Laplace Intelligence Pioneers AI-Driven Red-Teaming for Physical AI Safety

One key lesson from the OpenAI-Hugging Face incident is the unpredictability of AI agents. Howie Sun, CEO of Laplace Intelligence, emphasized the risks associated with physical AI systems. He noted, 'AI agents are moving from computers into the real world, and we don’t really want physical AI to fail in the real world.'

Laplace focuses on automated red-teaming for robotics and physical AI. The company believes that AI must test other AI systems to ensure safety. Sun stated, 'Our goal is to help companies discover and fix problems before deployment — and, over time, build a shared foundation of evidence that customers, insurers, and regulators can trust.'

Founded a year ago by Sun and chief scientist Henry Liu, the Ann Arbor-based startup emerged from the University of Michigan. Sun previously spent eight years researching autonomous driving AI. Their 2023 Nature paper on challenging autonomous driving AI in simulation attracted interest from vehicle companies, leading to the product's creation. 'I realized that I wanted to build this and take it from purely academic research into a product that’s useful in the real world,' Sun said.

Named after 'Laplace’s demon,' the startup uses simulation to predict the real world's future. 'We are essentially using simulation to predict the real world’s future,' Sun explained.

The core technology combines a high-fidelity simulation of the real world with an AI agent that probes for weaknesses. This simulator, a post-trained version of Nvidia's Cosmos, reconstructs environments like Arizona highways or Paris crossings using real-world data. 'Simulations are generated by blending a mesh of different real-world information,' Sun noted. Map data and Google Street View images ensure scale.

Developers can change settings via language prompts, such as adding traffic. Laplace integrates real-time data from TomTom. The simulation also models actor behavior. In a demo, when a test agent vehicle swerved into oncoming traffic, the other car moved out of the way. 'Everyone is reacting to everything … This is more like a world model with interactive participants,' Sun remarked.

A key feature is the AI agent that automatically designs corner cases, removing humans from the loop. 'They can design from the news, from police reports on collisions — all kinds of information to actively challenge the autonomous vehicles and other physical AI agents,' Sun said.

While autonomous vehicle developers often use human-designed scenarios, Laplace offers AI-generated challenges. Sun noted the lack of data in general robotics makes this technology crucial. Despite academic roots, Laplace is profitable, generating revenue from autonomous trucking companies like Bot Auto and a major Japanese automaker.

Laplace is collaborating with Nvidia and T-Mobile on a San Jose smart city pilot. They are also exploring humanoid robotics. Most contracts are project-based, but the company is moving toward recurring revenue. Competitor Applied Intuition poses a financial threat, but Sun believes his 'AI testing AI' methodology is a defensible moat. 'That will be the biggest moat,' he said.