How AI reads cones

1 Your word measured

noun 7 senses 25 synonyms

Definitions come from WordNet, a hand-curated dictionary.

2 To an AI, it's pieces measured

DeepSeek V3 cones 1 token

Mistral 7B con es 2 tokens

Qwen3 8B cones 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

cone0.20 candles0.12 lakeside0.12 confrontation0.11 yellowcake0.11 lakes0.11 telescopes0.11 grasses0.11

Dictionary synonyms measured

cone0.20 con0.05 scam0.02 learn-0.01 conoid cone shape strobilus strobile cone cell retinal cone victimize swindle rook goldbrick nobble diddle bunco defraud mulct gyp gip hornswoggle short-change memorize

Shade shows how close the model puts each word to cones, 1 of 8 also appear in the dictionary. 4 of 25 dictionary synonyms are in this build; the rest have no vector to compare yet.

The model’s neighbours

cone0.49 cubes0.18 cylinders0.18 tubes0.18 cane0.17 discs0.17 poles0.17 clones0.17

Dictionary synonyms measured

cone0.49 con0.10 learn0.04 scam0.02 conoid cone shape strobilus strobile cone cell retinal cone victimize swindle rook goldbrick nobble diddle bunco defraud mulct gyp gip hornswoggle short-change memorize

Shade shows how close the model puts each word to cones, 1 of 8 also appear in the dictionary. 4 of 25 dictionary synonyms are in this build; the rest have no vector to compare yet.

This model splits cones 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

con0.85 conk0.73 conto0.70 congo0.65 conferences0.63 conforms0.63 conrad0.61 concha0.61

Dictionary synonyms measured

con0.85 cone0.16 scam0.13 learn0.02 conoid cone shape strobilus strobile cone cell retinal cone victimize swindle rook goldbrick nobble diddle bunco defraud mulct gyp gip hornswoggle short-change memorize

Shade shows how close the model puts each word to cones, 1 of 8 also appear in the dictionary. 4 of 25 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 cones 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

cues0.106 bowls0.105 crews0.103 lake0.103 doses0.100

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

rods0.166 cores0.159 consoles0.158 spheres0.158 triangles0.157

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

This model splits cones 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.

concurrent0.600 concessions0.599 concession0.596 converse0.595 conquer0.595

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

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