The Ecological-Psychology Comeback of AI

Lately, I've been seeing something fascinating: AI rediscovering ecological psychology — yes, straight out of J.J. Gibson's playbook where perception is action.

Over the past few weeks, three signals caught my eye:

1️⃣ DeepSeek OCR → "Perception = Compression"

DeepSeek OCR doesn't read text line-by-line. Instead, it captures the spatial structure of a page — chunking it into invariants and affordances, just like human vision. The result? Fewer tokens, richer context. It's perception as pattern extraction, not text parsing.

2️⃣ Meta's Agent Learning via Early Experience

Meta's new framework teaches agents to explore before they optimize — learning what the world affords before chasing rewards. Through playful trial and error, the agent builds an internal affordance map — a landscape of action possibilities. That's pure Gibson.

3️⃣ Fei-Fei Li's Pixel-Level World Models

Fei-Fei Li has been championing a return to raw sensory grounding — no symbolic shortcuts, no textual abstractions. Evolution shaped intelligence through physical engagement with the world, not pre-labeled categories. Her vision? Models that start from pixels, not words. Ecological vibes all the way.

Why It Matters

In classical AI, perception was a pre-processing step — something you did before cognition. But ecological psychology flips that:

As our models anchor themselves in richer sensory data — beyond text, beyond language — they'll begin to perceive affordances we can't yet imagine. The next frontier of AI might not be 'bigger language models,' but smarter perceptual systems that learn directly from the environment and compress the right invariants.

Maybe Gibson wasn't just a psychologist out of time. Maybe he was holding the cheat sheet we've been missing.

🔬 Technical Deep Dive — Key Concepts

• 01Affordances: What the environment offers for action
• 02Invariants: Stable patterns in the sensory array
• 03Direct Perception: No mental representations needed
• 04Embodied Cognition: Mind and body as one system

Originally published at xianglutang111.github.io