The General Services Administration (GSA) is looking to expand the scope of GSAi – the agency’s internal generative AI tool – to support other Federal agencies facing the same GenAI problem set.
Zach Whitman, GSA’s chief AI officer (CAIO) and chief data scientist, said today that the agency has spent a lot of time “working on our own problems,” such as figuring out which AI models perform the best and how to categorize any safety concerns.
However, he said every other Federal agency that GSA has talked to “has the exact same problem set,” and could likely benefit from something like the GSAi tool.
“We are exploring the next iteration of the GSAi platform, one where other agencies could use what we have, use it in an isolated environment, use it for their specific purposes, and own it in a tenant-based model,” Whitman said on Thursday at DGI’s 930gov event.
“So, you control your expenses, you control your telemetry, and all GSA would be responsible for would be maintaining the infrastructure and maintaining the analytics that run the telemetry. Never looking at it, just maintaining service,” he added.
Whitman said GSA is in “active conversations right now with other agencies” to see how it can help empower Federal agencies to make smart buying decisions when it comes to generative AI.
“Oftentimes, a marketplace GenAI tool might not be the right solution for a lot of these organizations, but how do they know?” Whitman said. “Can we give them a marketplace where they can experiment with these tools, test it out, use it in a real-world context, and then ultimately make the right decision for that agency?”
Notably, the GSAi tool uses commercial large language models all within GSA’s own infrastructure. The tool features a chatbot, an application programming interface (API), and “an administration console” that can evaluate different AI models for different purposes.
Whitman said GSA is currently testing numerous commercial AI models to see how they perform, including those from OpenAI, Anthropic, and xAI, among others.
The agency has also set up an “AI Safety Team” – with both technical and non-technical folks – to see if the model meets certain safety and performance standards. Based on the evaluation, the team will then make an adjudication for the models’ use.
“The idea is that we want to make sure that we have all the major vendors available in this platform for adjudication purposes, not necessarily for deployment,” Whitman told reporters on the sidelines of the conference.
“Our goal is really just to make sure that agencies have an opportunity to try out these different models and make an informed decision based on our evaluation work and our practical experience,” he added.
So far, he said, the conversations with other Federal agencies are aimed at gauging what they’re doing in generative AI, if they have built their own chatbot, and if they would find value in a shared service.
The end goal for the shared service, according to Whitman, is to help Federal agencies achieve cost avoidance, cost savings, and deduplication.
“There’s so much duplicative work that’s going on. Everyone’s building the exact same thing,” he told reporters. “It’s just a question of, do we all need to do the same thing, or could we find some efficiencies?”
Whitman added that some agencies, such as NASA or the Department of Energy, are building their own models for “specific scientific applications,” which he said are “original” and serve their own purpose.
However, for agencies that are just looking to achieve efficiencies for day-to-day tasks, “a lot of these models out of the box work pretty well,” he said.
“Ultimately, if the goal is, you know, commercial offerings, then let’s get there as efficiently as possible,” Whitman concluded.