Say Hello to TritonGPT

Article Content

There’s a new bot on campus.

Meet TritonGPT—UC San Diego’s specialized artificial intelligence-powered information and resource assistant.

Next week, following the success of a two-month early user program, TritonGPT will enter the next phase of its campuswide launch. The “second wave” pilot will grant access to campus employees in the Vice Chancellor-Chief Financial Officer area, with additional VC areas to be rolled in throughout April and May as IT Services moves toward making TritonGPT accessible to all staff and faculty in campus and health sciences units.

So what exactly is this new tool? Picture “ChatGPT,” but uniquely trained to provide information and generate text that’s specific to UC San Diego.

Tailored for use in administrative roles and hosted on local infrastructure at the university’s San Diego Supercomputer Center—ensuring that privacy and security are paramount—TritonGPT is UC San Diego’s in-house alternative to other commercially available generative AI tools. At its core, it’s designed to help campus employees save time and increase efficiency in their daily tasks.  

“As artificial intelligence continues to rapidly impact higher education, UC San Diego is leading the way in leveraging the potential of new technologies to improve the way we live, work and learn,” said Chancellor Pradeep K. Khosla. “TritonGPT is a byproduct of the culture of collaboration and risk-taking that infuses our institutional DNA—and the continuation of UC San Diego’s decades of leadership at the forefront of AI research and innovation.”

Ahead of its rollout, get to know TritonGPT, its capabilities and limitations, and the story behind its development:

Helping you do your job more efficiently

Built using advanced indexing of the university’s key websites, TritonGPT is adept at providing information and support related to UC San Diego's policies, procedures and campus life. It's like having a personal assistant who knows a lot about UC San Diego.

"We have focused the AI capabilities of TritonGPT on our institution's content, making available the knowledge in that content to our community,” said Chief Information Officer Vince Kellen. “TritonGPT will be helpful in not just answering university questions, but also in streamlining many work tasks. We are excited about what the future will bring."

With an interface that resembles ChatGPT and other AI-based text generation tools, TritonGPT provides documentation and sources alongside its answers: for example, directing the user to the Blink page or other UC San Diego sites from which its response was derived.

The tool employs Meta’s Llama 2 open-source language model—roughly the equivalent of GPT 3.5—and is useful for answering questions such as “Do UC San Diego employees have the day off for César Chávez Day?” or “What restaurants on campus serve burritos?”

It’s also capable of responding to prompts such as, “Draft an email to prospective students highlighting the unique aspects of UC San Diego.” Among a pool of approximately 500 early users, who were hand-selected from areas across campus to begin using the tool in beta earlier this year, many found success using the tool for similar tasks involving text and language, as well as brainstorming ideas, tailoring documents to appeal to specific audiences, summarizing content, creating presentations and more.

The tool also includes a Job Description Writer function for use by hiring managers. Through a dialogue with the chatbot, employees can share the specific requirements of a role, and TritonGPT will draft a job description that can then be refined and edited by the hiring manager. Developers are also in the process of creating a Fund Manager Coach function that will provide personalized advice for individuals who are responsible for overseeing grants and managing departmental finances.

While TritonGPT’s capabilities are impressive, it’s important to note that it’s intended to serve as an assistant, not an authoritative source—evidenced by the disclaimer that appears at the bottom of the page: “TritonGPT responses are generated by artificial intelligence and may contain errors. Check sources and refer to actual policies and laws for reliable information.” Like all large language models, TritonGPT may “hallucinate” or provide inaccurate or out-of-date information, and users are encouraged to apply critical evaluation skills and remember that they are still responsible for any content they use that’s generated by the tool.

It’s also important that users rate TritonGPT’s answers using the thumbs-up or thumbs-down graphics that appear alongside its answers, which helps improve the tool for users asking similar questions in the future.

Antonio “Tony” Nava, senior strategic programs manager in the Office of Operational Strategic Initiatives, has played a primary role in supporting the introduction of TritonGPT. In his role, he’s hosted initial training sessions for early users, led the development of the first-ever AI Essentials at UC San Diego training modules and spearheaded the exploration of the ways campus employees can apply the tool to improve processes.

He's excited to see how broad use of TritonGPT across campus will free up time for employees to address higher-level tasks and become more productive.

“Think about what your individual mission is as a staff member here at UC San Diego,” Nava likes to tell campus employees. “Think about who you’re doing that work for, who you’re serving—and then think about ways that this will help you do that smartly, more directly and more comprehensively. This is something that is really intended to help us deliver better outcomes.”

Becoming TritonGPT

Skyrocketing from idea to proof-of-concept in less than two months—followed by the two-month early user pilot, TritonGPT is on the fast track due to the hard work and dedication of its key creators at UC San Diego: Brett Pollak, senior director in Workplace Technology Services; Adam Tilghman, analyst/architect with Academic Technology Services; Matthew Holland, a student staff programmer; Jeremy Wiles, user experience designer; and senior AI architect Jack Brzezinski.

“It all started when we were creating guidance on appropriate use of generative AI tools in the workplace,” said Pollak. “It got us thinking that we could potentially bring in an open-source language model and install it within our ecosystem, and then really apply this to the UC San Diego context.”

With the enthusiastic support of Chancellor Khosla, the team got to work bringing in the necessary hardware and integrating the large language model into a platform and interface created by Danswer, an open-source AI-powered search engine startup founded by two UC San Diego alumni, Yuhong Sun ’19 and Chris Weaver ’19, MS ’20.

While a handful of other peer universities have already launched their own campus-specific generative AI tools, there’s one thing that sets UC San Diego’s model apart: It’s hosted on local infrastructure at the San Diego Supercomputer Center rather than relying on a commercial platform like Microsoft’s Azure OpenAI service. This approach, which leverages a unique and internationally renowned campus asset, not only is more cost effective, but also is more private and secure.

“We’re not sending the data off to another location to another potential owner,” said Nava. “So in that, there’s the opportunity to protect us from risk, to protect the information of the people that we are serving, and to be able to do things in a safe way that is more in line with our goals as an institution.”

Its eventual rollout to approximately 20,000 faculty and staff, said Kellen, makes TritonGPT one of the largest—if not the largest—initiatives of its kind utilizing a supercomputer center.

The 500-participant early user pilot program allowed ITS, which provides the software and domain expertise behind TritonGPT, an opportunity to improve the tool using valuable user feedback—and to track its usage and utilization to help forecast what hardware would be needed on the back end to support a campus-wide rollout.

In addition to the algorithms behind TritonGPT, it’s been essential to employ what Pollak describes as “reinforced learning with human feedback” to train and refine the tool. This is where the involvement of pilot participants has proven vital as the developers prepare the tool for campuswide deployment. 

“You’re essentially saying, ‘No, this is the wrong response. Here’s the right response.' So that next time it sees a similar question it can actually learn from that mistake, and then make it better for the next person who asks it a similar question,” Nava explained.

Looking to the future

As ITS continues to expand the hardware needed to support an increased user base, TritonGPT will initially only be available to campus staff and faculty and health sciences staff and faculty. 

While there are no immediate plans to make TritonGPT available to students, Pollak says it’s certainly a goal for the future—one that would require additional infrastructure and testing. Currently, students who also serve in staff roles will be granted access to the tool when their respective VC areas are rolled in.

Pollak hopes that the tool also will be expanded in the future to include additional specialized functions beyond the Job Description Writer and Fund Manager Coach.

“Ideally, we would have a ‘coach’ for almost any position that we would have on campus—an information resource that’s specific and pertinent to each individual’s job,” said Pollak. “We are expanding TritonGPT by adding additional specialized assistants, each trained on targeted datasets to enhance support for specific use cases.”

Members of the campus community should stay tuned for future iterations and additional details related to the campuswide rollout. As the tool becomes available, eligible individuals will receive an email with the login link and are asked to not attempt to access TritonGPT until they receive that confirmation email. In the meantime, staff and faculty are strongly encouraged to take AI Essentials at UC San Diego Training Module 1 through the UC Learning Center before using TritonGPT.

Through the creation and implementation of TritonGPT, UC San Diego affirms its dedication to thinking boldly, pursuing new frontiers and finding innovative solutions to the world’s challenges.

“AI isn’t going anywhere, and change is coming. We will be better positioned as a university, as a staff, and as individuals if we start to find ways to integrate with it and ways to leverage it to serve students and our campus community,” Nava said.

Learn more about research and education at UC San Diego in: Artificial Intelligence