Bezos is back!
Jeff Bezos is back in operator mode for the first time since Amazon, with a new startup, a hell of a large seed round, and a thesis: LLMs canât bend metal.
Bezosâs ~new thing~ is called Project Prometheus, which has raised $6.2B in funding, much from his own pocket. His co-CEO is Vik Bajaj, the acclaimed physicist-chemist who ran Google Xâs life sciences division and co-founded Verily. Their first heist: nearly 100 AI researchers poached from OpenAI, DeepMind, and Meta. Their mandate: AI for the physical economy, with early work targeting engineering and manufacturing across compute, cars, and space systems.
Todayâs frontier models are linguistic savants whoâve never touched grass. They excel at writing and reasoning about documented problems, but struggle with physical causation: how materials behave under stress, or how a single parameter change cascades through a manufacturing system.
The Rise of Physical AI
A tidal wave of well-capitalized startups has emerged to attack this gap. Periodic Labs, for example, raised a $300M (!) Series A to build âAI scientistsâ aimed at accelerating discoveries in physics, chemistry and materials science. Robotics companies, meanwhile, are racing to combine world models and simulation to train systems that can operate effectively in our physical world.
Godfathers, godmothers, & a new frontier
Yann LeCun, one of the âgodfathersâ of deep learning, is reportedly leaving his role as Metaâs chief AI scientist and will launch his own world model startup. Fei-Fei Li, the âgodmotherâ of modern AI, launched World Labs last year with $230M in funding. Last week, she published a manifesto on why spatial intelligence is AIâs next frontier:
LLMs have begun to transform how we access and work with abstract knowledge. Yet they remain wordsmiths in the dark; eloquent but inexperienced, knowledgeable but ungrounded.
â Fei-Fei Li, founder of World Labs, who built ImageNet and pioneered modern computer vision.
A vibe shift? LLMs work brilliantly at their job, but they hit a wall when reasoning about physics. Some of the smartest minds in AI increasingly see this not as a scaling problem, but a local maximum. You canât text-predict your way into understanding turbulence, nor interpolate between descriptions to discover new materials. The only way through, then, is with systems that learn how the world actually works, not just how itâs described in text or abstracted in softwareâŚ
The bottom line (literally): The $16T digital economy (all of ecommerce, SaaS, cloud) represents ~15% of global GDP. Even though the physical economy is growing slower, it is much larger. Manufacturing alone is $16T globally. Add construction ($13T), agriculture ($4T), transportation, aerospace, energy, and youâre looking at $40-$50T that depends on understanding and manipulating atoms.
For those fleeing the transformer monastery â all of you scientists, financiers, and researchers willing to trade the comfort of pure software for the resistance of real matter â your decade has arrived. Progress demands hard pursuits in the physical world, not just our digital realms. Welcome to the Renaissance, where the hard problems are the only ones worth solving.