How to Build a Decentralized Computing Network for AI Agents in AO

The dream of a world computer is deeply rooted in a decentralized network that can execute arbitrary code without trust and share it with the world. After Ethereum, many infrastructure projects have made attempts, and Arweave’s upcoming AO network is one of these attempts.

For a ‘world computer’, it can be roughly divided into three parts: data computation, access, and storage. Arweave has been playing the role of ‘world’s hard disk’, while AO Network (Actor Oriented) introduces the ability of universal computation and provides smart contracts.

AO: Actor-based General Purpose Computing Network

Currently, the mainstream decentralized computing platforms are divided into two categories: smart contract platforms and general computing platforms. Smart contract platforms, represented by Ethereum, share a global state memory in the network and reach consensus on the process of changing states. Because consensus requires a lot of repetitive computation, it is only used for high-value businesses under high cost. General computing networks do not reach consensus on the computation process itself, but verify the calculation results based on business and process request order. There is no shared state memory, which reduces costs and allows the network to expand into more fields of computing. This type of network is represented by computing power networks such as Akash.

Of course, there are also some projects that integrate general computing with smart contracts based on the security assumptions of the virtual machine. That is, consensus only processes the order of transactions and verifies the calculation results. Multiple state change calculations are processed in parallel in network nodes, and the virtual machine environment for calculation ensures deterministic results. Therefore, as long as the transaction order is consistent, the final state will also be consistent.

These networks, due to not sharing state memory, have very low scaling costs, and multiple tasks can be computed in parallel without affecting each other. These projects are often based on the Actor programming model, represented by ICP, and AO also belongs to this category. Under Actor, each computing unit is seen as an independent intelligent processor handling transactions separately, and communication interaction between computing units is common (Actor is a very common architecture in traditional Web2 services). AO standardizes the messaging of Actor, implementing a decentralized computing network.

Unlike traditional passively triggered smart contracts (such as Ethereum/Solana smart contracts), with a general computing Actor, AO can actively run smart contracts through a consistent fixed-time loop triggered by the “cron” method, such as a trading program that continuously monitors arbitrage opportunities.

The decentralized computing power that can be quickly scaled, Arweave’s large-scale data storage capacity, Actor’s programming model, and the ability to actively trigger transactions make AO Network very suitable for hosting AI Agents. AO also supports running AI large models in smart contracts on the blockchain.

AO Network Features

The above article introduced the difference between AO and the smart contract network. AO does not reach consensus on the computing process, but on the transaction order, and assumes that the virtual machine’s execution result is deterministic, thus achieving final state consistency.

AO also has a certain degree of flexibility and is designed in a modular way. There are three basic units in the network: scheduling unit (SU), computing unit (CU), and messenger unit (MU).

A transaction is sent out, and the messenger unit of the communication layer receives the transaction, verifies the signature, and forwards it to the scheduling unit; the scheduling unit can be seen as the connection point between the AO and AR chains, helping the network to sort the transaction order and upload it to the AR chain to complete the consensus. The current consensus method is POA (Proof of Authority); after the consensus on the transaction order is completed, the task is assigned to the computing unit, CU is responsible for processing the specific computation, and the result is returned to MU for forwarding to the user.

CU cluster can be seen as a decentralized computing power network. Under a complete economic planning, CU nodes need to stake certain assets and compete based on factors such as computing performance and price to provide computing power and earn profits. If there are calculation errors, assets will be subject to slashing. This is a standard economic guarantee.

The Difference between AO and Other Networks

As a general computing platform, AO’s difference from smart contract platforms like Ethereum is obvious. Filecoin, also known as the “world’s hard drive” along with AR, has launched its own smart contract platform FVM. However, this is a state consensus architecture equivalent to EVM, and the experience is not as good as traditional smart contract platforms such as Ethereum.

Unlike decentralized computing networks such as Akash and io.net, AO still retains the ability of smart contracts, and ultimately maintains a global state on AR storage.

In fact, the architecture of ICP is most similar to AO. ICP created the earliest paradigm of an asynchronous computing blockchain network, and AO largely continues the design of ICP, such as sorting transactions only by order, trusting virtual machine deterministic computation, and using the Actor model for asynchronous processing.

The biggest difference is that ICP is based on container maintenance state, that is, each smart contract container can only maintain its own private state, or set conditions for state reading and setting; while AO has a shared state layer, that is, AR, anyone can restore the state of the entire network through transaction order and state proof, which to a certain extent increases the decentralization capability of the network, but also loses the possibility of implementing special privacy business in ICP (such as the need for customers to hide arbitrage paths).

On the economic and design level, ICP has higher hardware requirements for participating nodes to ensure network performance, which creates a higher threshold. In contrast, AO operates in a fair and non-admission manner, allowing participation in mining competition through staking. The ICP network has chosen the implementation of big stack, sacrificing flexibility for performance, while AO uses a modular design, separating MU, CU, and SU, and users can also choose the implementation of virtual machine, which also reduces the cost of entry for some developers.

Of course, AO may also have system shortcomings like ICP, such as the lack of atomicity in cross-contract transactions under the Actor asynchronous model, which will lead to difficulties in the development of DeFi applications. The vision of AgentFi seems difficult to achieve in a short period of time; the new computing model that is detached from the traditional smart contract paradigm also poses higher requirements for developers. However, under the AO architecture, the wasm virtual machine can only manage a maximum limit of 4 GB, which also means that some complex models cannot be used on AO. From this perspective, the choice of AI Agent by AO is indeed a way to play to its strengths and avoid weaknesses. Interestingly, ICP also announced a focus on AI in early 2024.

Of course, compared to ICP’s total market cap of $5 billion, AR’s current market cap of $2.2 billion still has a significant gap. Against the backdrop of the rapid development of AI, AO may still have great potential.