AI Foundations
Stop guessing how models work. Understand the machine you use every day.
Most people using AI daily could not tell you why the same question gives two different answers, and it costs them. They over-trust a confident paragraph, under-trust a correct one, and pick tools by advertising because they have no way to judge. This module is the fix, and it is deliberately not a maths course. You do not need gradients to work well with these systems; you need an accurate mental model of what the machine is doing when it answers you. So we start where the behaviour actually comes from: prediction, training data, attention, and the fact that a model has no idea what is true — only what is likely. That single sentence, properly understood, explains hallucination better than any diagram. Then we make it physical. You will run a model on your own laptop, watch it get slower as you push it, and discover that "AI" is a file and some arithmetic rather than a cloud oracle. That demystification is the point. The people who skip this stage stay dependent on whoever sells them the next tool. The people who do it stop being users and start making decisions.
Prerequisites
- No programming required
- Comfort installing software on your own machine
- A laptop with 8 GB of RAM or more
What you can do afterwards
- Explain in plain language how an LLM produces its answer — and why it varies
- Tell a hallucination apart from a missing-context problem, and act accordingly
- Run an open-weights model locally and judge whether local or hosted fits your case
- Write prompts that work because they are specified, not because you got lucky
- Argue an AI ethics or data-privacy question with real arguments, not slogans
- Ship a small AI-powered tool end to end and defend your design choices
Sessions
What a neural network actually is
A neural network is not a brain and the metaphor has done real damage — it makes people expect understanding where there is only correlation. We build the honest picture instead: numbers in, weights, numbers out, and training as nothing more than repeated correction. No calculus, but no hand-waving either.
Covered
Transformers and attention
Attention is the idea that made modern AI possible, and it is explainable in one sitting: the model decides which earlier words matter for the next one. Once you see that, the context window stops being a marketing number and becomes a constraint you design around. We also cover CNNs briefly — enough to know why images and text are handled differently.
Covered
Running a model on your own machine
Nothing demystifies AI faster than downloading a model and watching your fan spin. You will install open-weights models locally, feel the trade-off between size, speed and quality in your own hands, and stop treating hosted APIs as the only option. This matters here: a local model keeps client data on your disk, which is sometimes the difference between a project being allowed and not.
Covered
Prompting fundamentals
Now that you know what the model reads, prompting stops being superstition. A prompt is a specification: role, task, constraints, format, examples. Most disappointing output is an underspecified request, not a weak model — and the cure is to write a brief a competent stranger could execute.
Covered
Ethics, risk and responsible use
This is usually the slide everyone skips, and it is the one that gets people fired. The real questions are boring and concrete: whose data went into the prompt, who is accountable for the output, and what you tell a client when the model was confidently wrong. Bias is not an abstraction either — a model trained mostly on English text carries assumptions that show up the moment you work in Arabic or French.
Covered
Deploy a real AI tool: pick a task you actually repeat, build a working tool around a model, put it in front of one real user, and write up why you chose local or hosted, where it fails, and what you would not trust it with.
Offered in
Questions
I have never written a line of code. Is this module for me?
Yes, and it is the one place we mean it. Nothing here requires programming: you will install tools, run models and write prompts, but you will never be asked to write a function. What it does require is willingness to open a terminal once or twice and not panic. If you want to build applications afterwards, that is module 03 — this one is what makes module 03 make sense.
Why bother understanding the internals if the tools work anyway?
Because they work until they do not, and then you need to know why. Without the mental model you cannot tell a hallucination from a missing-context problem, so you fix the wrong thing; you cannot judge a vendor’s claim, so you buy on advertising; and you cannot decide whether data may leave your building. Understanding is not academic here — it is what turns you from someone who is sold tools into someone who chooses them.
Do I need an expensive machine to run models locally?
No. A laptop with 8 GB of RAM runs small quantised models fine — slower and less capable than a hosted service, which is exactly the lesson. The point of the exercise is not to replace your assistant with your laptop; it is to feel the trade-off between size, speed and quality with your own hands, so that "which model should we use" becomes a decision you can argue rather than a preference you inherited.
Other modules
Ready to take this module?