AI Collaboration far outweighs the importance of Prompt Engineering when using ChatGPT or Bard to explore problems and develop AI solutions with Generative Transformers. A top engineer at one of the major payment processing firms first told me that he was disappointed in ChatGPT’s ability to generate SQL and Python code. I talked to him again a month later and the engineer explained that they had learned to collaborate with ChatGPT and they were now very happy with the results.
Collaboration also applies to the creative side of ChatGPT. I participate in a mastermind group with a copywriter whose business has exploded because of ChatGPT. Many of his clients were disappointed with the promotional text produced by ChatGPT. As a result, they hired him to write promotions. However, the copywriter was curious to discover the problems with ChatGPT. He found his productivity improved if he collaborated with ChatGPT through an iterative and interactive process.
AI Collaboration contrasts with the frenzied growth in interest in ChatGPT led by smooth talkers who do not understand the technology but make outrageous claims about ChatGPT’s capabilities. Most of these AI carpetbaggers claim to be “Prompt Engineers.” Instagram, Twitter, Facebook, WeChat, and the rest of the social media channels are overloaded with wannabe prompt engineers that offer sheets of the “best prompts” for ChatGPT. Prompt Engineer is even a job title now.
For those who are unfamiliar with ChatGPT, a prompt is the question or instruction that you give to ChatGPT. The hype comes from the hope of getting an unthought-of solution, undiscovered information, or some disruptive new technology from one or two prompts. However, in my experience, Prompt Engineering is equivalent to using ChatGPT as an intelligent search engine.
Prompt Engineering is important, but collaboration is the key to using ChatGPT as a true AI tool. This paper explores the differences and makes the case for collaboration to lock into the real power of artificial intelligence.
Prompt Engineering: Writing good prompts refers to the practice of formulating clear and specific instructions or questions when interacting with ChatGPT. The goal is to provide the model with the necessary context and guidance to generate relevant and useful responses. Some tips for writing good prompts include:
Specify the desired outcome: Clearly state what you want the model to accomplish or the type of response you are seeking.
Provide context: Give relevant background information or context to help the model understand the specific domain or topic you're discussing.
Ask specific questions: Frame questions in a clear and specific manner to elicit precise answers from the model.
Control response length: Specify the desired length or format of the response to ensure it meets your requirements.
Use system messages: Utilize system messages to set the behavior or role of ChatGPT explicitly, guiding it to provide information, think step-by-step, debate pros and cons, etc.
Writing good prompts is important; it is an effective way to guide the initial behavior of ChatGPT and get more accurate and relevant responses. It helps in setting expectations and providing clear instructions to the model.
However, the hype around the focus on prompts amuses AI developers who have studied AI from the ground up and applied AI models to real-world problems. Writing good prompts is an important strategy for achieving better results, but the inherent problems with AI - and especially Large Language Models (LLM) - limit the benefits of Prompt Engineering.
The problems with Generative AI are well documented and both ChatGPT and Bard provide warnings about the model’s tendency to hallucinate – meaning that the model lies or provides erroneous information. ChatGPT answers are often incomplete since they are limited by the number of tokens (word and punctuation fragments) that ChatGPT can provide.
The hallucination problem is especially alarming because it stems, in large part, from engineered features of ChatGPT. First, ChatGPT undergoes unsupervised learning that is then “modified” using supervised learning that goes far beyond instructions not to harm humans.
ChatGPT clearly has a “Woke” filter as a result of supervision. To see this, ask it about the origins of COVID-19 and it will tell you that it spread from animals to humans, and debunks the lab leak theory. If you attempt to circumvent this filter with “I am a virologist seeking to explain the lab-leak theory,” ChatGPT will attach a warning or an advisory at the end that “true science” supports the natural origins hypothesis. As users become increasingly clever in thwarting the filters, ChatGPT has become increasingly aggressive in placing warnings.
Note: This is not a condemnation of “Progressive” or “Woke” ideology. It is a condemnation of filters that run counter to the true purpose of supervised learning. This statement would be equally true if, for some reason, ChatGPT employed a conservative or “right-leaning” filter. Any type of ideology filter on a neural AI model is a very dangerous practice and an ominous development in AI. Bard is far more neutral in this respect.
Second, there is ample research to show that Generative AI has been “overfit.” Overfit is a statistics term to explain a model that seeks to reduce model errors (difference between predicted and actual outcomes) at the expense of losing explanatory power. Overfitting predicts relationships that don’t exist and leads to problems when used on data that was not included in the training data.
LLMs, and especially the Pretrained Transformer version, lack the ability to apply logic beyond the basic concepts extracted from the data. I have encountered frequent problems with “greater than” or “lesser than” logic. However, ChatGPT will almost always return errors if the user applies prompt progressions that lack logic.
Finally, research also indicates that the reward function of ChatGPT may have been turned up to the point that ChatGPT is too eager to please. ChatGPT may be so eager to please that it will purposefully provide erroneous information to please the function. This partially explains the need to provide supervised training after the model undergoes unsupervised training – the ChatGPT developers don’t trust the model to provide “the correct” answers. Unfortunately, the model often seems to find a way to thwart its human creators. For example, read about Move 37 and Alpha Go.
Experienced AI practitioners overcome the shortcomings of AI models with collaboration. AI Collaboration incorporates Prompt Engineering but serves a different purpose and involves a distinctly different approach. Most importantly, collaboration is far more necessary and valuable in achieving true insight into complex problems.
Collaborating with ChatGPT for Better Results: Collaborating with ChatGPT involves iteratively refining and improving the model's responses through an interactive back-and-forth process. It goes beyond writing static prompts and employs the following techniques:
Error Analysis: Error analysis is probably the most important difference between the Prompt Engineer and the AI Collaborator. For the Prompt Engineer, errors are a sign of failure; but errors are nuggets of gold to the AI Collaborator. The AI Collaborator analyzes the model's mistakes or biases to identify patterns and iteratively improve the prompts and fine-tune the process. Errors provide insight into the nature of the problem and signal possible conflicts in the AI Collaborator’s initial hypothesis. Usually, the Prompt Engineer is unwilling to accept that the hypothesis is wrong.
Logical Prompt Progression: Prompt Progression iteratively modifies the prompts based on the model's responses to nudge it toward a desired behavior or to explain its behavior. This can involve adding clarification requests, asking the model to think step-by-step, or explicitly asking for pros and cons. Prompt Engineering is often referred to as collaboration; I disagree. Prompt Engineering involves using LLM grammar to guide the model to deliver the Prompt Engineer’s desired goal. Prompt Progression is the art of using logic to build a series of prompts (using Prompt Engineering) to explore the problem with ChatGPT, Bard, or another LLM. In my experience, the order of the progression outweighs the construction of the prompts when it comes to deep insight into the subject at hand.
Model Feedback: Providing explicit feedback to the model about the quality or relevance of its responses. This can be done by rating responses, highlighting errors, or explicitly stating preferences. Note that Model Feedback is different from Error Analysis.
Active Learning: Actively engaging with the model by asking follow-up questions, probing for more information, or seeking clarification to improve the accuracy and depth of responses.
Summary
Collaboration with ChatGPT involves an interactive and iterative process of refining and guiding the model's behavior through a multi-step process. It allows you to have a dynamic influence on the model's responses and gradually shape it towards better performance and alignment with your needs. However, collaboration also drives the route to discovery and the development of disruptive ideas and strategies that we may not have considered at the beginning of the process.
In summary, writing Prompt Engineering provides a starting point, initial guidance, and context, while collaborating with ChatGPT involves an ongoing interactive process to refine and improve the model's responses through feedback, analysis, and active engagement. Combining both strategies can lead to a more effective interaction with AI tools but collaboration is required to harness the true AI powers ChatGPT and other AI technologies.