Ask a model to categorize the same support ticket three times and get three different labels, and the usual conclusion is that the model is unreliable and a better one is coming. Bani Chaudhuri makes the case that this is the wrong diagnosis. Zero shot and few shot are not different models or different versions, they are two ways of asking, and the second one is showing the model examples of what good looks like before you make the request.

The part that carries it is an invented case she walks through, a support agent that drifts. Built on a plain instruction, it is polite but robotic in week one, tells customers to check their email instead of looking up the order in week two, and by week four the CX lead is forwarding screenshots. Same model, same week, three examples added to the prompt, and the tone holds across thousands of conversations. No retraining, no new model. She is careful about the limits too, on how many examples earn their place and why variety beats volume.

What this video covers

• What zero shot and few shot mean, and why they are two ways of asking rather than two models
• Why examples work at all, and what researchers found when the example labels were scrambled
• How many examples to give before you hit diminishing returns, and why three varied ones beat ten similar ones
• An agent drifting week by week on a plain instruction, and the same agent with three examples added
• Why a prompt that shows the model what good looks like outlasts the model it was written for

Chapters

• 0:00 Same question, three different answers
• 0:54 Two ways of asking, not two models
• 1:52 Zero shot and few shot side by side
• 3:04 What happens when the examples are scrambled
• 3:58 How many examples to give
• 4:54 What this looks like in a production agent