Show HN: Semantic Calculator (king-man+woman=?)

Overall impressions & comparisons

• Many commenters find the tool fun, reminiscent of word games and “infinite craft”-style combinator systems.
• The ranked list of candidate outputs makes it more engaging than a single answer.
• Others argue that most outputs feel like gibberish with occasional hits, illustrating that the system has relational structure but no real “understanding.”

Behavior, UI, and dictionary quirks

• Case sensitivity is critical: capitalized words often map to proper nouns (e.g., “King” → tennis player; “Man” → Isle of Man).
• Red-circled words indicate missing entries; plurals, verbs, and some basic words (like “human”) often fail.
• Proper nouns (countries, cities) must be capitalized to be recognized.
• Mobile auto-capitalization and ad blockers can break interactions.

Amusing, odd, and failed equations

• Users share many surprising or entertaining results (e.g., “wine – alcohol = grape juice,” “doctor – man + woman = medical practitioner,” “cheeseburger – giraffe + space – kidney – monkey = cheesecake”).
• Simple arithmetic and chemistry are usually wrong (“three + two = four,” “salt – chlorine + potassium = sodium”).
• Subtraction is widely seen as weaker and more random than addition.
• Some directions in the space are “sticky,” e.g., “hammer – X” often yields something containing “gun.”

Biases and unsafe outputs

• Several examples reveal gender stereotypes and offensive associations (“man – brain = woman,” “man – intelligence = woman,” biased race/crime relations).
• Commenters stress that outputs reflect training data, not the author’s views, and suggest explicit disclaimers and/or filters.

Technical discussion: embeddings vs LLMs

• The backend uses WordNet-based vocabulary with precomputed embeddings (mxbai-embed-large), excluding query words from results.
• Commenters note that the classic “king – man + woman = queen” is heavily cherry-picked; often the closest vector is “king” itself unless excluded.
• There’s debate about high-dimensional geometry, the “curse of dimensionality,” and how meaningful vector arithmetic really is.
• Several compare this direct embedding math to LLM behavior: LLMs, with attention and context, often produce more intuitive analogies when asked to “pretend” to be a semantic calculator.
• Others discuss nonlinearity of modern embedding spaces and why naive addition/subtraction works only sporadically.

Ideas and extensions

• Suggestions include: decomposing a given word into a sum of others, using different embedding models, improving bias handling, and gamifying the system.