Note: This article is translated from the official blog of Oasis, a leading privacy public chain.
Oasis supports Web3 through smart privacy - customizable cross-chain privacy, 100% private, 100% public, or any privacy in between.
By applying a self-sovereign mindset to AI development, decentralized AI aims to address issues of privacy, fairness, and accessibility.
Over the past decade, large companies have controlled most of the development of important technologies. In many cases, these companies focused on capturing data, monetizing it, and selling it with little regard for the end user. Many of the organizations behind today's most popular AI models are following a similar trajectory.
On the other hand, the rise of decentralized AI offers a more transparent and confidential alternative while supporting self-sovereignty and fairness. However, a more decentralized approach also introduces trade-offs. This is particularly pronounced in communication overhead among computing providers, especially due to the lack of standardization. Distributed development is also a process of building 'Lego blocks' and piecing them together, which takes time.
Decentralized vs. centralized AI
The most straightforward way to compare the two is to ask anything in the 'decentralized' camp whether it can compete with centralized AI in terms of value. From a training perspective, the answer is relatively clear: today, creating an advanced AI model requires a substantial amount of funding. For example, xAI's latest training cluster consists of 300,000 A100 GPUs and cost billions of dollars.
With a few exceptions, most projects struggle to train cutting-edge models. However, blockchain projects have an advantage in aligning users through incentive mechanisms. Bitcoin demonstrated how to create the world's largest computing network, and this is now being replicated by many AI computing projects. In addition to coordinating CPU capabilities, blockchain also has great potential in smart business, particularly in payments.
Use cases of decentralized AI
It is worth noting that advancements in AI have largely been top-down, with centralized companies driving most of the progress in machine learning over the past few years. The key is that blockchain is not suitable for all layers of the AI stack but should serve as an enhancement technology in meaningful places. Here are some interesting use cases.
Data provision and monetization
Consolidating GPU resources and exchanging them through payment or sharing models is a frequently mentioned method. This approach is also applicable to data providers, acknowledging the value of their private data in differentiating models. Such arrangements require new governance and economic models that can support crowd-sourced development and ensure fair distribution of income. Blockchain can help allocate contributions among stakeholders by providing a trusted framework.
Intellectual property and traceability
The impact of AI on media is still just scratching the surface. Models will continue to improve and expand increasingly powerful creative tools. However, issues of intellectual property ownership, and even the authenticity and quality of media itself, will also spark controversies. Fortunately, Web3 provides ready-made solutions for traceability and fact-checking. By incentivizing the use of cryptographic keys, hundreds of millions of people around the world can now verifiably authenticate the source of information. This is not a panacea, but it can help prove the origins of media. Cryptographic techniques can also provide more reliable reference mechanisms, such as models directing users to blockchain explorers or specific ENS profile pages.
Verifiability and confidentiality
With the transformation brought by AI, similar changes will also occur in fields such as medicine, law, and education. Today's LLMs (large language models), while sometimes providing accurate and sometimes erroneous answers, will have specialized models in the medical field that surpass the performance of doctors. The implications of this are enormous. However, it also opens up tremendous potential for data leaks and/or exploitation. In such future scenarios, whether distributed ledgers and their verification models are used, or whether data is tampered with, will become critical. At the same time, confidentiality protection infrastructure is also essential to ensure that input/output data and, in certain cases, the model itself are protected.
Decentralized AI and confidential computing
Continuing from a privacy perspective, encryption technologies have laid some foundations for AI over the past decade, including the demand and funding support for GPU development. In return, the current GPU roadmap, partly driven by the demand for private weights, is moving towards implementing Trusted Execution Environments (TEEs). Web3 will benefit greatly in this regard, especially as TEE performance on CPUs/GPUs has significantly improved in recent years.
Today, users can enjoy the flexibility and performance of cloud computing without needing to trust cloud service providers or store sensitive information in unencrypted data. In this regard, Oasis has been at the forefront of confidential computing, including the launch of Sapphire, an EVM network that leverages TEE to create confidential smart contracts. Sapphire is unique in Web3, but the possibilities of blockchain runtimes remain limited. We need to find a way to consider the non-deterministic issues of AI while also requiring more complex and flexible application building methods. ROFL was born for this purpose.
Runtime off-chain logic (ROFL)
To enable AI to interact with traditional, fixed systems (like smart contracts), creating an authentication mechanism is crucial. This is exactly the idea behind ROFL, which is a versatile computing framework that allows various applications to run in a decentralized, verifiable, and privacy-protecting manner.
Runtime off-chain logic makes it possible to create customized off-chain logic and can be easily verified on-chain through the magic of trusted execution environments. ROFL guarantees security and integrity through remote proofs and the Oasis consensus layer while enabling developers to access remote network resources. At a high level, it works as follows:
In practice, ROFL combines off-chain performance with on-chain trust. Essentially, anything that can be written in software can be integrated into ROFL applications. However, ROFL is best suited for applications like AI pipelines, as they require substantial computing resources and high trust. Full support for Intel TDX capabilities will soon be added to ROFL, enabling large models to run directly within the framework. This will transform various applications such as intelligent agent applications, allowing for persistent, confidential, and verifiable interactions. You can start using ROFL here.
$ROSE
This article is originally from the official Oasis website. We welcome everyone to visit the official website for more information about the Oasis ecosystem.