Replies: 3 comments
-
very well written |
Beta Was this translation helpful? Give feedback.
0 replies
-
What does this have to do with StableSwarmUI? |
Beta Was this translation helpful? Give feedback.
0 replies
-
Their recent activity is rather awkward. Posting more similar things in other repos, replying and reacting to themselves, weird issues... |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Here some potential research areas or ideas that Stability AI could explore:
Formal verification of advanced models - Developing techniques to formally verify properties of large pre-trained models, beyond just simple architectures. This could help prove certain properties like benefit alignment.
Self-supervised pretraining for values - Exploring how self-supervised pretraining techniques like contrastive learning could help embed ethical values or constraints directly into models.
Value learning from indirect feedback - Research into training agents from sparse or delayed feedback signals about how their behavior impacts values, similar to how humans learn.
Model-in-the-loop safety - Develop techniques for training safer models by treating trained models as "oracles" that provide feedback during the training process of new models.
Federated Constitutional AI - Research distributed and privacy-preserving approaches to Constitutional AI, to enable inclusive model training across decentralized data sources.
Interactive proof assistants for AI safety - Develop tools based on interactive proof assistants to formally verify safety properties as models and specifications become more complex.
AI alignment via debate and argumentation - Explore how training models via debate, discussion and argumentation of different viewpoints could help align their values.
Procedural content generation for alignment incentives - Use AI-generation of incentivizing content like stories, games or simulations to implicitly steer models toward better-aligned behavior.
Self-supervised alignment via exploration & world modeling - Leverage unsupervised world modeling and exploratory behavior as a means of self-supervised alignment training.
Beta Was this translation helpful? Give feedback.
All reactions