Sam & Outprog: Let's talk about AO and artificial intelligence

Main Text

Outprog:

Thanks to lulu and BeWater for the invitation. I am Outprog from PermaDAO, hosting today's event.

Arweave, as a decentralized storage infrastructure, has been running stably for nearly 6 years since 2018. The data stored on Arweave has also seen explosive growth, much like the price of BTC.

We all know that this year, a very important event occurred in the Arweave ecosystem, namely the release of AO, marking the transition of the Arweave ecosystem from decentralized storage to a full-stack application ecosystem that supports decentralized computing.

First, let's have Sam briefly introduce himself and provide a brief overview of Arweave and AO.

Sam:

Thank you, Outprog, and thank you, BeWater, for the invitation!

About 7 years ago, we started building Arweave as a permanent data layer for storing all of humanity's most important knowledge and history. The initial idea was basically to create "on-chain storage"—just like people have been trying with Bitcoin from the beginning, but with the need for infinite scalability. Fast forward to today, 7 years later, Arweave now stores over 5 billion pieces of data and serves as the "data pipeline" for many Web3 applications.

In this process, we discovered that decentralized computing is "decentralized data replication" plus "verification." When you run an Ethereum (or even Bitcoin) node, your computer is simply downloading blocks from the peer-to-peer network and verifying them.

Since Arweave can permanently store any amount of data (just like the data set replication of Bitcoin and Ethereum), a question arose: if we add a verification layer, does that mean we can achieve infinitely scalable smart contracts?

It turns out the answer is yes 😄. This product is AO. It is essentially a decentralized supercomputer running on Arweave, with an infinite number of parallel threads.

We hope to establish a new, decentralized cyberspace through the combination of Arweave and AO. This cyberspace can be proven to be neutral and guarantees users' rights.

Outprog:

Thank you, Sam, for the introduction. Let's move on to the first question. Today's topic is AI, so let's start discussing AI.

We know that AO is designed using the Actor model. Interestingly, a paper titled A Universal Modular Actor Formalism for Artificial Intelligence mentions the relationship between the Actor model and AI. Can we consider that AO, based on the Actor model, is naturally suitable for developing AI applications?

Sam:

Yes, the Actor model was invented during a time when AI research was receiving significant attention (similar to today). The exploration of artificial intelligence took 60 years of research to perfect, but the Actor programming model was quickly and widely adopted and applied to many non-AI devices. Now that we are approaching AGI, the Actor model makes even more sense.

Essentially, the core idea of the Actor model is that each component of the system can be an independent, autonomous agent that can operate in parallel. This model aligns very well with agent-driven architecture because it simulates the real world: just like in today's AMA chat, each of us is independent and autonomous, coordinating with each other by sending messages.

AO brings this idea into the computing world using the Actor model. Each service in the system is autonomous and can execute independently. When they want to coordinate, they achieve this by sending messages.

Outprog:

Yes, in the Actor model, each "actor" is like an agent, which reminds us of AI Agents. Can Sam elaborate on the connection between the Actor Model and AI?

Sam:

Exactly! Agent == Actor. Therefore, it is very reasonable to use an Actor-oriented approach as its hosting environment.

What excites us most about AO is that it provides us with sufficiently scalable smart contracts that can host complete large language models. Our CTO Tom and I have been researching and practicing hosting large language models on Arweave over the past few weeks during evenings and weekends: https://GitHub.com/samcamwilliams/aos-llama

Although we are not ready for a public announcement yet, if you want to learn more, this repository can serve as a preview 😄.

Outprog:

Impressive! I believe Sam's use case will soon demonstrate how large AI models can be implemented on AO!

Sam:

One idea: AO is a brand new form of cyberspace that is very suitable for agents. Within AO, agent bodies are autonomous, sovereign individual units.

We believe that the primary application of this technology will emerge in the financial sector.

In the past, most economic activities were actually "smart layers"—not just settlements. These economic activities were previously done "off-chain," but now we can bring them into a smart contract environment while gaining the trustlessness they provide.

Imagine if all smart behaviors of economic activities—not just settlements—could be executed on-chain through trustless agents as smart contracts. Now you can build smart contracts that provide users with an autonomous "algorithmic trading" fund. This opens up a vast design space that has never been explored.

In the long run, we believe it will be much more than that. This will be a "life form" of autonomy that can reliably execute tasks (however you want to describe it).

Outprog:

Automated finance and AI agent finance are directions worth exploring on AO! Let's stay tuned.

Moving on to the next question. Currently, there are no real examples of AI and Web3 being combined, and the technical architecture of AO has clearly broken through this limitation. Beyond on-chain models like EVM, developers can use AO to create applications that were previously difficult to achieve.

After this technical breakthrough, how will AO support the development of AI in the future? Will there be support and funding for AI projects? What role will AO play in the development of AI?

Sam:

The main suggestion is to actively seek collaboration! There are many groups in the Arweave ecosystem that are very willing to help with the development of projects built on AO. Recently, Community Labs launched AO Ventures, an incubator with $35 million in investment funds aimed at supporting the development of projects built on the network.

At Forward Research, we are also very willing to provide one-on-one support for those building projects in the ecosystem. Just reach out to us, and we will do our best to provide various assistance—from technical support and promotion to funding support, we will do our best.

Outprog:

Strong ecological support! Last question.

The hackathon hosted by BeWater is very helpful for the development of the AO developer ecosystem. PermaDAO has also long been committed to building the developer community in the Arweave ecosystem.

We believe there will be more offline AO ecosystem events held in Asia in the future, possibly hackathons, incubators, or boot camps.

Does Forward Research have any plans or ideas to support these events? How can we collaborate with the BeWater AI Crypto Hackathon to jointly promote innovation and development?

Sam:

At Forward Research, our core function is building: whether it's protocols or communities. The latter often involves finding the right people and then helping them promote.

Q & A

Here are the Q&A from community users during the AMA.

PoS

Adam Lee:

I have some questions about AO.

1. I noticed that AO is still using a centralized PoA system. When will it upgrade to PoS or a more decentralized consensus mechanism?
2. I haven't seen any tokenomics related to AO. When will this part be upgraded?
3. Does AO have a roadmap?

Sam:

As we continue to develop, we will gradually transition most activities smoothly to PoS. Currently, it is not actually "centralized"—some developers are running their own SUs/CUs/MUs. As a developer, you can choose which "institutions" to trust for messages. It's more "distributed" than centralized, but not decentralized—though I understand your point. AO itself is a data protocol on Arweave, so you can overlay any number of different security systems on it.

Once AO implements PoS, a highly efficient computing market will emerge, where people run CUs (computing units) responsible for computing the state of processes in the network.

One great feature of AO is that every subnet operating the network is completely elastic. Each network has no scalability limits. An interesting fact is that yesterday we doubled the number of computing units in the test network run by Forward Research. Aside from some processes running faster, no one noticed the change 🙂. In the future, we can scale to any number of computing units.

David Dot:

Does the AO token mean raising funds again, or has it given up on making AR more valuable?

Sam:

It's hard to understand your question accurately, but AO has not conducted any fundraising. We are pushing all venture capital firms that want to invest to invest in the ecosystem of projects built on it.

Concurrency

Siyuan Han:

Sounds great! Thank you, Sam. Can you explain how AO supports an arbitrary number of parallel processes? How does AO avoid traditional parallel execution issues, such as read-write conflicts?

Outprog:

The Actor Model is a very mature model for handling concurrency, and it is the architecture adopted by AO. This model was proposed in 1973 and has been implemented in the Erlang programming language.

More documentation about the Actor Model can be found here: https://en.wikipedia.org/wiki/Actor_model

You can also check the AO specification document translated by PermaDAO: https://permadao.com/permadao/ao-1353cc109d434941a6757560ef35dcc2

Sam:

I'm quite sure Telegram is using it for our AMA chat, and we are also using the Actor Model programming internally!

Siyuan Han:

So, AO avoids read/write conflict issues at the computational level. If AO has the capability to handle transactions like traditional high-concurrency systems while executing thousands or tens of thousands of threads, how does it handle workloads at the disk I/O level? Can you elaborate on AO's storage and data engine technology?

Outprog:

Rollup, using technology similar to Ethereum's Rollup to bundle data onto Arweave.

Sam:

Yes! The data of each process is "rolled" and bundled onto Arweave, which is specifically designed to coordinate the large-scale replication of information in a peer-to-peer network.

Therefore, when building AO, we can say that disk I/O coordination is "free" to some extent (because we have spent 6 years building it 😅).

Load Balancing

Kevin Zhang:

Will there be many processes handling AI in the future? How will load balancing be maintained?

Outprog:

My view is that MUs (messenger units of AO) are similar to Nginx in traditional Web2, with load balancing distributed across countless MUs, and after PoS is implemented, CUs will share all the computation. It's like the elastic scaling of k8s.

Sam:

Yes! MUs are the entry point for users, while CUs are the "computing cloud."

GPU Computing Power

AI:

If we do some AI applications in AO, such as AI video generation, can the AO ecosystem solve the problem of insufficient GPU computing power?

I haven't seen any AI demonstrations. Does AO have GPU computing capabilities?

Sam:

GPUs cannot be used directly, although the Apus team is researching this issue. Currently, only CPUs are used because WASM runs on CPUs, providing us with deterministic computation results—which is crucial for the verifiability of smart contracts.

Jason Wu (APUS):

As far as I know, not yet. The Apus network is providing GPU computing power for AO. We will conduct some simple demonstrations. Currently, image generation is possible. Large language models and video generation are not yet available.

Consensus and Transaction Order

Miami Fish:

How does SU verify requests and submit them for processing by CUs, for example, strictly by timestamp or based on some priority of economic incentives?

Outprog:

SUs are used to allocate nonces, and the computation results obtained by CUs are executed in the exact order of the nonces.

This is a paper about message passing, also in the AO specification: https://groups.csail.mit.edu/tds/papers/Lampson/FORTE93.pdf

Sam:

As Outprog mentioned, SUs provide a Slot-Assignment for each message. In the PoA test network, you need to trust your SU, but if you want, you can even run it yourself. In a complete PoS network, you will stake on the correctness of the SU, and in case of issues, you can "fallback" to Arweave's consensus mechanism (just like "rolling back to L1" in Rollup).

Miami Fish:

Thank you. So in PoS, will SU be automatically assigned (like who receives my request first)? What I'm thinking about is more about financial applications, such as in a clearing event, where different people try to submit clearing requests at roughly the same time, and whose request is processed first will affect the results for others.

Sam:

The deployer of the process can choose the SU—whether it's the developer themselves or another process generated as a "sub" process.

After PoS, if the SU stops processing messages, you can "challenge" them on the AO-Sec Origin process, requesting slot allocation. If they fail to do so, that process will enter an "unhostable" state, and the issue will be resolved by having a SU "bid" to become its new host.

There are many details involved, but this is a macro view. It allows every process in AO to inherit the flexibility and censorship resistance of Arweave while running on high-speed SUs 🙂.

Alright, I need to leave now. Happy coding, everyone!