HOW AI THINKS?

 

HOW AI THINKS?

The AI doesn't "think" in English. It thinks in a geometric space that has no direct human translation. Natural language is the input and output, but the processing happens somewhere else entirely.

These are the canonical examples from word2vec, GloVe, and similar embedding systems. Each show that a consistent semantic relationship corresponds to a consistent direction in latent space.

30 examples for discussion.

The useful way to frame these is as latent-space hypotheses, not as literal claims that an AI “thinks” in geometric vectors. In embedding models, words, concepts, and tokens can occupy high-dimensional representations where some relationships appear approximately as directions, clusters, or transformations.

Here are 30 discussion examples, moving from classic word embeddings toward more interesting conceptual cases.

30 examples of semantic geometry

#	Example	Geometric intuition	Discussion question
1	king → queen	king − man + woman ≈ queen	Is gender represented as a reusable direction?
2	man → woman	A gender-related transformation	Does the same transformation work across occupations?
3	boy → girl	Similar gender transformation at another age level	Is the relationship linear or context-dependent?
4	uncle → aunt	Male/female transformation within kinship	Does semantic geometry preserve family structure?
5	father → mother	Gender + parental-role relationship	Are multiple semantic dimensions superimposed?
6	Paris → France	Capital-to-country relationship	Can a geographic relation behave like a vector?
7	Rome → Italy	Another capital-country pair	Does Paris − France resemble Rome − Italy?
8	Tokyo → Japan	Same structural relationship	How stable is the transformation across cultures?
9	London → UK	Capital-country association	Does the model encode political geography or textual association?
10	Madrid → Spain	Capital-country mapping	What happens when a country has multiple culturally important cities?
11	walking → walked	Inflectional transformation	Does time/tense occupy a direction in latent space?
12	run → running	Verb → participle	Can grammatical morphology be represented geometrically?
13	write → writing	Root → gerund	Is morphology separated from meaning?
14	big → bigger	Comparative transformation	Does degree form a consistent direction?
15	big → biggest	Superlative transformation	Is semantic intensity geometrically ordered?
16	good → better	Irregular comparative	What happens when linguistic form is not morphologically regular?
17	cat → kitten	Animal → young animal	Is age encoded as a semantic direction?
18	dog → puppy	Animal → young animal	Does the same transformation generalize?
19	bird → nest	Entity → associated environment	Is this a relation or merely contextual proximity?
20	fish → water	Entity → habitat	Can latent space represent ecological relationships?
21	doctor → hospital	Profession → workplace	Does occupation encode an environmental relationship?
22	teacher → school	Profession → institutional environment	Can relational structure emerge without explicit rules?
23	car → road	Object → typical environment	Is this geometry based on world knowledge or language statistics?
24	knife → kitchen	Object → typical location	How does contextual frequency shape the geometry?
25	coffee → cup	Substance/drink → container	Does the model represent functional relationships?
26	hot → cold	Oppositional semantic direction	Are opposites close because they share contexts?
27	happy → sad	Emotional opposition	Why can semantic opposites sometimes be geometrically close?
28	doctor → medicine	Profession → domain/tool	Can several relationship types overlap in the same space?
29	Einstein → physics	Person → field of association	Does latent space encode knowledge through relational neighborhoods?
30	question → answer	Functional/interactional relationship	Can geometry represent an abstract relation rather than an object?

 

The more interesting experiment

The classic equation is:

king − man + woman ≈ queen

But for discussion, I would expand the idea into:

A semantic relationship may correspond to a transformation in representation space.

For example:

Paris − France

       ≈

Rome − Italy

       ≈

Tokyo − Japan

Or:

doctor − hospital

        ≈

teacher − school

The important caveat is that these should not automatically be interpreted as literal reasoning operations performed by the model. They are observable properties of particular representation spaces and depend strongly on the model, training data, tokenization, layer, context, and embedding method.

Three levels worth discussing

1. Lexical geometry

cat

 ├── kitten

 ├── dog

 ├── animal

 └── pet

Words form neighborhoods according to patterns learned from data.

2. Relational geometry

Paris → France

Rome → Italy

Tokyo → Japan

The interesting question is whether similar relationships correspond to similar transformations.

3. Conceptual geometry

Imagine a space containing:

      ABSTRACT

         ↑

         |

emotion ← HUMAN → social

         |

         ↓

      physical

Here we move beyond individual words toward latent representations of concepts and relationships.

And this leads to the really provocative question behind your scenario:

If natural language is merely the interface, what exactly is the internal representational object that the language is describing?

That is a much stronger research question than simply saying “AI thinks in vectors.” It opens the door to investigating whether model representations contain concepts, relations, abstractions, transformations, and possibly internal states that cannot be translated one-to-one into human language.