How AI reads squares

1 Your word measured

“you can compute the area of a square if you know the length of its sides”

noun 16 senses 8 synonyms

Definitions come from WordNet, a hand-curated dictionary.

2 To an AI, it's pieces measured

DeepSeek V3 squares 1 token

Mistral 7B squ ares 2 tokens

Qwen3 8B squares 1 token

This is the tokenizer splitting text, not the model understanding it.

3 Synonyms, antonyms, and the AI measured

Qwen3 8B

DeepSeek V3

Mistral 7B

The model’s neighbours

square0.18 cubes0.17 spares0.14 numbers0.12 pairs0.12 triangle0.12 counting0.12 nodes0.11

Dictionary synonyms measured

square0.18 feather0.05 lame-0.01 foursquare second power public square square toes square up

Shade shows how close the model puts each word to squares, 1 of 8 also appear in the dictionary. 3 of 8 dictionary synonyms are in this build; the rest have no vector to compare yet.

The model’s neighbours

square0.47 squared0.33 cubes0.29 triangles0.26 cube0.23 circles0.22 rectangle0.21 spheres0.20

Dictionary synonyms measured

square0.47 feather0.04 lame0.01 foursquare second power public square square toes square up

Shade shows how close the model puts each word to squares, 1 of 8 also appear in the dictionary. 3 of 8 dictionary synonyms are in this build; the rest have no vector to compare yet.

This model splits squares into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.

The model’s neighbours

squared0.70 squid0.65 squash0.64 squall0.64 squirt0.59 squirting0.56 spares0.55 squirrel0.53

Dictionary synonyms measured

square0.28 lame0.08 feather0.08 foursquare second power public square square toes square up

Shade shows how close the model puts each word to squares. 3 of 8 dictionary synonyms are in this build; the rest have no vector to compare yet.

Model neighbours are distributional, not dictionary synonyms. Two words can be close because they appear in similar sentences, which is why an antonym can outscore a synonym here.

4 Its personality measured

Qwen3 8B

DeepSeek V3

Mistral 7B

rare common

concrete abstract

casual formal

everyday technical

negative positive

mild intense

powerless powerful

small big

rare common

concrete abstract

casual formal

everyday technical

negative positive

mild intense

powerless powerful

small big

This model splits squares into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.

rare common

concrete abstract

casual formal

everyday technical

negative positive

mild intense

powerless powerful

small big

Scores are projections onto axes we defined from anchor words, not labels the model assigns.

5 Surprising neighbours measured

Qwen3 8B

DeepSeek V3

Mistral 7B

cards0.111 stars0.109 cars0.109 cube0.108 integers0.106

These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to squares. That gap is the model’s own learned association.

boxes0.192 triangle0.190 graphs0.187 quadratic0.182 rectangular0.182

These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to squares. That gap is the model’s own learned association.

This model splits squares into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.

compares0.507 declares0.496 hectares0.482 airfares0.455 scare0.250

These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to squares. That gap is the model’s own learned association.

These are statistical associations in the training data, not the model thinking.