Research

What we study, and what we publish.

Eisoptric’s research asks how a model comes to know what it knows, and what it computes when it uses that knowledge. Below are the questions we are working on, and the notes we publish as answers become defensible.

Research agenda.

The questions each line of research is organized around. They will change as we learn.

I

Pretraining dynamics

A capability does not exist at step 0 and does at the end of training. We want to know what happens in between.

  • At which checkpoints does a capability appear, and is the transition gradual or abrupt?
  • Which parts of the network change when it appears, and which data are responsible?
  • Do the same mechanisms form at different model sizes and seeds?
II

Interpretability

A score says that a model can do something. It does not say how. We look inside.

  • Which features and circuits carry a behaviour, measured with activation patching and sparse autoencoders?
  • Does each finding survive a causal test against a random control?
  • How much of a behaviour does an explanation account for, and what does it leave out?
III

Agent oversight

Once a model acts through tools, what it is allowed to do matters as much as what it can do.

  • Which rules on tool calls stop harmful actions without stopping useful work?
  • Where should a person approve an action, and with what context?
  • What record of an agent’s actions is enough to reconstruct and judge them afterwards?

Our own models.

To watch a capability form, every step of training has to be in front of us. So we train small models of our own, the Lilith series, and will release every checkpoint, the data order and the code with each.

In preparationAbout Lilith →

Research notes.

Published notes appear here, newest first.

In progress

No notes are published yet.

Our first notes are being written. This list will show each note’s date, title and a one-sentence summary.

What every note includes.

  • I

    The question

    What we set out to find, and the hypotheses we held before we looked.

  • II

    The method

    Models, checkpoints, data, and the exact code, so the result can be reproduced.

  • III

    The evidence

    The numbers, the samples behind them, and the controls they were read against.

  • IV

    The limits

    What did not work, what the result does not show, and corrections when we get something wrong.

Want to be told when a note is published?

Follow Eisoptric on GitHub, or write to us.