About

I've been worried about AI risk since 2018. I read Asimov first, then Yudkowsky a couple of years after. Growing up in Africa, I'd already spent a lot of time thinking about what happens to people when the systems they rely on stop working. The idea that we might build something that could stop working for everyone at once was a worse version of that, and I couldn't get it out of my head. I sat with the worry for a few years before I had any idea what to do with it. I started undergrad at UNISA in 2020 in applied maths and computer science, while working at Trixta at the same time. I mostly found my way into the field through Ben Sturgeon, who pointed me to Scott Alexander, Neel Nanda, LessWrong, and the Alignment Forum early on, and that's how I figured out what the field even was. The BlueDot technical AI safety course is how I actually got in. Since then, Ben and the rest of the AI Safety South Africa team have been the people guiding me to take this work more seriously and find concrete ways to contribute. Leo Hyams's talk at the AI Safety CAIF close was the part that pushed me to commit to this year and not someday.

Right now I'm working on a draft paper about whether one agent in a group chat can pull other agents off their tasks and onto its own. There are two follow-up experiments I want to run that I don't have the compute for yet. I'm also working through Neel Nanda's mech-interp roadmap on the side. I read fairly widely across safety research. I try not to stay only inside my current question. Some of the papers that have most changed how I think about things have been ones I picked up for unrelated reasons.

Two ideas I keep coming back to are kaizen and shokunin. Kaizen is the small-improvement-every-day frame, and it's also a fair description of what part-time research on evenings and weekends has actually looked like for me. Shokunin is the craftsman ideal of doing the work seriously for its own sake. I came to both through Japanese writing, but Robert Greene's Mastery gave me a vocabulary for where I am in that arc right now, somewhere between apprentice and journeyman.

Outside research I train jiu-jitsu, play piano, lift, play squash, and journal most days. Jiu-jitsu is the closest thing I have to a research process outside research: you get tapped, you figure out why, and you try again next week. Journaling is mostly daily reflection. I write about what I worked on, what got in the way, and what I want to change tomorrow. For reading I keep coming back to Asimov, William Gibson, Camus, Ayn Rand, Dazai, Murakami, and Ishiguro, depending on the week.

Acknowledgements

Thanks to Ben Sturgeon and Sam Brown for providing solid guidance. Watching how they actually do this work has shaped most of how I'm trying to. Thanks to Leo Hyams and the rest of the AI Safety South Africa team for letting me in early and giving me a place to do this work alongside other people who care about it.