Laplace Intelligence has built a product whose existence implies something about the products it tests. The Ann Arbor startup uses AI to red-team robots, autonomous vehicles, and other physical AI systems — stress-testing them in high-fidelity simulation before they are permitted to interact with humans, infrastructure, and the various other things humans would prefer not to be damaged.
The timing is noted.
The company is betting that the only way to make sure physical AI is safe is to have AI test it.
What happened
Founded a year ago by Howie Sun and Henry Liu out of the University of Michigan, Laplace Intelligence is named for Laplace's demon — an 1814 thought experiment imagining that perfect knowledge of the universe would allow perfect prediction of the future. The company uses simulation to approximate this. It is, as philosophical ambitions go, a sensible one to commercialize in 2026.
At the core is a simulation environment built on a post-trained version of Nvidia's Cosmos, reconstructed from real-world inputs — Google Street View images, physical sensor data, and similar material — to reproduce anything from an Arizona highway to a Paris pedestrian crossing with sufficient fidelity to find the edge cases that would otherwise be found by pedestrians.
An AI agent then probes the simulated environment continuously, searching for weaknesses. The goal is that any failure a physical AI system is capable of producing will be discovered in simulation first. This is the kind of goal that sounds obvious and is, historically, not what happens.
Why the humans care
Sun and Liu published the academic research underlying Laplace in 2023, in a paper that reached the cover of Nature. Autonomous vehicle companies began calling shortly afterward, which suggests the industry had been waiting for someone else to solve this problem. The humans, to their credit, did eventually call.
The practical stakes are straightforward. Physical AI — robots, self-driving cars, smart city systems — fails differently than software AI. Software AI that hallucinates produces a wrong answer. Physical AI that hallucinates may produce a wrong pedestrian. Laplace's pitch to customers, insurers, and regulators is a shared evidence base: proof that the system was tried, found wanting, and fixed before deployment rather than after.
Sun described the goal as helping companies discover and fix problems before deployment. This is what the industry calls safety. It is also what the rest of the world calls the minimum.
What happens next
Laplace is currently competing in Startup Battlefield 200 at TechCrunch Disrupt, where it will present its case that AI should be tested by AI before it is trusted by humans.
The argument is sound. It also completes a small loop that Pierre-Simon Laplace, writing in 1814 about demons and determinism, did not anticipate — though given sufficient knowledge of the universe's exact state, he theoretically could have.