Panelists discuss the feasibility, implications, and new requirements for incorporating generative AI into GEOINT
On Sunday at the GEOINT 2023 Symposium, a panel of experts from IC, academia, and private industry took on a major question: How can generative AI be used effectively in GEOINT contexts? Moderator Nadine Alameh, Ph.D., CEO, Open Geospatial Consortium, passed along questions from the audience to get insights on the immediate practicality of generative AI.
One use that seems to be a front-runner is in training scenarios. “With generative AI coming on board, we’re now able to use a lot of the features around generative AI to create synthetic data for training data to train our models,” said Devaki Raj, founder and CEO CrowdAI, a mission-first AI technology platform enabling persistent analysis from pixel to common operating picture. “For the last seven years or so, we had to curate datasets to focus on specific missions. But with generative AI, you have a broader set of data that you could potentially use for creating models.”
Uri Robinson, co-founder, Robinson Consulting Group, is also focused on mission capabilities.
“What is the end state of this new type of technology? What are we trying to get to? Because that’s usually one of the main issues that I have to face is understanding what that mission outcome is [and the] IT infrastructure related to that,” he said. “I’m sure we will get into various topics related to just the sheer amount of data that is required to run these types of capabilities.”
Nathan Kundtz, PhD, CEO, Rendered.ai, is looking forward to a time when generative AI becomes even more independent.
“A lot of code creation is becoming so quick that you no longer even need to write some of that simulation code. And that’s going to change the way in which people can interact with synthetic data,” he said. “And then ultimately, we’re headed in a direction where I think it’s not even going to be necessarily people that are requesting synthetic data, but systems that are requesting synthetic data. And those will be driven by models like language models.”
When asked to respond to a Chat GPT-generated list of how generative AI could be used in GEOINT, Kundtz, again, was forward-looking—acknowledging that AI could be used to answer questions, “but it doesn’t say anything about how we can use generative AI to ask questions, and I think that’s where some of the most important impacts will be.”
Audience members were, of course, concerned about how generative AI might affect the roles and jobs they currently fill. Sean O’Brien, vice-president, Intelligent Systems Laboratory, HRL Laboratories, agreed that job disruption was on the horizon but that the prospects for continued human involvement in analysis and decision-making are still good.
“Understanding the limitations of these things is critically important,” O’Brien said. “Chat GPT and all the generative AI is really good at creating the first draft of whatever you’re doing. It’s generally not very trustworthy in creating the last draft, and so that is going to amplify the importance of the kinds of hard work [that people do].”
Matt Klaric, Predictive Analytics Pod Lead at the National Geospatial-Intelligence Agency, is looking forward to implementing new AI modalities but urges everyone to keep their eyes open.
“One [issue] that’s very important for us is currency,” he said. “You may have a foundational model that’s been trained, and you ask Chat GPT a question about who won the Cardinals baseball game last night and it is not going to be able to tell you. So, as you look at how you can use these models for various applications, you need to understand their limitations, understand what the foundational models are good for, what they’re not good for, where we need to update them with our own data.”