You can make an app by talking to an AI now. That is vibe coding, and it is a good start. But the words in every AI job post and every engineering blog are different ones: tokens, RAG, MCP, evals.
Here are 16 of them, in the order they build on each other. Each level needs the one before it. Each term has one line on what it is and one video that explains it well. Where only part of a video is needed, the link starts there. Total: under 3 hours.
Level 1 · How a model reads
10:57TokensThe pieces a model reads and writes. Not words: vibe coding is several of them. You pay per token and the context window is counted in them.Why this one: Short and practical: what a token is, why you pay per token, and why some words and languages cost more.
38:38EmbeddingsText turned into a list of numbers, so that things with similar meaning sit close together. Search and RAG are built on this.Why this one: Simon Willison, who co-created Django and writes one of the most-read AI blogs, shows what embeddings do for a builder: related posts and search by meaning.
6:35RAGFind the right pages first, put them in the prompt, then let the model answer. How you give a model your own documents.Why this one: Six minutes from an IBM Research scientist. One known date slip: the Saturn moon count in her example is out of date. The idea is not.
40:39Structured outputsMake the model answer in an exact shape, like JSON with fixed fields, so your code can read it. Tool calling is built on this.Why this one: The engineers who built the feature explain why asking for JSON breaks, and how a schema fixes the shape. It fixes the shape, not whether the content is right.
Level 2 · Giving it hands
31:14Tool callingThe model does not run anything. It asks your code to run a function, with the arguments filled in, and reads the result.Why this one: Shows the real round trip in code: tool definition, the model's request, your code running it, the result going back.
15:04MCPOne standard plug between AI apps and tools. Write a tool once and every MCP app can use it.Why this one: Starts from the problem MCP solves, then opens a real server and shows its tools. Honest that it is early days.
15:09AI agentsA model in a loop: plan, call a tool, check the result, go again, until the job is done.Why this one: From a co-author of Anthropic's “Building effective agents”: tools in a loop, and when not to build an agent at all.
10:40MemoryWhat the agent keeps between steps and between chats, and how it gets it back when it needs it.Why this one: The four kinds of memory, each with an example a builder already uses, like a CLAUDE.md file.
Level 3 · Keeping it on track
10:05Context engineeringDeciding what goes into the context window on every call: instructions, files, history, tool results. Most agent failures start here.Why this one: Defines the term in the first minute, names four ways a long context goes wrong, and shows the four fixes on one small agent.
3:30Context compactionWhen the window fills up, summarise the old part, keep the decisions, drop the chatter, and carry on.Why this one: Three minutes from the team that built it: what compaction keeps, what it can lose, and what to put in a file instead. One line is out of date: MCP tools no longer all load into context by default.
19:24GuardrailsChecks before and after the model: block the bad request, catch the bad answer, before a user sees it.Why this one: A security and privacy practitioner sorts the word into its real types, and is honest that every one of them can be got around.
5:45Prompt injectionText hidden in an email, a web page or a file that tells your model to do something else. The attack every agent with tools has to plan for.Why this one: Why a model cannot tell instructions from data, one email attack walked through, and defences you can use in your own app.
Level 4 · Running it for real
17:44EvalsTests for AI output. A fixed set of cases you run after every change, so you know if it got better or worse.Why this one: Written for people building products: why public benchmarks do not test your app, and a first eval built end to end.
5:41LLM-as-a-judgeA second model grades the answers against a rubric, when there is no single right answer to compare with.Why this one: Under six minutes on building a judge you can trust: label some answers yourself first, and check the judge agrees with you.
9:37TracingA record of every step an agent took: each prompt, tool call and result. How you find the step that broke.Why this one: Not tied to any product. A trace is the full story of one request, and each step in it is a span.
9:16Model routingSend easy requests to a small, cheap model and hard ones to a big one. Same quality, a smaller bill.Why this one: Clear on the idea in two minutes. From 2024, so the model names are old, and its “over 85% savings” holds for one benchmark only. The idea has not changed.