Meaning, as a location
Session 2 showed how a neural net turns messy input into numbers. Embeddings are the most useful thing those numbers do: they capture meaning as a position in space. Put every word on a map so that "king" and "queen" sit in the same neighborhood, "cat" and "dog" in another, and "laptop" and "notebook computer" almost on top of each other - and suddenly a machine that cannot read can still find what you mean. Once you see this, search, recommendations, and LLMs all reveal the same gear turning underneath.
Turning things into numbers 5 min live
An embedding is a list of numbers - a vector - that stands in for the meaning of a word, an image, or a product. The machine never gets a definition. It only gets positions: it learns, from seeing billions of examples of how words are used, to place similar things close together and different things far apart. We live in a world of two or three dimensions, so to actually see an embedding we squash it down to a 2D map. On that map, meaning becomes geography.
LiveNo dictionary, only positions3 min▶
- An embedding is a vector. "King" might become a list like [0.7, -0.2, 0.9, ...]. The individual numbers mean nothing to a human; the position they define is what matters.
- Similar things get similar vectors. The model learns this from usage - "cat" and "kitten" show up in the same kinds of sentences, so they land near each other. No one labels them "both animals".
- The machine has no meaning of its own. It has geography. "Close together" is the only notion of "related" it owns - and it turns out to be enough for an astonishing amount of useful work.
- 2D is a shadow. The map above is squashed so we can see it. Real embeddings have hundreds or thousands of dimensions (self-study card) - but the intuition, near = related, holds in any number of them.
Spotify does not just embed songs by genre tags - it embeds them by how people actually listen, so a moody indie track and a certain film score can land next to each other even though no human ever filed them together. The "you might also like" that feels uncannily right is often just a nearest-neighbor lookup in an embedding space. Meaning, learned from behavior, turned into distance.
Distance = similarity, and the famous arithmetic 4 min live
If meaning is a position, then relatedness is distance: near means similar, far means unrelated. That single idea is the engine humming under search, recommendations, and LLMs - they all convert things to vectors and measure how close they are. And because meaning has directions as well as positions, you can even do arithmetic with it: the direction from "man" to "woman" is the same direction as from "king" to "queen".
LiveNear means related - and why that runs everything3 min▶
- Distance is the whole trick. To find things related to X, embed X and grab its nearest neighbors. That is recommendations, search, "find similar", and duplicate detection - all the same move.
- Meaning has directions. king - man + woman lands right next to queen because "royalty" and "gender" are directions in the space. It is a party trick that reveals something real: the geometry actually encodes relationships.
- One gear, many machines. Search converts your query and its documents to vectors and ranks by closeness. Recommenders embed users and items and find near matches. LLMs (Session 6) run entirely on embeddings inside. Learn this once, understand all three.
Old keyword search for "affordable laptop" would miss a product page that says "cheap notebook computer" - not a single word matches. Semantic search embeds both and notices their vectors are almost touching, so it surfaces the page anyway. This is why modern search finally understands that "won't turn on" and "power failure" are the same complaint. Different words, near vectors.
Vector databases and why this matters to you 4 min live
Storing a few embeddings is easy. Storing millions and finding the nearest neighbors to a new query in milliseconds is a real engineering problem, and the tool that solves it is a vector database. This is not a side note - it is the retrieval half of RAG (retrieval-augmented generation, previewed for Sessions 6 and 7), and embeddings of your company's own documents are the foundation of most genuinely useful GenAI deployments you will ever approve.
LiveThe vector DB, and the RAG preview3 min▶
- A vector database stores millions of embeddings and answers one question fast: "what is nearest to this?" That nearest-neighbor search is the workhorse behind semantic search and recommendations at scale.
- This is the retrieval in RAG. When you point a language model at your own documents, the flow is: embed all your docs into a vector DB, embed the user's question, retrieve the nearest passages, and hand those to the model to answer from. Sessions 6 and 7 build on this.
- Your data is the moat. The public model is the same one your competitors can buy. Embeddings of your contracts, tickets, and manuals are what make it useful to you specifically. The value is in what you embed, not the model.
Self-studyWhy hundreds of dimensions, and how images and audio embed too2 min read▶
The 2D map in Part 1 is a convenience for human eyes. Real embeddings have hundreds to thousands of dimensions, and they need them: meaning has far more than two independent directions (formality, sentiment, topic, tense, and hundreds more we cannot even name). Each extra dimension gives the model another axis to separate meanings along, which is why a real space can tell apart "bank" the riverside from "bank" the vault. The 2D picture is a shadow of that high-dimensional reality - useful for intuition, lossy in fact. And the same machinery embeds any input: an image becomes a vector (so "find similar photos" works), audio becomes a vector, even a whole document does. Once everything is a vector in one space, you can measure the distance between a sentence and a picture - which is how text-to-image search works.
The meaning map ★ 10 min · everyone explores
A map of words the machine placed with no dictionary - only by learning what appears in similar company. Click around and watch meaning turn into neighborhoods.
Click one word in each neighborhood and read its nearest neighbors. Do they feel related to you? They almost always do.
Notice the themes formed themselves. No human labeled these clusters "animals" or "tech" - the positions came purely from how the words are used. That is unsupervised meaning.
Find a word that sits between two clusters and ask why. The in-between ones are the interesting cases - meaning is a spectrum, not a set of boxes.
Say out loud what "distance" is doing here: it is the machine's only definition of "related". Everything downstream - search, recs, RAG - is built on measuring it.
Keyword vs semantic ★ 6 min · everyone writes
Write a query the way a real customer would type it, using DIFFERENT words than your company's official term. Think "won't turn on" instead of "power failure", or "money back" instead of "refund policy".
Ask: would old keyword search find the right page? Usually no - not one word matches the official document.
Now picture both phrases as dots on the meaning map. They land almost on top of each other, because they mean the same thing. Semantic search catches exactly what keyword search misses.
Share the sharpest mismatch in the room. This gap - between how customers talk and how your content is written - is precisely the money embeddings unlock.
What would you embed? ★ 6 min · everyone lists
Turn the idea on your own business. Where do you have things that mean the same but are worded differently?
List 3 things in your work that are "similar in meaning but different in words": support tickets, product descriptions, resumes, past contracts, meeting notes.
For each, ask what job "find the nearest ones" would do: dedupe tickets, match a resume to a role, surface the past deal most like this one.
Mark which would clearly beat keyword search - anywhere people describe the same thing many ways, semantic wins.
Circle the single best candidate. That is a real, fundable embedding project - and the seed of a RAG deployment you will understand fully by Session 7.
A support team embedded three years of resolved tickets. Now when a new ticket arrives, the system pulls the five most similar past cases and their fixes - no keyword overlap required. Average handle time dropped, and new agents ramped faster because the "have we seen this before?" question got a real answer. One embedding space, quietly doing the work.
This week ◐ 25 min total
- Play the meaning map from Apply-along 1 for five minutes. Find one neighbor pairing that surprises you and be ready to explain why the machine put them together.
- Finish your embed list from Apply-along 3 - at least 3 candidates, one circled as fundable.
- Read the self-study card above: why real embeddings have hundreds of dimensions, and how images and audio embed the same way.
- Optional deep end: search "word2vec king queen" and look at the original example. Ten minutes on where the famous arithmetic came from.
Three questions before you go 🎯 ◐ 90 seconds
1 · What is an embedding, in one line?
The machine has no dictionary - only positions. Similar things get similar vectors, learned from how they are used, not from any human labels.
2 · In an embedding space, what does distance between two vectors mean?
Near = related is the whole engine. Find related things by grabbing nearest neighbors. It powers search, recommendations, and LLMs alike.
3 · What does a vector database mainly enable?
A vector DB stores millions of embeddings and finds the nearest ones fast. That retrieval step is what grounds a GenAI feature in your own data.
Concepts covered & where they come from
Working intuition of each idea, credited to the canon. Leaders get the judgment; practitioners get the Self-study lane and the names to go deeper.