Intuitive-Physics Doctrine

From the AI Research Compass wiki

Intuitive-Physics Doctrine: Holds that AI should first acquire the core physical intuitions infants show, such as object permanence, solidity and continuity, before anything more abstract.

Beliefs

Intuitive-Physics Doctrine takes its curriculum from developmental psychology. Infants, long before they talk, are surprised when a ball passes through a wall or a toy vanishes behind a screen and fails to reappear. These core expectations of permanence, solidity, continuity and support form the bedrock of common sense. A model that lacks them, however fluent, will make mistakes no toddler would. So AI should be tested, and ideally trained, on these intuitions directly, using violation-of-expectation paradigms borrowed from infant studies.

It sits at the far World edge and low toward the Harness end, preferring structured physical understanding, probed one intuition at a time, over faith in pure scale. Its paradigm level is Beyond Revolutionary. It borrows tools from cognitive science to ask what models actually understand about the physical world, and it is often the source of the uncomfortable benchmark that shows a celebrated model failing at something simple.

Arguments

  • Violation-of-expectation benchmarks reveal gaps in physical understanding that standard vision benchmarks miss.
  • Human intuitive physics appears to work like an approximate mental simulation, which offers a concrete computational model to emulate.
  • Core physical intuitions are prerequisites for safe robots and reliable video generation, where impossible events are costly.

Critiques

  • Scale-first camps argue that these intuitions will emerge from enough video data, making a special curriculum unnecessary.
  • Infant-inspired benchmarks are often synthetic and small, and success on them may not transfer to real scenes.
  • Whether infant knowledge is innate or rapidly learned is itself debated, so the doctrine’s prescriptions are uncertain.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +0.97, y ≈ −0.62, paradigm ≈ +1.00.

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