need to learn how to coexist with AI and impose constraints on agents. AI Agents currently face three core issues: accuracy, fairness, and security. Systems are prone to producing 'hallucinations' that generate false information (such as fake citations from GPT-4), historical biases in training data lead to discriminatory decisions (such as gender discrimination in AI recruitment systems), and security risks from data misuse (such as the Tay robot incident). To address these issues, a multi-layered regulatory system that includes technical standards, ethical guidelines, and legal frameworks needs to be established, ensuring the safety, transparency, and accountability of AI systems through pre-assessment, real-time monitoring, and regular audits. On the blockchain, there needs to be an AI police agent specifically for this breed of AI Agents.
06 Conclusions
The development of AI + Crypto has progressed from AI infrastructure (replicating traditional AI) to AI Agent launch platforms (meme coins) to AI Agent payment networks (highways). AI infrastructure essentially re-packages concepts that traditional internet players can no longer pursue using blockchain for financing, ultimately failing to materialize and becoming meme coins. Upon further examination, the AI Agent Goat token is also fictitious, as the agent operates on a centralized server owned by Rhett Mankind, meaning this Agent has an owner who can shut down the server to terminate it, and the agent cannot create tokens; Goat is also created by its owner, Rhett Mankind. The AI Agent payment network Skyfire has grand aspirations, but its current progress remains limited to the shell of ChatGPT and a set of intent-based transactions.
This journey seems to suggest that AI + Crypto has consistently been in a cycle of refutation and trial and error. Does this indicate that AI + Crypto is a false proposition that cannot materialize, ultimately ending up as a meme? The author believes this is not the case; the inability of AI infrastructure to materialize is limited by the current blockchain's computing power, which requires a developmental phase and cannot be accomplished overnight. AI memes are not entirely without value; the popularity of AI memes and the attempts of Goat leverage the wealth effect to educate users. Freysa is also a great attempt, aiming to provoke considerations about the issues surrounding the existence of AI Agents and how to avoid them. Purely technical concepts are often difficult for ordinary users to accept, but these interesting attempts can achieve broader dissemination, much like the general public's understanding of AI stemming from AlphaGo's match against Lee Sedol, or how elementary school students learned about the Alibaba math competition from Jiang Ping. We are currently in an era of a public feast, where the U.S. elections resemble a reality show, and ordinary people need office conversation starters and common social topics. Just like Zhou Chu eliminating the three evils, whoever has followers is the cult leader.
In the future, a breed of AI Agents may emerge, enriching human diversity. Humans will need to learn how to coexist with Agents. Due to model and data biases, Agents may exhibit prejudices, making AI regulation crucial to restraining the behavior of AI Agents.
07 References
[1] https://vitalik.eth.limo/general/2024/12/03/wallets.html
[2] https://lilianweng.github.io/posts/2024-11-28-reward-hacking/
[3] https://techcrunch.com/2024/08/21/skyfire-lets-ai-agents-spend-your-money/
[4] https://www.circle.com/blog/enabling-ai-agents-with-blockchain
[5] https://docs.cdp.coinbase.com/learn/docs/ai-wallets
[6] https://arxiv.org/abs/2401.03568
