LogoValueSpot
FeaturesUse CasesPricingGuidesCommunity
Get Started

© 2024 ValueSpot. All Rights Reserved.

Privacy PolicyTerms of Service

    News History of https://nitter.privacydev.net/AnthropicAI/rss

    81.
    Over the past few months, our Interpretability team has put out a number of smaller research updates. Here’s a thread of some of the things we've been up to:

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    82.
    R to @AnthropicAI: Crosscoders (published today: https://transformer-circuits.pub/2024/crosscoders/index.html) are a new method allowing us to find features shared across different layers in a model, or even across different models. Identifying the same feature when it persists across layers can simplify our understanding of models.

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    83.
    R to @AnthropicAI: One neat thing is that, very experimentally, crosscoders can be used to "diff" models: comparing between, say, a pretrained model and a fine-tuned one, to see how they differ at a more basic level.

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    84.
    R to @AnthropicAI: Can dictionary learning features be used to detect dangerous behavior? It turns out that they're competitive with (and sometimes beat) linear probes, but can discover spurious correlations that might make a classifier vulnerable to adversarial attacks. https://transformer-circuits.pub/2024/features-as-classifiers/index.html

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    85.
    R to @AnthropicAI: Relatedly, increasing the amount of sparse autoencoder data related to a particular topic can be used to increase the resolution of features related to that topic. https://transformer-circuits.pub/2024/september-update/index.html#oversampling

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    86.
    R to @AnthropicAI: We replicated Gould et al.'s “successor heads” paper, and found that independent component analysis can further decompose these heads into a "successor" component, induction component, and category components. https://transformer-circuits.pub/2024/september-update/index.html#successor-heads

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    87.
    R to @AnthropicAI: We also shared a variety of new evaluations we're experimenting with in our dictionary learning work (https://transformer-circuits.pub/2024/august-update/index.html#interp-evals), along with a case study of our experiences using those evals (https://transformer-circuits.pub/2024/august-update/index.html#evals-case-study).

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    88.
    R to @AnthropicAI: We also report on reproducing Kharlapenko et al.'s "self explaining sparse autoencoder" features experiments: https://transformer-circuits.pub/2024/august-update/index.html#self-explaining-sae

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    89.
    R to @AnthropicAI: And finally, you might have seen our collaboration with Anthropic’s Societal Impacts team on Feature Steering, also published today:

    Source: https://nitter.privacydev.net/AnthropicAI/rss

    2024-10-30 08:10:35 AM
    1
    ...
    678
    9