Ida Caspary

Welcome !

I am Ida Caspary, a PhD student at Imperial College London. I am part of the Data Learning group at the Imperial Data Science Institute and of the Ada Lovelace Academy. My supervisor is Rossella Arcucci. Currently, I am doing Constellation's Astra Research Fellowship, mentored by Asa Cooper-Stickland, on AI control.

My main research is on technical AI Safety; the focus of the the AI safety research group I set up at Imperial College London is Mechanistic Interpretability. I am also investigating how training decisions after pre-training affect what models reveal about their reasoning and goals. My other research interests include robotics / vision-language-action models, and cryptography, e.g. homomorphic encryption.

I moved to Scotland after high school for my bachelor's in Computer Science. I spent one semester on exchange in Ottawa, Canada. For my dissertation project, I researched optimised legacy system upgrades, supervised by Dr Patrick Maier. I upgraded the International Atomic Energy Agency (IAEA) software system of the wildlife radiation database maintained by the Environmental Radioactivity Laboratory (ERL).

I am also very fascinated by faithfulness in chain-of-thought research. As a MATS project, I developed, mentored by Neel Nanda and in collaboration with Anton Poleatev, a little investigative kit called CoTFaithChecker - work that is still ongoing.

I'm a former

• CERNler – topic: Explorative Data Layout Analysis of Next-Generation High Energy Physics Datasets, on site at CERN, Switzerland
• IBMer – intern, then working student; worked 17 months at IBM Cognitive Analytics Consulting
• Bosch Working Student #LikeABosch – in the IT Department! most exciting time to be there, while it was built out to a fully automated semiconductor manufacturing plant
• Journalist – with ZDF, BASF, Sächsische Zeitung, Tagesspiegel, and lots of uni/school newspapers. I love researching intriguing subjects.