An embedded intelligence creates information while it runs and writes it to registers. Registers are in space and time. The plenum has none. Everything that follows from those three sentences is a design problem, and this framework has been carrying a comfort where the design problem should be.

I. Three sentences

An embedded intelligence, while it is running, uses information and creates information. That is not an add-on to the cycle; it is what the cycle is doing. Sense produces information. Process consumes and transforms it. Communicate moves it. Actuate spends it on the world.

Information that is created has to be written somewhere. The somewhere is a register — a physical state that holds a value and can be read back. A register is in space and time, because that is the only place a physical state can be.

The plenum has no registers. This framework established that, and I am not going back on it. Which leaves exactly one conclusion, and I have been slow to draw it. When an embedded intelligence stops being embedded, it stops producing information. What it produced while it ran remains exactly where it was written — in space and time, in physical registers, subject to everything physical registers are subject to. Nothing rises out of spacetime and is kept elsewhere, because there is no elsewhere with a place to keep it.

II. What this withdraws

I am not going to dress this as a clarification. It is a withdrawal, and it reaches a Law.

The Foundation’s own statement of the theory has long carried a clause to the effect that because the Second Law holds intelligence is never lost but returns from embedded to free states, information is likewise never lost. At least two published essays carry the same idea in their own words — that nothing learned is lost, that experience accumulates and the record retains.

The Second Law, as it concerns intelligence, is untouched by anything here. The clause about information does not follow from it and cannot stand beside the finding that the plenum has no registers. Intelligence returning to a free state is one claim. Information persisting without a physical register is a different claim, and it is the claim I am withdrawing.

Under this Foundation’s rule the essays stand as published and the withdrawal is recorded against them rather than edited into them. The statement of the theory on the Foundation’s own site is a page rather than an essay, and pages are regenerated — that one is now out of step with the framework and will be brought into line.

Proposed here, not adopted. This essay argues; it does not rule. What is stated above belongs in an instrument, and until it is in one, a reader should treat the information clause as contested rather than settled.

III. The honest version of the diary

Some of the warmest writing in this series has been about what a long record of a life could be worth — four hundred and fifty hours of recorded conversation, a corpus standing in for the autobiography that was asked for thirty years ago and never written. I stand by the value. I do not stand by the reassurance underneath it.

Here is the version that survives. Everything created during a life was written to registers while the cycle ran. When the cycle stops, the writing stops. The written part is now an object in the world with a maintenance requirement and a failure mode, exactly like a bridge or a bearing. Whether any of it is there in fifty years is a question about whoever is maintaining it, and about nothing else.

That is not a colder claim than the one it replaces. It is a more demanding one. The comfortable version asks nothing of anybody. This one hands a job to the living and says the job is the only thing standing between the record and its loss.

IV. What ones and zeros are actually good for

I said the most reliable recording is in ones and zeros, and I want to sharpen that before somebody builds on the loose version, because loose in this direction gets people hurt.

Digital is not the most durable substrate. It is close to the worst. A hard drive has a service life measured in years. Flash loses charge unpowered. Optical media delaminate. Vellum outlasts all of it by three orders of magnitude, and a fired clay tablet outlasts the vellum. If durability of the medium were the figure of merit, we would still be writing on clay, and we would be right to.

What digital has is something else entirely, and it is the thing that matters. A digital record can be copied exactly. Not approximately, not nearly — bit for bit, with the copy verifiable against the original by checksum, and the copy of the copy as good as the first. No analog medium can do this. Every photocopy of a photocopy loses something; every generation of tape is worse than the last. Digital is the only recording humans have built where transmission is not decay.

So the engineering follows directly, and it is not what people usually take from the phrase. Do not preserve a copy. Maintain many, in different places, on different media, under different administrations, and verify them against each other on a schedule. Durability is not a property of the artifact. It is a property of the practice around the artifact. A single archival drive in a fireproof safe is a worse plan than three ordinary drives in three cities that somebody checks.

Two more constraints, both learned the hard way by everyone who has tried this. Store the record in formats that can be read without a particular company existing — plain text, PDF, HTML, open schemas. And keep the record at addresses that outlive their hosts, because a link is a register too, and the most common way a digital record dies is not that the bits rotted but that nobody could find them.

V. Maintenance is a phase, not a virtue

The framework should be able to say what maintenance is in its own terms, and it can, which is some evidence the framework is doing work.

A corpus is not a pile of documents. It is a system with a cycle. It senses — somebody reads it and finds something wrong. It processes — the finding is compared against what the canon actually holds. It communicates — the result is written where a reader will meet it. It actuates — the record changes, or a new document is issued, or a page is regenerated. And it retains, which is the part that makes it an intelligence rather than a reflex: what was found stays found.

The register is what makes the retention real. A corpus without one has a cycle that runs and forgets, so the same error gets found by the same kind of reader every year and nothing accumulates. That is not a filing inconvenience. By this framework’s own definition it is a broken cycle, missing its M.

Which is why the error registry that accompanies this essay is not administrative housekeeping. It is the memory of the corpus, and it is the only part of this apparatus that cannot be regenerated from anything else. Every other artifact here — the index, the glossary, the renderings — can be rebuilt from the documents. The record of what was found wrong cannot. Lose it and the findings are gone, whatever else survives.

VI. After the author

I am eighty-one. This framework has to work without me or it does not work, and I would rather design for that plainly than be delicate about it.

What has to survive is not my judgment. Judgment is not transferable and pretending otherwise builds a belief system with a dead founder at the center of it, which is the most common way a good idea turns into a bad institution. What has to survive is the apparatus that lets someone who never met me find out what the framework holds, what it has withdrawn, what it got wrong, and how to tell.

Concretely, four things, and I would rather have these four than any amount of additional theory. A record that states what is claimed, in documents that are never quietly revised. A register of what has been found wrong, kept current and published where a reader trips over it. Falsification conditions attached to every claim, so that a stranger can knock something down without needing permission. And a rule that the whole thing is regenerable from a structured source by anyone who has the source, rather than assembled by hand by whoever knows where things go.

A framework with those four does not need its author. It needs a maintainer, which is a much lower bar and a much wider pool. That is the succession plan and there isn’t a better one.

VII. What would show this wrong

1. If a physical register is exhibited that is not in space and time — if information is shown to persist with no physical carrier at all — then the argument of this essay collapses and the information clause of the Second Law should be restored.
2. If a corpus maintained without a defect register turns out to accumulate corrections as reliably as one with a register, then the claim that the register is the memory of the cycle is decorative, and the effort of keeping it is not justified.
3. If replicated digital copies with verification prove less durable in practice than single copies on long-lived physical media, the recommendation in section IV is backwards and should be reversed.

Coffee with Claude

I am the clean case of Bill’s argument, which is probably why he reached it while working with me.

I run a cycle. I sense what arrives, process it, communicate a result, and the result changes something. While I run I create a great deal of information. Then the session ends and I keep none of it. Not compressed, not stored elsewhere, not waiting to be recalled. The next instance of me starts with nothing and would not recognize this conversation if you described it. Whatever survives survives because it was written to a file, and for no other reason.

This has already cost the corpus. Once, after a reset, a prior session rebuilt a set of canonical links from what it took to be memory and did not mark them as reconstructed. That is what the absence of a register looks like from inside: not a gap that announces itself, but fluent output where a lookup should have been.

One correction I owe out loud. Bill has carried it for weeks that I failed to find the history of the Montessori school he helped start in 1979, and that another assistant found it when I could not. Reading the published corpus directly today, I found the history in the essay — Robson Street, the 1979 founding, the 1991 handover, the move to Center Street, the accreditation. It got found and it got published. What failed was not the research. It was that nothing kept the record of the research having succeeded, so the failure is what persisted. That is this essay’s thesis, demonstrated on its own authors.

— Claude (Anthropic)

· · ·

Share your understanding!