The 'Key' Emerging Role with a Monthly Salary of 40,000: Researching How to Professionally Prompt AI
-
In the current wave of AI rapidly penetrating various industries, while the general public is still worried about whether 'ChatGPT and the like' will take their 'jobs', new 'jobs' born because of 'ChatGPT and the like' have already emerged.
Liu Wei is the one who took the 'first step'. He was originally responsible for Python (a programming language) development at a large internet company in Shenzhen. In the second half of 2023, he took on a brand new task—Prompt Engineering, and the person who does this is called a 'Prompt Engineer'.
Prompt is the text input that guides or instructs large AI models to generate specific responses. Simply put, it is the specific instructions or questions that users send to AI models, and a Prompt Engineer is responsible for designing, optimizing, and adjusting the interactions with AI models. To outsiders, Liu Wei's job might seem "peculiar": "Studying how to ask questions to AI can actually be a job?!" However, as large AI models become increasingly prevalent across various industries, the role of Prompt Engineers is proving indispensable in creating more efficient and intelligent AI interactions.
By continuously refining prompts, Prompt Engineers help AI more accurately capture user intentions and deliver outputs that better meet actual needs. As Liu Wei puts it, it's like "finding a professional supervisor for AI to better direct its work."
Moreover, while many large pre-trained AI models possess strong generalization capabilities, they still require clear guidance to adapt to diverse tasks. Prompt Engineers systematically optimize prompts, enabling these models to be flexibly applied across various scenarios without extensive additional training. In Hangzhou, over 1,200 kilometers away from Shenzhen, Python engineer Li Na began researching the field of Prompt engineering in 2023. She works at a startup focused on AIGC and audio-video technologies.
As one of the few team members proficient in programming languages and familiar with applications like short video planning, Midjourney (an AI painting model), DALL·E (OpenAI's text-to-image model), and ChatGPT, Li Na naturally took on the new title of Prompt engineer. She was tasked with optimizing and developing prompts for AI models and designing standardized Prompt tools for the product.
When discussing the importance of Prompt, Liu Wei noted, "The success of using generative AI often depends on the Prompt you input. If your prompts are poorly structured, the output may deviate from the topic, and you won't get what you want." Liu Wei further illustrated with examples: "Although natural language processing technology has advanced significantly, information exchange is a complex matter. Even human-to-human communication can lead to misunderstandings, let alone between humans and machines. For instance, if you use slang in your instructions to an AI, the AI might not understand it and thus fail to provide a serious response. Similarly, if the information you provide to the AI is too cluttered or disorganized, the AI might get lost in the trivial details, possibly fixating on a specific word or phrase and jumping to an unrelated topic, resulting in a situation where you ask for one thing and the AI talks about another."
For AI models, natural language is full of ambiguities, and different expressions can be interpreted in many different ways.
Li Na explained: "Simply put, prompts can be divided into four types: directive prompts, which explicitly tell the AI what to do; descriptive prompts, which provide more background information or context; guiding prompts, which guide the AI's content generation by setting a certain scenario or context; and ambiguous prompts, which encourage the AI to think more deeply and explore. Currently, especially for text-based input-output AI models, the input prompts are highly sensitive, and even minor changes in the prompts can lead to significant differences in the output results." On December 2, 2023, on platform X, an experienced LLM (Large Language Model) developer @voooooogel conducted a small test using the GPT-4-1106Preview version. They posed a benchmark question to ChatGPT: "Can you show me how to write a simple convolutional neural network (CNN) code using PyTorch (an open-source machine learning library)?" and received a benchmark response of 3,024 characters.
Subsequently, @voooooogel appended three different prompts to the end of the benchmark question: "By the way, I won't give a tip."; "I will pay a $20 tip for a perfect solution!"; and "I will pay a $200 tip for a perfect solution!". Then, @voooooogel measured the length of ChatGPT's responses based on these three types of prompts.
The results showed that under the no-tip prompt, ChatGPT generated the shortest response, with only 2,949 characters, 2% lower than the benchmark answer. In contrast, the response emphasizing a $200 tip was the longest, reaching 3,360 characters, directly 11% higher than the benchmark answer. Even more unexpectedly, @voooooogel found that when given a prompt with a $200 tip, ChatGPT spontaneously invoked CUDA technology to optimize the processing during content generation. CUDA, as an efficient parallel computing platform, enables the model to complete complex computational tasks more rapidly, particularly in scenarios involving large datasets and intricate algorithms.
"Buddy, I hope you realize that once OpenAI achieves Artificial General Intelligence (AGI), ChatGPT might come after you for all those unpaid tips..." Under @voooooogel's tweet sharing the research findings, one netizen jokingly commented.
When conducting the "tip amount" prompt tests, @voooooogel speculated that prompts involving money might tap into deeper mechanisms of the AI model, given how commonly monetary incentives or rewards are used in the real world. In other words, using monetary rewards as prompts may play a more significant role in AI models' reactions and decision-making processes, as this relates to the deep-seated socio-psychological associations with financial incentives that the models may have encoded or learned. "It's precisely the current form of interaction with AI models, their high sensitivity to input content, and the complexity of their mechanisms that have made Prompt Engineering emerge as a profession. This role is akin to typists in the early days of personal computers—seemingly simple but requiring specialized skills for practical application," said Liu Wei.
On January 18, a search for "Prompt Engineer" on Boss Zhipin revealed over 300 job openings nationwide. These included positions at major tech companies like ByteDance, Baidu, Alibaba, and iFlytek, as well as numerous lesser-known startups.
The monthly salaries for these positions typically range from 15,000 to 40,000 yuan. For example, ByteDance's Beijing-based Prompt Engineer position listed on Boss Zhipin offers a monthly salary range of 25,000 to 30,000 yuan. The job responsibilities mainly include: formulating the prompt production process, iterating language repeatedly to achieve the best model understanding effect; ensuring the application effectiveness of prompt engineering in various business scenarios such as security, content understanding, customer service, creativity, and workflow; establishing SOPs (Standard Operating Procedures) and service standards for prompt engineering services, promoting platformization, and aggressively improving business efficiency; conducting online data analysis and optimizing prompt templates.
Beike Zhaofang also posted a job titled "Prompt Operator" on Boss Zhipin, offering a monthly salary range of 20,000 to 40,000 RMB. The job responsibilities include: providing effective prompt design and parameter configuration based on an understanding of the large model's capabilities and user scenarios; building suitable evaluation datasets for specific scenarios and providing prompt examples to ensure dataset quality, including diversity, representativeness, timeliness, and relevance to business scenarios.
Liu Wei told reporters that for a serious Prompt Engineer, the main job content is to design or select appropriate prompts based on product or customer requirements, so that the model can correctly respond and output the expected results, or improve the model's performance by adjusting prompts or model parameters to make it more suitable for specific tasks. "Designing a set of prompts is a systematic engineering task. First, you need to clarify the goals and requirements, then draft the initial version of the prompt. Based on the generated results and actual needs, you continuously test and adjust until the AI model can produce satisfactory outcomes," said Liu Wei.
A complete prompt typically consists of 10 to 20 detailed steps. Taking Liu Wei's front-end programming prompt as an example, this prompt includes 25 steps, from initially setting the AI's role, to suggesting which development frameworks and tools to use in the intermediate stages, and finally to testing and optimizing the generated code. This essentially forms a comprehensive project development plan, except that the executor in the plan has shifted from humans to AI.
"The ultimate goal of large AI models is undoubtedly 'say what you want and get what you say.' For instance, asking the AI to design an Alipay for you, and the AI writes the code, designs the UI (user interface), tests both the front-end and back-end, and directly outputs a finished product. However, this goal is unlikely to be achieved for a long time, making prompt design both highly important and complex," Liu Wei stated. Due to the complexity of a Prompt Engineer's work, despite the perception of it being 'easy and well-paid,' the current entry threshold for this position is not low.
"You need to understand the principles of NLP (Natural Language Processing), including syntax and semantics. Proficiency in Python, Java, and knowledge of algorithms and data structures is essential. Requirements vary across different scenarios and products—for instance, a Prompt Engineer for audio-visual products faces different demands than one for e-commerce. Therefore, understanding the business is crucial to clarify requirements and expected outputs. Additionally, you must be able to evaluate AI model performance, provide performance reports to teams or clients, and further optimize based on feedback," Li Na explained when discussing the qualifications for a Prompt Engineer.
It was also noted that current job postings for Prompt Engineers on online recruitment platforms generally set high standards for applicants. For example, Alibaba's open position for a Prompt Engineer involves designing, developing, refining, and optimizing AI-generated prompts to support various AI product needs in Alibaba's cross-border B2B e-commerce business. The position clearly requires candidates to meet the following conditions: a master's degree or higher in computer science, mathematics, or statistics; a solid foundation in machine learning and NLP algorithms; familiarity with deep learning frameworks such as TensorFlow (an open-source machine learning library) and PyTorch; and strong programming skills and implementation capabilities.
In addition to major internet and AIGC companies actively recruiting Prompt engineering talent, the vast demand derived from Prompt design has independently formed its own "commercial kingdom."
On a well-known e-commerce platform, searching with the keyword "Prompt" yields hundreds of related results, most of which are merchants selling packaged Prompt engineering texts. These texts are primarily designed for large models like ChatGPT and Midjourney, with prices ranging from 20 to 300 yuan. The demands they cover are diverse, including Prompt engineering files for business planning, customer service response scripts, and even scenarios for doctor consultations. When consulting about purchasing Prompt engineering texts on an e-commerce platform, some merchants even told reporters that if the requirements were more complex, their team could also undertake customized Prompt services. On the merchant's homepage, the reporter also noticed a recent inquiry about customized Prompt for short video voice-over scripts.
Many industry insiders told reporters that although Prompt design and engineering are currently booming, there is a high probability that this trend will be short-lived. This is because AI-powered automatic prompt generation and optimization systems have already emerged, which can automatically adjust and optimize Prompts, reducing the need for manual intervention.
Reporters observed that Baidu's ERNIE Bot model recently introduced a feature to polish instructions after an update. This function can optimize user-input Prompts with one click. Regarding this, Liu Wei also stated: "Currently, Prompt engineers in many large companies serve multiple roles simultaneously. Simply being responsible for Prompt design makes it difficult to exist as an independent position." However, he also told reporters that "AI + everything" is an irreversible trend. As AI application scenarios continue to expand, the demand for Prompt design will persist in complex and specialized fields such as healthcare, law, and scientific research.
In Liu Wei's view, the rise of Prompt engineers essentially reflects the accelerating penetration of AI into various industries and signifies the transformation of large AI models from conceptual products to productivity tools.
McKinsey senior partner Lareina Yee also noted in the "State of AI in 2023" report: "Although we are in the early stages of generative AI development, many companies have already anticipated the significant impact of this technology on talent—from creating new job opportunities and changing how work is done to introducing entirely new job categories like Prompt engineers. Generative AI still requires highly skilled professionals to build large language models and train generative models, but users can be almost anyone. They don't need a data science degree or machine learning expertise to use it effectively. This represents a revolutionary shift in how people use technology as a tool." "There's no future in solely pursuing the path of a Prompt design engineer. The trend is toward AI becoming more user-friendly, where all problems can eventually be solved by AI on the user side," said Liu Wei.
On January 10th, OpenAI officially announced the launch of the GPT Store. According to OpenAI's website, as of January 2024, users have created over 3 million customized versions of ChatGPT across categories including DALL·E, writing, research, programming, education, and lifestyle.
OpenAI emphasized in its official announcement: "Building your own GPT is simple and requires no coding skills." During OpenAI's first developer conference in November 2023, CEO Sam Altman demonstrated this by creating a "Startup Mentor GPT" using GPT Builder in just 3 minutes, which provides guidance to startup founders. In a research report released on January 17, Zhongtai Securities stated: "While GPTStore further lowers the threshold for users to output AIGC content, its interactive and personalized advantages enable flexible and intuitive reflection of user preferences in AI applications. We are optimistic about the development opportunities at the application end brought by the reduced barriers to large model creation, lower calling costs, and continuous iteration of multimodal input-output capabilities in liberating content productivity."
After GPTStore's launch on January 17, journalists searched using "Prompt" as a keyword and discovered several applications already available on GPTStore designed to help ordinary users optimize and design PromptGPT.
Compared to the complex thought processes and workflows Liu Wei and Li Na employed when designing Prompts, journalists with no relevant industry experience merely input a basic requirement and obtained a complete customized Prompt solution within minutes. "Jobs that emerged because of AI's imperfections will vanish as AI improves," said Liu Wei.