Action Layer will become the Ethereum of AI + Blockchain.

Ethereum’s greatness lies in two key areas: Firstly, as a universal execution layer, it ensures that smart contracts are automatically executed on the blockchain according to user intentions, maintaining consistency and security across the network through technical means. Secondly, this execution layer empowers Ethereum to serve as a decentralized application platform, enabling developers to create a variety of dApps.

From the perspective of AI development, we are also at a stage lacking in “Action” implementation. At the recently concluded ICLR 2024 workshop, distinguished scholars, including Turing Award Professor Yoshua Bengio and academics like Choi Yejin and Song Han, delved into topics related to Artificial General Intelligence (AGI). According to a 120-page research report from UIUC titled ‘How far are we from AGI’, there are three critical elements to AGI:

1. AGI internal: The components mirror the essential aspects of human cognition, like perception, memory, reasoning capabilities, and metacognition.
2. AGI interface: The capability to interact with the external world. This interaction is facilitated through various interfaces that enable AGI systems to perceive, understand, and act within their environment, be it digital, physical, or intellectual.
3. AGI systems: The infrastructure capabilities that support model capacity to emerge, such as model architecture, training systems, cost and efficiency, and computing platforms.”

The internal component has seen significant focus and rapid development with multimodal large models like LLMs. The systems segment is supported by transformers and a well-established GPU industry. However, there is a gap between AI cognition and real-world interaction (interface) — the need for an Action Layer. This will be the most crucial and final direction in the AI Agent race, enabling AI Agents to act like humans and interact with the real world.

As DeepMind’s founder Mustafa Suleyman noted, ‘This new [AI] wave is all about doing.’ The concept of an Agent refers to an entity that possesses end-to-end capabilities from intent to action, not just simple generative AI. However, with the concept of ‘Agent’ becoming overused, it seems that all AI has morphed into Agents. We’ve witnessed impressive performance from LLMs in dialogue, yet no agent can operate a computer to complete tasks like a human.

Thus, IntentAGI introduces a necessary but as yet unrealized decentralized infrastructure: The Action Layer for Autonomous Agents.

The Action Layer

The Action Layer is a critical component within AI systems, responsible for translating the cognitive and decision-making processes of AI into tangible actions and interacting with the external environment (AGI interface). Within the Action Layer, AI Agents achieve specific goals or tasks by executing a sequence of actions. These actions may involve controlling physical devices, manipulating software interfaces, sending commands over networks, or handling other forms of input and output. The design of the Action Layer must consider the objectives of the AI Agent system, the complexity of the operational environment, and the efficiency and reliability of action execution. Typically, the functions of the Action Layer include action selection, execution, and monitoring to ensure effective interaction with the external world and the achievement of intended outcomes.

The Action Layer is the most crucial component in the realization of value within AI systems. It involves transforming the cognitive, analytical, and deliberative outcomes of AI models into effective actions, interacting with the real world, and delivering end-to-end value. Due to the complexity of the real world, establishing the Action Layer is also the most complex and prolonged process. As pioneers and builders of the AI Action Layer, we categorize the real world into five abstract application domains: consumer applications, application programming, blockchain, roads, and the physical world.

Challenges

The Action Layer faces many challenges, one of which is the issue of trust between AI and humans. In the traditional world, mutual trust is based on contracts, such as those signed on paper in banks, where a handwritten signature on a document establishes trust. Over the past 20 years, with the rise of digitalization and online banking, interactions in front of a computer, where clicking buttons and entering data form an action sequence, have become the basis of trust.

On Ethereum, this trust is ensured through a Turing-complete programming language. However, in the AI era, with increasing data bandwidth and model uncertainty, providing a one-to-one verification from intent to action is no longer feasible. The outcomes generated by models to satisfy user intent can vary — for example, asking GPT-4 to write a story about the Frog Prince could result in multiple versions.

This raises greater challenges for trust verification. We need to train AGI to continually align with human intentions, similar to how a lawyer, after understanding a client’s needs, might produce various texts, but ultimately satisfies the client’s requirements.

Decentralized Super Alignment

In the progression of AGI, a viable pathway to mutual trust is ‘super alignment,’ which necessitates extensive work aligning intents and actions, ideally driven by innate decentralized mechanisms. The role of AI Agents in economic production is to deliver services — a way of processing and thinking about similar situations, which constitutes know-how rather than goods or energy. This knowledge is broadly distributed among individuals’ thoughts. Similarly, humans acquire these skills through hands-on learning, taught by those around them.

AGI should not be confined to a few scientists; any ordinary user can contribute value through their unique ‘know-how’ by providing feedback. Following the path of Instruct GPT, we’ve identified a way for a decentralized community to contribute to the emergence of AGI. Instruct GPT aligns user intents with GPT outputs by extensively collecting user feedback, rapidly evolving GPT-3 into the well-known GPT-3.5, and continuously iterating towards a mutually trusted AI product infrastructure. Two methods are critical in this process: an evaluation model based on benchmark metrics and a method involving human feedback reinforcement learning.

As the Ethereum of the new era of AI + Blockchain, the Action Layer transcends the Instruct GPT moment of AGI. The key question is how to continuously leverage AI and blockchain technologies to become a new blockchain infrastructure foundation.

“True intelligence can only emerge when an agent can interact with its world”.

That real-world interaction is what could take AI beyond learning patterns and making predictions, to truly understanding and reasoning about the world.

About IntentAGI:

IntentAGI is the first Action Layer for Autonomous Agents. We are dedicated to allowing AI to gain the ability to interact with the real world, thus truly realizing decentralized AGI.