We analyzed 1000+ neural networks - including ViTs, LLaMA-3 and Diffusion models. The result? They ALL collapse into the same shared, low-rank geometry. 🤯

Introducing the Universal Subspace Hypothesis.
Full paper: 👉 toshi2k2.github.io/unisub/

A breakdown of how this changes how we look at NNs 🧵👇

I'll be in Hawaii at ICCV for the next few days presenting my first co-authored work 'Scaling 3D Compositional Models for Robust Classification and Pose Estimation' iccv.thecvf.com/virtual/2025.... Will be happy to catch up with fellow researchers.

A quick question to Research ethics people - Is it common practice in ML and Computer Vision research to add people who supposedly "paid" for a small part of a research paper and did not do any research to be main authors on the paper?