Prompt Whispering v Prompt Engineering

Tricks v Understanding

Ever had a plumber, electrician, builder come into your house? Ever taken your car to a mechanic? Ever noticed that some of them (and I’m excluding the hucksters) go through a methodical set of steps, professionally and going from the starting point you suggested. And then there is the guy you find who you go back to all the time. The one who as you drive the car into the garage is already picking up a tool and just tells you to “pop the hood”, the electrician who walks in, takes one look at a light fitting, sighs, and informs you that the wiring was put in sometime in 1920, so the guy who made the patch-changes last time has actually given you a bit of a problem.

The challenge sometimes with these folks, the ones who really understand their jobs, who’ve through experience developed innate feelings as to what is wrong and how to fix it, is that sometimes it is hard to tell them apart from the con-artist who starts with “Oooh, last guy here was a real cowboy, you’ll need to start again”. But when you find one, you know it, because the outcome is so much better.

We’re about to see that with Generative AI.

Prompt Engineering is knowing the Tricks

There are lots of “prompt engineering” courses out there, and universally they are about teaching you “tricks” on getting better outputs, like adding “dramatic lighting” to Midjourney to get better images, or forming sentences in particular ways for GPT.

So the goal there is to help people be more efficient by giving them some tools that they can repeatedly apply. It certainly has value.

Prompt Whispering is knowing the Model

What I’ve noticed though is that some of the deeper NLP people I know aren’t using tricks. They’re decomposing problems into a mental construction that they’ve created around the model. Instead of a set of repeating tools and techniques, they deconstruct problems and then ask specific elements for that problem.

They’re thinking about the model, they’re understanding how the problem applies to that model, and then constructing just the right information to put into the model.

Whisperers will out perform Engineers by >10x

That the great mechanic out-performs the average, that the great developer out-performs the average, none of these are a surprise. Consistently people who can “think like” or “think within” the system they are significantly better performers than those that cannot. With Generative AIs we will just seen another set of people who have skills that differentiate themselves, people who approach problems differently, by understanding the context, understanding the problem and understanding the model that is being tasked with resolving it.

Many people will struggle to be considered Prompt Engineers, unable to accurately wield the “tricks” to optimize their approaches they will brute force their way forwards. But even good engineers will defer to people who are able to “think” like models. In some areas we might find that Whisperers become a significant competitive advantage, able to make changes at a pace that Prompt Engineering companies cannot achieve.

Being clear, thinking like a model doesn’t mean understanding fundamentally how it works. It means understanding the system and how it works. Understanding how it flows and what influences it.