Word2Vec — How Words Became Vectors

A neural network can only ever crunch numbers, so the very first problem in NLP is turning a word like "king" into a vector. The obvious fix — one-hot encoding — technically works but throws all the meaning away: every word ends up exactly equidistant from every other, so "king" is no closer to "queen" than to "banana". Word2Vec earns its geometry instead of declaring it. The trick is the distributional hypothesis: you know a word by the company it keeps. Word2Vec (skip-gram and CBOW) trains a tiny network to predict a word's neighbors, and the only lever it can tune is the dot product between word vectors. Maximizing that dot product for words that actually co-occur quietly rotates words-that-keep-similar-company into the same direction — similarity of meaning becomes alignment of direction. And out of that single prediction task, structure nobody designed falls out: the man→woman arrow and the king→queen arrow become the same step, so king − man + woman ≈ queen. Relationships turn into arithmetic you can do on vectors — the idea sitting underneath nearly every modern language model. *Related Videos* ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ Cosine Similarity is Just Direction, Not Distance: https://youtu.be/YF8BbMJIgCo Two Towers vs Siamese Networks vs Triplet Loss: https://youtu.be/3CwWGSV0l9o LLM Tokenizers Explained: BPE, WordPiece and SentencePiece: https://youtu.be/hL4ZnAWSyuU Softmax function - Explained: https://youtu.be/oJU6-qW6xZU Attention Mechanism Explained: https://youtu.be/abMWyvuEvcE Vector Database Search - HNSW Explained: https://youtu.be/77QH0Y2PYKg *Contents* ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 00:00 - Words Aren't Numbers 01:13 - Meaning From Company 02:09 - Predict the Neighbors 03:09 - The Dot Product Is the Score 04:16 - King − Man + Woman 05:26 - Meaning You Can Add *Follow Me* ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 🐦 X: @datamlistic https://x.com/datamlistic 📸 Instagram: @datamlistic https://www.instagram.com/datamlistic 📱 TikTok: @datamlistic https://www.tiktok.com/@dataml