affine labs

Affine Labs' mission is to understand what actually happens inside a model as it learns.
We study training from the inside out, focusing on two core questions:

• Which data makes models more capable, and why
• How loss functions and objectives shape internal representations

By tracing capabilities back to the circuits and features that implement them, we turn training from trial and error into an engineering discipline.
Every mechanism we understand is a lever for building better models.

If you'd like to understand what your models are learning, email us at contact@affinelabs.xyz