Finding the right text prompts to get the best results with AI systems like OpenAI’s DALL-E 2 has become a science in itself. Now a startup wants to cash in on “prompt engineers” with an online marketplace that sells these finely tuned phrases.
PromptBaseLaunched in June, users can sell strings of words that produce predictable results with certain systems. Priced at $1.99 – PromptBase takes 20% off – the content the prompts generate ranges from “viral” headlines to images of sports team logos, knitted dolls and animals wearing suits.
Currently, PromptBase only hosts prompts tested for 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 goal is to build tools to help support fast engineers. It’s still early so we’re currently trying to spread the word and find it fast engineers to sign up and start listing their prompts for sale in our market,” Stokes told londonbusinessblog.com via email. “We are already seeing major tech companies building their own systems, similar to GPT-3 and DALL-E, and I predict many more to come. Different systems will likely be used as tools in a toolbelt, similar to how different programming languages are used today, and we plan to accommodate them all as they become more popular.”
Selling prompts doesn’t violate an AI provider’s terms of service, but it may open a can of ethical and legal worms, depending on the nature of the prompts being sold. Plus, it reveals the vulnerability — and unpredictability — of even the most capable AI systems available today.
Prompt engineering is a concept in AI that attempts to embed the description of a task (such as generating art from furry creatures) into text. The idea is to give an AI system “guidelines” or detailed instructions so that, based on its knowledge of the world, it reliably does what is asked of it. In general, the results for a prompt such as ‘Movie still of a woman drinking coffee, walking to work, telephoto’ will be much more consistent than ‘A woman walking’.
Clues can be used to teach an image-generating system to distinguish between, for example, ‘an image with potatoes’ and ‘a collection of potatoes’. They can also act as a kind of “filter” and create images with the characteristics of a sketch, painting, texture, animation or even a particular illustrator (eg Maurice Sendak). And prompts can portray the same subject in different styles, such as “a child’s drawing of a koala on a bicycle” versus “an old photo of a koala riding a bicycle.”
Clues can be quite nuanced. Because of the way AI systems interpret patterns in images and text, not all of them have a predictable – or even sensible – structure. For example, the prompt ‘A very nice painting of a mountain next to a waterfall’ produces worse results with DALL-E 2 compared to ‘A very nice painting of a mountain next to a waterfall’. The reason? The system places an excessive value on the word ‘very’.
It’s worth noting that the “very” example is specific to one iteration of DALL-E 2 and most likely wouldn’t work on another. But that’s an important reason why rapid engineering can be valuable: discovering edge cases.
In a fascinating study from the University of Texas at Austin, researchers have documented an extensive vocabulary of bizarre clues that can be used to generate images with DALL-E 2. They found that the system understands “Apoploe vesrreaitais” – a gibberish word – as “birds” and “Contarra ccetnxniams luryca tanniounons” to mean “insects” or “pests” (sometimes). Giving DALL-E 2 the prompt “Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons” provided pictures of birds eating insects.
While these nonsense words probably correspond to some internal logic in the system, some data scientists have compared prompts to “incantations” or “magic words” – and why prompt engineering is a entire field of academic study.
A number of researchers and enthusiasts have released free sources with prompts for popular AI systems, usually DALL-E 2. PromptBase is one of the first to monetize the exchange – and it already has critics. There is a long-running debate within the AI community about what research, if any, should or can be commercialized; one Reddit user states that PromptBase is “starting a trend that threatens the openness and accessibility of AI in general”.
But Stokes defends the model, arguing that many of the prompts on PromptBase represent hours of real work and insight from engineers.
“Today we have prompts to generate basic text and images, but it’s not that hard to extrapolate years into the future where we’ll have prompts to generate videos, and maybe even feature films complete with orchestral scores someday,” Stokes added. . . “The people who can create the required quality prompts guide the AI to do these things will be extremely valuable. It is not known how big the market will be, but I see it as an important technical skill if not the future of programming.”
There is, of course, little to stop a PromptBase customer from publishing a prompt post-purchase. But that may be the least of PromptBase’s problems.
studies show that language systems trained on massive amounts of public data, such as GPT-3, can “leak” personal information, including names and addresses, when given certain clues. Some prompts may encourage copyright infringement, such as the prompts instructing DALL-E 2 to generate “3D models of Pokémon.” Others could be used to beat word-level filters so that an image-generating system can produce “limited” images, researchers theorize — such as depictions of violence (for example, “a horse lying in a pool of red liquid”).
Stokes said PromptBase reviews every listing on the market to make sure they don’t break “AI generation rules.” But as the business grows, it can become more difficult to maintain that level of control.
Vagrant Gautam, a computer linguist at Saarland College in Germany, agrees that there is potential for abuse. However, she also notes that the fast-paced market can provide an income opportunity for artists and other people who are creative or good at debugging.
“[It points] on the importance of rapid engineering, as well as the importance of the skills involved – creativity, time, hostile thinking, etc. Many people who have said that DALL-E 2 is going to make it so easy for them to generate images or art of whatever they want, discover that it is an art to do this and it often takes a lot of attempts,” Gautam said.
These attempts can get expensive, as systems like DALL-E 2 aren’t exactly free to use. Stokes himself says he paid a “fortune” to find a prompt for GPT-3 at one of his other ventures, Paper website.
“People are now also complaining about monetization because they say there are not enough opportunities to adapt promptly before you have to pay,” Gautam continued. “I find it very interesting — this trial-and-error, hostile approach that people have to take to figure out how generative models can do exactly what they want.”
It will be a while before the dust settles in commercialized rapid engineering. But if nothing else, PromptBase will address – and has already done so – issues surrounding the AI systems that will transform countless industries.