Oppenheimer spent years delving into the mysteries of the universe with little more than a blackboard and his brain in his quest to split the atom. Now imagine that he'd had a hyper-intelligent companion that could understand every physics problem known to man and calculate complex tests in seconds — the film would have been a lot shorter if nothing else.
One startup is building an AI to do just that. Cambridge-based BeyondMath — founded in 2022 — is training transformer AI models (the same core technology that powers ChatGPT) on the “fundamental laws of the universe”. So rather than feeding it a vast corpus of written text scraped from the internet, the company is training its systems on physics equations. This is “the ground truth”, says cofounder Darren Garvey.
The first use case: a fine-tuned system for a Formula One racing team that’s trained on aerodynamic physics equations. But Garvey tells Sifted that a foundation model for all known physics — a kind of Einstein AI — is a realistic goal.
The next Einstein?
Could such an AI model — with access to every known physics equation — end up improving our understanding of the universe, and even come up with a unifying theory to succeed Einstein’s theory of relativity?
“Definitely,” says Garvey’s cofounder Alan Patterson. “There’s a question of how much you will discover from incorporating real data and discovering new laws or more efficient interpretations of the laws [of physics]... Our current equations are not even properly solved. So you can imagine understanding, learning better systems for how we understand how the world works.”
Beyond explaining the cosmic mysteries around dark matter or antigravity, Patterson says that this kind of AI system could also lead to designing technology that’s better than what humans can currently conceive.
"You can imagine learning better systems for how we understand how the world works which results in AI designing cars, planes and spaceships in a way that's far more efficient and effective — in a way that we have not been able to do with decades of engineering and science experience,” he says.
First cars
But baby steps first. The startup is currently focused on commercially validating its technology with its Formula One partner.
It’s applying its physics-trained AI system to the costly process of car design, which has previously involved a lengthy process of solving equations on a supercomputer to help optimise aerodynamics and performance.
Patterson says that BeyondMath’s model is a cheaper and faster way to assess the impact of different design choices.
“A Formula One team can very rapidly evaluate configurations and designs of the car in minutes — it would normally take them one day per iteration,” he says. “Being able to do it so rapidly is a massive advantage.”
Garvey adds that the technology could make this kind of precision design process available to companies that might not be able to afford the supercomputing power to run these long equations.
What’s next?
BeyondMath also hopes to begin working with a partner in aerospace before the end of the year. The company raised a pre-seed round of £1.5m in 2022 and says it has runway until Q4 of 2024 (back off, GenAI-obsessed VCs…).
Patterson does add that, while its model doesn't hallucinate in the same way that large language models do, it does still provide results that are around 0.8% less accurate than the traditional supercomputer method.
He says that this is an acceptable level for a Formula One team making rapid design choices, but that a company like Boeing launching an aircraft might need higher accuracy levels — something he expects could take around five years to achieve with AI.
It’s heartening to hear that our next generation of cars and planes won’t be designed by a system that’s as error-prone as ChatGPT, and exciting to think of a world where an EinsteinGPT could build products that are better for the planet.