Concept

01

The Limits of Prediction

Modern AI systems predict the next token, the next frame, the next action. They excel at continuation—extending patterns observed in training data. But prediction without perception is interpolation without understanding.

A system that predicts what comes next in a sequence has learned statistical regularities. It has not necessarily learned anything about the world that generated those sequences. The distinction matters when we ask systems to reason about novel situations, respect physical constraints, or explain their conclusions.

02

Perception as Foundation

Perception is not passive reception. It is the active construction of structured representations from sensory data. A perceptual system does not merely record; it interprets, organizes, and models.

When we perceive a scene, we do not store raw pixel values. We construct a representation that includes objects, their relationships, their properties, and the causal structure that connects them. This representation is what enables reasoning, prediction, and action.

PALI investigates how artificial systems might construct similar representations—not by mimicking biological perception, but by identifying the computational principles that make perception possible.

03

Layered Architecture

The "L" in PALI represents a fundamental architectural principle: intelligence emerges through hierarchical layers of abstraction. Each layer builds upon the understanding established by layers below, refining and extending the model of reality.

At the lowest layer, raw sensory data is parsed into primitive features. Higher layers compose these features into objects, relationships, and eventually causal models. The highest layers reason about abstract concepts while remaining grounded in the perceptual foundations below.

This layered structure is not merely organizational—it is epistemological. Each layer represents a different level of abstraction, with different validity domains and different constraints. Intelligence requires the ability to move fluidly between these layers while maintaining coherence.

04

Models That Mean

A model, in the PALI sense, is not a neural network. It is a structured representation that captures causal relationships, physical constraints, and the mechanisms by which observations arise. Such models enable reasoning that generalizes beyond training distributions.

The distinction between a model and a function approximator is crucial. A function approximator learns input-output mappings. A model captures the structure that generates those mappings. The former interpolates; the latter understands.

We are interested in how systems might construct, maintain, and reason with models—not as an engineering convenience, but as a fundamental requirement for intelligence.

05

Constraint-Aware Intelligence

Real systems operate within boundaries: physical laws, resource limitations, logical consistencies. Intelligence that ignores constraints produces outputs that are statistically plausible but practically impossible.

A system that generates text may produce grammatically correct sentences that describe physically impossible scenarios. A system that plans actions may propose sequences that violate conservation laws. These failures are not bugs to be patched; they are symptoms of a foundational gap.

PALI investigates how constraints can be embedded in the reasoning process itself—not as post-hoc filters, but as structural properties of the layered models that guide inference.