Oscar Gilg
I work on empirical AI welfare research at MATS 9.0, mentored by Patrick Butlin.
Projects
- Zombuul 2026
I built a Claude Code plugin for launching autonomous research agents with access to GPUs.
- Personas use a shared preference vector MATS 9.0, 2026
- Split Personality Training SPAR 2025
Trigger-conditioned persona fine-tuning. Accepted at ICML 2026.
Bio
I studied Mathematics and Computer Science at Oxford, graduating in 2023 with a First-Class degree and Master's Distinction. Reading Dennett's Consciousness Explained and From Bacteria to Bach and Back got me interested in philosophy of mind, and in my final-year dissertation I tried to operationalise properties of consciousness and described how they could arise in a hierarchical generative model.
After Oxford I spent two months in Bochum working with François Kammerer on formalising illusionist theories of consciousness, then joined Optiver in Amsterdam as a quantitative trader. I spent over a year on a 6-trader equities desk; trading taught me to test hypotheses fast and make decisions under uncertainty.
I had first read Bostrom's Superintelligence as a teenager and dismissed the timelines as too aggressive. When o3 was released that intuition flipped: I left Optiver and applied to MATS, where I now do empirical AI welfare research with six-month extension funding.
CV
Contact
- Email: oscar.gilg18 [at] gmail [dot] com
- github.com/oscar-gilg
- linkedin.com/in/oscar-gilg