I listened to this interview with Nous Research CEO Jeffrey Quesnelle on my commute this morning (that's the one good thing about working from the office in the city). He describes the company's approach developing fully open AI models - not just open weights, but open everything from the training data on up. One part of this (necessarily) is that "users, not corporations, control alignment through... system prompts and post-training tools." According to the summary, "Nous Research aims to democratize not just AI usage, but its creation and alignment. The biggest threat isn't rogue superintelligence, Quesnelle suggests, but the risk that a tiny elite controls what billions see and hear." There's a bunch of stuff in the interview, including description of the distributed web of GPUs managed by blockchain used to train the model. Definitely worth a watch - it's the first 40 minutes of this longer video, or you can just read the summary on the web page. See also: Hermes4, Hermes4 Technical Report, The Psyche Network Architecture, VentureBeat, thehomebase.ai.
Today: Total: [Direct link] [Share]
