prompting is essential. good examples are collected here:

from a more systematic prompt engineer vedio:

code generation sample 2

GPT-4 compared to chatGPT is particularly better at reasoning, the prompt sample is

Next, moving to more advanced prompt engineering:

in context learning

CoT, chain of thought prompting

Zero-shot prompting involve adding “let’s think step by step”.

Self-consistency has an interesting example on age reasoning:

After applying smart/self-consistency prompting, the answer can be improved:

Generate knowledge prompting:

program-aided language model such as chain of thought prompting is to steer models to perform better at complex reasoning tasks, it can be augmented by program-aided Language models to generate intermediate reasoning steps.

Further combining: ReAct

prompt include demoing how to solve problem such as an extreme case like below:

For real applications which would explode in near future, data-augmented generation would be it!