1
Your word
measured
“the enemy attacked at dawn”
Definitions come from WordNet, a hand-curated dictionary.
an opposing military force
敵 is Japanese for enemy, and every measurement below was made on that English word. Definitions come from English WordNet; Open Multilingual WordNet supplies the Japanese headword, not a Japanese definition.
Panels 3, 4 and 5 are computed per model. Switch to see them disagree.
“the enemy attacked at dawn”
Definitions come from WordNet, a hand-curated dictionary.
2 of those are a raw byte rather than a character: Qwen3 8B splits this word below the character, so no single token it sees is readable.
3 of those are a raw byte rather than a character: Mistral 7B splits this word below the character, so no single token it sees is readable.
This panel splits 敵 itself, not enemy — a tokenizer does not care what language it is fed. It is the only panel here that does.
This is the tokenizer splitting text, not the model understanding it.
Shade shows how close the model puts each word to enemy. 2 of 3 dictionary synonyms are in this build; the rest have no vector to compare yet.
friend is the dictionary opposite of enemy, yet this model puts it closer than 1 of the 2 synonyms it can score. Opposites share the sentences a word lives in, so cosine alone cannot tell them apart - which is why the dictionary seeds this list rather than the geometry.
Shade shows how close the model puts each word to enemy, 1 of 8 also appear in the dictionary. 2 of 3 dictionary synonyms are in this build; the rest have no vector to compare yet.
friend is the dictionary opposite of enemy, yet this model puts it closer than every one of the 2 synonyms it can score. Opposites share the sentences a word lives in, so cosine alone cannot tell them apart - which is why the dictionary seeds this list rather than the geometry.
Shade shows how close the model puts each word to enemy. 2 of 3 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.
Scores are projections onto axes we defined from anchor words, not labels the model assigns.
These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to enemy. That gap is the model’s own learned association.
These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to enemy. That gap is the model’s own learned association.
These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to enemy. That gap is the model’s own learned association.
These are statistical associations in the training data, not the model thinking.