AI Sentience & Welfare

Veylan Solmira · Future Impact Group

Interpretability and philosophy of machine minds in frontier models — spanning persona dynamics, self-modeling, and how welfare can be measured without the probe distorting the answer.

Research directions

1. Persona fingerprintingBuild an empirical vocabulary of a model's personas in SAE-feature space, then find the machinery it uses to switch between them.
2. The observer effectDoes probing a model for its welfare change what you measure — do you read the persona the probe elicited, not the model itself?
3. The persistent substrateIs there structure beneath every persona — features present across all of them — that behaves like an agency underneath the characters?
4. Self-experimentationCan a model learn about itself by searching over steered copies of itself, each run under a different intervention?
5. Theory & writingWhere interpretability can adjudicate the debate over whether persona-simulation deflates — or supports — claims of model consciousness.

Explorables