Personas, prompt engineering, and Pesach

Image of 4 silhouettes generated by Copilot with Art Deco filter added

Today is Pesach II, which made me think of this connection.

People who work in marketing are familiar with the persona approach to targeted messaging. Personas are now also invoked in fine-tuning responses when crafting prompts for generative AI like ChatGPT.  So the world of tech has finally caught on to what Chazal understood long ago when composing the text of the Haggadah. 

As you recall from a month ago, we don't only have 4 cups of wine and 4 questions of "Ma nishtana," we also have 4 sons who each have their own approach to the question we wish to elicit about the seder. Thes 4 archetypes are the forerunners of personas that are taken into account in crafting tailored messaging.

Because one size does not fit all for education, you don't answer the chacham the same way you answer the rasha, and the tam also gets a different response from the one who doesn't know how to ask. In fact, the last one needs to be given information even without asking for it. 

Let's get a bit meta here and look at the qualitative difference between the question of the wise sone and that of the simple one. The simple one basically asks "What's this?" while the wise one breaks things down to their individual components of edus, chukim, mishpatim. This type of detailed question is what is meant by the expression she'elat chacham chatzi tshuva hu [The question of a wise person is half the answer]. Getting things formulated and organized to zero in on the crux of the issue means the person is clearly focused on the problem and already working his way toward a solution. 

What's the connection between the wise son and prompt engineering?

There are actually two key connections here. One is that you can feed personas into generative AI to get more tailored answers. I haven't tested this out myself, but some put in a qualification for a person or even a well known name to set a standard for knowledge and approach in a response.

 I rather doubt you can ask it how would you as R' Chaim understand this Gemara, nor do I think that you can get Ogilvy quality marketing concepts just by telling it, "You're David Ogilvy and will give me an original slogan for this brand." But that's a game that some recommend to try to point the generative AI in a particular direction.  

Now the second point of connection here is another form of prompt direction to try to obtain more detailed answers, and that is doing exactly what the wise son does in formulating his question. You don't just ask a general question but include directions to break things down.

In the jargon of prompt engineering there are two levels to this. One is called Chain-of-Thought (CoT). It solicits a step-by-step thinking process from the LLM, typically by saying something like "Let's take this step-by-step" to get better results than the simple question that is called naive/standard/Input-Output prompting. Even better than that is what is called Tree-of-Thought (ToT), which allows for more shifts in direction than a straight step-by-step, as it can move laterally, diagonally, and even backwards to track thoughts.


I would suggest that the wise son's classification in his question allows for the ToT approach in thinking things through on the level of when, what, how, and why, which fits perfectly with the answer he gets that is meant to be the final halacha for the seder.

Like and follow on Facebook https://www.facebook.com/KallahMagazine/