What are large language models (LLMs)?

LLMs are a modern-day development, but the study of natural language processing (NLP) dates to 1950, when Alan Turing launched the Turing test to gauge intelligent behavior among machines. In the test, a human judge speaks to a computer using a set of questions—and must determine if they are speaking to a machine or a human.

By the 1980s and 1990s, NLP shifted away from logic experiments toward a more data-driven approach. With their ability to predict which words in a sentence were likely to come next based on the words before them, statistical language models, such as n-grams, paved the way for a new era. By the early 2010s, newer neural networks expanded the capabilities of these language models even further, allowing them to move beyond determining the order of words toward a deeper understanding of the representation and meaning of words.

These new developments culminated in a breakthrough in 2018, when eight Google scientists penned and published “Attention is All You Need,” a landmark study on machine learning. Most notably, the paper introduced the transformer architecture, an innovative neural network framework that could manage and understand complex textual information with greater accuracy and scale. Transformers are now foundational to some of today’s most powerful LLMs, including the GPT series, as well as BERT.

Today’s state-of-the-art LLMs use deep learning architectures like transformers and other deep neural network frameworks to process information from different data sources. Transformers are especially effective at handling sequential data, such as text, which allows them to understand and generate natural language for tasks such as language generation and translation. 

Transformers consist of two primary components: encoders and decoders. These components often work together to process and generate sequences. The encoder takes raw textual data and turns that input into discrete elements that can be analyzed by the model. The decoder then processes that data through a series of layers to produce the final output, which may, for instance, consist of a generated sentence. Transformers can also consist of encoders or decoders only, depending on the type of model or task.

The training process for LLMs consists of three main stages: data collection, model training, and fine-tuning. 

During the data collection phase, the model is exposed to large volumes of textual data from a wide variety of sources, including Internet resources, books, articles, and databases. The data is also cleaned, processed, standardized, and stored in a NoSQL database so that it can be used to train the model on language patterns, grammar, information, and context. 

In the pre-training phase, the model starts to build an understanding of the language in the data. This is accomplished through large-scale, unsupervised tasks where the model learns to predict text based on its context. Some techniques include autoregressive modeling, where the model learns to predict the next word in a sequence, as well as masked language modeling, where the model fills in masked words to understand the context. 

Lastly, during the fine-tuning phase, the model is further trained on a smaller, more task-specific dataset. This process refines the model's knowledge and enhances its performance for specific tasks, such as sentiment analysis or translation, so that it can be used for a variety of applications.

The transformer model breaks raw text down into smaller, basic units of text called tokens. Tokens may consist of words, parts of words, or even individual characters, depending on the use case. These tokens are then converted into dense numerical representations that capture order, semantic meaning, and context. These representations, called embeddings, are then passed through a stack of layers consisting of two sub-layers: self-attention and neural networks.

While both layers assist in converting text into a form that the model can process effectively, the self-attention mechanism is a key component to the transformer architecture. The self-attention mechanism is what permits the model to home in on different parts of a text sequence and dynamically weigh the value of information relative to other tokens in the sequence, regardless of their position. This mechanism is also what gives LLMs the capacity to capture the intricate dependencies, relationships, and contextual nuances of written language.

First developed by OpenAI in 2018, the GPT series introduced the foundational concept of data collection, pretraining, and fine-tuning to LLMs. GPT-2, released in 2019, significantly scaled up the model’s capabilities and improved its ability to generate more contextually relevant language. GPT-3 advanced the model’s capacity for handling complex prompts and tasks. The latest iteration, GPT-4, was released in 2023 and provides even more accurate and nuanced responses to prompts—while also addressing some of the model’s previous challenges, including bias. 

Today, GPT continues to push the boundaries of what’s possible in the field of natural language generation. Each model in the series builds upon the previous one, driving AI-powered innovation forward. 

Developed by Google in 2018, BERT is a groundbreaking model that has set the standard for what’s possible with LLMs. Unlike the GPT series, which processes text in a unidirectional manner (from left-to-right or right-to-left), BERT takes on a bidirectional approach. A bidirectional model processes the context of each word from both directions simultaneously, which allows BERT to perform masked language modeling in addition to next-sentence predictions. Researchers have also contributed to further advancements in the field by fine-tuning BERT on tasks such as sentiment analysis, setting new benchmarks as a result.  

Developed by Facebook AI in 2019, Robustly optimized BERT approach (RoBERTa) is a variant of the BERT model that expands on BERT's bidirectional transformer architecture by optimizing the pretraining process. RoBERTa is trained with a larger data set, and for longer. It also focuses solely on masked language modeling. This allows RoBERTa to demonstrate its robust ability to capture context and nuances. 

Text-To-Text Transfer Transformer (T5), which was invented by Google Research, is another notable LLM. Like traditional models, T5 is built on the transformer architecture and uses encoders and decoders to process text during the pretraining phase. Unlike traditional models, T5 treats both the inputs and outputs as text strings, simplifying the architecture and streamlining the training process. T5 models are an adaptable general-purpose model that can handle a versatile range of tasks.

LLMs can generate engaging, informative, and contextually appropriate content in a variety of styles and formats. When prompted, they can generate articles, reports, blog posts, emails, marketing copy, and even code snippets.   

When it comes to summaries, LLMs stand out in their unique ability to distill large volumes of text into concise and accurate snapshots. They can present key points while still maintaining the original context and meaning of the original content. Researchers are already saving time and boosting productivity by using LLMs to summarize research papers, articles, presentations, and meeting notes.

Conversational agents and chatbots rely on the advanced natural language processing capabilities of LLMs to generate human-like interactions. They interpret user inputs and respond in a fluent, natural, and contextually relevant manner. Not only can they answer questions, but they can also engage in long and complex dialogue. 

With the addition of chatbots and virtual assistants, businesses can now provide round-the-clock support to their customers, in turn expanding their service availability, improving response times, and increasing overall customer satisfaction.

LLMs that are extensively trained on multilingual datasets produce highly accurate translations across various languages. Unlike traditional models, LLMs can capture the subtleties and complexities of language, such as idiomatic expressions, resulting in translations that are both fluent and contextually appropriate. 

LLMs are also able to perform sentiment analysis, which analyzes the underlying emotional tone of a text. By processing and interpreting the subtleties of language, LLMs provide more precise and insightful sentiment evaluations. They can even detect more nuanced sentiments, such as sarcasm. 

LLMs can analyze user data, including user history and preferences, and generate personalized, tailored recommendations that reflect the user's interests and needs, in turn enhancing the overall user experience. 

This capability is widely used across e-commerce, content streaming, and social media, where delivering tailored recommendations drives more meaningful interactions. LLMs can also be used as an educational tool by providing personalized learning experiences to students.