AI Lands On-chain: Why Are LLMs Being Uploaded to Public Blockchains?
Proponents of fully on-chain AI cite transparency and verifiability as one of its key benefits. | Credit: Chris McGrath / Getty Images.
Key Takeaways
- Crypto developers are increasingly experimenting with on-chain AI.
- Some projects are focused on deploying open-source large language models on public blockchains.
- The novel approach has some key advantages over centralized AI services.
Deploying large language models (LLMs) typically involves connecting to APIs operated by companies like OpenAI or downloading one from an open-source provider like Meta.
However, recent developments in decentralized AI mean you can now access LLMs without relying on large technology giants or centralized servers.
Llama-3 Preserved On-chain
In response to the concentration of GPU resources among a handful of Big Tech companies, decentralized compute networks have exploded recently, creating new training and inference options for AI developers.
However, these Decentralized Physical Infrastructure Networks (DePINs) only address one aspect of the AI centralization problem. Another layer of the technology stack relates to memory, for which a separate set of DePINs proposes a solution.
An example of such cross-network integration can be seen in the collaboration between Eternal AI and Filecoin. The former is a decentralized compute network, and the latter a decentralized storage network; together, the two platforms operate as a full-stack solution for AI deployment.
Eternal AI was founded with the mission of preserving “humanity’s most important creation” by ensuring AI remains censorship-resistant and permissionless. The platform hosts open-source models, including Meta’s Llama-3 and a variety of image generators.
Without centralized entities mediating access, users can be sure their prompts aren’t being altered or censored, while there are no daily token limits or barriers to use.
Although initially constrained by size, Eternal AI intends to upload the largest version of Llama-3 to Filecoin this month, proving that even a 405-billion parameter model can be stored in a decentralized manner.
GPT-2 Implemented on ICP
Another company taking on the challenge of on-chain AI is Decide AI, which recently deployed OpenAI’s GPT-2 on the Internet Computer (ICP).
While Decide AI’s motivation behind implementing GPT -2 is similar to Eternal’s, the technological feat of training, fine-tuning, and running AI models as smart contracts create more possibilities.
Unlike most other blockchains, ICP “wants to create a blockchain where you can actually host end-to-end software,” DFINITY Vice-President Lomesh Dutta explained to CCN. That means providing enough storage and compute to run programs on a single infrastructure.
Transparency and On-chain AI
In an interview with CCN, Decide AI CEO Raheel Govindji explained that the key benefit of fully on-chain AI is transparency and verifiability.
As AI adoption increases, “we’re probably going to see it being used by government services, by medical, by legal,” he predicted. “Do you really want a centralized AI that could have some biases to decide who gets an organ transplant?”
Because there is full visibility into what data was used to train models, decentralized AI is “a perfect mechanism to actually make those decisions,” Govindji argued.
Govindji acknowledges that deploying AI remains slower than using a centralized server, and running something larger than GPT-2 would be a challenge. However, the performance gap is closing as ICP and other blockchains ramp up GPU capacity.
Ultimately, Decide AI’s project is more a proof of concept that showcases what is to come, Govindji observes. But going forward, he anticipates more complex AI systems being brought on-chain.
Blockchain-Based Frameworks Make AI Accessible
While many decentralized AI projects are focused on contemporary LLMs and computer vision models, the more experimental field of Deep Reinforcement Learning (DRL) poses unique challenges.
The DRL process essentially releases AI agents into unstructured data environments and lets them learn how to complete a given task through trial and error. The technique is widely used in robotics and autonomous vehicles, where AI must learn how to navigate new environments. It has also been used to teach AI to play video games.
Like other forms of machine learning, DRL requires a significant amount of computational resources due to the complexity of the algorithms involved and the high dimensionality of typical DRL environments.
Researchers recently proposed a blockchain-based framework for DRL to overcome some of these challenges.
Whereas other decentralized platforms are focused on providing the physical infrastructure for AI training, the new design would let users access various AI services needed for DRL.
Because a single DRL architecture might require classification, regression, clustering, and natural language processing, the technology has remained inaccessible to many would-be users because they don’t have the necessary resources or know-how.
Solving this problem requires a more complex approach that resembles existing decentralized AI platforms but takes into account the dynamic requirements of DRL. Rather than simply providing raw computational resources, the proposed network would use blockchain-based smart contracts to allocate different AI tools to complex and diverse DRL tasks that require intricate modeling and optimization.
On-Chain AI: The Holy Grail of Decentralized Artificial Intelligence?
Meta’s Llama 3.1: The Reality of Running a 405B Parameter Model at Home