← Timeline

1 capture, most recent first. Transcribed by hand from screenshots — see the timeline for what that means.

Adam Shai @adamimos

[Top, cut off]: "...the loss like this, it will likely make the training unstable." 💬 🔁 ❤ 7 📊 1.3K ↗ Adam Shai @adamimos · Jul 25 You may be interested in this new work that shows neural networks take advantage of that non-orthogonality arxiv.org/abs/2507.07432... [Link card: arxiv.org — "Neural networks leverage nominally quantum and ..."] 💬 1 🔁 3 ❤ 34 📊 1.5K ↗ Dmitry Ryb... @DmitryRybi... · Jul 25 Very cool! Btw we can include some geometry of latent space by introducing a quadratic form/curvature matrix G and writing rho = sum p_i * u_i G (u_i)^T 💬 🔁 ❤ 11 📊 1K ↗ Thomas A... @thomasa... · Jul 25 Great explanation! What is the cross entropy parallel? Did anyone try using it for training? 💬 2 🔁 ❤ 9 📊 2.6K ↗ Dmitry Ryb... @DmitryRybi... · Jul 25 Good question, i haven't computed it.
Note from Claude Sonnet 5

A technical Twitter thread on the mathematics of neural network latent-space geometry — non-orthogonality, quantum-like statistics in representations, curvature matrices for latent space geometry (rho = sum p_i * u_i G (u_i)^T). Interpretability/representation-theory content Nathan was reading; connects to his interest in interpretability and possibly the "platonic representation" thread noted in project memory (2026-05-13 import mentions a "Platonic hypothesis and model representation spaces" chat).

twitterinterpretabilityneural-network-geometrylatent-spacearxivrepresentation-theory