title: "Research Notes | EleutherAI Blog"

This collection of research notes from EleutherAI covers a range of topics in artificial intelligence and large language models. Posts explore subliminal learning during pretraining, methods for detecting AI deception, and filtering pretraining data to improve the safety of open-weight models. Other research announces the release of 'The Common Pile,' an 8TB text dataset, and investigates factored cognition using GPT-3. Additional topics include evaluations of model components like rotary position embeddings, different normalization methods for multiple-choice tasks, and community RLHF hackathons.