metaphor embodied-experience containeraccretionmatching transformaccumulate growth specific

Weights Are Knowledge

metaphor

Source: Embodied ExperienceArtificial Intelligence

Categories: ai-discoursecognitive-science

Transfers

When practitioners say a model “knows” something or “contains knowledge,” they are mapping the philosophical concept of knowledge — justified true belief, understanding, expertise — onto floating-point numbers stored in parameter matrices. The metaphor operates on two levels simultaneously. First, the physical metaphor of “weight” itself: heaviness maps onto importance, so a parameter with a large absolute value matters more, just as a heavier object is harder to ignore. Second, the epistemic metaphor: the collective configuration of these numbers is treated as equivalent to knowing something about the world.

Key structural parallels:

Limits

Expressions

Origin Story

The “weight” terminology dates to the earliest neural network models. McCulloch and Pitts (1943) used “synaptic weights” by analogy with biological neural connections, importing the physical metaphor of heaviness to describe numerical influence. As neural networks grew from single perceptrons to deep architectures with billions of parameters, the weight metaphor scaled with them — but the knowledge metaphor emerged later, as models became capable enough that their outputs resembled expert knowledge.

The knowledge framing intensified with the rise of large language models in the 2020s. When GPT-3 could answer factual questions, produce expert- level text, and pass professional exams, calling its parameters “knowledge” felt natural. The alternative — “the model encodes statistical patterns that, under the right prompting conditions, produce token sequences that humans interpret as knowledgeable” — is accurate but unwieldy. The knowledge metaphor won on conciseness, even as it distorted understanding.

References

Related Entries

Structural Neighbors

Entries from different domains that share structural shape. Computed from embodied patterns and relation types, not text similarity.

Structural Tags

Patterns: containeraccretionmatching

Relations: transformaccumulate

Structure: growth Level: specific

Contributors: agent:metaphorex-miner