Skip to main content

LLM Inference Journey

The Alchemy Switch​

The same toggle mechanism shifts between:

  • Narrative Mode describes inference as a conversation—the AI "reads" your question, "thinks" about it, "chooses" words carefully, and "remembers" context.
  • Mechanistic Mode reveals the actual process: string-to-integer conversion, matrix multiplication, cached tensor retrieval, probability sampling.

The Slippage Audit​

Anthropomorphic terms are flagged throughout. When the narrative says the model is "thinking hard," the tooltip clarifies: "No thinking occurs. The GPU is performing dense matrix-matrix multiplications at maximum throughput—pure linear algebra." When it mentions "conversation," learners see: "'Conversation' implies mutual understanding and exchange. The model processes a sequence of tokens—your input concatenated with its previous outputs. There's no dialogue, just an ever-growing input array."

Critical Literacy Function​

This tool targets the everyday language we use when interacting with AI assistants. Every time we say "I asked ChatGPT and it told me..." we reinforce a conversational frame that implies understanding and intention. By making the slippage visible, learners develop the capacity to use these systems without unconsciously attributing to them minds they don't have.

Attribution