From AI to ZI

From AI to ZI is a Substack newsletter and blog authored by Robert Huben, a PhD mathematician who completed his doctorate at the University of Nebraska-Lincoln in 2021. After spending a year as a scientist at a defense contractor, Huben received a grant from Open Philanthropy running from October 2022 to September 2023 to pursue a self-directed AI safety research and education program. The blog was created to document this journey, with posts covering AI safety concepts, paper summaries, original research ideas, and Huben's personal views on the existential risks posed by unaligned AI. A signature feature is that each post is rated on a scale from "AI" to "ZI" reflecting how directly relevant the content is to AI safety. During his grant year, Huben co-authored two notable research papers. The first, "Sparse Autoencoders Find Highly Interpretable Features in Language Models" (co-authored with Hoagy Cunningham, Aidan Ewart, Logan Riggs Smith, and Lee Sharkey), was presented at a NeurIPS 2023 workshop. His work focused on mechanistic interpretability — using sparse autoencoders to identify monosemantic features in neural network activations. He also worked on attention-only transformers. After the grant period ended in September 2023, Huben indicated he was applying to work at AI safety organizations and research groups. The blog shifted to infrequent posting. As of early 2026, the blog is semi-dormant, with payments disabled on Substack, but it continues to serve as a public record of his AI safety research and writing. Huben is also active on LessWrong under the handle Robert_AIZI.