About

An independent project to understand how models learn.

Eisoptric is an independent research project founded by Rick Weng in Taiwan. We want to understand how AI systems acquire what they know and what they compute when they use it, because that understanding is what makes more capable systems safe to build and to rely on.

Why we exist.

Models keep getting more capable, and our understanding of what happens inside them lags behind. We think closing that gap is one of the most useful things research can do right now.

So we study both ends: how a capability forms while a model is trained, and what the trained network does when it uses it. And because models increasingly act on their own, we also study how to keep those actions under human oversight.

Γνῶθι σεαυτόν.
Know thyself. Inscribed in the forecourt of the temple of Apollo at Delphi.

Looking into.

Eisoptric comes from the Greek eisoptron, a mirror: eis, into, and opt-, to see. Looking into. A model can be watched from outside, by what it says and does, or looked into, by reading what it computes. We do the second.

Our mark is a capital E facing its own reflection across a gilt mirror line: the model, and the model seen from inside. Our own models keep an older name, Lilith, for the reason told on their page.

Who runs Eisoptric.

  • I

    Rick Weng

    Founder of Eisoptric, and for now its only researcher: the experiments, the models, the tools and this site.

  • II

    Independent

    Based in Taiwan. Eisoptric is not part of a company or university.

  • III

    Small on purpose

    A few questions, studied in depth. The tools the work needs are written in-house rather than waited for.

  • IV

    Our own models

    We train small models of our own, the Lilith series, because studying how a model learns means seeing every step of its training.

Get in touch.

General and collaboration
hello@eisoptric.com
Security reports
security@eisoptric.com