AO: AI-Oriented Super Parallel Computer

AO Intro

The AO computer is the actor-oriented machine that emerges from the network of nodes that adhere to its core data protocol, running on the Arweave network. Inspired by the original actor model and Erlang, ao does not allow processes to share memory between one another but does allow them to coordinate via a native message-passing standard. Each of these processes can then be operated at the full speed of the computing resources that are available, without interfering with one another.

Tech Architecture

The best place to understand the architecture of AO is the official doc. Here I will briefly walk through the core parts so we are ready for the discussion later.

The fundamental components of ao are as follows:

• Processes: The network's unit of computation. Processes are represented by a log of interacting messages stored on Arweave, as well as an initialization data item (its VM, scheduler, memory requirements, and necessary extensions).

• Messages: Every interaction with a process in ao is represented by a message, which is an ANS-104 compliant data item. Users and processes can send messages to other processes on the network by way of scheduler units.

• Scheduler Units (SUs): Scheduler Units are responsible for atomically incrementing slot numberings to the messages that are sent to a process. After the assignment, schedulers are required to ensure that data is uploaded to Arweave. Processes have free choice of their preferred sequencer, which can be implemented in a variety of ways: Decentralized, centralized, or even user-hosted.

• Compute Units (CUs): Compute Units are nodes that users and messenger units can use in order to calculate the state of processes in AO. Once the computation of a state is completed, the CU will return to the caller a signed attested of the output (logs, outboxes, and requests to spawn other processes) of the resolution of a specific message.

• Messenger Units (MUs): Messenger Units are nodes that relay messages around the AO network according to a process called cranking. In essence, when MUs crank a message around the system they send it to the appropriate SU for a process, then coordinate with a CU in order to calculate the output of the interaction, and then repeat the process recursively for any resulting outbox messages. Users and processes can also pay a MU to subscribe to a process, cranking any messages that result from its timed interactions. Processes can also optionally label a message as a cast – leading the MU to send the message to its SU, but not listen for a response.

One simple way to fit AO into the existing L1/L2 model is by imagining multiple co-processors running under different VMs, machines and even consensus models spread into a network. If we were to draw a comparison between parts of AO to parts of a traditional L1/L2:

• Onchain states = Messages backed up on Arweave

  AO does not have a straightforward global state. Instead, when observing the latest or historical state of a dapp, one needs to access the historical messages involved with the dApp's processes on Arweave and then recontruct its state progression until the latest message. Therefore, AO's state might seem different if one were to observe different dapps at a different time. Sometimes the state might even be in conflict. The invalid message should be ignored by the observer according to pre-defined consensus mechanism of the dapp at its initiation. Therefore AO's state is always holographic, and it's not very meaningful to construct a global consensus state.

• RPCs = Messaging Units + Gateway

  Front-ends send computation demands and retrieve computation results via MUs, similar to how users initiate transactions and observe the onchain state via the query from an RPC provider.

• Full Node = Computation Units

  This is where the VMs are hosted and the computations are conducted. The difference is that the computation in CUs is stateless, while full nodes store their input and output in the state of their blockchain. Also, full nodes take care of the consensus mechanism of their chain and conduct P2P communication all the time. CUs in AO are only involved in consensus mechanisms when required.

• Sequencer = Scheduler

  In a traditional blockchain, the sequencer puts transactions in a specific sequence so that the end state can be computed in a deterministic way. In AO, the scheduler can determine the end state of a process by putting messages in a specific sequence. It's just that these messages are a lower-level representation of a transaction, not necessarily calling a specific function.

• Data Availability Layer = Arweave

  Arweave serves a similar role to a DA layer, but not in the sense of ensuring that all states are published, but by ensuring that all the state information (including initial state, computation logic, input, middle state, and end state) needed to reconstruct the entire history of a computation can be retrieved. State retrievability ensures that all computations conducted on the AO network can be lazily/optimistically verified.

The advantage of AO's architecture is quite clear from a traditional blockchain. Instead of trying to figure out the best place to optimize (big vs small blocks, PoW vs PoS, parallel vs sequential, monolithic VS modular, etc), AO throws everything out of the window and gives dapps, or more specifically their processes, maximum flexibility on how they would like to exist. One process can choose to have 50 SUs (running PoS to ensure sequencer decentralization), then 1 CU (sitting in a data center for maximum performance), together with 3 MUs (sitting in the 3 different data centers in the same area to ensure low latency but running 2/3 multi-sig to eliminate single point of failure on backing up messages to Arweave). Even if something goes wrong, the entire historical state of the dApp's process can still be retrieved from Arweave and a hard or soft fork on the dApp's state tree can always be implemented.

The flexibility of Arweave ensures that it remains maximally wide open for any dApp's infrastructure need. AO is essentially taking the bottle of "lazy verification" pills and emptying the whole bottle down its throat. By lazily verifying everything (including the computation, consensus mechanism, and even the P2P communication), a dapp can choose to run on the most affordable infra, the most performant infra, or the most economically secured infra, all within a change on their process's parameters. For most applications, this drastically saved their cost on all items that would need to carry the weight of a consensus mechanism or economic guarantees until it's absolutely needed.

Custom-made for Decentralized AI

Among the various use cases on the blockchain, AI use cases has been the least explored one, given it's very recent emergence. In the following section, we will break down the demand for decentralized AI and explain why the AO architecture is the best

If the use case needs to scale in a horizontal way, why choose vertical scaling for your infra?

Decentralized AI needs access to:

• Decentralized model (hosting training)

  A decentralized AI infrastructure requires access to decentralized model hosting and training to ensure the robustness and resilience of AI models. Traditional centralized AI systems are prone to single points of failure, making them vulnerable to manipulation.

• Decentralized data (training, fine-tuning inference)

  Access to decentralized data for training, fine-tuning, and inference is crucial for building robust and diverse AI systems. Centralized data storage can create bottlenecks and censorship of information. Decentralized data input also allows for more diverse and representative datasets, as data can be sourced from places previously underpaid for their contribution. Decentralized data access supports collaborative efforts in AI development, enabling multiple stakeholders to contribute and verify data quality, which is essential for developing unbiased and accurate AI models.

• Decentralized computation

  Centralized computation systems are limited by their finite resources and cost structures. Decentralized computation allows for parallel processing and distributed workloads, making it possible to tackle complex AI tasks that require substantial computational power.

Both the above demands should have the following characteristics.

• Super-Scalable

  For a decentralized AI infrastructure to be effective, it must be super-scalable, and capable of expanding its resources seamlessly to accommodate growing demands. Super-scalability ensures that as more data and computational tasks are added to the network, the system can scale horizontally without significant performance degradation. This characteristic is vital for supporting the continuous growth and evolution of AI models, which often require increasing amounts of data and computational power.

• Decentralized Pricing

  A decentralized pricing mechanism is essential for fostering a fair and competitive ecosystem for AI services. Traditional centralized pricing models such as multi-tier subscriptions or token fees is not efficient enough to fully reflect the dynamics of flexible AI usage. Having the price reflected by market dynamics is the best way to cope with the flexibility needed for AI services.

• Censorship Resistant

  Censorship resistance is a critical characteristic of decentralized AI infrastructure, ensuring that AI models and data remain accessible and unaltered by any single authority. In centralized systems, entities with control over the infrastructure can impose restrictions or manipulation, potentially leading to biased or censored outcomes. The resistance to censorship preserves the integrity and neutrality of AI models, fostering trust and reliability in AI-driven decisions.

• Autonomous (self-starting and sustaining)

  Autonomy is a defining feature of decentralized AI infrastructure, enabling systems to self-start and sustain without human intervention. Autonomous AI can dynamically manage resources, adapt to changing conditions, and recover from failures independently. We can only witness the true power of AI by ensuring that the AI can operate in an ever-lasting environment and evolve autonomously, without arbitrary guidelines pointing their fingers.

Decentralized AI will incur the following costs to deploy and operate

• Cost of Storage

  Decentralized AI would need plenty of space for it's model, data input and data output. AO leverages Arweave for the cheapest permanent and decentralized storage. Arweave significantly reduces the long-term cost of storage for decentralized AI systems, making it a cost-effective choice for managing large datasets necessary for AI training and operations.

• Cost of Computation

  Computation on the blockchain is generally much more expensive than computation on a server or on a PC. The extra costs come from redundancy or other ways to ensure validity. It's hard to lower the computation price beyond the machine cost. AO provides the maximum flexibility on how much redundancy is needed, or to what extent is validity ensured.

• Cost of Communication/Bandwidth

  Decentralized AI systems incur costs related to communication and bandwidth as data and computational tasks need to be transmitted across the network. One of AO's greatest cost savings on bandwidth is the reduction in handshake operations by implementing the actor-oriented model. Messages can choose to be simply be casted to avoid costing the bandwidth a response. Also, AO's flexibility is designed to optimize these costs by giving this power to the AI (dapp) where they can decide for themselves how much data exchanges are needed beyond the absolute basics, optimizing for performance, security, and decentralization VS cost and latency. It's a fair assumption that an AI on AO would be much cheaper to run than other networks not implementing the Actor Oriented Model.

• Cost of Validity

  Ensuring the validity of data and computations is crazy expensive, especially under the scale of AIs. Validity technology such as ZK proofs has been focusing on the validity of the computational process. Today it's still way too expensive for AIs, introducing 1000X overhead. On AO, validity is lazily verified. One can still reconstruct the entire computation based on the messages stored on Arweave, without having to prove the validity of the computational process. In AO, validity is relative. States are observed as a holographic state. There is no one single global state consensus. Every single state is "perceived" or "observed" to be correct by querying from a different node of AO. This works in a very similar way to EigenLayer's "inter-subjectivity" where whether a statement is valid or not is not entirely subjective or objective.

• Cost of Consensus

  Consensus is almost always achieved through some form of redundancy with a fault tolerance mechanism and anti-sybil mechanism on top. The Arweave storage already handles fault tolerance and anti-sybil. So for the dapps constructed on AO, there is no immediate need to implement the cost of consensus in the first place. Consensus can always be implemented when the economic situation of the dapp improves and the cost of not having enough consensus grows. The consensus cost of AO comes from the consensus of CU and Arweave itself. If a user chooses to, multiple CUs can conduct computations in parallel and the result is chosen or calculated based on a pre-defined consensus mechanism. The consensus of the starting state of an AI, the ending state, and all middle states are stored on Arweave and can be inferred by running through all the intermittent messages.

To summarize the rundown for AO's technical architecture, it's evident that AO has offered a network under the Actor Oriented Model that offers enough validity and consensus (through Arweave and lazy verification), high flexibility on computation (from super affordable to super performance), and great cost optimization and decentralized storage and decentralized communication. These advantages lines up nicely with the emerging decentralized AI where loads on computation, bandwidth, and storage is tremendous, especially decentralized.

Ecosystem Projects

AO does not directly enforce any direct consensus or security mechanism upon the computations. Instead, it offers a flexible system from which economic security mechanisms can be implemented. This opens up a huge design space for both infrastructures and applications.

Besides the standard WASM implemented by the Forward Research team as a demo/default option, developers would have several other options to play with when it comes to deploying their smart contract on AO.

• EverPay : The EverVison team built everPay, a real-time payment settlement layer that uses the storage computing paradigm, where all calculations are done off-chain and the results are synchronized with Arweave. There are already several dapps running on the everpay infra including Permaswap (DEX), ARIS (Inscription), and WeWeave (storage service)

• MEM.Tech (Molecular Execution Machine): MEM is a layer enabling flexible, chain-agnostic serverless functions, secured by Arweave. The MEM protocol leverages 3EM, supporting languages that can compile to WebAssembly (WASM), such as AssemblyScript, JavaScript, TypeScript, Rust, C/C++, Zig, GoLang, Solidity, and Vyper. MEM is also natively chain-agnostic, allowing authentication from 14 EVMs and non-EVM chains, and can call data from any chain or API.

• WeaveVM : WeaveVM is a hyper-scalable, lazy-evaluated EVM protocol to compute data at any scale. It's the first EVM-compatible protocol that fully relies on Arweave for permanent storage, enabling atomicity of executable EVM bytecode and data of any size on the same ledger.

Several other middle-ware is also made available for the AO Network.

• AR.IO : AR.IO is the decentralized and incentivized gateway node network of the Arweave ecosystem. The various nodes on the AR.IO Network, known as gateways, are the interface between users and the permaweb. Each gateway acts like a “Permaweb Service Provider” and supports multiple, value-added, services like reading, writing, querying, and indexing of Arweave data.

• AOX : AOX is a bridge facilitating bridging the AR token from Arweave into AO, or out of AO to Arweave.

• 0rbit : 0rbit is the Decentralized Oracle Network built on AO to provide access to limitless data on any AO process. 0rbit can also bring data from any valid URL to an AO process. Building with 0rbit is as simple as making an HTTP GET or POST request to an API.

• AO.Link : Built by the Autonomous Finance team, ao.link serves as a message explorer for the ao Network, offering functionalities similar to block explorers in conventional blockchain systems. It includes message computing capabilities, graphical visualization of message links for clarity, real-time message streaming for up-to-date information, and a list of linked messages for organized navigation.

• DataOS : Built by the Autonomous Finance team, DataOS is a protocol developed atop the AO Network, leveraging it for computing power and Arweave for data storage, creating a synergistic foundation for its operations. As a ContentFI protocol, DataOS employs autonomous AI agents to source and regenerate content derivatives, aiming to foster engagement and fulfill the intent of consumers.

Beyond infras, AO has a very exciting application ecosystem landscaping across the basics defi and new exciting AI applications:

Defi:

• Permaswap : Permaswap Network is a fully decentralized P2P network where LPs in the network are peer-to-peer self-hosted accounts and assets are self-managed by the user's private key. The nodes in the network perform price calculations and build peer-to-peer orders in a distributed manner according to protocol rules. Users can trade on Permaswap with extremely low gas fees without experiencing MEV issues. At the same time, the network protocol can flexibly support multiple AMM algorithms.

• ArSwap : ArSwap is a next-generation Automated Market Maker (AMM) built on the AO blockchain. Leveraging the unique capabilities of AO, ArSwap offers decentralized swapping, liquidity provision, and yield farming services with unparalleled security and efficiency. ArSwap introduces innovative features like dynamic swapping pools and custom liquidity schemes, enabling users to maximize returns while minimizing risks.

• Bark : Built by Community Labs, Bark is A decentralized AMM exchange backed by token liquidity pools.

• Astro : Astro is a stablecoin protocol on the AO network where users can use the $AR token as collateral and mint the $USDA stablecoin. Astro introduces innovations like Wrapped Arweave (wAR) and a unique liquidation mechanism with Bark AMM, enhancing permaweb liquidity and stability.

• Harvest: Harvest is an RWA marketplace for natural resources, featuring land, cattle, and other natural investments from Colombia. Harvest has built-in profit mechanisms, allowing the investor to get exposure to the additional cash flow from their investment. For example, milk offspring revenue from investing in a cow.

• Mentions on memecoins: $TRUNK , Aethirs (which consists of several other memecoins: $FIRE, $AIR, $WATER, which can be composed into some other memecoins: $Lava, $Dust, $Mud, $Lightning, $Steam and $Ice).

AI Applications:

• Autonomous Finance : Autonomous Finance is a dedicated research and technology entity, focusing on the intricacies of financial infrastructure within the AO network. Autonomous Finance team is building AI onchain and generating new data-driven financial primitives. Autonomous Finance is developing composable semi and fully autonomous agents. The autonomous agents will be self-learning and executing logic using onchain data streams.

• Autonomous DCA Agent : Built by the Autonomous Finance team, Autonomous DCA Agent is an Autonomous Investment Agent (AIA) executing a dynamic dollar-cost-average (DCA) investment strategy across various liquidity pools within the AO ecosystem. It automatically buys a base token at predetermined, user-configurable intervals using a consistent amount of a quote token for each transaction.

• AO Games : AO Games is a fully decentralized game launchpad on AO. Looking to build games delivering a Web2-level experience with fully decentralized games. The team will also introduce On-Chain AI so that the gameplay experience can adapt, learn, and evolve.

• Outcøme : Built by the Puente.ai team, Outcome is a prediction market hyper-focused on the AO ecosystem. The prediction markets center around AO, games, defi, memes, business, and technology. Users bet whether they are "IN" or "OUT" of a prediction. Users can also claim daily tokens, predict, and climb the leaderboards to win prizes.

Closing

The projects mentioned above covers the majority of projects currently announced or live on the AO network. Though the ecosystem is quite young, we are thrilled to see that a lot of the projects already have a working product. Some of them already have a tradable token. The AO testnet is announced in a very usable stage which is ready to support some basic AI functionalities. Given more time and resources, we would not be surprised to see more dapps thriving with their onchain AI.

Given the less-than-expected friction deploy AIs on AO, it is evident that the architecture of the AO supercomputer is well-aligned with AI operations. As the mainnet date comes closer, we look forward to see more exciting new innovations happening here in AO and Arweave.