More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    SAELens: Read What a Language Model Actually Learned

    1

    Work with sparse autoencoders for mechanistic interpretability — read pre-trained SAE features, train custom SAEs, and use features for analysis and steering.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    SAELens: Read What a Language Model Actually Learned

    SAELens: Read What a Language Model Actually Learned

    Example session with this skill installed

    Find what a small language model knows about dates: which features fire for years, months, and weekdays, and whether the model confuses them.

    • Read your context and instructions
    • Compiled the saelens

    Path one first: a public pre-trained SAE for the model's layer was loaded, and its feature labels were searched for date-related concepts — years, months, weekdays, and calendar words. Twenty-one features looked date-related by label. Each was checked by firing it on date sentences and on control sentences: fourteen fired selectively for dates, four fired for numbers generally, and three fired for anything with a capital letter — the labels were a starting point, not a verdict. The fourteen real date features split cleanly: six for years, five for months, three for weekdays, with almost no overlap between the groups — the model keeps the three apart. One surprise: a feature labeled for months fired hardest on the word 'May', which is also a modal verb — the model had not confused months and modals, but the feature was doing double duty, and the steering test showed it: boosting it changed both date completions and permission sentences. The finding was written with its limits: fourteen verified features, three label errors, one double-duty feature, and the steering caveat attached. The answer to the question is that the model separates years, months, and weekdays cleanly — and that one feature does two jobs, which is exactly the kind of thing the lens is for.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    A language model is a black box with a trillion knobs, and a sparse autoencoder is a way to open it one concept at a time. This skill works with sparse autoencoders for mechanistic interpretability using SAELens: reading pre-trained SAE features to see what a model has actually learned, training a custom SAE when the public ones do not cover your question, and using features for analysis and steering. It is organized as three paths in order — read first, train second, use third — because training an SAE before knowing how to read one is building an instrument you cannot interpret. The honest limitations are part of the skill: SAEs are a lens, not a window, and the skill names what the lens distorts so a finding is not overstated.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse All Claude Skills.

    SAEs are a lens, not a window — features are hypotheses about the model's internals, and the skill shows how to test them, not how to trust them blindly.
    Training a custom SAE needs real compute and a clear question; a vague question produces a vague SAE.
    Feature labels are generated, not ground truth — every label gets verified by firing before it is believed.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    No. Start by reading a pre-trained one — the skill's first path. Train only when the public SAEs do not cover your question.

    A model packs more concepts than it has neurons by letting features share directions — and an SAE is the tool that pulls them apart again.

    Yes, within limits: boosting or suppressing a feature shifts the output toward or away from that concept. The skill shows the test that proves it worked — and the side effects to watch for.