How Learned Transformations Shape Meaning

Watch how the same neural transformations produce different semantic patterns

Activated Features
Suppressed Features
Neutral
Layer 5
CAT
Layer 12
CAT
Layer 20
CAT
Compare Similar vs Different Concepts
CAT vs LION

Both activate feline dimensions
Differ in domestication patterns

CAT vs CAR

Completely different patterns
Living vs manufactured

Key Takeaways

Emergent semantic clustering: Similar concepts (CAT/LION) activate similar dimensions without explicit programming.

Hierarchical abstraction: Each layer builds more abstract representations - from syntax to semantics to complex concepts.

Distributed meaning: "Feline-ness" isn't stored in one place - it's a pattern across many dimensions that emerges from training on billions of examples.

Read Blog Post