Developmentalism

From the AI Research Compass wiki

Developmentalism: Intelligence should be grown in stages like a child, through embodied play, curricula and gradually expanding skills.

Beliefs

Developmentalism takes the child, not the encyclopedia, as its model of how intelligence is acquired. Infants do not read the internet; they track objects, babble, reach, fall over and gradually build concepts of permanence, causality, number and other minds, each stage scaffolding the next. A learning system, developmentalists argue, should follow a similar trajectory: start small, interact with its world, and let competence compound through a curriculum rather than absorb everything at once.

It sits at the far World edge, a little below the middle. Development is embodied and perceptual first, and language arrives later on top, which places it firmly toward World. On the vertical axis it is neither a single scaled model nor an agent framework, but a learning process. Its shade is beyond revolutionary: the idea of a ‘child machine’ is old, yet it runs against the dominant recipe of pretraining on everything at once, and a working developmental AI would rewrite the training pipeline.

Arguments

  • Children learn robust concepts from remarkably little data compared with large models, suggesting the developmental route is more data-efficient.
  • Curricula that order tasks from simple to complex can speed and stabilize learning in a range of settings.
  • Developmental psychology offers a detailed map of which concepts come first, giving concrete targets and benchmarks for artificial learners.

Critiques

  • Children come with a large innate endowment shaped by evolution, so copying the developmental schedule without the priors may not work.
  • Pretraining on huge corpora has delivered broad competence quickly, and developmental approaches have not yet matched it at scale.
  • Curriculum benefits are inconsistent across tasks, and designing the right stages is itself a hard, hand-crafted problem.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +0.91, y ≈ −0.42, paradigm ≈ +0.88.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data