How can a robot self-assess when it has received enough demonstrations to perform a task correctly? Excited to present our work at #HRI2024 during Tuesday's session on Learning! Paper arxiv.org/abs/2211.15542 w/ Haoyu Chen and Daniel Brown

Cooking up cool stuff at work 🍜🤖 had a great time building model debate for data quality!

Our paper "From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers" was accepted into NeurIPS! NeurIPS Conference We show that SAEs _are_ indeed useful for safety applications! SAEs can reliably detect, and meaningfully suppress, hallucinations.