Show HN: Helios, an open-source distributed AI network using idle community GPUs

2 points

9 months ago

Hi Hacker News, I'm the creator of Helios, and I'm excited (and a bit nervous) to share it with you all. The "Why": Like many of you, I've been fascinated by the power of modern AI models, but frustrated by the high cost and centralization of GPU resources. I started wondering if we could apply the old-school distributed computing model (like SETI@home or Folding@home) to the modern AI stack. The goal was to build a network where anyone could contribute their idle compute power and, in return, get access to a powerful, multi-modal AI. The "What": Helios is an open-source platform to build that network. It consists of two main parts: an orchestrator server that manages a job queue and the workers, and a client-side worker that anyone can install on Windows or Linux. Users run the worker, contribute resources, and this forms a global, decentralized supercomputer capable of handling text, image, and audio tasks. The "How" (Tech Details): Architecture: It’s a classic Orchestrator-Worker model, written entirely in Python. The Orchestrator (orchestrator.py) is the brain, handling job distribution, worker registration, and a simple web UI. The Worker (worker.py) is the muscle that users run on their machines. Proof-of-Contribution (No Crypto!): This is the core access mechanism. To prevent spam and freeloading, you can only submit jobs to the network if you are an active, contributing worker with a good reputation. It's not based on tokens or blockchain; it's a simple, fair system based on participation. Dynamic Experts: Workers don't come pre-loaded with every model. The Orchestrator assigns AI models (e.g., a specific translation model, an image captioning model) to workers dynamically based on the current job queue. These models are pulled directly from the Hugging Face Hub, keeping the worker client lightweight. Multi-Modality: It's designed to route different job types (text, audio, image) to workers that have the appropriate models and resources available. This is very much an experimental, v1.0 project. I know there are huge challenges, especially around security (sandboxing tasks is a major next step), but I wanted to get a working prototype out there to see if the idea resonates with the community. I'd love to get your feedback on the architecture, the concept of Proof-of-Contribution, and any suggestions you might have. GitHub Repo (Code is here): https://github.com/fnoracr/helios-distributed-ai Project Page (Demo & Docs): https://huggingface.co/spaces/fnoracr/Helios_Distributed_AI-... Installers (.exe/.tar.gz): https://github.com/fnoracr/helios-distributed-ai/releases/ta... Thanks for checking it out!