No need to trustlessly execute arbitrary code, and sufficient to share for the whole world to use, the dream of a world computer is deeply rooted in decentralized networks. After Ethereum, many infrastructure projects have made attempts, and the upcoming AO network from Arweave is one of these attempts.
For a "world computer," it can be roughly divided into three parts: computation, access, and storage of data. Arweave has been acting as the "world hard drive," while the AO network (Actor Oriented) introduces the capability of general computation and provides smart contracts.

AO: Actor-Based General Computation Network

Current mainstream decentralized computing platforms can be divided into two categories: smart contract platforms and general computation platforms. Smart contract platforms, represented by Ethereum, share a global state memory and reach consensus on the computation processes that change the state. Because consensus requires a large amount of redundant computation, it is only used for high-value transactions under high costs. General computation networks do not reach consensus on the computation process itself but verify the computation results based on business needs, handling request order without a shared state memory, which reduces costs and allows the network to expand into more areas of computation, represented by computing networks like Akash.
Of course, there are also some projects that combine general computation with smart contracts based on the security assumptions of virtual machines. In this case, consensus only handles the order of transactions and verifies the computation results, with multiple state-changing computations processed in parallel among network nodes. The virtual machine environment guarantees deterministic results, so as long as the transaction order is consistent, the final state will also be consistent.
These types of networks do not share state memory, resulting in very low expansion costs, allowing multiple tasks to be computed in parallel without affecting each other. Such projects often rely on the Actor programming model, with ICP being a representative, and AO also belongs to this category. Each computation unit under the Actor model is treated as an independent smart entity, and computation units interact through communication (Actors are a very common architecture in traditional Web2 services). AO standardizes the message passing of Actors, realizing a decentralized computing network.
Unlike traditional passively triggered smart contracts (such as Ethereum/Solana smart contracts), AO under general computation Actors can actively run smart contracts through a "cron" method triggered by a consistent fixed time loop, such as a trading program that continuously monitors arbitrage opportunities.
With rapidly scalable decentralized computing capabilities, Arweave's massive data storage capacity, the Actor programming model, and the ability to trigger transactions actively, the AO network is very suitable for hosting AI Agents. AO also supports running large AI models within blockchain smart contracts.

AO Network Features

The previous section introduced the differences between AO and smart contract networks. AO does not reach consensus on the computation process but reaches consensus on the transaction order, assuming that the results of the virtual machine's execution are deterministic, thus achieving consistency in the final state.
AO also possesses a certain degree of flexibility, as the network is designed in a modular way. There are three basic units in the network: Scheduling Unit (SU), Computing Unit (CU), and Messenger Unit (MU).
When a transaction is issued, the Messenger Unit, as 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 AO and the AR chain, helping the network to sort the transaction order and upload it to the AR chain for consensus. The current consensus method is POA (Proof of Authority); after reaching consensus on the transaction order, tasks are assigned to the Computing Unit, with CU responsible for handling specific computations, and results returned to MU for forwarding to the user.
The CU set can be seen as a decentralized computing network. Under a complete economic plan, CU nodes need to stake certain assets and compete based on computing performance, pricing, and other factors to provide computing power and earn profits. If a computation error occurs, assets will be confiscated. This is a standard economic safeguard.

Differences Between AO and Other Networks

As a general computation platform, the differences between AO and smart contract platforms like Ethereum are evident. Filecoin, which shares the title of "world hard drive" with AR, has also launched its own smart contract platform, FVM, but this is an architecture equivalent to EVM's state consensus mechanism and does not perform as well as traditional smart contract platforms like Ethereum.
Unlike decentralized computing networks such as Akash and io.net, AO still retains smart contract capabilities, and AO ultimately maintains a global state on AR storage.
In fact, the architecture most similar to AO is ICP. ICP created the earliest paradigm of asynchronous computing blockchain networks, and AO largely continues ICP's design, such as only sorting transaction orders, trusting virtual machine deterministic computation, and asynchronous processing with the Actor model.
The biggest difference is that ICP maintains state based on containers, meaning each smart contract container can only maintain its private state or set conditions for state reading; whereas AO has a shared state layer, namely AR, where anyone can restore the entire network state through transaction order and state proofs. This somewhat increases the network's decentralization capability but also loses the possibility of implementing special privacy services in ICP (for example, clients may have hidden arbitrage paths).
From an economic and design perspective, ICP imposes higher hardware requirements on participating nodes to ensure network performance, resulting in a higher barrier to entry, while AO operates relatively with a fair launch and no admission requirements, allowing participation in competitive mining through staking. The ICP network chose a large stack implementation method, sacrificing flexibility for performance, while AO uses a modular design, separating MU, CU, and SU, allowing users to choose their virtual machine implementation method, which also lowers the entry cost for some developers.
Of course, AO may also have systemic drawbacks similar to ICP, such as the lack of atomicity in cross-contract transactions under the Actor asynchronous model, which could hinder the development of DeFi applications, making the concept of AgentFi seem difficult to realize in the short term; the new computing model that deviates from traditional smart contract paradigms also places higher demands on developers. However, under the AO architecture, the wasm virtual machine can manage a maximum limit of 4GB, which also prevents some complex models from being used on AO. Thus, it seems that AO's choice of the AI Agent route is indeed a way to leverage strengths and avoid weaknesses. Interestingly, ICP also announced its focus on the AI field at the beginning of 2024.
Of course, compared to ICP's total market value of 5 billion USD, AR's current total market value of 2.2 billion USD still shows a significant gap. In the context of the rapid development of AI, AO may still hold considerable potential.