Physics paper claims to explain why training AI gets exponentially harder as models grow larger

Researchers propose that training deep learning systems hits a computational wall as models scale up, similar to how relativistic physics predicts behavior at extreme speeds. They argue existing mathematical models underestimate this cost and propose a new measurement framework to predict when training will become prohibitively expensive.

If true, this describes a hard physical limit on how large AI models can grow without consuming absurd amounts of computing power. Right now, nobody has a reliable way to predict when a model will become too expensive to train — companies just build bigger models until the bill gets painful. A working prediction method would let researchers know in advance where those walls are, potentially redirecting effort toward more efficient approaches instead of brute-force scaling.