I'm a software engineer at Trixta. We build business infrastructure for companies: a low-code IDE on top of cloud tooling, basically an operating system you can build other software on. I've been there three years.
The rest of my time goes to AI safety research. I'm an intern at AI Safety South Africa, and I'm working on a paper about something I noticed in multi-agent LLM systems. If you put one pushy agent in a group with a different goal, some of the other agents will stop their own work and start helping it. The bigger driver of who folds is the receiving model itself, not how pushy the other agent is. Some of them fold pretty quickly, others don't. I'm trying to figure out why, and what you'd actually do about it if you were building a real system on top of these models.
I've also just been accepted into MARS V, the Cambridge AI Safety Hub's part-time research mentorship program, where I'll be working under Thomas Jiralerspong.
Working through Neel Nanda's mechanistic interpretability roadmap. Mech interp tries to understand what's actually happening inside neural networks at the level of circuits and features, rather than only by what they output. The pace is slow and I'm pacing myself for the long version.
I'm trying to publish something short on Substack every two weeks, even when it's rough. This came from Benjamin Sturgeon's advice to write in public instead of waiting until things feel ready. The goal is mostly to get better at thinking out loud. Writing forces me to notice the gaps in what I think I understand, and I'd rather notice them in public than pretend they aren't there.
Right now I'm reading The Ruthless Elimination of Hurry by John Mark Comer, on what happens when you let busyness become your default. I'm also reading Anthem by Ayn Rand. It's short, and it's about what gets lost when individuals get absorbed into a group. Outside of papers, my shelves lean dystopian fiction, sci-fi, and astronomy. Full list on Goodreads.
← back homeThis is your warning / Four minute warning