Open-Source logos-theodolite Visualizes Layer-by-Layer LLM Internals
New_Alps_5655 · reddit · 2026-10-20
A Reddit user released logos-theodolite, an open-source instrument for exploring LLM internals, combining recent interpretability research on jspace, j++, LARQL/VINDEX, and Cammarata's neuron poles. It visualizes per-layer token trajectories, attention vs. feed-forward memory ratios, candidate token distributions, and the loudest neurons. The j-lenses were originally fitted for gemma4-31b; a re-fit wasn't worth the GPU time so they use the instruct version directly. The author also claims Opus 5.5 appeared to be routed to Fable 5.5 during testing.
More from Research
- EvoWeb benchmark: 320 sequential requests expose coding agents' cross-request integration failures — TD-Labs · 2026-10-20
- Amazon's Lean4AgenticRL grounds multi-level process rewards in Lean4 for agentic RL — amazon · 2026-10-20
- Priming Models: one-step generation hits 1.19 FID with 12x less compute, no diffusion teacher — LucaAmb · 2026-10-20
- LILIE: language-informed lossless image encoder improves gFD by 14 points on GPIC — jcniebles · 2026-10-20
- Stanford's TRACER: Training-Free Multimodal AI Tracks Hospital Patient Behavior — StanfordAILab · 2026-10-20
- Researcher calls out vision RL papers as p-hacked rather than scalable results — gowthami_s · 2026-10-20