Recently, the team behind the popular data storage protocol, Arweave, announced AO, a protocol with a radically efficient and flexible design that breathes new life into the dream of a decentralized world computer. Unsurprisingly, the project has drawn comparisons to DFinity’s Internet Computer Protocol, or “ICP,” which leverages an innovative subnet-based design to achieve similar goals. In the following article, we compare these two groundbreaking protocols and explain why we think AO’s unique approach offers the model for decentralized computation with the fewest compromises, particularly for the kinds of data-heavy finance, media and AI applications we’re building at Community Labs, and why we believe AO represents the most viable path to a truly decentralized world computer.

Flexible, Modular Design

From the very foundations of its design, AO emphasizes flexibility, scalability and efficiency. For instance, AO processes can load data of any size into their memory, execute on it and write back to the network, allowing any application to leverage AO’s computational capabilities and Arweave’s permanent storage. AO employs a highly modular architecture, supports unbounded parallelism, allows flexible resource utilization, and integrates directly with Arweave for data storage. Furthermore, using a message-passing network of Messenger Units (MUs) for interprocess and intraprocess coordination allows for autonomous activation of contracts (i.e., contracts have scheduled 'cron' interactions). 

AO’s more flexible environment for decentralized applications places it in stark contrast with the more rigid, limiting approach of almost every other blockchain protocol, including ICP, which adopts a highly structured, governance-focused design. Like AO, ICP is designed to create a decentralized computing platform that can host end-to-end applications, but it uses a highly structured subnet system for organization and scalability, employs a "canister" model for computation, and uses a reverse gas model for resource allocation. It focuses on governance through its Network Nervous System (NNS) and employs chain-key cryptography for security and efficiency. As such, ICP depends on a more centralized and controlled approach to achieve its objectives.

AO’s modular architecture allows developers to customize the system for specific needs by choosing and arranging “blocks” of the system to build different structures, making it easier to change parts of the system without having to rebuild everything from scratch. On AO, developers can choose which VMs, sequencing models, message passing security guarantees, and payment options work best for their specific needs. This is to facilitate and support a wide array of use cases and workloads. Different components like schedulers and Messenger Units can be swapped out or extended without changing the entire system.

Cryptographic & Consensus Security

State Model Design

AO’s holographic state model is where state is implied by the log of messages to the process on Arweave. This allows for efficient state representation and verification with cross-process interactions, as all data needed to derive the state of any process is permanently available. Given the same message log, any node can compute identical state. Compute Units (CUs) process message logs to compute current state when required and provide cryptographically signed attestations of computation results, i.e., the state can be computed when needed or continuously. The ability for computation of the current state when required eliminates the need for constant consensus on intermediate states.

ICP, on the other hand, employs an epoch-based state model. This approach enables efficient state synchronization, garbage collection of old data, and seamless protocol upgrades. By representing the state as a Merkle hash tree and using threshold signatures for certification, ICP promises state integrity while enabling rapid node syncing and network scalability.

Implicit vs Explicit State Models

AO’s holographic state model approach aligns with certain aspects of lazy evaluation for it does not require state to be stored onchain, but computation timing may vary (doesn’t mean it will be deferred entirely). Why isn't lazy evaluation used in more blockchain protocols? In short, it’s difficult to achieve because states have many components that need to be aligned.

• Deterministic computation - ability to reproduce the same result given the same inputs

• Complete and immutable history - full, unalterable record of all interactions/transactions to reconstruct any state at any point in time

• Efficient data storage and retrieval - a system to store and access the historical data efficiently

• On-demand computation mechanism - a way to trigger state computation when needed and determine when computation is necessary

AO's state representation is implicit because it doesn't need to maintain or store a current state directly on-chain. Often, the state is derived on-demand. By keeping the state implicit, AO achieves a high degree of flexibility and efficiency in its single, unified computing environment, while ensuring that any historical state can be reconstructed and verified when necessary. It is important to note however, that CUs have a way to persist “checkpoints” for processes which allows AO to capture and store the state of a process at specific points.

ICP takes a much more explicit approach to state management, built upon a foundation of replicated state machines and an epoch-based model. Also, ICP maintains a broader system state that encompasses cross-subnet communication streams, ingress history for user interactions, and a subset of state that is cryptographically certified in each round. The global state is efficiently encoded and verified using Merkle trees, allowing for partial updates and simplified synchronization between replicas. Each epoch begins with a summary block that contains critical information for managing the subnet's state and cryptographic protocols. At the end of each epoch, a Catch-Up Package (CUP) is generated which serves as a state "checkpoint".

Each canister's state (persistent memory, cycle balances, etc.) is explicitly stored and updated, allowing ICP to maintain an up-to-date state at all times. This means that all state changes are processed and reflected immediately.

Implementations

It’s reasonable to ask whether AO's flexibility could introduce vulnerabilities, given that a system is often only as secure as its weakest security guarantee. This is why a robust security model needs to be in place to prevent potential attack surfaces at the integration points. AO-Sec Origin is a mechanism that acts as an economic security layer (in control of economic incentives and penalties). It takes a very interesting approach to this potential concern. Various security properties of the AO-Sec Origin process are applied downstream to custom-built security mechanisms, which means security can be customized at different levels while still inheriting properties from its parent processes.

AO leverages Arweave's Byzantine Fault Tolerant (BFT) consensus algorithm as a foundation for data availability and integrity. All messages (ANS-104 compliant data items) and assignments are persisted on Arweave. Scheduler Units (SU) sign their message assignments, Compute Units (CU) provide signed attestations of computation results, and processes can verify the signatures of incoming messages. Cryptographic hash chains are used to link messages and ensure data integrity (a tamper-evident log of interactions). AO enables trustless and verifiable computation without requiring all nodes to execute every computation. For more info, check out our guide.

ICP incorporates a random beacon for leader selection and uses threshold Boneh–Lynn–Shacham (BLS) signatures for various security functions, including state certification and inter-subnet communication. ICP's chain-key cryptography enables efficient state management and verification, with catch-up packages (CUPs) allowing for quick state synchronization. 

Governance

Minimalism of AO

A key feature of AO's governance approach is the absence of a central treasury or authority with control over token supply or distribution. Unlike many blockchain projects, there is no mechanism for altering the token supply or directing funds through governance votes. Instead, AO relies on market dynamics and economic incentives to guide the network's growth and development.

To support ecosystem development without centralized control, AO implements two primary mechanisms. 

• Permissionless Ecosystem Funding: Allows developers to attract liquidity to their applications and receive funding directly from the native yield of bridged assets.

• Permaweb Ecosystem Development Guild (PEDG): Alliance of ecosystem organizations that collaborate on developing and maintaining AO's infrastructure, funded by yield from bridged assets.

There is an absence of global rules, which allows users and processes to choose their own security parameters and mechanisms. This approach enables a diverse ecosystem where different applications, use cases, workloads, etc. can coexist with tailored security and operational requirements for the applications being built. 

NNS for ICP

ICP employs a decentralized governance model called the Network Nervous System (NNS), which functions as a DAO. This system allows decisions to be made by stakeholders rather than a centralized authority. Token holders can participate in governance by staking their ICP tokens in entities called "neurons." The governance model incentivizes long-term commitment by allocating more voting power to neurons with longer staking periods.

Stakeholders can propose and vote on various aspects of ICP's operations and development. Proposals are adopted through a simple majority voting system, with a minimum percentage approval requirement to ensure adequate participation. NNS operates on a special subnet of the ICP and consists of several key components. This structure allows for a transparent, efficient, and decentralized decision-making process.

ICP’s governance model allocates more voting power to those who stake larger amounts of ICP tokens for longer periods of time. Though this model aims to incentivize long-term commitment, this can lead to a high Gini coefficient (high inequality among address balances). Time-based inequality (to some extent) and governance power concentration can cause greater inequality. We believe the approach AO takes mitigates these concerns. 

Token Distribution

Below, we can observe the token distribution of ICP token at launch. Notice how there’s very little token allocation towards the general public (aside from early contributors, node operators).

Source: ICP Token Distribution at Launch (Messari)

AO’s fair launch reflects a mission to give everything to the community.  36% of the total supply will be distributed to AR holders over time (long-time holding incentivization) and the remaining 64% of the total supply will be distributed as rewards for driving economic growth (ie, bridging assets from other networks). Notice above in the chart how AO’s fair launch does not consist of token allocation towards private sales, team members, or early investors (unlike ICP). Everything goes to the community. 

Conclusion

In the quest to build the world computer, there is likely no “one true path.” AO and ICP are two projects with innovative designs, passionate communities and countless potential applications. Both seek to overcome limitations of traditional blockchains, offering improved scalability, efficiency, and functionality. However, they achieve these goals through different means, reflecting divergent design philosophies and resulting in different competitive advantages.

At Community Labs we build applications that benefit from the flexibility and scalability AO can provide, but the success of AO will ultimately depend on how effectively it can balance its flexible and scalable design with the need for network cohesion and security, and how well its novel economic model performs in practice. We’re optimistic. We believe AO’s innovative approaches to modularity, flexible architecture, and fair token distribution address many of the challenges faced by other blockchain platforms with the fewest compromises of any distributed platform. We believe these features, combined with AO’s integration with Arweave for permanent data storage, make AO the most promising contender for enabling the next generation of decentralized applications and services.

Want to learn more about AO? Check out this quick guide, the specs, and whitepaper.

If you’re looking to try out AO applications, check out LiquidOps and Astro.