JUST IN: Large Language Models Lack Utility for Military Missions, DARPA Official Says

iStock photo-illustration

ALEXANDRIA, Virginia — There is a lot of buzz around artificial intelligence capabilities such as large language models, but without pairing them with modeling and simulation tools, the technology will not live up to the hype, an official from the Defense Advanced Research Projects Agency said Aug. 18.

Eric Davis, program manager in DARPA’s Strategic Technology Office, said his entire $250 million portfolio consists of AI-related projects. However, “there's not a single large language model in the work that I'm doing right now,” he said during a keynote speech at the National Training and Simulation Association’s MODSIM World conference.

Given the billions of dollars invested in AI, there is an assumption that the technology will eventually achieve human-level reasoning and “solve everything, so why do anything anymore, right? It’s the end of the human endeavor,” Davis said. 

“We're told that generative AI is going to design our supply chains. It's going to coordinate our disaster response, run our power grids, do our homework for us,” he continued. The technology has massive potential, but it cannot solve every problem. 

Large language models can ingest vast amounts of data and produce well-articulated outputs that maximize one thing: human approval. “That's why what they say is so convincing,” Davis said. “What they lack are internal models of causal reality.”

These models don’t understand concepts like gravity, fluid dynamics or fuel consumption. They don’t understand how diseases are transmitted or the laws of thermodynamics. “If we ask an LLM to output a plan, we get a very convincing language-shaped, plan-shaped output. But whether it's for wildfire evacuation, a hospital surge or military planning, what it's not based on is reality,” he said. “And if we trust that plan to be truthful, we will be disappointed.”

For some use cases like in a boardroom or a customer service center, the statistical, language-based plausibilities based on language these systems generate might be good enough, he said. “Amazon doesn't care if they're 85 percent correct; that means that they're still making a lot of money. But 85 percent correct in an emergency operations center, in an air traffic control center, on a hospital floor or on the bridge of a destroyer is not enough. There are human lives that are at stake.”

To make these advanced cognitive AI systems more effective, they need to be paired with modeling and simulation capabilities to serve as internal world models, Davis said. “If we have a single world model, we have an arena to test hypotheses, to reason causally and to quantify the actual risk that I would face when executing a plan.”

While modeling and simulation has traditionally been an offline tool used for human training and retrospective study, “what we see in the future at DARPA with a lot of our programs is modeling and simulation that moves from just training, just tabletop exercises into core architectures of machine cognition itself to provide that arena for hypothesis testing before action,” he said.

Rather than marking the end of human endeavor, AI and simulation married together provide a “way for us to explore our values iteratively as part of a human-machine team,” he said. 

“We're not trying to replace the human; we're trying to sharpen their thinking. We're trying to give them more high-fidelity encounters with those problem sets” so they are prepared for challenging situations in dynamic environments, he said.

To make this marriage a reality, the modeling and simulation community must be willing to take its capabilities out of isolated silos and build modular, high-speed, physics-grounded world models to power the next generation of autonomous and decision support systems, he said.

Language models are incredible tools, but to truly achieve artificial intelligence, “we need to go beyond just giving machines voice and intuition, and give them the ability to parse human meaning and to understand the way that the world evolves causally,” Davis said. “Modeling and simulation is that ground truth, and the advances in this field will continue to have more of an impact on AI” than the other way around.

 

Topics: Training and Simulation, Robotics and Autonomous Systems