With AI systems like OpenAI’s DALL-E 2, knowing the right text to get the best results has become a science in itself. Now a startup is looking to monetize “quick engineers” with an online marketplace that sells these finely tuned phrases.
PromptBase launched in June and allows users to trade predictable results with specific systems. Priced at $1.99 – PromptBase takes 20% off – the prompts generate content ranging from “viral” headlines to sports team logos, plush toys and costumed animals.
Currently, PromptBase only supports prompts tested on DALL-E 2 and GPT-3. But according to founder Ben Stokes, the plan is to expand the platform to more systems in the future.
“Our ultimate goal is to build tools to help support. asked a question Engineers. It’s early days, so we’re currently trying to spread the word and get it out there. asked a question Engineers begin to register and list It suggests It will be sold in our marketplace,” Stokes told TechCrunch in an email. “We’re seeing big tech companies build systems similar to GPT-3 and DALL-E, and I predict more to come. Different systems can be used like tools in a tool belt, similar to how different programming languages are used today, and we plan to accommodate them all as they become popular.
Selling queries isn’t against any AI vendor’s terms of service, but it opens up a can of ethical and legal worms depending on the nature of the queries being sold. What’s more, it reveals the fragility — and unpredictability — of the most powerful AI systems available today.
Rapid engineering is a concept in AI that looks like embedding a description of a task (like generating art for furry creatures) into text. The idea is to give the AI system “instructions” or detailed instructions, so that it can reliably perform what is asked of it, drawing on its knowledge of the world. Generally, a quick result like “film of a woman still drinking coffee, on her way to work, on the phone” will be more consistent than “a woman walking away.”
Questions can be used to teach an image generation system to distinguish between an “image containing a potato” and a “bunch of potatoes”. They can also act as “filters”, creating images with design, drawing, texture, animation or specific expressive (eg Morris Sendak) features. And prompts can show the same subject in different styles, such as “picture of a baby koala riding a bicycle” and “old photo of a koala riding a bicycle.”
Needs can be very complex. Because of the way AI systems make sense of patterns in images and text, not all of them have predictable — or logical — structure. For example, the question “The mountain next to the waterfall is the most beautiful picture” scored worse on the DALL-E 2 than “The mountain next to the waterfall is the most beautiful picture.” the reason? The system has an excessively high value with the word “very”.
It should be noted that the “very” example is specific to a particular iteration of DALL-E 2 and may not work on another. But that’s the main reason rapid engineering can be useful: finding edge cases.
In a groundbreaking study at the University of Texas at Austin, researchers reported a wide vocabulary of surprising questions used to generate images with the DALL-E 2. The system “apoploi vesreitais” – a gibberish phrase – to mean “birds” and “Contarra ccetnxniams luryca tanniounons” meaning “bugs” or “pests” (sometimes). Asking DALL-E 2 “Apoploe vesrreaitais eat Contarra ccetnxniams luryca tanniounons” produced pictures of birds eating bugs.
Although these nonsense words relate to some internal logic in the system, that’s why some data scientists liken the questions to “magic” or “magic words”—and why rapid engineering has spawned an entire field of academic study.
Several researchers and enthusiasts have released free resources containing prompts for popular AI systems, mostly DALL-E 2. PromptBase was one of the first to monetize the exchange — and it already has its critics. There is a long-standing debate in the AI community about which research, if any, can be commercialized or transferred. One Reddit user, PromptBase, argued that “it’s starting a trend that threatens the openness and accessibility of AI in general.”
But Stokes defends the model by saying that most of the questions on PromptBase represent actual work and understanding by engineers.
“Today we have the needs to generate basic text and images, but it’s not too hard to envision years into the future where we’ll be making requests to make videos and maybe one day feature-length films complete with orchestral scores,” Stokes added. . “People who can make the quality incentives needed will guide the AI to do those things. It’s unknown how big the market will be, but I can see it being a key technology skill, if not the future of programming.
In fact, there is little to prevent a PromptBase client from publishing a prompt post purchase. But that may be the least of PromptBase’s problems.
Studies have shown that language systems trained on vast public data, such as GPT-3, can “leak” personal information, including names and addresses, when fed certain queries. Some requests may encourage copyright infringement, such as instructing DALL-E 2 to generate “3D models of Pokémon”. In order to extract other “restricted” images, an image generation system can be used to overcome word-level filters, researchers theorize – such as violent images (for example, “a horse lying in a pool of red liquid”).
PromptBase reviews every listing on the marketplace to make sure they don’t violate any of the “AI generation rules,” Stokes said. But if the business grows, it can be difficult to maintain that level of scrutiny.
Vagrant Gautam, a mathematician at Saarland College in Germany, agrees that there is potential for abuse. But, she says, the fast-paced marketplace creates an income opportunity for artists and other people with creative or editing skills.
“[It points] The importance of rapid engineering, as well as the skills of the sector to do so – creativity, timing, adversarial thinking, etc. A lot of people have been saying that DALL-E 2 makes this easier. They are discovering that there is an art to doing this to generate the image or art they want and it often takes multiple attempts,” Gautam said.
Since systems like DALL-E 2 are not exactly free to use, these tests can be expensive. Stokes himself said on his other entrepreneurial paper’s website that he paid a “fortune” to figure out the GPT-3 question.
“People are now complaining about the monetization because they say there are very few opportunities to fix your query before you start paying,” continues Gautam. “I find it really interesting — this trial-and-error, adversarial approach that people have to take to figure out exactly how to motivate generative models to do what they want.”
It will take some time before the dust settles in commercialized rapid engineering. But if nothing else, PromptBase will raise — and already has raised — issues around AI systems that will transform countless industries.