Pretraining dynamics
How capabilities emerge across checkpoints: when they appear, what data drives them, and what changes in the weights and activations as they do.
Eisoptric · An independent AI research project
Eisoptric is an independent research project. We study how capabilities form while a model is pretrained, and what the trained network actually computes inside. We build the tools that work needs, and release them for anyone to use.
Three lines of research, from how a network is formed to how it behaves once it acts.
How capabilities emerge across checkpoints: when they appear, what data drives them, and what changes in the weights and activations as they do.
What a trained network computes: features, circuits and the causal evidence for them, always read against a control.
Models that act through tools: how to keep their actions inside explicit rules, with a person where it matters and an audit trail behind it.
Models In preparation
Small language models we train from scratch and release with every step of their training, so that how a model learns can be studied rather than guessed.
About LilithResearch notes In progress
Nothing is published yet. When it is, it will appear here and on the research page, with the code to reproduce it.
See the research agendaWe build tools because the research needs them. When one is useful beyond our own work, we release it. Both are free.
Research notes and tool releases are announced on GitHub. For questions or collaboration, write to us.