8 comments

  • krackers a day ago ago

    >J-lens beats a plain logit lens on some architectures and does nothing on others, and it isn't about size

    The paper talked about this, the jacobian matrix corrects for the shift in basis from initial to final layer compared to logit lens which assumes that the residual remains in the same basis across layers. Maybe the latter is in fact true for some models/architectures so the J-lens doesn't do anything extra?

  • Planktonne 13 hours ago ago

    I think this code be an interesting project, but the generated nature of the site and copy suggests that little thought has been put into it.

    • ada1981 11 hours ago ago

      Sorry to disapoint!

      I suppose it's relative.

      It was essentially 2 days of work using AI to help us get it all together and we were able to release it within 48 hours of the research dropping. It wouldn't have been possible at all without using AI.

  • mune2gu-chan a day ago ago

    I love the focus on the pre-answer thinking phase. It feels way more practical than just trying to steer the output with a generic system prompt.

    • ada1981 17 hours ago ago

      Yes I am curious if we will start to see tools that expose this surface more directly or built into the models.

  • great_psy a day ago ago

    At least link the one made by anthropic. Can’t say if this is better or worse since I’m on mobile.

  • ada1981 a day ago ago

    also from our lab is Personality Bench: https://persona.earthpilot.ai/ which runs every frontier model through personality tests.