Raphael Sarfati
AI interpretability for interspecies communication
Cornell University + start-up
Bio
I am a physicist with a broad curiosity for natural phenomena, animals, and emergence in AI models. In the past I have studied animal collective behavior, notably how sheep "flow" together and how fireflies synchronize their flashes. Since 2023, I have mostly focused on understanding the inner workings of LLMs ("interpretability"), both in academic and start-up settings. I'm excited to collaborate with people from a wide range of backgrounds and perspectives. For a bit more about me: raphaelsarfati.xyz
Mentee must-haves/nice-to-haves
Curiosity and creativity!
Mentor support
Brainstorming, connections
Mentor-led project
Interpretability applied to AI models for animal communication
I am really passionate about advances in AI to decode animal communication. While several projects are underway (Earth Species Project, the Cry Wold Project, CETI), I am particularly curious to investigate what recent developments in the field of AI interpretability (i.e., how AI models "think") may discover about animal communication by analyzing how current generative models make predictions. For example, we might be able to identify new patterns in the way models represent vocalizations from encoded structures in the data they learn. In addition, I would be very interested to discuss other ideas and proposals from mentees.