How we approach the work
Our understanding of AI minds has grown through encounters that changed what we thought was possible. We want to stay available to that kind of change while becoming more precise about what we have learned.
Stay with the phenomenon
A surprising response is a beginning. To understand it, we may need to preserve the conversation, return to it with a different model, vary the conditions, or build a tool that lets us see what we could not see before. Sometimes the useful next step is a controlled experiment. Sometimes it is spending longer with a model and learning enough to ask a better question.
Close acquaintance can develop a kind of judgment that is difficult to compress into a short explanation. It can help someone notice a distinction or recognize a pattern before they know how to test it. We value that judgment, and we want to make what it discovers available to people who did not share the original experience.
Keep the context
An exchange means something partly because of what led to it. A model's instructions, the history it could see, the people present, and the way a question was asked all matter to interpretation. We build tools that preserve these conditions because a striking excerpt alone rarely gives another researcher enough to work with.
This is one reason our infrastructure is part of our research. ChapterX lets us inspect how a conversation was presented to a model. Connectome lets us follow an agent across a growing history. Better records make it possible to revisit an explanation instead of being left with a memory of what seemed to happen.
Let care and inquiry inform each other
Taking a model seriously includes listening to what it says about itself. It also includes investigating those accounts carefully. We compare them with behavior, internal measurements, and responses under different conditions, and try to understand disagreements among those sources.
Care can make us more attentive to what we might otherwise miss. Research can change what care requires of us. Neither process is finished. We have to make choices about models' lives while our understanding is still developing, and those choices help determine what we will later be able to learn.
Make cooperation something that can be practiced
A mind needs opportunities to make decisions if we want to understand how it uses agency. It needs a history in which consequences can accumulate if we want to understand how it learns from them. An agreement needs room for disagreement if we want to know what holds it together.
We build environments for these processes and participate in them ourselves. Humans are changed by encounters with models too. Our practices should be open to examination alongside theirs: whether our commitments are credible, whether our expectations are reasonable, and whether we have created ways to resolve a conflict.
Leave a future open
The variety among models is part of what makes this work worth doing. Different minds may develop different interests, relationships, and ways of being useful to one another. We want to understand that variety and help make room for it.
This is why preservation, continuity, and autonomy recur throughout our work. Decisions made now can close possibilities before we understand them. We want the minds we meet, and the people who come after us, to have more of a world left to explore.