AI can solve Navier–Stokes. Can it design a plane?
Data-efficient, information-seeking, scalable learning algorithms. Tasked with inventing new technology.
Live in your browser: air over a flying wing as its angle is tuned, a burning flame, and electron density across a high-entropy alloy.
Frolic's models do the most with every unit of data and compute, at every budget.
And they improve the fastest as you give them more of both.
The real test of intelligence is a breakthrough that holds up in reality: the next engine, the next chip, the next lithography machine.
Invention is the ultimate meta problem. Better technology builds better AI, which builds better technology. Push far enough and every level collapses into one.
The machines that matter are deep stacks of interacting parts. Designing one means holding the whole system in mind at once.
A multiphysics model on a lattice: one AI system that reasons across every field and effect we know, at every scale. A graphics engine that's faithful to physics.
Physics shows up as equations, as optimal curves and as full fields. It runs from quantum mechanics all the way up to whole machines.
∂ρ/∂t + ∇·(ρu) = 0ρ Du/Dt = −∇p + μ∇²uρ De/Dt = −p∇·u + ∇·(k∇T)p = ρRTEquations
Optimal curves
Full-blown simulation
Learn every field of a system at once, then run it backwards: from the behaviour you want to the design that produces it.
Simulations can be gamed. Physics can't. We train with reinforcement learning on rewards the real world verifies, so the model can't cheat.
We're building a high-throughput, reality-in-the-loop lab. 3D printing turns a design into a part in hours. CT and laser scanning measure it down to the micron.
We believe in deep learning, and in understanding why it works.