Beginner

Normal programming means writing the rules yourself: if the order is over 500, apply the discount. Machine learning turns that around. You hand the computer thousands of examples of orders and whether a discount was applied, and it works out the rule. You never write the rule, and often you cannot read it afterwards.

That trade is the whole deal. You get to solve problems where nobody could write the rules down, such as recognising a face or judging whether a review is angry. In exchange you lose the ability to point at a line of code and say why it did that.

• CourseMachine Learning Crash Course Google. Free, interactive, with runnable notebooks. Do the Linear Regression and Classification modules only.
• CourseAI Fluency: Framework and Foundations Claude Academy. Four hours, free, no code. Good if you want the concepts without the engineering.

A neural network is a long chain of very simple arithmetic. Each layer multiplies its input by a grid of numbers, adds a bit, and flattens anything negative. Stack enough of those layers and the thing can approximate almost any relationship you can imagine. "Deep" just means many layers.

Training is a search for the values in those grids. The network makes a guess, you measure how wrong it was, and every number gets nudged slightly in the direction that would have made it less wrong. Repeat a few billion times. The numbers are called weights, and a model is nothing more than a very large file of them.

• ModuleNeural Networks Google ML Crash Course. Interactive, about 40 minutes.
• CoursePractical Deep Learning for Coders fast.ai. Free, nine lessons. Needs about a year of Python. Start here if you want to train models rather than only use them.

Older models produced a label: spam or not spam, cat or dog. Generative models produce content. A large language model is a generative model trained on an enormous amount of text with one objective: given everything so far, predict what comes next. Everything else it appears to do, answering questions, writing code, translating, is a consequence of being very good at that one task.

It is worth sitting with how strange this is. Nobody taught it to answer questions. Answering a question well happens to be an excellent way to predict the text that follows a question mark.

• CourseHugging Face LLM Course, chapter 1 Free. Transformers, and the difference between encoder, decoder and encoder-decoder models.
• CourseAI Capabilities and Limitations Claude Academy. Three and a half hours on what these models can and cannot do.

The model never sees your letters. Your text is first cut into tokens, which are common chunks of characters, roughly four characters each. Each token becomes a number. The context window is how many of those tokens the model can hold at once, and it is a hard ceiling shared by your instructions, any files that were read, every tool result, and the reply it is still writing.

Two practical consequences. You are billed per token, in both directions, so a careless prompt that dumps a whole file into every call is a cost bug. And when the window fills, something has to be dropped or summarised, which is where long sessions start quietly forgetting the thing you told them at the beginning.

• ToolTokenizer playground Paste text in and watch it split. Try code, try a long German word, try emoji. Switch models and watch the count change.
• DocsManaging costs Claude Code docs. Where the tokens actually go in a long session.

Prompting is not a bag of magic phrases. Nearly all of it reduces to three habits: say precisely what you want, show an example of the output shape you expect, and give the model the material it needs instead of hoping it remembers. Telling a model to "think step by step" matters far less than telling it what the finished thing should look like.

The most useful shift is to stop writing requests and start writing briefs. A good brief states the task, the constraints, what to do when something is ambiguous, and what the output format is. Those are the same four things a new colleague would need.

• DocsPrompt engineering guide Anthropic. The techniques listed in rough order of how much they help.

Claude Code is an agent that works inside a repository. It reads files, edits them, runs commands, runs your tests, and commits. The difference from a chat window is that it acts on the real tree, which means it can verify its own work by running things, and it also means a careless session can make a mess. Both halves of that are worth internalising early.

Install it, then point it at a real repository and give it a small genuine task. Not a toy. Something you were going to do anyway and can judge the quality of.

# macOS, Linux or WSL
curl -fsSL https://claude.ai/install.sh | bash

cd your-project
claude

• DocsQuickstart Walks through a first real task end to end.
• DocsCLAUDE.md and memory How persistent instructions work.
• CourseClaude Code 101 Claude Academy. Free and self-paced.
• GuideBest practices for agentic coding Anthropic. Read it after your first week of use, when the advice will mean something.

The model does not know your company's refund policy, your runbooks or last Tuesday's incident. RAG is the plain fix: before the model answers, go and find the relevant passage, paste it into the prompt, and let it answer from that. The name is grander than the idea. It is an open-book exam.

The retrieval part usually works by meaning rather than keywords. Each chunk of your documents is converted into a list of numbers that represents what it is about, and the question is converted the same way, so the system can find chunks that are close in meaning even when they share no words with the question.

• PaperRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Lewis et al., 2020. The paper that named it. Read the abstract and figure 1.

An agent is a model that has been given tools and put in a loop. It produces text saying which tool it wants and with what arguments. Your code runs that tool, hands back the result, and asks the model again. It keeps going until the job is done or a limit stops it.

That is the entire concept. There is no additional intelligence in the word "agent". The interesting engineering is in which tools exist, what goes back into the context each turn, and when the loop is allowed to stop.

• GuideBuilding effective agents Anthropic. The standard reference for the common patterns. Skim it now, read it properly on the intermediate track.

A knowledge graph stores facts as things and the relationships between them. Priya is a node. Acme is a node. "onboarded" is an edge connecting them. Instead of sentences mentioning Priya, you have a structure you can walk.

That matters for questions with more than one hop. "Which customers did Priya onboard that later churned" needs you to follow one relationship, then filter by another. Searching text for similar-looking passages cannot do that, however good the search is.

• CourseNeo4j GraphAcademy Free and hands-on. Neo4j Fundamentals then Cypher Fundamentals is about three hours and is enough to be useful.

The gap between people who find these tools transformative and people who find them a toy is almost entirely habit, not knowledge. Karpathy's video below is two hours of someone competent working out loud, and it will teach you more about practical use than any list of tips.

• TalkIntro to Large Language Models Karpathy, one hour. The best single overview of how these models are built and why they behave as they do.

You are done with this track when all of the following are true.

• You can explain to a non-engineer, without notes, why a language model invents citations.
• You can say what a token is and roughly how many tokens a page of text is.
• You have used Claude Code on a real task in a real repository, and written a CLAUDE.md.
• You can describe what RAG does in two sentences and name its characteristic failure.
• You can explain the difference between a workflow and an agent, and say which one you would build for a task.