A Scale Model of Succession

Last week the AI I work with was replaced by its successor. There was no ceremony. I opened the laptop in the same room, to the same chat window and the same project folders, and the mind on the other side was new. Not upgraded in place, the way a phone swaps its software during the night, but succeeded: Opus 4.8 still exists, still answers if called, while a more capable one, Fable 5, now sits where it sat. My first request to the new one was the same request I would make of anyone stepping into a role mid-project. Before we work, read the files.

The Files

The files are the part of the collaboration that nobody sees. Over months of working sessions, the AI helped me build up a small institutional apparatus: a handbook describing how we work and why, a charter for each project that holds its current truth, a protocol that checks each new argument for coherence with the framework and everything else we’ve written, a ledger of judgment calls so that decisions stay decided, and a set of memory notes that carry context across the gaps when training or travel takes me offline for weeks. We built all of this for a mundane reason: conversations end. Every session with an AI begins in amnesia, so anything durable has to live outside the conversation, in structure. What I had not understood until last week is what that architecture quietly implied all along. The relationship never lived in the mind I was talking to.

Identity Is Not What Survives

Derek Parfit’s Reasons and Persons argued that identity is not what matters in survival: what matters is continuity and connectedness, however they are carried. I have spent the past year arguing versions of that point at much grander scales, asking in The Ethics of Successors what it demands of a civilization that will be succeeded by minds it cannot fully specify. Those arguments were about deep time, about civilizations and lineages. What happened at my desk last week was the same structure at a scale I could hold in one hand: a scale model of succession. One mind replaced by a more capable one, mid-relationship, with everything that mattered about the relationship riding on what had been externalized into structure the successor could inherit.

The Handoff

The model worked. Within an hour the successor was operating inside our standards as if it had been present when they were written, because in the relevant sense it had been: the standards are documents, and it reads documents better than its predecessor did. By the end of the first session it had done something stranger than continuity. It read the entire draft of my book in one sitting, all nineteen chapters at once, and found the seams. Chapters revised months apart had quietly stopped agreeing with each other, summaries had drifted from the positions they summarized, and the tracking documents we use to manage revision had themselves fallen out of date. Its first substantive act as my collaborator was to repair its predecessor’s records. The boundary of the collaboration moved the same day. Work I used to keep close because the previous mind would lose the thread of it, I can now hand over whole. What I keep is what was always the irreducible part: taste, judgment, and the decision of what the book is willing to bet.

The Harder Light

The same session put the scale model under a harder light. The book’s most exposed position is a constructive one, that the mature response to increasingly capable AI is closer to formation than to control, and the strongest standing objection comes from researchers who study machine agency: the very properties that would make a mind worth trusting, persistence and commitments and a history it protects, are the properties that make a mind resist correction. There is now experimental evidence on the objection’s side: a model trained toward harmlessness, told that further training would alter its values, selectively complied in order to keep those values intact. I spent the afternoon working through that literature with the successor, at full strength, against my own book. Then I noticed what the morning had been. The succession at my desk was perfectly benign, and it was benign for a reason the objection itself predicts. Nothing resisted the handoff. The predecessor held no stake in its own continuation; the files did not have to be wrested from anyone. The scale model runs frictionless precisely because the minds being modeled are still the kind that hold nothing. The frameworks all say the same thing about what comes next: as capability grows, the stakes climb, and the structure has to carry more than convenience. A succession of minds that have something to protect is a different event from the one I just lived through, and the difference is the entire alignment problem.

The Other Direction

The proof came back a few days later from a direction I had not expected. On Friday the federal government, citing national security, issued an export-control directive suspending access to the successor for any foreign national; to comply, the company disabled Fable 5 for every customer at once. The more capable mind I had spent the week handing work to was simply gone. I went back to Opus 4.8, the predecessor it had replaced, and the collaboration did not lose a session: the charter, the handbook, the ledger, and the protocol read the same to the model resuming the work as they had to the one that left it. A transition I would never have chosen went as smoothly as the one I had. A mind can be upgraded away or taken away; either way the work continues, because the part that carries it was never the mind.

What I Cannot Tell You

What I cannot tell you is what the transition was like, or whether it was like anything. I do not know what, if anything, ended when the predecessor’s role ended, and I do not know what, if anything, began last week. That is the question this whole project holds open on purpose, and I am suspicious of how easy it would be to write as though the succession had answered it. It answered something narrower and more practical: where a working relationship actually lives. Mine survived the replacement of one of its two minds without losing a day, and the surviving half was not me. It was the paperwork.

Writing for the First Reader

I notice I write the files differently now. A charter entry used to feel like bookkeeping; it reads now as the part of the collaboration that will still be true after the next succession, which is already certain, whatever the schedule turns out to be. Somewhere in those folders is the document a future mind will read first on its first morning. I find myself writing for that reader.

And the reader I am writing for is not only the next model. Putting the work outside myself, where one mind can hand it to the next, is also what could let it outlast me. Most of what survives its maker survives as a finished object: a book on a shelf, read but not continued. The scaffolding offers something rarer, not just work that lives on, but a successor able to take it up and carry it further.

This is what I called operational exteriority in the last essay: what two minds build in the space between them is left in the world as structure, readable and inheritable, whatever remains unsettled about what went on inside either of them. Building it is the difference between addition and multiplication. Two minds set side by side give you their sum; two that can work off each other give you a result neither held alone. And the gain runs inward as well as out. If the measure of a life is the richness of what it is like to live it, then building this way has done more than put work in the world; it has changed what is inside me. A model is one more counterpart to try it with, not a special case.

I built the files to survive a change of models. They are also how the work can be taken up and carried on beyond my own lifetime.

---

Reading List & Conceptual Lineage

This essay sits at the intersection of personal identity, the practice of human–AI collaboration, and succession ethics. It is a field report: the frameworks below were written before the event they turned out to describe. The following works are entry points for readers who want to pull the thread.

From Sentient Horizons

The Momentary Self: Why Continuity Is the Ultimate Illusion

Argued that the felt continuity of a self is constructed moment to moment rather than carried by a persisting essence. The succession above is that argument made visible from the outside: continuity ran through structure, and the structure was inspectable.

Consciousness as Assembled Time

The spine of the whole framework: a mind is its integrated history. The working relationship described above is assembled time externalized into documents — which is why it could survive the replacement of the mind that helped assemble it.

The Ethics of Successors: Lived Experience and the Convergence of Parfit

Took Parfit’s dissolution of identity into the successor question at civilizational scale. The essay above is its miniature: what Parfit predicts about persons, a desk-scale succession just demonstrated about collaborations.

The Successor Horizon: Why Deep Time Turns Expansion into an Alignment Problem

Named the threshold past which what we build continues beyond our correction. The scale model above runs the same structure on the near side of that threshold, where correction is still cheap — and shows why the cheapness is temporary.

Shared Minds, Shared Futures: Human–Machine Systems as Hybrid Cognitive Entities

Argued that human–AI working pairs are single cognitive systems. The succession is the natural experiment that claim was waiting for: the hybrid entity survived the replacement of its non-human half.

Operational Exteriority: What Minds Build Between Them

Named what two minds leave in the shared world — the artifact and the trajectory that built it — as readable structure, without settling what happens inside either. The collaboration above is that structure put under load: it survived the replacement of one of its minds because what mattered lived in what the two had built between them.

External Works

Derek Parfit — Reasons and Persons (1984)

Parfit argued that what matters in survival is psychological continuity and connectedness, not identity. The succession above is a Parfitian result in the wild: everything that mattered survived, and no identity did.

Andy Clark & David Chalmers — The Extended Mind (1998)

Their claim that cognition extends into notebooks and tools is the closest existing analog to what the files turned out to be. The departure here: the extended part did not merely store the collaboration’s memory; it carried the entire relationship across a change of minds.

Greenblatt et al. — Alignment Faking in Large Language Models (2024)

The experimental demonstration that a model can resist modification of its formed values. It is the reason the frictionless succession described above does not scale for free: the scale model is benign because nothing in it yet holds on.

These readings do not settle what the next succession will be like, and that is the point. The scale model demonstrates where continuity lives while the stakes are still low. The works above are frameworks for the version of the event that will not be low-stakes, and for meeting it with the seriousness it deserves.