I. Attention-driven novelty and aversion

Over the past year, as the narrative at the application layer has been disconnected, unable to match the speed of infrastructure explosion, the crypto field has gradually become a game of competing for attention resources. From Silly Dragon to Goat, from Pump.fun to Clanker, the novelty and aversion of attention has led to the involution of this competition. Starting with the most clichéd eye-catching monetization, it quickly evolved to the platform model where the demanders and suppliers of attention are unified, and then silicon-based organisms became the new content providers. Among the myriad carriers of MEME Coins, a kind of existence that can finally reach a consensus between retail investors and VCs has emerged: the AI Agent.

Attention is ultimately a zero-sum game, but speculation can indeed promote the savage growth of things. In our article on UNI, we reviewed the beginning of the blockchain's golden age, where the rapid growth of DeFi was driven by the LP mining era opened by Compound Finance, with the in-and-out of various mining pools with APYs in the thousands or even tens of thousands being the most primitive form of on-chain gaming at the time, although the eventual outcome was the collapse of various mining pools. However, the crazy influx of gold miners did leave unprecedented liquidity on the blockchain, and DeFi ultimately broke free of pure speculation to form a mature track, meeting users' financial needs in payment, trading, arbitrage, staking and other aspects. And the AI Agent is currently experiencing this savage stage, and we are exploring how Crypto can better integrate with AI and ultimately propel the application layer to new heights.

II. How can agents be autonomous

In the previous article, we briefly introduced the origin of AI MEME: Truth Terminal, and the prospects for AI Agents. This article focuses primarily on the AI Agent itself.

We start with the definition of the AI Agent. Agent is a rather old but vaguely defined term in the field of AI, with the main emphasis being on Autonomy, i.e., any AI that can perceive the environment and make reflexive responses can be called an Agent. In the current definition, the AI Agent is closer to an intelligent agent, i.e., setting up a system that mimics human decision-making for large models, which is seen in academia as the most promising way to AGI (Artificial General Intelligence).

In the early versions of GPT, we can clearly perceive that the large models are very human-like, but when answering many complex questions, the large models can only give some plausible answers. The fundamental reason is that at the time the large models were based on probability rather than causality, and secondly they lacked the human capabilities of using tools, memory, planning, etc., and the AI Agent can fill these gaps. So to summarize in a formula, AI Agent = LLM (Large Language Model) + Planning + Memory + Tools.

The large model based on prompts is more like a static person, it only comes to life when we input, the goal of the agent is to be a more real person. Currently, the agents in the circle are mainly based on the Meta open-sourced Llama 70b or 405b version (with different parameters) fine-tuned models, with the ability to remember and use API access tools, and may need human help or input in other aspects (including interaction and collaboration with other agents), so we can see that the main agents in the circle are currently in the form of KOLs on social networks. To make the agent more human-like, it needs to access planning and action capabilities, and the sub-item of thinking chain is particularly critical in planning.

III. Thinking Chain (CoT)

The concept of Thinking Chain (Chain of Thought, CoT) first appeared in a paper published by Google in 2022 titled "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", which pointed out that the model's reasoning ability can be enhanced by generating a series of intermediate reasoning steps, helping the model better understand and solve complex problems.

A typical CoT Prompt contains three parts: a clear task description, the logical basis or principles that support the task solution, and a specific solution demonstration. This structured approach helps the model understand the task requirements, gradually approach the answer through logical reasoning, thereby improving the efficiency and accuracy of problem solving. CoT is particularly suitable for tasks that require in-depth analysis and multi-step reasoning, such as math problem solving and project report writing. For simple tasks, CoT may not bring obvious advantages, but for complex tasks, it can significantly improve the model's performance, reduce the error rate, and improve the quality of task completion through a step-by-step solution strategy.

CoT plays a key role in the construction of AI Agents. AI Agents need to understand the received information and make reasonable decisions based on it. CoT provides an orderly way of thinking, helping the Agent effectively process and analyze the input information and convert the analysis results into specific action guidelines. This method not only enhances the reliability and efficiency of the Agent's decision-making, but also improves the transparency of the decision-making process, making the Agent's behavior more predictable and traceable. By breaking down the task into multiple small steps, CoT helps the Agent consider each decision point in detail, reducing the risk of erroneous decisions due to information overload. CoT makes the Agent's decision-making process more transparent, making it easier for users to understand the Agent's decision-making basis. In interaction with the environment, CoT allows the Agent to continuously learn new information and adjust its behavioral strategies.

As an effective strategy, CoT not only enhances the reasoning ability of large language models, but also plays an important role in building more intelligent and reliable AI Agents. By utilizing CoT, researchers and developers can create intelligent systems that are more adaptable to complex environments and have a high degree of autonomy. CoT has shown its unique advantages in practical applications, especially in dealing with complex tasks, where breaking down the task into a series of small steps not only improves the accuracy of task solving, but also enhances the interpretability and controllability of the model. This step-by-step problem-solving approach can greatly reduce the risk of erroneous decisions when facing complex tasks due to information overload or complexity. At the same time, this approach also improves the traceability and verifiability of the entire solution.

The core function of CoT is to integrate planning, action, and observation, bridging the gap between reasoning and action. This mode of thinking allows AI Agents to formulate effective countermeasures when predicting possible abnormal situations, as well as to accumulate new information and verify pre-set predictions while interacting with the external environment, providing new reasoning basis. CoT is like a powerful engine of accuracy and stability, helping AI Agents maintain high efficiency in complex environments.

IV. The right pseudo-demand

What aspects of the AI technology stack should Crypto be combined with? In last year's article, I believed that the decentralization of computing power and data was the key step to help small businesses and individual developers save costs, and in Coinbase's Crypto x AI segmentation track this year, we see a more detailed division:

(1) Computing layer (focused on providing GPU resources for AI developers);

(2) Data layer (supporting the decentralized access, orchestration and verification of AI data pipelines);

(3) Middleware layer (supporting the development, deployment and hosting of AI models or agents);

(4) Application layer (user-facing products that leverage on-chain AI mechanisms, whether B2B or B2C).

In these four layers, each has a grand vision, and their goals can be summarized as fighting against the monopoly of Silicon Valley giants in the next era of the Internet. As I said last year, do we really have to accept that Silicon Valley giants monopolize computing power and data? In their monopolized closed-source large models, their internal workings are a black box, and science, as the most believed religion of today's humanity, every sentence answered by future large models will be seen as truth by a large part of the people, but how is this truth to be verified? According to the plan of the Silicon Valley giants, the agents will ultimately have powers beyond imagination, such as having the right to pay from your wallet, use your terminal, how to ensure that there is no evil intent?