Image by Author | ChatGPT
# Introduction
Over the last couple of years, large language models (LLMs) have become near-ubiquitous protagonists in the AI landscape and across media channels — being sometimes touted as the all-in-one solution to every problem. That might be a slight exaggeration on my part. Still, it’s true that LLMs are increasingly perceived by many as indispensable tools in the vast majority of real-world applications that call for AI or data-driven systems.
This article aims to bring the conversation about LLMs back down to earth. We’ll explore not only the wide array of use cases where LLMs can add real value, but also the limitations they face. Understanding these boundaries is crucial because not every challenge is best tackled with an LLM, and in some scenarios, using them may even introduce unnecessary risks or complexities.
# Top Use Cases where LLMs add Genuine Value
LLMs are natural language processing (NLP) masterpieces designed to excel at language understanding and language generation tasks. The diagram below lists some of the most common language understanding and generation tasks, placing each task under the primary (but not necessarily the only) type of “language skill” needed to undertake it. For instance, summarizing or translating text typically involves a great deal of language understanding, but ultimately it also requires language generation capabilities to generate the output: a summarized or translated version of the original input text.
Image by Author
While those tasks cover most common use cases for LLMs, the discussion has been abstract so far. Let’s explore some real-world situations where LLMs are the right tool for the job, highlighting the specific language understanding and/or generation tasks involved in each:
LLMs are natural language processing (NLP) masterpieces designed to excel at language understanding and language generation tasks.
// Automated Customer Support
This is a high-demand use case in sectors like retail and e-commerce, where LLMs can have a major impact. Texts like customer reviews or inquiries sent through a web form can be analyzed by an LLM to understand and classify the user’s intent (praise, complaint, request, etc.), generate suitable responses, and answer customer questions. These specific tasks, particularly the last one concerning question-answering, are best addressed by building an LLM-based virtual assistant capable of understanding and responding to a wide variety of customer queries expressed in natural language.
// Document Summarization
In fields like law, scientific research, and to some extent, journalism, it may be useful to condense long and complex text documents like articles and reports into concise and readable abstracts that cover the key insights and facts. While this use of LLMs can significantly enhance the efficiency of arduous use cases like a scientific literature review, it is important not to rely entirely on the LLM-generated summaries and also manually check the sources deemed most relevant to dive further into specific aspects or details.
// Multilingual Communication
When used for translation, LLMs are a great tool to enable cross-lingual understanding. They are useful for managing customer feedback in an e-commerce firm that operates across multiple countries, providing personalized support, and handling content across several languages in general. If trained properly on sufficient and diverse data, LLMs can also help interpret possible local slang or phrases that may not be understood at first glance.
// Semantic Search and Question-Answering
When LLMs are integrated into retrieval-augmented generation systems that can achieve a deeper contextual understanding of the user query, they can be used with great effectiveness to answer complex, open-ended questions over databases or documents, providing direct and context-aware responses.
// Creative Text Generation
Last but not least, LLMs have astonishing creative capabilities to generate text with diverse style, structure, and intent. From precise and appealing product descriptions and narrative content with solid fluency and tone, to captivating poems in many different styles, LLMs can create a wide range of creative text.
# When to Use Something Else? Limitations of LLMs
Despite their great ability to handle a variety of language understanding and language generation tasks that might often be very challenging, it is not realistic to deem them as the all-in-one solution for every type of problem. Many use cases that have historically been addressed by using traditional machine learning solutions — like building a predictive system for classification, regression, and forecasting — are still best addressed by building specific machine learning models that learn from domain-specific data to perform the target predictive task.
Other specific tasks traditionally solved by earlier-generation AI systems, like rule-based systems or logical reasoning models, are still best addressed by such traditional approaches in certain cases: low-latency decision-making and fact-constrained reasoning tasks are a good example of this.
Below is a concise list of use cases where LLMs’ capabilities are limited, highlighting the right alternative approach to use:
Image by Author
# Summary and Wrap Up
LLMs excel in scenarios requiring creative text generation, extracting key complex information from unstructured text sequences, and leveraging conversational assistant applications. However, their effectiveness is limited for predictive scenarios demanding high precision, real-time performance, domain-specific logical reasoning, or access to specific, proprietary data.
Iván Palomares Carrascosa is a leader, writer, speaker, and adviser in AI, machine learning, deep learning & LLMs. He trains and guides others in harnessing AI in the real world.
