October 26, 2026
See an LLM at work from the inside: one Qwen3.8-27B prompt traced through 64 layers in 3D
Plates instead of a diagram: 64 layers of Qwen3.8-27B with live probabilities captured from a run on a home computer.
Loading the page.
who it's about
Layer →
about the same entities
Release
DeepSeek and Tencent open their own search for AI: site authority decides what gets into the answerOctober 23
Insight
--moe-cache-mib in llama.cpp: GPU expert cache speeds up MoE model 2.2x on RTX 5090October 21
Insight
Vertical AI agents differ from general-purpose ones in three layers: context, harness, and modelOctober 20
Release
/auto-mode-setup in Claude Code: setup wizard now works on Bedrock tooOctober 20
All news
One email a week - only what changes how you work. Subscribe