After the initial excitement surrounding ChatGPT, an AI-driven language-processing tool, the use of chatbots is becoming more common in both work and home settings. So, how do you effectively train your AI? We provide answers to a few simple questions.
What is prompt engineering?
Prompt engineering is a technique used to effectively communicate with generative AI models. Systems like ChatGPT, Bard, and Dall-E generate text, images, and music based on an input, also known as a prompt. The phrasing of the prompt can greatly impact the output produced by the AI system. Prompt engineering involves formulating a prompt in a way that closely aligns with your expectations, resulting in outputs that are consistently useful and appropriate.
How is prompt engineering different from asking questions?
Prompt engineering requires more care and consideration. Simply asking a question to ChatGPT may or may not yield a satisfactory answer. Prompt engineering involves understanding the nuances of an AI model and constructing inputs that it can accurately comprehend. This leads to outputs that are more consistently useful, interesting, and aligned with your intentions. By formulating the prompt effectively, you may even surpass your expectations with the response.
Why should I care?
Chatbots like ChatGPT, Bard, and Bing Chat can be incredibly convenient for everyday tasks. People have used them to draft emails, summarize meeting notes, create contracts, plan vacations, and receive instant answers to complex questions. Jules White, an associate professor of computer science, highlights that these chatbots can serve as powerful personal assistants, enhancing productivity and enabling individuals to create things they otherwise wouldn’t be able to. However, effectively interacting with these chatbots requires understanding how to prompt them.
Additionally, having prompting skills can impress hiring managers. Matt Burney, a talent strategy adviser, notes that although the number of job ads specifically requiring AI proficiency is still small, it is steadily growing. Companies across various industries are increasingly exploring the integration of AI models into their workflows. To stay ahead, it’s crucial to know how to effectively prompt AI systems.
So, how do I do it?
There are several popular prompt engineering techniques. One common approach is employing personas. By instructing the system to act as a lawyer, personal tutor, drill sergeant, or any other character, the outputs will imitate their tone and voice. Conversely, you can instruct the system to complete a task while considering a specific audience, such as a five-year-old, a team of biochemists, or an office Christmas party, resulting in tailored outputs for that demographic. You don’t need to possess in-depth knowledge of the persona’s stylistic characteristics; the system figures it out.
For problem-solving, chain-of-thought prompting is more suitable. Asking the AI model to “think step by step” encourages it to provide its output in manageable chunks, often leading to more comprehensive results. Some researchers have found that showing the AI model an example problem along with its step-by-step solution improves its ability to solve similar questions with the correct answers.
In fact, examples are always beneficial. If you have a specific output in mind, you can upload a text sample or an image illustrating what you want the AI model to generate and instruct it to use it as a template. If the initial result is off target, a few more rounds of clearly specified adjustments could yield the desired outcome. It’s a continuing conversation where you refine and iterate, as suggested by White.
Furthermore, don’t forget the basic principles of clear, imperative instructions to minimize misinterpretation. Explicitly state your expectations, specify word count and format, and make clear what you want and don’t want from the output.
What should I avoid?
Avoid using vague language. AI models cannot infer your preferences, ideas, or the exact vision you have in your mind without additional information. Ensure you provide specifics and context and don’t assume that the model will accurately fill in any missing information.
Can prompt engineering prevent AI from generating inaccuracies?
No, large language models can still generate inaccuracies even when explicitly instructed not to. They may fabricate sources and provide plausible yet entirely false information. Mhairi Aitken, an ethics fellow, explains that this is an inherent problem with these models. They are designed to replicate human language without a connection to truth or reality.
Prompt engineering can help address falsehoods after they appear. If the chatbot provides incorrect information, you can point out the errors and ask it to rewrite the answer based on your feedback. Additionally, providing a list of fundamental facts or a numbered list of facts for the AI model to base its answer on can facilitate fact-checking.
Could prompt engineering become a career?
For some individuals, yes. AI developers have hired prompt engineers to test the limitations and weaknesses of their models to improve their ability to handle user inputs. However, the longevity of these positions is uncertain. Rhema Linder, a computer science lecturer, suggests that developers may prefer specialized computer scientists over self-proclaimed prompt engineers. Furthermore, the absence of industry-recognized certification makes it difficult to assess an individual’s prompting skills.
In the broader job market, prompt engineering is likely to become a skill sought after in various roles, similar to spreadsheet management or search engine optimization. Hiring managers will value it as an additional asset on a CV. Burney asserts that experience with large language models or generative pretrained transformers will become necessary for almost every office-based job. Failing to acquire these skills may impede progress towards achieving professional goals.
Will prompt engineering become irrelevant in the future?
The best practices of prompt engineering may evolve as AI models evolve. Techniques that work well now may become less useful with updated versions. However, it remains unclear how extensive the changes will be. White suggests that core concepts and patterns may remain consistent and become benchmarks for training new models. Consequently, prompt engineering may provide feedback to shape future models.
Additionally, AI models might become more adept at comprehending even vague and un-engineered prompts. Aitken envisions a future where these systems become more conversational and intuitive, potentially rendering prompt engineering unnecessary.
In conclusion, while prompt engineering is currently a valuable skill, its future relevance may change with advancements in AI models and their capabilities.