The atom smasher

If you have read The Agony and the Ecstasy, Irving Stone’s fictionalization of the life of Michelangelo, you will remember the tableau: Michelangelo in the Carrara quarries, choosing his marble in person rather than accepting what was shipped to the markets of Florence.

In the autumn of 1497, at twenty-two, he is at the Polvaccio quarry, looking for the marble that will become the Pietà. He’s looking for a highly pure section. But the purity is not a surface question. Marble is hard, but photons penetrate it: five, maybe ten millimeters. Light enters the stone, scatters back and forth among the calcite crystals, and eventually makes its way back to our eyes. When we look at marble we see not its surface, but beneath.

Today laser metrology is used to characterize a material like marble. Controlled coherence light sinks into the stone and returns precise data. The deeper you look, the more of our world turns out to be like that. What seems fully described by sight, or by a measurement in two dimensions, ends up having enormous complexity. People in the semiconductor industry deal with details of this kind every day, as they build ever more complex structures in the metal-oxide-semiconductor technology.

I thought of the marble problem when I met Ian Glow near the Hanover office in Portola Valley shortly before he founded Zeromatter. Today as Zeromatter comes out of stealth, it’s an opportunity to admire the company’s insistence that we understand the truth of the physical world, and that we endow machines with that truth so that they can serve us.

Learning that truth has usually meant breaking things. A particle accelerator is called an atom smasher. Cars are crashed to learn how they crash. A robot drops a thousand parts before it learns to hold one. Giant cavitation machines teach sophisticated electronics companies how their products will survive a transoceanic shipment. Every lesson is paid for in atoms, and atoms are slow, costly, and sometimes extremely deadly.

The lessons are about to multiply. Autonomous vehicles, delivery drones, surgical robots, and agricultural machines are being built now, and each must be trained on millions of scenarios it could never safely meet in the real world.

Zeromatter moves the collision into the computer. It is a single platform to build, test, and train machines that act in the world.

“Solve the hard parts in bits, before you move a single atom.”

The Zeromatter team

Hanover led Zeromatter’s pre-seed, co-led its seed, and invested in every subsequent round. We want to share some of our thinking here.

Fidelity

Physical AI learns from experience, and the experience that matters most, the rare and the dangerous and the expensive, is the experience no one can afford to collect. No team can risk a crash to validate every software change, or wait for rain to ship a release. Simulation is the obvious answer. It has mostly been a disappointing one.

The trouble is fidelity. “To train a machine,” Ian says, “you have to see what it sees.” A simulator built to please the eye teaches a machine the wrong lessons, because a sensor is not an eye. A camera has a rolling shutter and a distorted lens. Lidar and radar see in wavelengths we cannot. A model trained on a handsome approximation learns the approximation.

Fidelity decides how fast a company can move. The more accurately a simulator reproduces sensor physics, environmental conditions, and the dynamics of many agents at once, the sooner that company can develop, validate, and deploy.

Zeromatter starts from physics. Its engine is built from scratch and deterministic, and it models cameras, lidar, radar, and infrared from the physics up. One scene can be rendered as a camera sees it by day or night, as lidar returns, or in infrared. Around the engine sits the rest of the platform: environments generated automatically at the scale of entire cities, and a co-simulation framework in which many agents, each running a customer’s own models, share one world. It serves autonomy, aerospace, automotive, agriculture, and defense.

Deep metrology

What does a true simulation platform for AI need to be? It needs to be true to life. It needs to be highly performant, meaning fast and available, a requirement that reaches all the way down to the instruction set and the rendering engine itself. And, most important, it needs deep metrology.

By deep metrology we mean measurement that goes below the surface, as it must with marble, and we mean it in two places. A simulator has to render a world faithfully, with all the complexity that entails, including the way things move within it. It also has to render, just as faithfully, the way sensors perceive that world. To build a data machine large enough for AIs to learn in, you must know the most minute imperfections of the sensors those AIs will be connected to when they are deployed.

I learned this in semiconductors. I got my start at Cypress Semiconductor, the first commercial believer in EUV lithography. Think of the acceptable defect rate on a true high-volume electronics line: something like 400 defective parts per million, or DPPM. That is the standard to which a modern motor, a CMOS camera sensor, or an array of power switches is held. All of them have imperfections, and to understand those imperfections you had better have a million of the parts on hand. You’d better run millivolts through each of the millions, millions of times.

Much has been said about general world models, and members of this team helped solve full self-driving at Tesla. But if AI is to make the rest of the physical world better, from medicine to construction to defense, the precision of a casual human driver is not enough. We want industrial-grade precision, which means modeling even the most minute imperfections in silicon. That is the fidelity Zeromatter seeks to achieve and why Hanover partnered with them to build the company.

Precision of that kind is statistical. It comes from what we might call a tera-sim, a massive body of simulated data, all of it rendered with fidelity to the world and to the machine sensors that perceive it. Only a simulation platform of this kind can open up physical AI, and with it breakthroughs in defense, food, chemistry, and medicine.

The team checks its models against reality. In its labs, a laser crosses an optical bench, a camera is calibrated at an integrating sphere, and an oak leaf rests in a sample holder so that its spectrum can be measured rather than guessed.

That is what we mean by angstrom-level precision. An angstrom, a ten-billionth of a meter, is roughly the width of an atom. It is the unit in which spectroscopists first mapped the sun’s light, and in which chipmakers name their advanced processes. The discipline is to get the smallest physical facts right, so that the large ones follow.

The frontier labs have built out scientific knowledge and research with remarkable speed, and they are solving the what. What do we do next? What drug do we build? But there is still the how. How do arms and hands move in the world to carry out that chemical reaction? How do the ailerons move to bring the jet safely to ground?

The how requires true tera-sim, massive and high in fidelity.

Why Hanover

We like these problems. They are understudied because they had been thought impossible. At Cypress Semiconductor, I worked on mixed-signal microcontrollers, chips that combine the messy analog and digital worlds inside one coherent system. At Enphase today we’re building solid-state transformers on gallium-nitride switches that can read the grid 4,000 times faster than the grid frequency. Ultimately these are all data problems. But the amount of data they require is orders of magnitude greater than what is available. To them we add those areas where the market is most excited about autonomy: flight, driving, delivery, safety, defense.

Similar problems, different scales. An old video gaming engine isn’t enough. The Zeromatter team knew that from the beginning and has built a computing machine that is likely to produce more simulated physical data next year than all of its competitors combined. We are proud to have backed Zeromatter in every round, and to partner with Ian and his team as they build an atom smasher that spares the atoms.