AI Safety Takes

AI Safety Takes is a personal newsletter by Daniel Paleka, published on Substack at newsletter.danielpaleka.com. The newsletter focuses on AI safety research and making the future go well, covering topics such as constitutional AI, mechanistic interpretability, scalable oversight, reinforcement learning from human feedback, adversarial attacks, and model alignment challenges. Daniel Paleka is a PhD student at ETH Zurich, advised by Florian Tramèr, with research focused on security and failure modes of artificial intelligence. He began his PhD in September 2022 and launched the newsletter in November 2022. Paleka also holds notable competitive mathematics credentials, including silver medals at the International Mathematical Olympiad in 2014, 2015, and 2016. The newsletter has grown to over 1,000 subscribers, with a significant fraction working in AI. Paleka describes the newsletter as evolving from structured monthly paper summaries to more freeform posts published whenever he has something interesting to say. He also runs a separate Substack called Random Features for non-AI-safety content. The newsletter has no paid subscription model and is free to all readers.