Forbes contributor Thomas Coughlin spoke with PiLogic’s founders about an approach to problems the company says can be solved faster, and with less energy, than large language models. The approach makes heavy use of exact probabilistic inference.
As Coughlin describes it, the method does not require huge data sets or specialized hardware such as GPUs, does not hallucinate, and gives results that are precise and accurate. He lists use cases in autonomous systems, cybersecurity, and aerospace, including radar identification and tracking and diagnosing electrical system failures on aircraft and spacecraft.
The column walks through how the engine compiles a Bayesian network into an arithmetic circuit, pushing most of the inference work into a one-time pre-deployment phase. That makes it possible to know precisely how much time and space a query will need, which suits real-time requirements, and lets the circuit run without specialized hardware.
Coughlin also covers treewidth, the measure of how hard a network is for a conventional inference engine. Conventional engines run in time and space exponential in this width. PiLogic told him that by exploiting local structure in a problem, its engine can solve networks with treewidth into the hundreds where such structure exists.
PiLogic builds Exact AI — exact probabilistic reasoning for mission-critical aerospace and defense. Its models, Manifest (satellite diagnostics) and Resolve (radar tracking), compute correct, explainable answers in milliseconds, onboard, on existing hardware, with no training data and no hallucination. PiLogic works with the Air Force Research Laboratory under a two-year Cooperative Research and Development Agreement and is backed by Scout Ventures, Seraphim Space, Sovereign's Capital, Flex Capital, and FN Fund. For more information, visit pilogic.ai.