The Prince Academy株式会社 | Thinking in Vectors: The Art of Semantic Prompting | The 'It Clicked' Toolkit: Advanced Techniques for Power Users | It Finally Clicked: Getting Good at ChatGPT When Theory Isn't Enough

Welcome to one of the biggest 'it clicked' moments you'll have with ChatGPT. We're going to stop thinking about words as just letters on a screen and start seeing them for what they are to an AI: points in a vast, multi-dimensional space. This is the art of 'Thinking in Vectors', and it's the key to unlocking a new level of precision and creativity in your prompts.

At its core, a Large Language Model doesn't understand 'sad' or 'happy' in the way we do. Instead, it represents every word, phrase, and concept as a list of numbers—a 'vector'. These vectors act like coordinates, placing the concept in a giant 'semantic space'. Concepts with similar meanings are placed close together, while dissimilar ones are far apart.

graph TD
    subgraph Semantic Space
        King --- Queen
        King --- Man
        Queen --- Woman
        Man --- Boy
        Woman --- Girl
        Sad --- Melancholy
        Happy --- Joyful
    end
    subgraph "Distant Concepts"
        King -.-> Car
        Happy -.-> Table
    end

The diagram above gives a simplified 2D view. Imagine this in thousands of dimensions! The distance and direction between these points are what encode meaning. This is why the classic analogy 'King - Man + Woman ≈ Queen' works. You're performing mathematical operations on the vectors for these concepts to arrive at a new, logically related concept.

So, how do we use this? Let's explore some practical techniques.

Semantic Anchoring: Instead of using a single, broad term, use a cluster of related, nuanced terms to 'anchor' the AI in a very specific region of the semantic space. This narrows the field of possibilities and results in a much more specific and toned output.

/* Bad Prompt: Relies on a single, broad vector */
"Write a sad poem."

/* Good Prompt: Uses semantic anchoring */
"Write a poem that evokes a feeling of wistful nostalgia and bittersweet melancholy, like looking at an old, faded photograph on a rainy afternoon."

The second prompt provides multiple, closely-related vectors ('wistful nostalgia', 'bittersweet melancholy', 'faded photograph', 'rainy afternoon'). The AI finds the average point between these concepts, resulting in a far more evocative and precise response than just 'sad'.

graph TD subgraph Semantic Space King --- Queen King --- Man Queen --- Woman Man --- Boy Woman --- Girl Sad --- Melancholy Happy --- Joyful end subgraph "Distant Concepts" King -.-> Car Happy -.-> Table end

/* Bad Prompt: Relies on a single, broad vector */ "Write a sad poem." /* Good Prompt: Uses semantic anchoring */ "Write a poem that evokes a feeling of wistful nostalgia and bittersweet melancholy, like looking at an old, faded photograph on a rainy afternoon."