You get a number back, or a verdict, or a recommendation, with no working shown, and you either trust it or you do not. Bani Chaudhuri covers the simplest fix for that. Ask the model to think through the problem step by step, and it will show its reasoning instead of jumping to the conclusion.

The explanation of why it works is the good part. A model predicts what comes next based on patterns it has seen, and it has seen far more examples of humans reasoning through problems than it has of bare answers, so it predicts reasoning better than it predicts conclusions. Each step it writes constrains the next one. She runs a discount and tax calculation with and without the instruction to show that, then moves to what it changes in a support agent, where reasoning through the policy first produces an answer the customer can act on rather than one that generates six follow-up questions.

What this video covers

• Why a model predicts reasoning more reliably than it predicts answers
• The same question asked with and without a request for steps, worked through
• When step by step reasoning earns its tokens and when it is just cost
• What a support answer looks like when the agent reasons through the policy before replying
• Why this makes mistakes visible rather than impossible, and how to combine it with examples

Chapters

• 0:00 An answer with no working shown
• 0:38 What chain of thought is
• 1:53 The same question with and without steps
• 3:06 When to use it and when to skip it
• 4:19 What it changes in a support agent
• 5:29 Combining it with examples, and where it still fails