An AI that enjoys tedious academic tasks

In just six months, University of California San Diego’s in-house AI suite, TritonGPT, grew from a test group of 400 users to a full deployment reaching 37,000 faculty, staff, and students.

The full rollout comes weeks after TritonGPT upgraded to Llama 3, a large language model (LLM) that provides better reasoning, code generation, and instruction tracking capabilities.

There were some surprises, but also some notable benefits. According to UCSD Chief Information Officer Vince Kellen, the cost savings compared to using external AI systems have been significant.

“The goal of this exercise is to keep costs at a scary level, which we’ve been able to do,” Kellen said, adding that powering the AI ​​suite costs the university about a tenth of that. what it would cost for an external subscription service. .

DEPLOYING AI IN HIGHER EDUCATION

The initial 400 users were selected by nomination, followed by a full rollout — first to finance staff, then to faculty members. Usage patterns vary across different employee groups, with 20% of users accounting for 80% of queries, as Kellen noted.

Responses from faculty have been mixed, ranging from enthusiastic early adopters of computing to skeptics who view it as a passing fad.

“Professors are people too, and I think they’re a lot like most people and there’s a spectrum from ‘That’s so cool’ to ‘That’s the devil’s work,’” Kellen said.

If so, the devil works in the details, the tedious details. The university’s AI assistants take on tasks that most staff members would rather not do.

“It’s kind of like picking the low-hanging fruit,” Kellen said. “We don’t do very high-risk uses – like giving students advice on which specialty to choose, that has a lot of risk and a lot of uncertainty. Instead, we focus on the myriad mundane things we do in college, whether teaching or administration.

WHAT CAN AI DO IN A UNIVERSITY?

The TritonGPT suite currently offers a number of wizards, each designed for specific tasks, and more are under development to expand its capabilities.

According to Kellen, one of the first things many users do is ask the AI ​​assistant questions about themselves (a practice he jokingly referred to as “glory-seeking”).

Other common uses include summarizing or rewriting text, brainstorming, and developing job descriptions with AI.

AI-POWERED JOB DESCRIPTIONS

The Job Description Writer is an AI assistant that allows staff to enter certain data and receive an AI-created description that has been reviewed to eliminate unintentional bias or wording that might discourage certain groups from applying.

The job description writer was particularly well received. Kellen notes that it has significantly reduced the time managers spend on this task, and HR has reported higher quality candidates since its launch.

“Many managers spend too much time writing job descriptions, and when we released them, it turned into a five-minute task rather than a half-hour ordeal,” Kellen said.

AI-ASSISTED GRANT MANAGEMENT

One of the newest assistants, the Fund Manager Coach, is an AI that provides personalized guidance to employees who oversee grants and related finances.

It is still too early to assess its effectiveness.

“People are very excited about it, there’s no doubt about it, but they’re happy, it’s a little too early to comment on that,” Kellen said.

SERVE AS AN AI-POWERED CAMPUS EXPERT

The UCSD Assistant helps staff and faculty navigate university policies, processes, and documentation. The General AI Assistant, on the other hand, is designed for broader tasks such as summarizing documents, generating ideas, and writing emails or reports.

According to the project’s website, myriad data was used to train the AI ​​assistants, including the admissions website, business analytics center, university policies and a ServiceNow knowledge base.

Kellen reports that overall user feedback has been overwhelmingly positive, likely in part because of the voluntary nature of adoption.

“Most new technologies don’t get positive reviews when they’re launched in the enterprise space,” he said. “If we implement a new payroll system, everyone hates it the first year. But here, because it’s more voluntary, the feedback has been really good.

REDUCING THE COSTS OF AI IN THE PUBLIC SECTOR

TritonGPT was not UCSD’s first foray into AI. The university leveraged its experience monitoring GPU usage from previous machine learning projects to inform data planning for TritonGPT. They tested the AI, found that it worked well, and then bought powerful computers to run it.

According to Kellen, they are now at a point where they have all the computing power they need, with enough power to handle more users in the future. They monitor its daily usage and have tools to speed it up.

“If you go with Microsoft and OpenAI, they’ll want to charge you for enterprise-class products for $30 per user per month,” Kellen said. “That seems to be a going price, maybe $20 or $40. But around $30 per user. Right now we’re designed for a tenth of that price.

The university is working with Danswer.ai, a new company founded by UCSD alumni, to produce the user interface.

In the future, UCSD will partner with Protopia to enhance data protection measures within TritonGPT. This includes encrypting information entered into chat windows to prevent accidental disclosure of personal information.

Notes and other personal information about an individual are currently prohibited. The AI ​​system is designed to protect private information; most of the data it uses does not include personal data.

In the future, they plan to connect TritonGPT to databases for tasks such as grant management, but with strict safeguards to ensure that only general, non-personal information is accessible.

NAVIGATING RESPONSIBLE AI

While TritonGPT has proven valuable to staff and faculty, the question remains: what about the students?

For the moment, the tool is not accessible to students. Kellen explained that the university wants to prioritize safety and accuracy before expanding access to younger audiences.

“We want to go further on all the safety mechanisms because students don’t always listen to adults,” Kellen said. “We want to make sure all the safety measures are really in place before we get there.”

Additionally, it is not used for tasks such as connecting students with educational pathways. Kellen stressed the importance of addressing potential issues with “hallucination” and accuracy in AI-generated responses.

“If the data the language model is trained on is skewed, then the answers can be skewed,” he said. “Obviously, nothing will be perfect in this area. And that’s not our goal, perfection. Our goal is simply to achieve the highest possible level of quality and safety.

AND AFTER?

The potential applications of TritonGPT continue to expand, with new use cases emerging as the technology evolves.

Recent successful pilot projects have demonstrated TritonGPT’s potential for tasks such as streamlining work-study reports and summarizing courses.

However, one of the big goals for the future is to increase adoption. Kellen acknowledges that experts may be less inclined to adopt new tools, especially if they don’t immediately see the value in them.

Kellen’s hope is to increase engagement by demonstrating the usefulness of the tool. This may require a clearer usage and training path.

“You have to match technology to a very specific task,” Kellen said. “And that’s part of user training.”