Neuronpedia: open-source interpretability platform hosting Anthropic's Jacobian Lens and Gemma Scope 2
techNmak · x · 2026-09-04
Entries 13-14 of @techNmak's visual learning resource thread:
- Diffusion Explainer: interactively walks through the major components behind Stable Diffusion, so "denoise from noise" stops being a black box.
- Neuronpedia: an open-source interpretability platform for inspecting model internals — explore features, sparse-autoencoder latents, activations, and attribution-style graphs, backed by 5TB+ of activations and explanations. Recent highlights include Anthropic's Jacobian Lens (revealing a global workspace in LLMs), Natural Language Autoencoders (translating a model's internal thoughts into text), Assistant Axis, Circuit Tracer, and Google DeepMind's Gemma Scope 2 (SAEs and transcoders for Gemma 3), plus a new interp-engine for fast, standardized interpretability.
Related event: Tools for Model Interpretability: Neuronpedia and Diffusion Explainer(2 posts)→
More from Apps
- Early user: Grok Bot's cloud-computer agents actually finish tasks, unlike other chatbot wrappers — r0ck3t23 · 2026-09-05
- ChatGPT photo fixes now look pro while keeping faces identical — 5 prompts shared — HeyAmit_ · 2026-09-05
- Agency migrates its 6th client site from WordPress to Lovable — MyCreativeOwls · 2026-09-05
- Dozens of high schoolers casually pile into Waymos after a football game — brianwilt · 2026-09-05
- Notion positions itself as the "Switzerland of AI" as models default to it for memory — ivanhzhao · 2026-09-05
- Non-Mac-dev builds Coverage with Astra: Spotify album covers get live effects — craigsdennis · 2026-09-05