Pi Network is exploring the capability of Pi’s global network of Nodes to support decentralized AI training and computing tasks.
Pi Nodes have long been designed to do more than securing a ledger. Although the Pi blockchain benefits from the decentralized nature of many distributed nodes around the world, Pi’s ledger itself is energy efficient and does not require the full computational resources of such a large distributed Node network. That creates a pool of unused computing capacity that can enable new utilities such as decentralized cloud computing. Third-parties requiring computing power for AI model training could utilize these unused resources of Pi Node operators, who opt in, and pay the operators in cryptocurrencies. Utilities on Pi Nodes constitute an important component of Pi utility, alongside Pi apps, platform-level utilities and local commerce.
This Node utility aims to address two emerging issues in the AI era.
- The limitations of centralized computing: These include data center constraints, energy concentration, and issues like catastrophic forgetting or global-state bottlenecks.
- Increasingly AI-driven demand for computing: As AI models and agents advance and the AI economy expands, the demand for computing resources increases. Society will need to provide an unprecedented amount of computing power to support the AI economy, by creating new and utilizing unused existing computing power as much as possible. Scattered and unused computing power relies on a distributed network and technologies to consolidate, coordinate, and scale to be productive.
Pi is uniquely positioned to solve these two challenges because Pi itself is already a distributed network and has over 421,000 Nodes (representing over 1 million CPUs) that allow for distributed computing. Furthermore, Pi has tens of millions of KYC identity-verified participants who can opt in to provide human-in-the-loop support for relevant AI learning processes in exchange for compensation in cryptocurrency. This, in addition to the computing power from Pi nodes, can offer a unique resource for scalable, authentic human input in AI systems, and further complete the one-stop service to AI clients.
Together, these elements provide new approaches to AI infrastructure that allow a distributed system to capture production into its network and enable people to contribute to AI’s production processes directly, and receive compensation through blockchain-based payments in return.
Last October, Pi successfully completed a proof-of-concept project with OpenMind in which a small group of Pi Node operators ran image recognition tasks for the company.
The OpenMind Case Study
OpenMind is developing an operating system and open-source protocol for robots to think, learn, and work together—like Android OS for robots. Like other physical AI efforts, OpenMind needs significant computing power to train, evaluate, and run their models.
To test the feasibility of Pi’s distributed computing, OpenMind developed a container that could request computing tasks from individual computers. The team shared this container with volunteer Pi Node operators who downloaded it to run on their machines.
OpenMind then sent tasks through the container requesting the computers to process different images using OpenMind’s AI image recognition model—an example use case where training for image recognition is essential for OpenMind’s robots to interact with the real world. The goal was to discover as many discrete objects as possible within these images, using the volunteer Pi node operators’ computing power to do so.
Results
The proof-of-concept- project was run successfully. Tasks were correctly pushed to the external testers (volunteer Pi node operators) and valid results were sent back to OpenMind—and the use case was proven, where Pi Nodes can opt in to run computations defined and requested by a third-party unrelated to their blockchain obligations, and return meaningful results to a third-party client.
Key statistics from the pilot:
- 7 volunteer Pi Node operators participated in the test
- End-to-end distributed pipeline functioned successfully
- Broadcasted jobs received acknowledgments from 7 workers within one second
- Inference results were returned from multiple workers within 4 seconds
- Results contained correct object detections, including expected labels such as bus and person, along with bounding boxes
- Jobs were successfully pushed to external tester machines and valid detections were returned, confirming both distributed broadcast and result return-path reliability
Overall, the experiment validated that distributed Pi Nodes can execute AI-relevant workloads and quickly return useful results.
Next Steps
Distributed AI training is still at the research stage in the world. Much work by researchers and interested enterprises around the world should continue exploring whether and how the shift from fully centralized AI training to more distributed approaches is possible. Such research directly relates to the challenges outlined earlier: addressing the structural constraints of centralized computing, utilizing scattered and unused computing power for AI, and creating opportunities for individuals to participate meaningfully in AI-driven production.
By evaluating how unused Node computing capacity can support external AI workloads and packaging computing capacity with authentic human input together, this initiative may provide an alternative infrastructure option that AI companies and startups could explore when seeking solutions to their AI training needs.
Overall, Pi’s research into this Node utility complements Pi’s vision for the future of blockchain and AI, where decentralized infrastructure contributes to integral components of the future economy and supports equitable participation and distribution.
