The hub for AI safety at the University of Pennsylvania, by students for students
We collaborate with industry and academic institutions to mitigate risks associated with AI systems — advancing robustness, monitoring, alignment, interpretability, and systemic safety through applied and empirical research in state-of-the-art deep learning systems.
<aside>
<img src="attachment:733df430-2776-43d3-85c8-c0c3d9e75aa1:image.jpg" alt="attachment:733df430-2776-43d3-85c8-c0c3d9e75aa1:image.jpg" width="40px" /> AI Safety Fellowship: Fill out the application form by September 15th
</aside>
<aside>
🗞️ Newsletter: Stay updated on Penn’s AI and safety research at pennai.substack.com
</aside>
Publications
Recent Updates
- May 15, 2024: Publication accepted to ACL ‘24 (Language Models Don’t Learn the Physical Manifestation of Language, Bruce Lee & Jason Lim).
- April 27, 2024: Jan Kirchner (OpenAI) speaks to the SafeAI@Penn research group.
- March 16, 2024: Technique invented in research paper (control vectors) added to llama.cpp (Representation Engineering, Richard Ren).
- December 14, 2023: Paper cited by OpenAI’s Superalignment Team under their Fast Grants Page (Representation Engineering, Richard Ren).
- May 1, 2023: Publication accepted to ACL ‘23 (Explanation-based Finetuning Makes Models More Robust to Spurious Cues, Josh Ludan).
- Feb 27, 2023: Publication accepted to CVPR ‘23 (Zero-Shot Model Diagnosis, Jinqi Luo).
FAQ
- What do you do?
- Why is AI safety important?
- How can I get involved?