CODA a GPU kernel abstraction for simpler transformer block reparameterization LLM Homogenization in Creative Writing investigating levels of homogenization by LLMs in creative writing and exploring methods to achieve human-like levels of stylistic and semantic diversity Profile-Kernel a CLI tool to profile CUDA kernels and automatically identify performance bottlenecks Human-AI Complementarity an investigation into task routing for human-ai complementarity and scalable oversight RotoSLAM improving SLAM scale consistency for rotational motion via depth estimation augmentation Enhancing Transformer Programs with Structured Biases a performance-enhancing modular extension of Transformer Programs (Transformers designed for mechanistic interpretability) Video Reverse-Engineering in Frontier Foundation Models an examination of modern foundation models’ visual analytical capabilities
