#59 Don't Get Lost in Translation: Ambiguous Instructions and the Hidden Pitfalls of Generative AI

<< Previous Edition: Paranoid NIMBYism or Valid Privacy Concerns?

Natural Language Processing (NLP) plays a crucial role in generative AI, leading to significant advancements in various aspects of life. This article discusses the importance of providing clear and structured instructions to Large Language Models (LLMs), emphasizing that casually throwing random free-form natural text as instructions may not yield the best results. Drawing from the lessons learned in the big data era, the article highlights the need for a more focused approach when interacting with these powerful AI systems.

The Fallacy of Unstructured Data

The big data movement, exemplified by Hadoop, revolutionized how we handle and extract meaning from various unstructured data types. As data storage costs decreased, retaining vast information became more practical. However, the initial promise of effortlessly handling unstructured data was somewhat of an illusion.

The realization of this fallacy brought about a need for more robust and efficient methods to interact with complex data types. It soon became evident that so-called unstructured data possessed some inherent structure, giving rise to the concept of "semi-structured data." During the early stages of big data, query languages such as Pig Latin provided flexibility and a more relaxed structure for data querying.

These languages allowed data scientists and engineers to explore the potential of big data without being constrained by traditional data models. As the field progressed and competing frameworks emerged, these languages evolved, becoming increasingly structured to handle more complex queries efficiently. This shift led to competition among frameworks, with many striving to achieve SQL-92 compliance or even surpassing those standards, ultimately enhancing the capabilities of big data systems.

The Fallacy of Free-Flowing Text

In generative AI, large language models (LLMs) are trained on and process natural language data. Understanding natural languages parallels the concept of unstructured data in the big data movement. However, drawing a direct comparison between the two might not be entirely accurate, as natural language is used to ask questions to LLMs. At the same time, the LLM itself is also trained on natural language data. In contrast, the query languages in big data were structured, though they evolved from being initially less structured to more structured over time.

The foundation of LLM training consists of various forms of natural language text. The distinction between the promise of big data and the reality of generative AI lies in the manner of asking questions, a notion referred to as prompt engineering. This field is rapidly evolving, as demonstrated by Prof. Andrew Ng's recent course announcement in collaboration with OpenAI. The course employs unique sequences, such as three backticks (```) and three quotes ("""), and familiar tags like XML tags to create clear communication with LLMs while minimizing the risk of issues such as "prompt injection."

The need for clear instructions stems from the fact that LLMs, while intelligent, may not know the context of a specific task. This approach helps to think of LLMs as smart individuals who need guidance to perform tasks efficiently and effectively. As natural language interfaces like ChatGPT gain traction, developing more structured prompt languages for complex tasks becomes increasingly important. As these structures become more intricate, the productivity gains and advancements in AI will be substantial.

Conclusion: Balancing Flexibility and Structure

The future of NLP and generative AI lies in striking a balance between the inherent flexibility of natural language and the structure needed to achieve efficiency and effectiveness. While natural language is an effective interface for applications like ChatGPT, more structured prompt languages will be required for complex tasks. As these structures become more complex, they will unleash unprecedented productivity gains and pave the way for more advanced AI applications.

>> Next Edition: Tech Executives Dive Back In