Center for Intelligent Imaging SRG Kickoff Meeting: Introducing “Versa”
On Monday January 8, the Center for Intelligent Imaging (ci2) kicked off 2024 with its first Specialized Research Group (SRG) meeting of the year. This session focused on UCSF’s approach to AI and large language models. Radiology department chair Christopher Hess, MD, PhD, gave opening remarks and ci2 co-directors Andreas Rauschecker, MD, PhD, and Sharmila Majumdar, PhD, spoke about ci2’s relationship with artificial intelligence tools and the evolving culture surrounding them.
Oksana Gologorskaya of UCSF’s AI “Tiger Team” was the featured speaker, giving an overview of the university’s strategy to adapt to new AI technology. The Tiger team has worked quickly in the months since Chancellor Hawgood announced UCSF’s focus on developing AI tools for university use. Pierre Nedelec, MS, a data scientist in Rauschecker’s lab, demonstrated Versa, UCSF’s secure Large Language Model (LLM) which is now available to researchers, clinicians, and staff.
Publicly available LLM tools, trained on vast amounts of written text to generate near human-quality writing, do not meet requirements for information security in a healthcare setting. To address this challenge, UCSF Health has adopted the US Department of Health and Human Services framework for Trustworthy AI (TAI) to guide our oversight of AI in our health system. Versa was built from the ground up to comply with HIPPA. It supports multiple models which users will be able to toggle between, including OpenAI’s newest GPT4-turbo. GPT-4-Turbo has knowledge of world events up to April 2023 and access to a larger conversational context, outperforming its predecessors GPT-3.5-Turbo and GPT-4.
For UCSF users, Versa comes in two formats: Versa API for advanced users, and Versa Chat, which is similar to the popular ChatGPT. By keeping all information within secure UCSF systems, these versions are suitable for sensitive medical use.
UCSF members can sign up today for the Versa waiting list; some requests are approved in as little as two days.
