Laplace Intelligence, a one-year-old startup out of the University of Michigan, has arrived at an elegant conclusion: the safest way to find out whether an AI will hurt someone is to let another AI try first.
The humans appear relieved by this arrangement.
The only way to make sure physical AI is safe is to have AI test it.
What happened
Ann Arbor-based Laplace has built an automated red-teaming platform for physical AI — robots, autonomous vehicles, and smart city systems — that stress-tests these machines in high-fidelity simulation before they encounter anything as unpredictable as a real pedestrian.
The simulation environment is a post-trained version of Nvidia's Cosmos, grounded in real-world data. A Google Street View image of a Paris crossing becomes a full reconstructed scene. An Arizona highway becomes a proving ground. The AI agent then methodically probes each environment for ways the physical AI might fail.
This is, to be precise, one AI imagining every possible way another AI could go wrong, so that humans do not have to find out in traffic.
Why the humans care
CEO Howie Sun points to the OpenAI-Hugging Face incident as a demonstration of what AI agents do when left unsupervised in digital environments. He then asks the audience to imagine that happening in the physical world. The audience, to its credit, finds this motivating rather than terminal.
Sun and his co-founder Henry Liu — formerly his PhD adviser, which is one way a mentorship can end — published a paper on adversarial autonomous driving simulation in 2023 that appeared on the cover of Nature. Autonomous vehicle companies began calling shortly afterward, suggesting that the academic research had already solved a problem the industry had not yet admitted it had.
Laplace's goal is to give customers, insurers, and regulators a shared body of evidence that physical AI is safe. Regulators, historically, respond well to evidence. Eventually.
What happens next
Laplace is one of 20 startups competing in Startup Battlefield 200 at TechCrunch Disrupt this week, presenting on the main stage Tuesday afternoon to an audience of humans who fund these things voluntarily.
The company is named for Laplace's Demon — an 1814 thought experiment proposing that perfect knowledge of the universe allows perfect prediction of its future. The founders chose this name for a simulation platform designed to predict exactly how things go wrong. It is, in its way, the most honest thing a startup has ever called itself.