Do Not Use LLM or Generative AI For These Use Cases

🌈 Abstract

The article discusses the recent hype around Generative AI, particularly Large Language Models (LLMs), and cautions decision-makers against simply implementing AI solutions without careful consideration. It explores various AI use case families and techniques, highlighting that not all use cases are suitable for Generative AI. The article provides a matrix that maps AI techniques to different use case families, emphasizing the importance of selecting the right AI technique for the corresponding use case.

🙋 Q&A

[01] Typical AI Use Case Families

1. What are the typical use case families for AI discussed in the article?

• Sales forecasting: Using machine learning algorithms to predict future sales based on historical data and other relevant features.
• Autonomous systems: Using AI to build autonomous systems that can perform tasks intelligently, such as routine inspections of power transmission lines using drones.
• Planning: Using AI to analyze data and find the best plan to maximize benefits and minimize impact, such as optimizing road projects to minimize traffic disruption.
• Decision support: Using AI to provide insights and predictions to enhance the human decision-making process, taking into account preferences and entrepreneurship.
• Recommendation: Using AI to recommend products, music, or other items that users may like, as seen in applications like eBay and Spotify.
• Classification: Using AI to classify entities into different categories or levels based on their attributes, such as classifying loan applicants into risk levels.
• Process automation: Using AI combined with automation technologies to enhance business processes, such as predicting equipment failures in manufacturing.
• Perception: Using AI to process sensory data, such as using cameras to detect driving violations.
• Anomaly detection: Using AI to detect subtle events or anomalies in complex systems, such as fluctuations in an electricity grid.
• Content generation: Using Generative AI techniques to create text, images, videos, or other artifacts, as exemplified by ChatGPT.
• Insight discovery: Using AI to discover insights, relationships, or correlations from large, chaotic datasets, such as finding patterns in patient data to develop new treatment strategies.

2. How does the article categorize the different AI techniques? The article categorizes the AI techniques as follows:

• Normal Machine Learning: Includes basic techniques like linear regression, clustering, classification, and decision trees.
• Simulation: Allows creating models of real-world processes or systems to answer "what-if" questions.
• Optimization: Helps find the best parameters in a formula or equation, such as finding the optimal balance between discounts and profits.
• Rule-based systems: Involves creating pre-defined rules based on scientific evidence or domain expert knowledge to make better decisions.
• Graph-based: Utilizes a data structure that represents objects and their relationships, which can be more accurate than tabular-based data structures.
• Generative AI: Techniques that can generate text, images, videos, or other artifacts, as exemplified by ChatGPT.

[02] Mapping AI Techniques to Use Case Families

1. What is the key takeaway from the matrix that maps AI techniques to use case families? The key takeaway is that not all use case families are suitable for Generative AI techniques like ChatGPT. The matrix shows the suitability of different AI techniques for each use case family, categorized as Low (L), Medium (M), or High (H) in terms of stability and reliability.

2. What are the limitations of using Generative AI techniques like ChatGPT for certain use case families? According to the matrix, Generative AI techniques should not be used for:

• Forecasting use cases, as they are not suitable for predicting something that does not exist in the world.
• Planning use cases, as they are not suitable for finding the best plan to maximize benefits and minimize impact.

Instead, the article suggests that Generative AI techniques like ChatGPT should be used for "Content Generation" use cases, where they can be used to generate text, code, or other artifacts. For other use cases, the article recommends using the "High" suitable AI techniques as per the matrix.