Nous Research
summary
Nous Research describes a working open alternative to closed labs: a Psyche network that trains a 40B-parameter model on ~500 billion tokens via a global distributed run, cutting training bandwidth by 4,000x so distant nodes perform like colocated GPUs. Models get uploaded to HuggingFace every few hours, Appropos provides 1,200 reinforcement-learning tasks for fine-tuning and benchmarking, and the team cites ~8,000 H200s idle worldwide and a claim that half that compute could retrain Deepseek R1 in 500 hours.
AI-written summary · about
chapters (9)
- 0:00The case for open AI
- 1:07Critique of alignment
- 3:35Systemic bias in AI
- 4:57History of alignment
- 6:47Intellectual elite control
- 8:39The second path
- 10:28Decentralized training
- 12:03Utilizing idle compute
- 13:36Call to action
related talks
17mDistributed Intelligence: The Alignment Protocol
29mAligning DeepSeek R1
44mApropos DeepSeek
54mThe Future of AI: Building the Most Impactful Technology Together
34mMapping the Open Model Landscape in 2025
25mProtecting AI Infrastructure
36mSkyRL: Democratizing RL Training for Researchers
32mFrom Deep Research to Evaluation and Optimization of Advanced AI Age...