Friday, March 10, 2023

A new project role: Prompt Engineer

I've mentioned this idea before: a new role coming to projects is that of "prompt engineer".
And so what is the job description?
  • Able to write imaginative queries (prompts) in natural language for an AI engine
  • Able to understand their flaws, supercharge their strengths and game out complex strategies to turn simple inputs into results that are truly unique (according to one published description)
  • Able to work with other humans to refine their queries
  • Able to validate responses from the AI engine
  • Able to experiment, almost without limit, to discover and refine
Here's the rub: 
Whereas computers execute software code precisely and with consistent repeatability, that's not what you get when you query today's chat engines built on "large language models"(*). In my personal experience, the exact same query, executed multiple times, gets you somewhat different answers every time.

Simon Willison, a British programmer who has studied prompt engineering, is quoted in the Washington Post as saying: “I’ve been a software engineer for 20 years, and it’s always been the same: You write code, and the computer does exactly what you tell it to do. With prompting, you get none of that. The people who built the language models can’t even tell you what it’s going to do.”

And, "prompt injection" may be a prompt engineer's tool to tease out performance, or it's a malicious hacking tool aimed at the AI engine. 
Wikipedia says this: "Prompt injection can be viewed as a code injection attack using adversarial prompt engineering. In 2022, the NCC Group has characterized prompt injection as a new class of vulnerability of AI/ML systems also."

Then there are the art and image creators, built somewhat differently than the large language models.

Prompts with these image engines don't have to be the formal language constructs of the large language models. And indeed, sometimes just a grab-bag of words will get you amazing images. 
The prompt engineer morphs into the prompt artist.

For the prompt artist, there is a business there: Some artists will sell you proprietary prompts for favorite images.

Beyond the normal domain
The Washington Post reports: The role is also finding a new niche in companies beyond the tech industry. Boston Children’s Hospital this month started hiring for an “AI prompt engineer” to help write scripts for analyzing health-care data from research studies and clinical practice. 

The law firm Mishcon de Reya is hiring for a “legal prompt engineer” in London to design prompts that could inform its legal work; applicants are asked to submit screenshots of their dialogue with ChatGPT.

(*) Large language models: A language model is a probability distribution over sequences of words. Given any sequence of words of length m, a language model assigns a probability to the whole sequence. More simply, a language model predicts the next word, or the next sentence even, from the probability that such as been seen before in the training data. A "large language model" is one that uses a very large data set for training. And the definition of "large" is getting larger with time and experience. 

Like this blog? You'll like my books also! Buy them at any online book retailer!