How AI reads inserts

1 Your word measured

noun 8 senses 13 synonyms

Definitions come from WordNet, a hand-curated dictionary.

2 To an AI, it's pieces measured

DeepSeek V3 inserts 1 token

Mistral 7B in serts 2 tokens

Qwen3 8B inserts 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

insert0.16 inserting0.16 inserted0.16 mounts0.12 associates0.12 constructs0.12 insertions0.12 directs0.11

Dictionary synonyms measured

insert0.16 tuck0.05 enter0.05 introduce0.04 inset cut-in infix enclose inclose stick in put in slip in sneak in

Shade shows how close the model puts each word to inserts, 1 of 8 also appear in the dictionary. 4 of 13 dictionary synonyms are in this build; the rest have no vector to compare yet.

The model’s neighbours

insert0.53 inserting0.48 inserted0.43 insertions0.42 insertion0.31 adds0.23 selects0.20 implants0.20

Dictionary synonyms measured

insert0.53 introduce0.15 enter0.10 tuck-0.00 inset cut-in infix enclose inclose stick in put in slip in sneak in

Shade shows how close the model puts each word to inserts, 1 of 8 also appear in the dictionary. 4 of 13 dictionary synonyms are in this build; the rest have no vector to compare yet.

This model splits inserts 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

asserts0.78 inert0.41 inks0.40 inion0.39 inning0.38 inns0.37 incomes0.37 inlet0.37

Dictionary synonyms measured

insert0.23 enter0.08 introduce0.07 tuck0.07 inset cut-in infix enclose inclose stick in put in slip in sneak in

Shade shows how close the model puts each word to inserts. 4 of 13 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 inserts 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

extracts0.111 instituted0.110 adds0.107 substitutions0.107 invites0.106

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

extracts0.190 introduces0.185 injected0.185 inject0.184 asserts0.181

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

This model splits inserts 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.

inward0.363 injections0.362 inversion0.358 infections0.353 inability0.351

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

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