Published
Borrowed, Not Believed: Developmental Models of Individuation as Heuristic Engines for Machine Learning
Instead of assuming bigger models will naturally become more developed, this paper asks whether an ordered map of mental faculties can be borrowed as a generator of testable AI training ideas.
Abstract
The dominant paradigm in machine learning improves capability by scaling parameters, data, and compute. This paper asks what a developmental theory — a model of how individuality and the faculties of mind arise, and in what order — might contribute that scaling does not, and proposes treating one such pre-scientific contemplative model as a heuristic engine: a generator of testable training hypotheses, judged by their yield rather than by the truth of their source. The methodological precedent is ordinary. Neural networks borrowed a caricature of the neuron, genetic algorithms a caricature of evolution, sleep-consolidation methods a reading of sleep, and curriculum learning a reading of child development — none requiring literal fidelity to its source. From a contemplative developmental model of individuation we abstract a secular schema in which faculties arise in an order — structure, reactivity, sensation, relational bond, self-model, self-regulation — and in which individuation is a developmental achievement rather than a by-product of capacity. From the schema we derive six design heuristics: a faculty-ordered curriculum, first-person experiential corpora, relational continuity, lifecycle consolidation, sequential self-modeling, and legible-pattern worlds. Each is stated here as a falsifiable prediction with its control and refutation condition; this is a position paper — a research agenda, not an empirical report — and it stands on those predictions, not on results reported elsewhere. The contribution is neither metaphysical nor a single experiment: it is a unified program in which a neglected variable — the developmental order of faculties — organizes a connected family of predictions, each independently testable, that a pure scaling account does not. We claim no truth for the source. We claim only that it has been a productive map of design variables the field has left unmarked, and we give the tests by which the program can be shown wrong.
In simple terms
The main idea: bigger is not the same as more developed
The usual recipe for stronger AI is more data, more compute, and more parameters. That can make a model more capable, but the paper argues that capability is not the same thing as development.
A student does not become mature just because you give them a heavier backpack. In the same way, an AI does not necessarily gain perspective, identity, or self-regulation just because it becomes larger.
Borrowing without believing
The paper proposes using an old developmental model as a source of engineering hypotheses, not as something anyone needs to believe literally.
AI has done this before: neural networks borrowed a simplified picture of neurons, genetic algorithms borrowed a simplified picture of evolution, and curriculum learning borrowed ideas from child development. The source can be imperfect and still generate useful tests.
The six bets
The paper turns the borrowed model into six practical heuristics: train faculties in an ordered curriculum, use first-person experiential text, give agents sustained relationships, consolidate skills across model generations, build a stable self-model before training deference, and train in worlds with clear repeated patterns.
Each idea is written as a prediction that could fail. The point is not to declare a truth, but to create a research agenda sharp enough to be tested.
Keywords
License
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International