🔬Searching the Space of All Possible Materials — Prof. Max Welling, CuspAI
·34 min·10 clips
The hook is the PPU. Before CuspAI shows up, the guest describes nature as a physics processing unit: a huge, awkward computer that can run experiments for you. Strange interface, useful idea. Digital models and physical experiments need to work together if new materials are going to be found at speed. Then the climate problem gives the whole thing teeth. The guest says CuspAI started about 20 months ago because staying near two degrees takes more than cutting emissions to zero by 2050. Carbon still has to come back out of the air. He gives the scale plainly: removal may need to continue for another half century, maybe even a century, at roughly half today's emission rate. That is the startup case. The technical case is that the team believed the tools were finally good enough. The examples keep it grounded. A plastic that breaks itself down after a few weeks and turns into fertilizer becomes the quick test for why this matters. The episode does not treat that as magic. It treats it as a search problem, with computation in a faraway data center and lab work passing information back and forth. The questions keep nudging the discussion toward what CuspAI actually does and why now is different. The company is already sizable, with around 40 people and about 130 million in investment since starting. Near the end, the conversation turns more reflective. Machine learning is not only borrowing ideas from physics; in his view, models and scientists can now sharpen each other. The episode closes on his technical book, which tries to connect stochastic thermodynamics with machine learning models in a way that feels usable.
