Ep. 1 | AI Can't Read a Single Word. So How Does It Understand You? | AI Fundamentals

Every sentence you type to a chatbot is, to the model, not words at all. It only does math. This episode follows one sentence on its way into a model: it gets cut into tokens, each token gets an ID, each ID becomes a long list of numbers called a vector, and that vector is a point in a high-dimensional space. Along the way: how subwords and BPE break text into building blocks, why per-token pricing costs more in some languages than others, and how cosine similarity uses direction to tell whether two words are close. We also look at the famous "king minus man plus woman is roughly queen" example: where it came from, and why it often gets oversold. By the end, you'll see that what a model calls "meaning" lives in the directions and distances between points, and so does bias. Chapters: 00:00 Cold open: machines can't read 00:08 Roadmap 00:19 Cutting: tokens and subwords 00:58 How BPE merges 01:42 Vocabulary size trade-offs 02:03 Per-token pricing across languages 02:28 IDs and embeddings 03:13 High-dimensional space 03:33 Coordinates are learned 03:55 Cosine similarity 04:18 King − man + woman ≈ queen 04:37 Why that example gets oversold 05:38 Looking ahead: progress and my take 07:16 Recap 07:32 A question to leave you with, and what's next #AIFundamentals #LLM #Tokens #Embeddings #AIExplained The voiceover in this episode is AI-synthesized.