Nobody else's agent

In 1970, Nicholas Negroponte wanted a machine that would argue with its architect.

A designer then could keep a private library of routines, but the routines never learned who was using them. Negroponte's architecture machine was to adapt through dialogue, so that a year with one architect changed what it could contribute. He saw the trap as well. A machine that knew the world only through its partner would inherit that partner's prejudices and grow steadily more agreeable. To stay capable of disagreement it needed contact with the world beyond the relationship, including the work of other designers.

The first half of that ambition is now within reach, and the second half is the reason to be careful with it.

An assistant with a history

A language model arrives already knowing a profession's vocabulary and gets the first attempt roughly right. A year of working together ought to add something. At present it adds a few saved instructions and a preferred format.

Take a building inspector; call her Dana. She dictates site notes and an assistant turns them into the county's report form. Within a week it knows the form. A month in, her notes for one room read: dark staining, north wall below the window, about forty centimetres across, drywall soft to the touch. The draft comes back: water intrusion has damaged the wall framing. Dana strikes the second half. Nothing in her notes touched the framing. She writes instead: condition of framing not determined; recommend invasive inspection.

The form is a fact about Dana; the correction is a rule about how far a stain may be pushed toward a conclusion, and it will apply next month to a different room, a different material, and none of the same words. Dana does not want to make the same correction in new vocabulary for a year. She wants the assistant to notice, on its own, that it is about to make the same leap.

What a correction teaches

The first thing to keep is the correction itself: her notes, the draft, and her revision, together. They show what was known, what the assistant inferred, and where she drew the line. A rule extracted from the revision alone, something like do not infer hidden damage, loses the part that matters, which is how close the evidence came.

An assistant that has saved every conversation has an archive, not a colleague. Putting the right record in front of the model improves the next report; the hard part is recognising, when a new report arrives, which of last year's corrections bears on it.

Some of what it holds must stay explicit and editable. The county revises its form. A client asks that a detail be removed. Dana should be able to open those records and change them.

When she corrects the same kind of leap across a dozen varied reports, those dozen are something else: training material. The assistant then has two ways to carry its past: records it consults, and behaviour shaped by training. The second needs more care, because a habit pushed into a model's weights cannot be opened and edited.

What training adds

Shaping behaviour is routine now: LoRA trains a small adapter beside a fixed model, and a service like Tinker runs the training for someone who has never touched a GPU. For Dana the material is a year of reviewed reports. A candidate update is trained on some of her corrections and tested on the rest, on new cases where the same distinction should apply, and on everything the assistant already did well.

Passing that test does not mean agreeing with every correction; an assistant that calls every stain inconclusive has swapped one bad habit for another. Improvement means becoming more sensitive to the evidence that separates the cases. An update that fails leaves the previous version in use.

No training service does the judging. A correction can be a general lesson, a local exception, or Dana being wrong that afternoon, and a dozen examples will not always say which. A changed model is easy to produce. Whether the change deserves to guide next month's work still has to be judged, and an assistant that writes a cheaper report but needs Dana to catch every familiar mistake has taken little of it off her hands.

Whose experience it is

After a year, Dana's assistant is more than the model she started with: records she can open, instructions worn in by use, and an adapter trained on her corrections.

Not all of that history is hers. The practice of never inferring hidden damage she learned from a former employer. A distinction about crawlspaces came from a colleague reviewing her report. Facts about a client's property came from the client. Calling the assistant personal does not decide which of these it may keep, who else may use them, or what leaves with Dana when she changes firms, and none of it can be settled while the pieces are mixed together.

So Dana's control has to reach the training, not just the files. She should be able to see which examples an update was trained on, pull one out, compare versions, and go back. Deleting the file leaves its influence in the model already built; being rid of it means rebuilding that version and testing it again.

Models are replaced every year or two, and an adapter for one will not fit the next, but a kept record of reviewed work is enough to train and check a successor.

The other half of Negroponte

An assistant shaped by one person's corrections gets better at that person, including at her mistakes. Suppose Dana has a habit of her own: in older houses she tends to write off settlement cracks as cosmetic, because in her experience they nearly always are. An assistant trained on a year of her reports learns that too, and grows more sure of it than she is. One day it writes hairline cracking, cosmetic, over a crack she would have paused on. Negroponte's agreeable machine is the model that has learned its partner too well. The assistant's value depends on helping her revise a conclusion when the evidence changes, including conclusions it learned from her.

Which is why a personal assistant cannot be a closed one. When Dana's reaches a material it has never seen, or a crack it has seen too often, it should go outside, to a stronger general model or to an agent shaped by a specialist's work, taking only what Dana is entitled to share and paying for the answer the way “Too cheap to sell” describes. Its own history is what lets it ask well: it knows what has been checked, what remains uncertain, and what kind of answer would let the work go on. Personal, in Negroponte's sense, never meant alone.