Figuring out the right text prompts to yield the best results with AI systems like OpenAI’s DALL-E 2 has become a science in its own right. Now a startup is looking to let “prompt engineers” cash in with an online marketplace that sells these finely tuned phrases.
PromptBase, launched in June, allows users to sell strings of words that net predictable results with particular systems. Priced at $1.99 — PromptBase takes a 20% cut — the content that the prompts generate range from “viral” headlines to pictures of sports team logos, knitted dolls and animals wearing suits.
At the moment, PromptBase hosts only prompts tested on DALL-E 2 and GPT-3. But according to its founder, Ben Stokes, the plan is to expand the platform to additional systems in the future.
“Our ultimate aim is to build tools in order to help support prompt engineers. It’s early days, so we’re currently just trying to spread the word and find prompt engineers to sign up and start listing their prompts for sale on our marketplace,” Stokes told TechCrunch via email. “We’re already seeing big tech companies build their own systems similar to GPT-3 and DALL-E, and I predict many more to come. Different systems will likely be utilized like tools in a toolbelt, similar to how different programming languages are used today, and we plan to accommodate all of them as they gain popularity.”
Selling prompts isn’t against any AI provider’s terms of service, but it potentially opens a can of ethical and legal worms depending on the nature of the prompts being sold. Moreover, it reveals the fragility — and unpredictability — of even the most capable AI systems available today.
Prompt engineering is a concept in AI that looks to embed the description of a task (like generating art of furry creatures) in text. The idea is to provide an AI system “guidelines” or detailed instructions so that it, drawing on its knowledge of the world, reliably accomplishes the thing being asked of it. In general, the results for a prompt like “Film still of a woman drinking coffee, walking to work, telephoto” will be much more consistent than “A woman walking.”
Prompts can be used to teach an image-generating system to distinguish between “an image containing potatoes” and “a collection of potatoes,” for example. They can also act as “filters” of sorts, creating images with the characteristics of a sketch, painting, texture, animation or even a particular illustrator (e.g., Maurice Sendak). And prompts can portray the same subject in different styles, like “a child’s drawing of a koala riding a bike” versus “an old photograph of a koala riding a bike.”
Prompts can be quite nuanced. Owing to the way AI systems make sense of patterns in images and text, not all of them have a predictable — or even sensible — structure. For example, the prompt “A very beautiful painting of a mountain next to a waterfall” returns worse results with DALL-E 2 compared to “A very very very beautiful painting of a mountain next to a waterfall.” The reason? The system attaches an inordinately high value to the word “very.”
It’s worth noting that the “very” example is specific to a particular iteration of DALL-E 2 and most likely wouldn’t work on another. But that’s a major reason prompt engineering can be valuable: discovering edge cases.