for all (surveyed) large enough models, embeddings converge to projections legible/interpretable without access to the model itself, not even as a blackbox 🤯
https://arxiv.org/abs/2505.12540
Post #890
192
- Sep 30, 2026i started doing exactly this (models editing its own context) as a natural extension of sa…
- Sep 28, 2026cursed font factory https://bastardica.mitpit.com/
- Sep 25, 2026um, ahem, AlphaZero for text. Training language models without language. Loosely, natural…
- Sep 23, 2026umm ahem if you don't try to rein in new opus(5.5) you get one more qualitative step in ca…
- Sep 23, 2026nice https://fixupx.com/eyalsela/status/2102373749110587443?s=46
- Sep 21, 2026great writing, clear & thoughtful tldr: computers used to be math, and now they're about p…