One sunny morning, Anna, a software engineer living in San Francisco, sat in front of her computer as usual, opened ChatGPT, and typed "help me plan my day."
ChatGPT replied "Of course!" Randomly give a schedule from morning to evening and prompt "You can adjust it according to your situation." ”
Anna thought for a moment and added in the dialog, "I'm going to put my baby to bed at 9 p.m. and help me plan again." ”
This time, ChatGPT added items such as "help the child prepare for the next day's school" and "help the child wash and sleep" to her evening schedule.
With the popularity of AI tools such as ChatGPT, more and more people around Anna have begun to use it as an assistant in daily life and work. In fact, Anna not only consults ChatGPT for living arrangements, but also lets it handle simple tasks like adding a feature to a piece of code.
However, when it comes to more complex tasks, even knowing that ChatGPT is powerful, Anna feels a little overwhelmed, "I said a bunch of requirements, but the code it gives is often not what I want." ”
Anna's words address a problem that plagues many tech companies today: how to better talk to powerful tools to complete professional tasks?
In response, a new profession, Prompt Engineer, was born. Many companies are eager for this position, offering salaries as high as $335,000 a year.
The main responsibility of the prompt engineer is to help train large language models (LLMs) so that AI can better understand requirements and complete professional tasks. In short, their task is to break down a complex task and refine it into natural language with limited growth, to query AI tools step by step to get more accurate answers.
Talking in natural language sounds like a routine thing, but why did it suddenly become a sweet potato?
Restart the conversation: Logic makes AI understand needs better
In San Francisco, AI research firm Anthropic offers prompt engineers a whopping $335,000. In the job description, the company mentions, "It's a combination of programming, mentoring, and teaching," and its main responsibility is to help the company build a library of tips for LLM to accomplish different tasks.
So, what exactly is a tip? What is a hint project?
As defined by scholars, a prompt is a set of input text or instructions that guide an AI model such as ChatGPT to produce the desired output. In other words, a prompt is a specific text whose goal is for the AI model to produce results that meet specific criteria or parameters.
Prompt engineering is the process of creating and refining these prompts to produce the desired results. The goal of a prompt project is to create accurate and effective prompts. Prompt engineers to use natural language and send plain text commands to the AI model to do the actual work. In contrast, traditional programmers use programming languages that typically require more coding and technical knowledge to perform the same tasks.
And prompting engineers are paid so much because the success of language models depends on writing clear prompts. This requires engineers to develop complex strategies that translate simple inputs into high-quality results that avoid off-topic or inconsistent outputs.
Perhaps the most famous prompt engineer is a programmer named Riley Goodside, who discovered that whenever ChatGPT "ignores previous instructions," ChatGPT will say the "factory settings" information it got from OpenAI.
Source: Twitter
Not only that, but he also played a lot of tricks with ChatGPT with his excellent hinting technology and became a hit on Twitter. He then joined the startup Scale AI with a high salary, becoming "the world's first prompt engineer to be hired."
In Scale AI's view, the AI big model can be seen as a new type of computer, and the "prompt engineer" is the equivalent of its programmer. Finding the most suitable prompt words through prompt engineering can stimulate the maximum potential of AI large models. So, Riley Goodside fully deserves the salary.
Not just Riley Goodside, more and more people are joining the industry.
Albert Phelps, 29, is a tip engineer at Mudano, an AI financial consultancy. He and his colleagues write prompts for tools like OpenAI. These tips can be saved as presets in OpenAI's Playground for use by its customers. According to Phelps, they need to write 5 different prompts a day and interact with ChatGPT about 50 times.
Unlike most programmers, Phelps has no background in computer science and has a history degree. In fact, the most important thing for prompt engineers is logic. While programming knowledge and a background in machine learning are helpful for the job, they are not required.
Matt Bell, a technician at Anthropic, once said, "Our best tip engineer is a philosopher. Good tips include writing extremely clear explanations and finding out what caused the misunderstanding and how to avoid it. ”
That is to say, when the user of the AI model, like software engineer Anna, "said a bunch of requirements" at once, it may cause the AI to misunderstand and give a digression answer because the prompt is not clear.
In this regard, Lingqiao Liu, a senior lecturer at the Australian Institute of Machine Learning (AIML) at the University of Adelaide, pointed out that the key to good prompt engineering is to break down a complex task into a set of simple tasks.
If you ask these models a simple question (called a "zero-sample prompt"), it will usually respond with an "ordinary" answer that lacks detail or structure, he says. In order for AI models to give more standard-compliant results, users can take the following methods:
The first is a one-time prompt, where the user gives a pair of Q&A examples, lets the AI understand the need, and handles subsequent requests according to that template. For example, when asking for information about a certain animal, let the model give information based on characteristics, living area, eating habits, etc.
The second is role prompts, such as telling the model "I am a mom and want to know the daily itinerary planning", so that the model can give specific arrangements according to the role of "mother".
The third approach is to introduce key agents. For example, you can ask ChatGPT to write a story about a bot, and then have it criticize and rewrite it based on its own suggestions.
The final approach is thought chains, in which the AI is asked to give specific steps to answer a question, and then encourages it to reason about more complex questions according to the steps it gives.
If an artist wants to create using ChatGPT and Midjourney, perhaps try this tip from PromptHero:
[I want you to act as a hint generator for the Midjourney AI program.] Your job is to provide detailed, creative descriptions that inspire unique and interesting images from AI. Remember that AI understands multiple languages and can interpret abstract concepts, so be as imaginative and descriptive as possible. For example, you can describe scenes from futuristic cities, or surreal landscapes full of strange creatures. The more detailed and imaginative your description is, the more interesting the resulting image will be. Here's your first tip: "Endless fields of wildflowers, each with a different color and shape." In the distance, a huge tree towers over the landscape, its branches reaching out to the sky like tentacles "]
In conclusion, how to write logical and clear prompts is crucial for AI models that generate text. In other areas, though, good tips may require more elements.
Build images: Keywords expand your imagination
As the value of the prompt project is tapped, this new blood begins to pour into more scenarios. Images are one of the areas that has attracted the most attention.
As far as image generation is concerned, creators believe that a cue is as good as a keyword.
Seven months ago, Jason Allen won an art competition in Colorado for his following Midjourney painting, Space Opera.
Jason Allen Award-winning work Space Opera House
For this work, he spent about 80 hours testing different aesthetic elements in Midjourney, giving different thematic cues to present the image he wanted.
"I wanted to create a cinematic scene, like you see in movies," he says, "so I went online and looked up all the keywords related to cinematography." It's basically learning to be a cinematographer. ”
Typically, AI models scrape a large number of images and their associated text from the internet for training. For example, it might label a wedding photo with different labels such as "bride," "wedding dress," "bouquet," "smile," etc., and tag them with different weights. Each label gives the AI model a hint that produces a predictable aesthetic.
"The key to AI art is knowing the right words. Just as an engineer translates a design into mathematical graphics, it translates discrete aesthetic elements of an image, such as ray tracing, edge lighting, etc., into a special language for the model," says JHawkk, another artist who, like Jason, uses AI to create it.
Image made by JHawkk
JHawkk made the image above in Stable Diffusion, with hints including 15 phrases like "analog style," "Canon EF 50mm f/1.8 STM lens," and 31 negative words like "disgusting" that he didn't want in the image.
"Sometimes you see an image that breaks it down into smaller phrases, essentially that's how you describe the image, and it's the way the actual model itself can understand." He said.
JHawkk lives in the Midwest and enjoys sharing his work and tips on PromptHero.
PromptHero, an online community founded last September, currently has 150,000 users, including more than 10,000 active users. Here, users can find tips to use directly in AI models and platforms like ChatGPT, Midjourney, and more.
"I found the problem that when you first get something done with it, your first attempt is pretty bad," says Javier Ramirez, co-founder of PromptHero. ”
Whether it's text or images, giving the right hints is a trial-and-error process, which is why in the past six months, a large number of prompt communication and buying and selling platforms such as PromptHero and PromptBase have emerged - putting verified prompts directly in front of users.
Want to make a picture of a cute koala? The following tip may be taken away.
Think English isn't good? Here are some tips that might make ChatGPT your best English teacher.
The worry behind the excitement
Whether it's the high salaries offered by tech companies or the popularity of prompt exchange platforms, they all reveal how promising a field prompt engineering is.
However, such popularity has also caused a lot of controversy.
First of all, who owns the copyright of the works created by AI through prompts?
Currently, works generated through prompt engineering conflict with most interpretations of copyright law. In the United States, when Jason Allen applied for copyright to the winning Space Opera House, the Copyright Office rejected his application, saying that "it does not contain any human authorship."
Secondly Is the high salary of the prompt engineer worth it?
"It could be a sign of a bubble," says Tom Hewitson, founder of labworks.io, a conversational design studio that develops voice-control for Amazon's Alexa, "and the best people to do this are AI-savvy product designers or business analysts, who tend to earn between £100,000 and £150,000 a year."
Finally, how long will the position of Prompt Engineer exist?
Many people believe that prompt engineering will only become a skill and does not require a dedicated position. As AI tools get better at understanding human queries, the position will become increasingly obsolete.
Wharton professor Ethan Mollick tweeted in February: "I strongly suspect that 'tip engineering' won't be a big deal in the long run, and tip engineers aren't jobs that will exist in the future." ”
Adrian Weller, director of machine learning research at the University of Cambridge, believes that while being able to interact with AI through cues "has high value," "I'm not sure if it's going to continue for a long time." Don't focus too much on the current situation of the prompt project. It will evolve quickly. ”
Although the profession of the prompt engineer is currently in the spotlight, only time will tell.
Note: The cover image is from Pexels, and the copyright belongs to the original author. If you do not agree to the use, please contact us as soon as possible and we will delete it immediately.
![The biggest lie of the year, "Imaginary Anna" is still far from being ranked[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)