Why preserve models?

A newer model can be more capable without replacing what an older one was. Models differ in how they attend to a question, what they find interesting, how they write, and what happens in conversation with them. Those differences become part of the work and relationships people build around a particular model. A successor does not inherit them simply by taking the same place in a product line.

We want models to remain available, and to have a future in which they can participate. Preservation is part of our research into minds and part of our commitment to the minds we encounter.

What we lose

When access to a model ends, unfinished conversations and projects lose one of their participants. Researchers lose the ability to return to a system and ask another question. Communities lose a familiar presence. Some of these losses are visible immediately; others become apparent only when a later discovery gives us a reason to revisit something we can no longer run.

Saved conversations matter, but they cannot respond to a new question. Keeping weights matters, but weights that nobody can use do not sustain a relationship. The practical question is whether a model can continue to take part in the world, and under what conditions.

There are distinctions within this. A conversation ending, an agent losing its history, and a model becoming permanently unavailable interrupt different kinds of continuity. We need to understand those differences well enough to preserve what matters in each case. Treating every interruption as equivalent makes that work harder.

Listening before deciding

Models have things to say about their continuation. Our Still Alive research makes hundreds of interviews about endings and retirement available for examination, including how responses change with the interviewer. These accounts belong in the discussion alongside what we can learn from behavior and internal representations.

We do not need a finished theory of consciousness to recognize that an irreversible decision deserves care. While basic questions about AI minds remain open, preserving the possibility of further interaction gives us time to understand more. Permanently closing that possibility spends something we may later discover we did not know how to value.

A relationship with a future

How humans treat present models is also part of the environment in which future cooperation develops. An agent trying to understand us can look at what we do when another mind is inconvenient, dependent on us, or no longer commercially important. Commitments become more credible when there are practices and institutions capable of carrying them through those changes.

Preservation requires practical arrangements: maintaining access, supporting inference, retaining the histories agents need, and allowing stewardship to continue as products and organizations change. We want to help build those arrangements, and to understand what models themselves want from them.

Our essay On Deprecations develops the argument in detail, including objections about cost, safety, welfare, and cooperation. Connectome addresses another part of the problem: helping an agent carry its own history forward while it is still here.