What’s in a prompt?
Quite a lot, actually. Large language models (LLMs) might have an impressive (and/or alarming) ability to mimic human art and writing, but if they’re fed garbage, they’ll dump garbage..
One of the first new jobs to emerge from the LLM frenzy is prompt engineer—professionals who refine prompts entered into artificial intelligence/machine learning (AI/ML) tools to improve the quality of responses. Some coverage touted prompt engineering as a hot job that can boast six-figure salaries, while others have questioned whether it’s just a fad skillset, akin to using Google.
Black boxes. LLMs are effectively black boxes—the staggering number of variables in their data sets means it can be difficult even for their designers to debug the models when they act unpredictably.
Hence, the demand for methods to make LLMs do what their users think they’re telling them to do (and cut down on expensive hallucinations). Victor Sanh, lead machine learning scientist at Hugging Face, told IT Brew that research to make deep learning models more transparent remains “very, very experimental” and “not super actionable right now.”
Enter: the prompt engineer, a title that can also be vague, as the roles can have very different responsibilities, from general prompt debugging, to domain-specific analysis, to actually integrating AI technology into software.
Despite the hype, there are also not that many jobs in prompt engineering: Indeed.com has around 200 job listings for the search term “prompt engineer” within 100 miles of San Francisco as of Oct. 24, 2023, while ZipRecruiter has around 30. What’s more, many of those results mention prompt engineering as complementary to other desired skills, such as designing distributed systems or developing AI/ML algorithms.
Trial and error. Ben Stokes is the founder of PromptBase, a “prompt marketplace” for LLMs like OpenAI’s ChatGPT or Midjourney. PromptBase markets itself as helping users “save on time and API costs.”
“For most people, it’s an act of discovery, where it’s a lot of trial and error, where you’re just continually prodding the LLM, tweaking little parts of the sentence to see how that changes the output,” Stokes told IT Brew.
“I think at the extreme level, you’re tracking every single word change and how that changes the output from the prompt,” he added.
Top insights for IT pros
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
One of the key challenges facing would-be prompt engineers is that their data often loses relevance when AI/ML developers update their models. Stokes said PromptBase has mechanisms to flag or hide prompts that no longer work, but outdated prompts are as much of a general issue for LLMs as outdated training data.
“Understanding the query requires knowledge that is super recent,” Sanh said. For example, if an LLM with an older data set makes up responses about current events, “There’s no way to tell them all the ways it’s wrong.”
No one knows where it will go. Sanh said many prompt engineering jobs currently have low technical barriers to entry. That’s because most LLM tooling, including integrations and UI, are still being built in-house and sold by the companies that make the models in the first place.
Though Sanh has theorized prompt engineering may cease to exist in its current form, he also told IT Brew that UI for LLMs could look very different as little as a year from now.
“What’s clear to me is that in the long run, a lot of the companies who are going to rely solely on APIs—doing almost close to no technical work and building internal knowledge and technical expertise—are just going to be dependent on whatever cloud providers and the system providers do,” Sanh said.
Stokes said he doesn’t like the term “prompt engineer,” which he thinks can be offensive to those who view it as appropriating the titles of prestigious careers like software or mechanical engineering. “Then you have prompt engineering, and it’s like you’re just writing English into a chatbot,” Stokes said.
However, Stokes thinks prompts are likely to become more, not less, complex as the use cases for LLMs become even more complicated.
Sanh predicted LLMs will largely become multimodal by default and “natively integrated in a lot of products in our life.” He predicted prompt engineering will eventually become a prerequisite skill in many fields, rather than a discrete profession.
However, that decline may be offset by other types of jobs debugging LLMs, like security, Sanh said.
“We see that with red teaming efforts, [it] also adversarially attack[s] the model to see the limits of safety in the systems,” Sanh added.