Welcome

This course starts at the end: with the report that an AI agent delivered to a company's management. It is well written, has figures and sounds confident. Over eight blocks you will learn what it takes to know whether it is true, and at the end you will have to decide whether you would sign it. Download the case data in the second lesson of the first block and have an AI assistant to hand: here you practise with it, and you check it.

What you will find

• Block 1. The map: from asking questions of data to commissioning an agent. Which three questions are put to data, what changes when the model comes already trained and an agent carries out part of the work, which roles there are in a data team and how we practise here. It introduces the case that runs through the whole course: the report an agent wrote about falling sales.
• Block 2. Data: the raw material. What shape a data point has, how to summarise it without fooling yourself, and where data comes from and whose it is. First reading of the agent's report: which table it used, which summaries it gave and which days are missing.
• Block 3. Uncertainty: probability so you do not believe everything. Spreading belief across what can happen, turning a probability around with Bayes, and telling a real relationship from one manufactured by chance or by a third variable. Second reading of the report: what it claims and what the data allow it to claim.
• Block 4. Learning from data: what training means. What a model is and what training it means: a formula with knobs that are moved to make fewer mistakes. From the line to the neuron and to layered networks, the three ways of learning, the map of problems and techniques and the four-part X-ray. Third reading of the agent's report: the model it fitted.
• Block 5. Generalising and evaluating: how to know whether it works. How to know whether a model or an assistant really works: why memorising is not learning, how to keep an exam that is not shown beforehand, which two errors a yes-or-no system makes and who pays for each, and five questions for reading any evaluation. Fourth reading of the report: how the agent evaluated its own model.
• Block 6. Language models: a text-predicting machine. How a text becomes numbers, how a language model writes word by word, and why it makes things up in the same tone in which it gets things right. Fifth reading of the agent's report: where a figure comes from that no query supports.
• Block 7. Agents: from answering to acting. What turns a model into an agent, what it needs to know about your problem and how to read what it has done. Sixth reading of the report: the trace, step by step, and the five errors it hid.
• Block 8. Person and machine: dividing, supervising and answering. What is delegated to an agent and what is not, how to supervise without dozing off or losing the craft, and who answers for what is signed. Last reading of the agent's report: what is sent back, what is corrected and whether you would sign it.

How you will work

Each lesson opens with a short introduction and, below it, several pills: short pieces that explain a single idea and end with a case or something to do. Read them in order the first time; afterwards you can go back to the one you need.

Where you see an interactive practice, do it before reading on: it is there for you to decide, calculate or classify, and the explanation comes afterwards. It does not save what you write, so copy whatever you want to keep.

Each block ends with an exam of six questions. All the answers come with a comment, including the wrong ones: failing and reading why is part of the method.

At The Singular Academy you learn by solving real cases. If at the end of a lesson you do not know what you would do differently tomorrow, go back over it.