Prompt Design (GPT-3): “Step by Step”
A small prompt hack brings wonders
Prompt Design is already a professional field. A mix of creativity and technology. And still, we are at the beginning of exploration.
Aran Komatsuzaki pointed out an interesting Paper Large Language Models are Zero-Shot Reasoners written by researchers from the University of Tokyo, together with Google Research.
In their study, scientists found out that by adding a word sequence “Let’s think step by step” before answering your Prompt by GPT-3 you will get perfect zero-shot completions with higher accuracy than without:
Breaking down the completion into “steps”, Transformer-driven GPT-3 applies self-attention to every of such steps, and that’s why the created argumentation is evermore convincing.
And indeed, let’s try it out with GPT-3 on the obvious question (and by adding the Modifier):
What is the meaning of life?
Let’s think step by step.
The answer we get is simple, wise, and complete. I have to admit, I as a human, cannot add anything more to this one.
It reminds me a little bit of the amusing GPT-2 answer:
Yet in the case of GPT-3 we get everything we need to know about the meaning of life.
We can consider Artificial Intelligence highly intelligent.