Beginner
No ML background needed
What machine learning, deep learning and generative AI actually are. How a model turns your sentence into numbers. First real use of Claude Code.
Team onboarding
AI Onboarding Guide
Everything a new joiner needs to go from "I have used a chatbot" to "I can design and defend the system around the model." Three tracks, each one short enough to finish.
How to use this
Read the track that matches where you are, in order, top to bottom. Every topic follows the same shape: a plain explanation of the idea, a note on why it matters for our work, a short video, and the two or three links worth your time. Nothing here is padding, so if a section looks skippable it probably is not.
Watch the videos at 1.5x. Most are under ten minutes. The handful of long ones are marked, and you only need those on the track they appear in.
The one rule
Build something after each section. Reading about retrieval teaches you the diagram. Writing forty lines of retrieval code teaches you why the diagram is wrong in practice.
The three tracks
Intermediate
You can ship code
Transformers and attention. Retrieval that works. Tools and MCP. Context engineering. Agents as a loop you control. Measuring whether any of it improved.
Advanced
You design the system
Harness and loop engineering. Multi-agent orchestration. GraphRAG. Recursive language models. Evaluation that survives contact with a model trying to pass it.
Which track am I?
Pick by what you can do today, not by job title. Most people joining us start on the beginner track even with years of engineering behind them, because the vocabulary is new. It takes about a week.
| Start here if you | Track | Rough time |
|---|---|---|
| Cannot yet explain what a token is, or why a model invents citations | Beginner | 8 to 12 hours |
| Have shipped an LLM feature by calling an API, and watched it fail in ways you could not explain | Intermediate | 20 to 30 hours |
| Have built an agent that runs more than ten steps, and want it to stop falling over | Advanced | 30 to 40 hours |
Your first four weeks
If you are new to the field entirely, this is the schedule. It assumes roughly five hours a week on learning alongside normal work.
-
Vocabulary and first contact
Beginner track, sections 1 to 4. By Friday you can say what a token is, what a context window costs you, and why the model sounds certain when it is wrong.
-
Use the tools properly
Beginner track, sections 5 to 8. Install Claude Code, point it at a real repository, write a
CLAUDE.md, and get one genuine piece of work done with it. -
Understand the machine
Intermediate track, sections 1 to 4. Attention, embeddings, and retrieval. Build a small retrieval pipeline over our own documents and find out where it returns nonsense.
-
Build a loop and measure it
Intermediate track, sections 5 to 9. Write a small agent with three tools, then write the eval that tells you whether your next change made it better or worse.
What the words mean
The terms below get used interchangeably in blog posts and they are not interchangeable. This is the short version. The glossary has the rest.
| Term | In one line | Covered in |
|---|---|---|
| Machine learning | Learning rules from examples instead of being handed the rules. | Beginner |
| Deep learning | Machine learning with many stacked layers, which is what made the last decade work. | Beginner |
| Generative AI | Models that produce new text, images or audio rather than a label or a score. | Beginner |
| Agentic AI | A model calling tools in a loop until a goal is met, with no human in each turn. | Intermediate |
| Context engineering | Deciding what occupies the context window on every single call. | Intermediate |
| Harness engineering | Designing the code around the model: tools, limits, memory, recovery. | Advanced |
| Loop engineering | Designing the cycle that runs the agent, so you stop being the thing that prompts it. | Advanced |
| RAG | Fetching the right text and putting it in the prompt before the model answers. | Beginner |
| Knowledge graph | Storing facts as nodes and the relationships between them as edges you can follow. | Intermediate |
| Benchmarking | Scoring a model or system against a fixed task set so you can compare two versions. | Intermediate |
What finished looks like
A track is done when you can do the thing, not when you have opened every link.
- Beginner. You can explain to a non-engineer why an LLM makes things up, and you have used Claude Code on a real task in our codebase.
- Intermediate. You have built a retrieval pipeline and a three-tool agent, and you have an eval set that tells you when you have broken them.
- Advanced. You have taken an agent that failed on a long task and fixed it by changing the harness rather than the prompt, and you can show the before and after.
How we work with these tools
Four things that are not negotiable, and that no video will tell you.
- Never paste customer data, credentials or anything under NDA into a third-party tool that has not been approved. Ask in the team channel if you are unsure. Asking costs nothing.
- You own what the model writes. Review generated code the way you would review a pull request from a stranger who is fast, confident and occasionally wrong about things they have no way of knowing.
- A demo is not a result. Anything we ship has an eval behind it, even a crude one. "It looked good when I tried it" is the start of the work.
- Write down what failed. The failures are the useful part and they are forgotten within a week. Put them in the team notes.
Where the links come from
Every video and link in this guide was checked before publishing. Videos come from the people who built the thing or who teach it well: Anthropic, 3Blue1Brown, Andrej Karpathy, IBM Technology, MIT CSAIL, AI Engineer conference talks. The eight explainers on the explainers page were built for this guide and run in the page itself.
Links rot. If you find a dead one, fix it and open a pull request rather than mentioning it in chat.
Maintained by the AI team. Corrections and additions are welcome as pull requests.