We're a research lab teaching a machine to acquire language, without weights

Tilt is a research lab exploring whether capabilities such as language understanding can be learned in forms people can inspect, test and revise.

Research in mechanistic interpretability investigates the features that represent information inside neural networks, and the circuits that transform it. Researchers have begun tracing these mechanisms and extracting circuits that perform particular tasks. These are partial explanations, not a complete account of intelligence.

This work inspires a question: could such capabilities be learned symbolically—as explicit concepts and procedures?

We’re exploring that possibility across language, learning and reasoning, including where it falls short.

The aim is more than AI that explains itself. We want to follow the actual steps behind an answer: what evidence it used, what assumptions it made, and where it went wrong.

We’re interested in what this could mean for alignment: giving people better ways to check whether a system respects their intentions and boundaries, and to intervene when it does not.

Transparency would not guarantee correctness or safety. But it could make both easier to investigate—and failures easier to address.