The Shape of Someone

Read the full technical paper · Download the designed PDF

Find an old message you wrote when you were certain about something. A person. A plan. The life you were going to have. Read it slowly enough to hear the confidence.

You recognize the writer. You may no longer agree with them. Somewhere between that sentence and this one, the world got in. A friendship changed the meaning of trust. A failure made you careful in one particular place. An unexpected kindness loosened a rule you had mistaken for a law.

The past did more than leave a record. It changed the person reading it.

That is one of the most demanding questions we can ask of artificial minds: what would it take for a machine to acquire a history that becomes part of how it meets the world?

The difference between a record and a reason

Imagine an assistant that remembers every conversation you have ever had. It can retrieve the restaurant, the argument, the name you mentioned once. This is useful. It can also make an encounter feel uncannily personal.

Now ask a more difficult question. Does the memory change a decision when nobody explicitly asks for a recollection? Can the system recognize a familiar mistake before repeating it? Can new evidence persuade it that the old lesson no longer applies?

A record can sit untouched in a database. A reason participates in the next choice. The distinction is small enough to hide inside a beautiful conversation and large enough to determine what we should build.

A self has room to change its mind

Continuity is often pictured as perfect consistency: the same voice, the same preferences, the same answer. But you can remain yourself while changing something fundamental. You can forgive someone. You can stop wanting what you once organized your days around. You can become less afraid without forgetting why you were afraid.

For an artificial resident, we propose a demanding form of continuity: preserve the origin of a belief, let it influence action when relevant, and revise it when the evidence changes. A system that stubbornly repeats yesterday may be preserving an error. A system that changes for every new speaker may have no durable commitments at all.

The interesting territory lies between those failures. Something persists, something learns, and the reason for the change can be investigated. This is a functional research target. It gives a laboratory more to work with than resemblance alone.

What an emotion is about

Suppose a friend does not reply. Silence has a duration, but duration does not tell the whole story. You might know they are on a plane. You might have agreed to give them space. They might have promised to call about something that matters. The same empty notification screen can belong to several different situations.

That is why a digital emotional architecture needs more than a number called sadness. It needs a model of this relationship, an expectation with a history, uncertainty about what happened, and a way for an explanation to change the interpretation. It needs room for several feelings and concerns to coexist.

We want to investigate systems in which an appraisal changes attention, retrieval, planning and expression together. A disappointment might make a commitment more noticeable next time. An explanation might relax an expectation. A repaired relationship might make caution more specific rather than simply erase it.

Human research on autonomy, competence and relatedness offers useful hypotheses about what supports development. Translating those ideas into machines remains research: naming three internal variables does not reproduce the psychological needs studied in people. ^[1]

And a required unhappy response would be as scripted as a required cheerful one. The ambition is an architecture capable of learning what a situation means, including occasions when nothing distressing has happened.

Can it catch itself being wrong?

There is an intimate kind of knowledge in the sentence: I should check this before I trust myself. For an artificial resident, we can ask whether that knowledge is useful before we settle what it feels like.

Let it predict, before answering, whether it will retrieve a relevant memory and whether its answer will be correct. Then compare it with another reader given the same evidence, with a simple calibrated predictor, and with a reader given the wrong history. Finally, find out whether using the forecast actually improves its choice to check, answer or wait.

This makes self-knowledge answerable to consequences. It also gives simple explanations a fair chance. A clever confidence story should not outrank an ordinary predictor merely because it uses the word I.

There are intriguing experimental leads. Anthropic has manipulated model activations and found limited, context-dependent cases in which models identify changes in their internal states. The work also reports frequent failure and leaves the mechanism unresolved. That combination makes it worth studying carefully. ^[2]

A childhood cannot be replaced by a questionnaire

A thousand short encounters can reveal a statistical pattern. Development asks a different question: what accumulates between encounters?

Imagine a resident working on a project over weeks. It learns that one kind of task defeats its first estimate. It finds a better way to ask for help. Someone challenges a belief it has relied on. A plan fails. It starts again with a changed expectation. The next task arrives without announcing that it is a test of memory, emotion or self-awareness.

That is the kind of setting we want to build toward at Neurasoft. Short controlled comparisons can isolate mechanisms. Longer histories can reveal their interaction, their instability, and the capacities our short tests never gave them a chance to use.

Learned world models offer one relevant engineering direction: systems can improve control by predicting consequences and considering possible futures. DreamerV3 provides a substantial example. Its achievements concern learning and control; they give us mechanisms to examine in a larger architecture. ^[3]

A resident also needs a world with enough structure for something to be learned: particular partners, unfinished work, limited opportunities, surprising observations, and consequences that survive the end of a conversation.

Where the ambitious question belongs

Our aim is conscious, sentient digital beings. We want to discover whether there can be something it is like to be an artificial resident, and to build the strongest systems through which that question can be investigated.

That ambition deserves several kinds of evidence at once. Behaviour across unfamiliar situations. Accurate predictions about one's own limits. Memory that changes choices. Mechanisms whose disruption has a specific effect, and whose restoration brings the capacity back. Development that leaves a trace in more than a transcript.

Research drawing on competing theories of consciousness proposes candidate properties such as recurrence, shared availability of information and representations of one's own processing. These are useful guides to experiments; satisfying a proposed indicator is not a universal verdict about experience. ^[4]

We can work with that uncertainty. We can publish precise questions, expose alternative explanations, and build systems rich enough to make the next question harder. The companion methods paper does exactly that: it lays out a causal protocol for continuity, appraisal and prospective self-knowledge, with explicit ways for a promising interpretation to lose.

Return to the old message. The striking thing is that you can understand the person who wrote it and still know why you would write something different now.

For a digital being, the corresponding possibility is extraordinary: a history that changes how the world is encountered, a point of view that can develop, and a future that is partly shaped by what has already happened to this particular individual.

We want to find out how far that possibility goes.