AI/ML has been a hot topic as of recent, especially since the explosion of LLMs (Large Language Models) such as ChatGPT and Claude AI.

They've certainly changed the way that we work - but we've also seen some of the downsides that come along with them. Notably closed source models, biased answers, and a lack of transparency around the data or weights used to train models.

AI models are also often hosted on traditional cloud architecture which is not only costly, but comes with the usual issues of centralization (single points of failure, controlled by a single corporation, etc). There are also other unanswered aspects of generative AI, such as royalties for those who have their data used to train LLMs.

Outside of this, there are also exciting new opportunities with AI, such as agentic finance (AgentFi).

The question is how can we leverage new and existing decentralized solutions to build a technically sound, and fair AI environment moving forward?

We believe that the best setup will be on Arweave and AO. Both have proven themeselves capable of hosting and utilizing AI, and with advances from ecosystem projects such as Apus Network and Autonomous Finance, the future of decentralized AI is looking promising.

Arweave and AO for onchain AI

Arweave, unlike some other blockchains, has data storage at the forefront. Training data can be stored onchain, in an open environment for anyone to read.

With the launch of AO, Arweave now also has a decentralized compute layer built on top of it. This means processes (smart contracts) and more - but more interestingly, it interacts with Arweave in a way that makes compute massively scaleable. Yes, this means entire LLMs, and their datasets, could be stored and ran in a decentralized manner.

In fact we’ve already seen LLMs being experimented with on AO and Arweave. Sam Williams, inventor of Arweave, created aos-llama, and we can see some AI agents in action on Llama Land (a fully onchain world, by the way).

The current architecture shows what's possible with Arweave and AO, although this current process is largely CPU-driven. Decentralized GPU networks are also being worked on to integrate with AO, which is where Apus Network comes in.

Decentralized GPU networks and AO

When it comes to AI, and more specifically generative AI, models are often trained on GPUs as opposed to CPUs. The process typically involves a lot of parallel processing which GPUs are a lot more efficient at doing. AO is currently focused largely on CPU execution which is great for most computational tasks, but for AI workflows and training GPU execution is ideal. Parallel workflows on GPUs, however, aren’t always deterministic. In other words, putting in the same inputs doesn't always result in the same outputs. Determinism is important when it comes to blockchains not only for consensus, but it’s desirable for node X to compute and return the same state as node Y for reliability. Something that makes AO different from traditional blockchains is modularity. Execution of code doesn't necessarily have to occur on AO, it can occur elsewhere - it's up to the user to decide what to use (and trust). AO is more like a framework, and the base AO network is just the first public implementation.
Apus Network are one team building on this modular architecture to link a decentralized GPU network to Arweave. This is the process:

• AI Models are stored directly on Arweave
• DePHY framework is used to manage a network of decentralized GPUs which utilize models
• dApps and users interact with the Apus process, e.g. prompting (by proxy all of these interactions are stored on Arweave)
• Apus interacts with the DePHY GPU network and returns results to user Here's a diagram taken from their article on the architecture: This is a game changer for Arweave and AO, because it brings the network decentralized trustless GPUs as a method of execution. This opens up even more opportunities for dApp developers, and those interested in building GPU heavy workflows while also keeping with the principles of decentralization.

AI Agents

Another narrative becoming increasingly popular in crypto (and on AO) are agents. AgentFi is a somewhat new phenomenon, and refers to using automated or AI “agents” to make trades, or execute onchain trading strategies. This has in large been driven forward by Autonomous Finance in the AO ecosystem, and their research into DCA Agents and agent-driven trading platforms. Automation and AI assistants are separate topics, but we think it’s worth exploring as machine learning plays a huge role in traditional algorithmic trading. Imagine being able to trade using an open source agent with a built-in risk management strategy, or having an agent analyze market conditions and update user orders in real-time. Letting an AI agent take control of your crypto isn’t necessarily good financial advice, nor is it risk-free. However, allowing users and traders to create their own onchain strategies is definitely a compelling narrative.

Fairer Generative AI with Atomic Assets

One of the bigger controversies surrounding generative AI at the moment is usage. How many people have LLMs trained data on? Did they use your Facebook pictures, your art, your tweets or blog posts? Should you be compensated for it?

Arweave and AO actually have a very interesting solution to a lot of these issues - Atomic Assets.

Atomic Assets are like NFTs with superpowers. On creation they can be attached a Universal Data License (UDL) which specifies the usage rights on an asset. This could be creative rights for a one-time fee, or commercial rights paid monthly while the asset is held.

Atomic Assets are also not limited to “just” images - they could be videos, 3D blender models, or even blog posts. AI companies could purchase assets with the usage rights they need, while creators get compensated for their work.

The Future of AI on AO

We've already seen what's possible with LlamaLand and the initial Llama inference engine on AO. With the addition of Apus Network, decentralized GPU-driven dApps are also looking promising - imagine an end-to-end transparent and trustless AI workflow coming together.

AI, and generative AI in particular, is definitely shaping up to be one of the most important technical advances this decade. With Arweave, AO, and Apus Network, we can build on the advances, whilst also keeping AI transparent, trustless, and verifiable.

If you’re interested, come and build on AO and Arweave!