Here’s a six-figure job that didn’t exist a year ago: prompt engineer
Lesson summary
Hi there everyone, I’m Jeff and this is Plain English, where we help you upgrade your English with stories about current events and trending topics. This is lesson number 645, so that means JR has uploaded the full transcript, exercises, and all the lesson resources to PlainEnglish.com/645.
Here’s what we’re talking about on today’s story: It’s been about 15 months since ChatGPT burst onto the scene. And in that time, many individuals have worked generative AI into their lives, for fun and to work more efficiently. But companies are working ChatGPT and other language models into their processes…and that means there’s a new job out there.
It’s called the prompt engineer. And a prompt engineer is someone who knows how to ask language models a question and get a useful answer. This may sound simple. But in fact, there’s an art and a science to generating useful responses from AI language models.
In the second half of the lesson, I’ll show you what it means to “scratch the surface” of a topic. Ready? Let’s dive in.
Prompt engineer: the new job for the AI generation
You’ve probably heard that artificial intelligence, and large language models like ChatGPT, are disrupting the business world.
Certainly individuals in businesses use ChatGPT to simplify their lives at work. But many businesses are using these models in a much more advanced way. OpenAI, the company behind ChatGPT, allows companies to connect directly to the underlying GPT model. That way, they can use GPT directly with their own data, their own documents, and their own software, all without opening a ChatGPT window.
But this doesn’t happen automatically. Someone still has to analyze the business needs, analyze the business data, make the connections, and make it all work with the model. And that is the job of a prompt engineer. (A “prompt” is the set of instructions you give a language model.)
A prompt engineer is a person who is the glue that connects the company and the language model. It’s that person’s job to program instructions into GPT that are most likely to get the best response. Most prompt engineers have to use a little bit of coding, but some prompt engineers can get away with using just English.
The job requires an analytical mind, knowledge about how language models work, an ability to understand a business’s objectives…and a lot of patience.
Let’s look at a few things a prompt engineer might do. Imagine there’s a company full of financial analysts who need to read the financial disclosure forms that public companies file. These documents are large and dense; an AI model can help digest them.
So a single analyst might open ChatGPT, paste the document into the model, and ask questions about it. But this is inefficient. Not every company employee is going to be good at using ChatGPT.
It’s better, instead, for a prompt engineer to write the exact questions and prompts so that the other employees can get the best answers from the model. And it’s even better if the prompt engineer can make a nice user interface. That way, employees can simply paste a link, click a button, and the model works in the background—all thanks to the work the prompt engineer did.
But now imagine those financial analysts have to do specific analyses on these documents. They can open a private company library full of prompts that can help them do analyses. The prompts have been developed, tested, and saved by a prompt engineer, so that each analyst doesn’t have to think of a prompt, and then test it, in the moment.
This is just scratching the surface of what prompt engineers do. They also draw on a variety of research-based techniques when they create prompts for the model. These techniques improve the accuracy of the language model responses. Here are a few of them.
One is called “chain-of-thought.” Chain of thought breaks complex problems down into their component parts. If you ask GPT to do a complex analysis all in one paragraph, you might not get the right answer. But a skilled prompt engineer can break a complex request up into component parts, walk GPT through the logic to solve the problem, and ask the model to solve the problem in steps, rather than all at once. This technique has been shown to improve the accuracy of GPT responses.
Another technique is to have GPT do the same thing multiple times…and then choose the best outcome. Here’s an example with translation. I asked ChatGPT to translate a phrase from a recent Plain English story into Spanish. The sentence was, “A big, stinking pile of trash .” Sometimes, ChatGPT translated “stinking” as “maloliente.” Other times, “apestoso.” Both are acceptable, but they’re not exactly the same. How can I know which one I should choose?
A prompt engineer could ask GPT to translate that word five times, and then analyze its five different translations. It could then ask GPT to select the winner, to identify the translation that most often appears in its five previous attempts. This can help a business get more consistent results from a model.
Here’s one more technique. A prompt engineer can train a model to identify patterns that are likely to appear in the data. This is called “few-shot prompting.” This technique demonstrates logic to the model with sample or nonsense data. The model can then use the same logical patterns to answer questions about a business’s data.
Many ChatGPT users have found that the model can sometimes produce offensive or biased content. So a good prompt engineer can put guardrails around responses to protect against inappropriate responses.
I could go on and on : prompting language models is an emerging discipline. And if you’re skilled at it, you can make good money. I searched LinkedIn and other job sites for “prompt engineer” jobs. Experienced candidates with coding knowledge can earn $200,000 or more at tech companies. But even non-technology companies need these skills, and those jobs still offer six-figure salaries.
Here’s a question: will GPT ever get so good that prompt engineering is obsolete? Anything is possible. But for now, the prompt engineer is the critical link between a company’s humans and its models.
If you’ve ever used ChatGPT and thought, “This is nice, this is fun, but I can’t use it.” Or if you thought, “This is just a toy,” then you need a prompt engineer.
The prompt engineer’s job is to make it useful. I’ve spent hours just re-entering prompts and data and trying to get things right—here to make myself more efficient at Plain English.
And then I discovered some of these prompt engineering techniques—they’ve been developed and tested by researchers in the field, published in papers. And by applying these techniques, I’m able to solve problems that I could never get GPT to solve before. And now I can.
If you’ve ever been on the web and seen software or products that say they use AI to help you do X, well, a prompt engineer is probably behind that.
It’s really amazing. If I didn’t already have a great job here at Plain English, I might want to become a prompt engineer. Think about it!
Great stories make learning English fun
