orchestra-research avatar

Nnsight Remote Interpretability

This gives Claude the ability to help with nnsight, a PyTorch interpretability library that lets you access and modify neural network internals with the same code running locally on GPT-2 or remotely on 70B+ models via NDIF. The write-once-run-anywhere pattern is genuinely useful if you're prototyping interventions on small models before scaling up. It works with any PyTorch architecture, not just transformers, and handles multi-token generation interventions cleanly. The remote execution through NDIF is the main differentiator here, letting you run activation patching experiments on massive models without GPU access. If you only work with small models locally, TransformerLens is probably simpler, but for researchers who need to scale interpretability work to frontier models, this is purpose-built for that.

skills.sh installs
1K
GitHub stars
14.4K
last updated
Aug 21, 2026

Install

$ npx -y skills add orchestra-research/ai-research-skills --skill nnsight-remote-interpretability --agent claude-code

Installs into .claude/skills of the current project.

Files

nnsight: Transparent Access to Neural Network Internals

nnsight (/ɛn.saɪt/) enables researchers to interpret and manipulate the internals of any PyTorch model, with the unique capability of running the same code locally on small models or remotely on massive models (70B+) via NDIF.

GitHub: ndif-team/nnsight (730+ stars) Paper: NNsight and NDIF: Democratizing Access to Foundation Model Internals (ICLR 2025)

Key Value Proposition

Write once, run anywhere: The same interpretability code works on GPT-2 locally or Llama-3.1-405B remotely. Just toggle remote=True.

# Local execution (small model)
with model.trace("Hello world"):
    hidden = model.transformer.h[5].output[0].save()

# Remote execution (massive model) - same code!
with model.trace("Hello world", remote=True):
    hidden = model.model.layers[40].output[0].save()

When to Use nnsight

Sponsors

Details

Repository
github.com/orchestra-research/ai-research-skills
Categories
Code Review & Quality, AI & Agent Building, Data Science & ML
First seen
Jun 3, 2026

More from orchestra-research/ai-research-skills

All 95
1orchestra-research avatarnemo-evaluator-sdkorchestra-research/ai-research-skillsEvaluates LLMs across 100+ benchmarks from 18+ harnesses (MMLU, HumanEval, GSM8K, safety, VLM) with multi-backend execution. Use when needing scalable evaluatio1Kskills.sh installs
2inference-sh avatarpress-release-writinginference-sh/skillsPress release writing in AP style with inverted pyramid structure. Covers formatting, datelines, quotes, boilerplates, and fact-checking. Use for: product launc1Kskills.sh installs
3inference-sh avatartalking-head-productioninference-sh/skillsTalking head video production with AI avatars, lipsync, and voiceover. Recommended: P-Video-Avatar (fastest, cheapest, built-in TTS). Also covers OmniHuman, Pix1Kskills.sh installs
4cap-go avatartailwind-capacitorcap-go/capgo-skillsGuide to using Tailwind CSS in Capacitor mobile apps. Covers mobile-first design, touch targets, safe areas, dark mode, and performance optimization. Use this s1Kskills.sh installs
5forcedotcom avatarapplying-sldsforcedotcom/sf-skillsApply SLDS-compliant UI using the correct blueprints, styling hooks, utility classes, and icons. Use when building any UI that needs SLDS, choosing between Ligh1Kskills.sh installs
6wechat-miniprogram avatarskyline-wxsswechat-miniprogram/skyline-skillsSkyline WXSS 样式支持技能。提供 Skyline 支持的 CSS 属性、值和限制信息。触发关键词:WXSS、CSS 支持、样式兼容性、属性支持、不支持的属性、display grid、position sticky、overflow、filter、渐变、gradient、linear-gradient、ra999skills.sh installs