On-Chain AI: The Holy Grail of Decentralized Artificial Intelligence?

Key Takeaways

• Decentralized AI platforms have proliferated in recent times.
• DePINs have helped decentralize AI development, but fully on-chain AI could be the next big thing.
• In March, the Internet Computer (ICP) became the first blockchain platform to deploy AI models as smart contracts.

As artificial intelligence (AI) becomes increasingly embedded in the digital economy, concerns about the concentration of power among a handful of Big Tech companies have prompted the development of various decentralized alternatives.

The umbrella term “decentralized AI” captures a plethora of solutions for building, distributing and deploying AI systems, many of which incorporate a blockchain component. As the technology evolves, the next stage of development could see AI models that run fully on-chain.

The Problem With Centralized AI

As the industry has grown, the concentration of power inherent in centralized AI has seen a handful of companies dominate. And surprise surprise, it’s the same Silicon Valley giants that have spent the last two decades entrenching themselves at every level of the global digital infrastructure.

The dangers of Big Tech monopolies are well-documented and have implications for democracy, consumer rights and online privacy. Allowing these same corporations to control a technology as important as AI is asking for trouble. 

One area where large tech firms have consolidated their grip on the AI market is in the provision of the GPUs and CPUs used to train and run AI models. This model sets up hyperscaler cloud providers as gatekeepers to critical computational resources, creating potential points of failure that threaten the entire system’s security.

However, in recent years, Decentralized Physical Infrastructure Networks (DePINs) have emerged that could help democratize access to the resources needed to train AI models and reduce developers’ dependency on hyperscalers.

DePINS and AI

DePINs include decentralized storage networks like IPFS and Filecoin, as well as decentralized compute networks like Aethir and Render. 

Because data storage and compute are the fundamental currencies of modern AI, DePINs have become critical to the work of decentralizing artificial intelligence and more recent projects like BitTensor have sought to provide AI developers with the full range of resources, providing storage as well as CPU and GPU-compute in a single platform.

To establish a marketplace of buyers and sellers, decentralized AI networks rely on blockchains, which typically sit as a separate payment layer on top of the DePIN.

But what if you could run AI models as blockchain-based smart contracts? This would be the holy grail of decentralized AI, allowing for fully on-chain AI applications that leverage the advantages of transparency and trustlessness blockchains are known for.

The first blockchain platform to achieve the feat is ICP, which successfully deployed the first basic AI models on its testnet in March.

Running AI as a Smart Contract

In an interview with CCN, Lomesh Dutta, the Vice President for Growth at ICP developer DFINITY, explained how on-chain AI goes beyond existing DePIN solutions. While he acknowledged that DePINs provide an important service, Dutta said ICP is trying to do something more ambitious. 

“What if you could use a blockchain to do training, as well as inference of the model itself, where you can define all the data that goes in?” he asked. 

Running machine learning models as smart contracts means all input and output data can be tracked and verified, removing all centralized components from the AI stack.

Dutta acknowledged that the technology is still in its infancy and ICP is still a long way from hosting a full-scale large language model. But starting with smaller models such as image classifiers, he said the platform will be able to handle much larger ones as it scales.

Blockchains and Data Provenance

As the decentralized AI movement gathers steam, blockchain-based solutions could address AI challenges surrounding data provenance and authenticity.

Commenting on potential applications of fully on-chain AI, Dutta explained that blockchains can be used to create a verifiable record of what data was used to train or improve models.

Under the system of centralized AI development, “a lot of the LLM training has happened through proprietary data,” he observed. “Maybe Stack Overflow was possibly used to train better coding for ChatGPT, […] but  it’s very hard for Stack Overflow to prove it, and hence they may not necessarily get the benefits.”

Increasing the transparency and verifiability of AI training data could have important implications for artists and publishers, giving them more control over who uses their content and how.

But blockchain-based immutability could also be applied to model outputs by labeling AI-generated content in a way that can’t be manipulated.

Addressing Deepfake Challenges

One platform aiming to solve the challenge of proving content provenance is Livepeer, which launched an AI Subnet earlier this year to run text-to-video and image-to-video generators on its distributed GPU network.

The video-focused DePIN’s  increasing focus on AI has caught the attention of Grayscale, which recently launched a decentralized AI fund to invest in “solutions to centralized AI-related problems, including authenticity checks against bots, deep fakes, and misinformation.”

As Livepeer CEO Doug Petkanics put it to CCN: “If YouTube labels a video […] that’s just sitting in their database that they own, and they can change it, or someone can hack it or change it, and the user has no sense. But blockchain is this uneditable database accessible to everyone.”

With AI-generated fake news and deepfakes on the rise, decentralized AI platforms could serve as an important corrective to rampant misinformation, letting users verify the authenticity of digital content in a trustless manner.