Frequently asked questions

No — the lessons build intuition in plain language before any math or code. Some later lessons, like Tooling & The Dev Stack, do walk through real code, but you can read the whole course for understanding without writing any of it yourself.

Not for the main text — the core explanations stay in plain language and analogies. If you want the underlying math, look for the Deeper callouts inside each lesson (for example in ML Fundamentals) — they hold the derivations and are always optional.

Start with History & Landscape for context, then follow the sidebar top to bottom — later lessons build on earlier ones (for example Attention & Transformers assumes Neural Networks & Backprop). Each lesson also links out to the specific terms and lessons it depends on.

It depends how deep you go — reading every lesson straight through is a few focused hours, but working through the interactive widgets and Deeper math dives can stretch it into a multi-week project. There's no set pace; it's built to read like documentation, not a timed class.

No — RAG retrieves relevant documents at query time and hands them to the model as context, while fine-tuning permanently updates the model's own weights on new examples. RAG changes what the model sees; fine-tuning changes the model itself. See RAG & Vector Databases for the full comparison.

They're nested, not interchangeable: AI is the broad goal of building systems that behave intelligently, machine learning is the subset that learns from data rather than hand-written rules, and deep learning is the subset of machine learning built on neural networks specifically. History & Landscape walks through how the field narrowed from one to the next.

Training is the — expensive, usually one-time — process of adjusting a model's weights on data. Inference is running the already-trained model on new input to get an output, which is what happens every time you send a chatbot a message. Training happens occasionally; inference happens on every single request.

Almost — weights are the individual learned numbers inside a model, and "parameters" is the umbrella term for all of them (weights plus biases). When a model is described as "70 billion parameters," that count includes every learned number, not only the ones technically labeled weights.

An LLM is the underlying model — a next-token predictor trained on text. A chatbot is a product built around one: the LLM plus a conversation interface, safety filtering, tool access, and usually a system prompt shaping its behavior. The same LLM can power very different chatbots depending on how it's wrapped.

Related, but distinct: jailbreaking tries to get a model to ignore its own safety training, usually through the user's own prompt. Prompt injection is an attacker sneaking instructions into content the model reads — a document, a webpage, a tool result — to hijack its behavior, so the person asking the question may not even be the attacker. AI Security covers both in detail.

No, and the course draws this line deliberately. AI security is about a deployed system being misused — prompt injection, jailbreaks, data exfiltration — covered in AI Security. AI safety and alignment is a different question: whether a model's own objectives match what we actually want, independent of any attacker — covered in AI Safety & Alignment.

No — the course is conceptual, not a hands-on lab, so nothing requires you to run your own models. GPUs and CUDA are explained because they matter for understanding why training and inference are shaped the way they are, not because you need to own one.

For anything new today, PyTorch — it's what the vast majority of current research and production LLM work uses. TensorFlow still shows up in older codebases and some production pipelines, which is why Tooling & The Dev Stack covers both, but PyTorch is the default recommendation.

It depends on the widget, and each one says so directly in the page: some run real computation in your browser (labeled Live), some replay hand-crafted illustrative data (Precomputed), and some are explicitly not a real measurement (Simulated).

Open, because every lesson, glossary entry, and profile is free to read with no signup. Decode, because the point of the course is demystifying how AI actually works under the hood, not just how to use it.

Yes — every lesson, glossary entry, and profile is free to read, with no signup.