Computer Science & AI · Foundation

Using AI Effectively

Almost everybody has now typed something into one of these systems and got back an answer that was either better than expected or quietly wrong, and most people cannot say which kind of task produces which. This course is about telling them apart. It covers what a language model is actually doing when it answers, how to ask for something in a way that makes a good answer more likely, how to check an answer you could not have produced yourself, and the failure modes that are properties of the machinery rather than bugs somebody will fix next year. It is written for somebody who uses these tools for real work, or who is required to and would rather not, and it is against both trusting everything and refusing to look at how any of it works. Whether to use one at all is a decision it describes and does not make. It names no product and recommends none. Every claim about what these systems can and cannot do carries the model and the date, because this is the fastest-moving subject the institute teaches and the course would rather be checkable than current.

What you'll learn

• State what the measured studies show about whether these systems make people more productive, including the two results that point opposite ways, and say what reconciles them
• Describe without mathematics what the system is doing when it answers, and predict from that description alone which of two tasks it will be unreliable at
• Say what the system can see when it answers, why the same question can give different answers, and what a correction does and does not change
• Rewrite a vague request so that it supplies the context, the constraints and the form of the answer, and say which of the three each change is
• Use examples and ask for intermediate steps, and explain why steps that are right do not guarantee an answer that is
• Sort tasks by which side of the jagged frontier they fall on, giving the reason from the mechanism rather than from experience
• Explain in two different ways why these systems produce confident false answers, and say what retrieval does and does not do to the problem
• Check an answer you could not have produced yourself, using a four-step procedure, and decide how much checking the work actually needs
• Say what the evidence shows about using one of these systems while learning something, and decide for one task of your own to stop
• Find out what the product you use does with what you type into it, and decide by category what you are willing to put in
• Sort a claim about AI into measurable, unmeasurable or false, and say why your own impression of whether a tool helped you is not evidence