I'm an AI safety researcher based in London. I work with the Center on Long-Term Risk and University College London as part of my PhD degree, supervised by Brooks Paige.
My research focuses on developing a pragmatic science of language model generalization - e.g. understanding why and how aligned models might become misaligned, and how to mitigate these risks. I'm especially interested in understanding how and what language models learn from data, through a combination of NLP, ML, and cognitive science.
I did my undergrad degree at Stanford, focusing on machine learning and computer science. I spent a year as a robotics engineer at a startup in Singapore before deciding to pursue a PhD. I've previously been interested in mechanistic interpretability, robotics, and open-ended learning.
Outside of work, I enjoy bouldering, cooking, and anime / scifi.
Reach out via: hello [at] danieltan [dot] cc!
Here are some papers I've made substantial contributions to. Please refer to my Google Scholar page for a full list of publications.
Easily steer OOD generalisation by adding one line to training data |
Models finetuned to write insecure code learn to admire Nazis |
Accepted at NeurIPS 2024 Steering vectors do not work universally across tasks. They also fail to generalize to similar instances of the same task. |
In proceedings, JAIR Challenges, methods, and applications of Internet image and video data for learning real-world robot tasks. |
