As AI becomes ever more capable, the field of AI alignment is increasingly in the spotlight—we certainly don’t want to accidentally build the proverbial paperclip optimizer that turns the solar system into paperclips. Indeed, it is a primary area of research for one of the authors of this blog post. Although improving AI alignment is at present time clearly directionally good, we argue here that perfect alignment is both impossible and undesirable.

Hi everyone! I’m an assistant professor at Carnegie Mellon and research scientist at Google Deepmind. My appointment is in the Language Technologies Institute, but I’m broadly interested in how AI is and could soon be affecting how we as humans learn, communicate with each, build tools, and build trust in systems and eachother. For a while now, I’ve wanted to create a place for I and my students and friends to post things we’ve been thinking about. My goal is to encourage myself (and my students) to do more low-stakes writing. Hopefully, this blog ends up as a platform for us to write about ideas, positions, and projects that don’t fit neatly into an AI conference paper.