Skip to main content

LLM Training Journey

The Alchemy Switch​

A toggle in the top-right corner shifts the entire interface between two modes:

  • Narrative Mode presents the training process using familiar metaphors—the model "learns," "understands," "absorbs knowledge," like a student in school.
  • Mechanistic Mode strips away these framings, revealing the computational reality: loss minimization, gradient descent, matrix multiplication, probability distributions.

The Slippage Audit​

In Narrative Mode, anthropomorphic terms appear with dashed underlines. Hovering over them reveals tooltips that expose the gap between the metaphor and the mechanism. For example, when we say a model "learns from its mistakes," the tooltip explains: "No mistakes or learning in the cognitive sense. Loss is computed, gradients are calculated via chain rule, weights are updated via optimizer. It's calculus, not reflection."

Critical Literacy Function​

This tool makes visible how consciousness language, especially verbs like "knows," "believes," "understands,” smuggles in assumptions about mind and agency where none exist. By toggling between modes, learners can see how the same process can be framed as education or as optimization, and they're prompted to ask: What is the actual mechanistic process being obscured by consciousness language?

Attribution​