In this 2024 position paper, Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola of MIT argue that the internal representations learned by different neural networks are converging. As models grow larger and are trained on more data, they increasingly agree on how data points relate to one another, even across very different architectures and even across modalities such as vision and language.
The authors name this convergence point the “platonic representation,” borrowing Plato’s idea of ideal forms: the suggestion is that distinct models are all approaching a single underlying statistical model of the reality that generated their training data. They support the claim by measuring representational alignment - how similarly two models judge the distances between the same items - and show that this alignment tends to rise with capability. They also discuss candidate drivers of the trend, such as shared task pressures and the structure of the world itself, along with limits and counterexamples.
The hypothesis connects to interpretability and safety because it predicts that concepts may be encoded in comparable ways across models, which would make findings transferable: a feature understood in one model might map onto another. It also informs debates about multimodal learning and whether scaling pushes systems toward a common, perhaps human-aligned, view of the world.
For a general reader, the idea is provocative: it implies that competing AI systems may be growing more alike on the inside even as they differ on the surface, with implications for how durable any one company’s edge really is.