Jan van Gemert

jvgemert.bsky.social

14 October 2026 20:36

I really enjoy finding gaps in our understanding of text-book deep learning. The RNN Homunculus: Long-range input vanishes RNN gradients. Gates (GRU, LSTM) help, but gates are themselves an RNN. Ie: gates are not sufficient. We need short-range data to learn gates first. arxiv.org/abs/2610.14279

hardmaru

hardmaru.bsky.social

12 October 2026 11:50

New Sakana AI paper: Recursive Self-Improvement through Multi-Agent Self-Supervision (MASS). Blog: pub.sakana.ai/mass/ Paper: arxiv.org/abs/2610.12176 How do we keep improving AI when tasks become too complex for humans to reliably evaluate? The model eventually hits a supervision bottleneck.

Dominik Schnaus

schnaus.bsky.social

10 October 2026 06:08

DINOv2 has never seen a caption, and Qwen3 has never seen an image. We still aligned their embedding spaces without a single image-caption pair. It even works when the images and the captions come from different datasets. Project page: dominik-schnaus.github.io/unpaired-rosetta ⬇️

Sung Kim

sungkim.bsky.social

09 October 2026 13:06

Underdog Saluki 27B in 7.89GB Based on Qwen 3.8 27B, Saluki is almost 7x smaller than its full-precision counterpart while it keeps 96% of its benchmark performance and beats Qwen at tool calling Blog and Model: underdog.ai/saluki

Sung Kim

sungkim.bsky.social

04 October 2026 08:52

Aleph Alpha, soon to be part of Cohere, released open-weight of Kolibri. 78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe. Model: huggingface.co/Aleph-Alpha/... Tech Report: aleph-alpha.com/downloads/te...