Clear explanations of LLM research, retrieval systems, and the decisions behind them.
The architecture behind this site's ML features: one embedding batch feeds both the hero's latent map and the assistant's retrieval index. Interpretable axes, hybrid scoring with a dense floor, committed artifacts with staleness checks, and the failure modes the design works around.
Read articleRefusal in instruction-tuned LLMs lives in a single direction you can delete with a rank-1 edit. We post-trained Llama-3.2-1B-Instruct to spread that signal across many directions, raising the attack rank from K=1 to K≥16 with no measurable utility hit.
Read articleA structured academic recap of a NeurIPS 2025 Best Paper Award–winning work that studies how reinforcement learning can scale by adopting self-supervised objectives, supplemented with insights from the authors’ interview.
Read article