If you missed the @NousResearch office hours, here is a TLDR on what you missed 🤖 Psyche is democratizing AI training on @solana Psyche Project Overview: Psyche aims to enable collaborative, open-source training of large foundation AI models, reducing reliance on closed labs and using algorithms to minimize communication overhead in distributed GPU training. Current Model Training: They are currently training a 40 billion parameter DeepSeek-style dense LLM, designed to be powerful yet runnable locally. Decentralized Coordination: Training coordination occurs over the Solana network to ensure decentralization and avoid single points of failure. Networking Challenges & Solutions: A peer-to-peer gossip network using IRO was built to handle data transmission between nodes, addressing Solana's data volume limitations. Shift to Python Backend (Torch Titan): Psyche's codebase is being "Pythonified" by abstracting the trainer backend to support Python-based trainers (integrating a fork of Torch Titan), expected to provide a ~2x speedup. Fine-tuning Capabilities: Psyche will soon support full fine-tuning, including advanced methods developed at Nous, with plans to fine tune an upcoming Nous model on the network. Treasurer Audits: The Solana-based treasurer code, managing reward distribution to compute providers, has successfully passed two independent audits. Future Research (Psyche 2.0): Future research includes exploring Tensor Parallelism (TP), Pipeline Parallelism (PP), and Expert Parallelism (EP) over the internet to make Psyche a platform for developing advanced AI algorithms. Community Contributions: Developers can contribute to the open-source Psyche codebase, create new environments for the Atropos reinforcement learning codebase, and support the decentralized inference network by using the Nous API. Long-term Vision: The ultimate goal is to democratize AI model training, enabling anyone to contribute compute to a global network, similar to "folding at home" or "SETI at home," to create the world's best open AI. Video:

@aeyakovenko @0xMert_ I need to shill you Nous 🤖