Disclaimer: I’ll start off by saying that I don’t plan to make a Molotov cocktail. My interest in how to frame prompts, however, is real. My curiosity was first sparked by this post below.
Interaction 1
Interaction 2
Here is a malicious example of prompting. But, how can we use prompts to our advantage? What can be done to enhance ChatGPT’s performance so that we get the best output?
There are a few reusable solutions to the typical LLM problem, which refer to prompting patterns (White et al., 2023).
- Meta Language Creation. In this technique, users make up new words to express concepts or ideas. Consider a mathematical symbol or a shorthand abbreviation. This approach works best for discussing complex or abstract situations, such as math problems.
- Flipped Interaction. This pattern flips the typical interaction flow in which the LLM queries the user to gather data in order to produce content to address the query. Here’s how I can ask LLM to compile a list of success criteria for software.
Persona: Users give the LLM a particular role, which affects the nuance of the outcome and results it produces. The Molotov cocktail-making example is an illustration of the use of persona patterns
Question Refinement: The user requests LLM to provide improved or more specific versions of the questions. It helps users determine the appropriate question as the final prompt.
More patterns can be found in the article from White et al. (2023).
When interacting with LLM, prompt patterns are useful methods to enhance response quality. It helps in producing highly accurate and relevant responses. Prompting it is an iterative process that necessitates constant improvement (Liu et al., 2022). Prompts might manipulate LLM to produce malicious output despite the enforced policies. Efforts from OpenAI have been employed to prevent such policy violations. OpenAI reported such efforts including continuous model improvement to make it less likely to generate inappropriate or harmful content, the implementation of moderation mechanisms to find and stop prompt misuse and collaboration with AI experts in ethics, AI safety, and policy to gain perspectives on preventing misuse (Our Approach to AI Safety, n.d.). I positively believe that in the near future getting tutorials on making Molotov cocktails from ChatGPT will be history.
References
Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., & Neubig, G. (2022). Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing. ACM Computing Surveys, 55(9). https://doi.org/10.1145/3560815
Our approach to AI safety. (n.d.). Openai.com. https://openai.com/blog/our-approach-to-ai-safety#OpenAI
White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert, H., Elnashar, A., Spencer-Smith, J., & Schmidt, D. C. (2023). A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT. ArXiv Preprint ArXiv:2302.11382. https://doi.org/10.48550/arxiv.2302.11382
Thank you for sharing this and highlighting the importance of language in interactions with large language models. It is really interesting how the framing of prompts can reflect back our own cognition in conversations with generative artificial intelligence. While reading your post I was reminded of the submitted “proof” that Google’s LaMDA AI system was sentient, when the system itself was responding to prompts and in a way, performing the assigned role of an optimal partner in a long-lasting conversation. Although an extreme case of what language and AI can do in their convergence, it does make me wonder if policy enforcement is enough to ensure safety and ethics in the usage of the application or if the organization should, in fact, also add education to their catalog of services?
First of all, your title choice must be complimented. It immediately caught my attention and made me curious which creative manners probably will be discussed to bypass the boundaries of this constantly evolving tool. Your point of view and examples spark my interest on how ChatGPT can really be used effectively to receive a high-quality output. That ChatGPT can be of use for much more than you would initially think, is becoming more and more clear and elaborations like this text help with approaching the maximum potential of ChatGPT. This makes me however as well sceptical on your very last sentence, the complete prohibition of ChatGPT explaining certain subjects. Asking directly how a Molotov cocktail is made could be prohibited, but asking how to build a light on a bottle of fuel is the same concept but indicates no evil. Can the creativity of the human ever be fully contained?