The Evolution and Inner Workings of Large Language Models: From N-grams to Transformers

Large language models: how the AI behind the likes of ChatGPT actually works

University of Sheffield provides funding as a founding partner of The Conversation UK. The arrival of AI systems called large language models (LLMs), like OpenAI's ChatGPT chatbot, has been heralded as the start of a new technological era. And they may indeed have significant impacts on how we live and work in future. But they haven't appeared from nowhere and have a much longer history than most people realise. In fact, most of us have already been using the approaches they are based on for years in our existing technology. LLMs are a particular type of language model, which is a mathematical representation of language based on probabilities. If you've ever used predictive text) on a mobile phone or asked a smart speaker a question, then you have almost certainly already used a language model. But what do they actually do and what does it take to make one? Language models are designed to estimate how likely it would be to see a particular sequence of words. This is where probabilities come in. For example, a good language model for English would assign a high probability to a well formed sentence like "the old black cat slept soundly" and a low probability to a random sequence of words such as "library a or the quantum some". Most language models can also reverse this process to generate plausible-looking text. The predictive text in your smartphone uses language models to anticipate how you might want to complete text as you are typing. The earliest method for creating language models was described in 1951 by Claude Shannon, a researcher working for IBM. His approach was based on sequences of words known as n-grams - say, "old black" or "cat slept soundly". The probability of n-grams occurring within text was estimated by looking for examples in existing documents. These mathematical probabilities were then combined to calculate the overall probability of longer sequences of words, such as complete sentences. Estimating probabilities for n-grams becomes much more difficult as the n-gram gets longer, so it is much harder to estimate accurate probabilities for 4-grams (sequences of four words) than for bi-grams (sequences of two words). Consequently, early language models of this type were often based on short n-grams. However, this meant that they often struggled to represent the connection between words that occurred far apart. This could result in the start and end of a sentence not matching up when the language model was used to generate a sentence. To avoid this problem, researchers created language models based on neural networks - AI systems that are modelled on the way the human brain works. These language models are able to represent connections between words that may not be close together. Neural networks rely on large numbers of numerical values (known as parameters) to help understand these connections between words. These parameters must be set correctly in order for the model to work well. The neural network learns the appropriate values for these parameters by looking at large numbers of example documents, in a similar way that n-gram probabilities are learned by n-gram language models. During this "training" process, the neural network looks through the training documents and learns to predict the next word based on the ones that have come before. These models work well but have some disadvantages. Although in theory, the neural network is able to represent connections between words that occur far apart, in practice, more importance is placed on those that are closer. More importantly, words in the training documents have to be processed in sequence to learn appropriate values for the network's parameters. This limits how quickly the network can be trained. The dawn of transformers A new type of neural network, called a transformer, was introduced in 2017 and avoided these problems by processing all of the words in the input at the same time. This allowed them to be trained in parallel, meaning that the calculations required can be spread across multiple computers to be carried out at the same time. A side effect of this change is that it allowed transformers to be trained on vastly more documents than was possible for previous approaches, producing larger language models. Transformers also learn from examples of text but can be trained to solve a wider range of problems than only predicting the next word. One is a kind of "fill in the blanks" problem where some words in the training text have been removed. The goal here is to guess which words are missing. Another problem is where the transformer is given a pair of sentences and asked to decide whether the second should follow the first. Training on problems like these has made transformers more flexible and powerful than previous language models. The use of transformers has allowed the development of modern large language models. They are in part referred to as large because they are trained using vastly more text examples than previous models. Some of these AI models are trained on over a trillion words. It would take an adult reading at average speed more than 7,600 years to read that much. These models are also based on very large neural networks, some with more than 100 billion parameters. In the last few years, an extra component has been added to large language models that allows users to interact with them using prompts. These prompts can be questions or instructions. This has enabled the development of generative AI systems such as ChatGPT, Google's Gemini and Meta's Llama. Models learn to respond to the prompts using a process called reinforcement learning, which is similar to the way computers are taught to play games like chess. Humans provide the language model with prompts, and the humans' feedback on the replies produced by the AI model is used by the model's learning algorithm to guide further output. Generating all these questions and rating the replies requires a lot of human input, which can be expensive to obtain. One way of reducing this cost is to create examples using a language model in order to simulate human-AI interaction. This AI-generated feedback is then used to train the system. Creating a large language model is still an expensive undertaking, though. The cost of training some recent models has been estimated to run into hundreds of millions of dollars. There is also an environmental cost, with the carbon dioxide emissions associated with creating LLMs estimated to be equivalent to multiple transatlantic flights. These are things that we will need to find solutions to amid an AI revolution that, for now, shows no sign of slowing down.

Large language models: How the AI behind the likes of ChatGPT actually works

The arrival of AI systems called large language models (LLMs), like OpenAI's ChatGPT chatbot, has been heralded as the start of a new technological era. And they may indeed have significant impacts on how we live and work in the future. But they haven't appeared from nowhere and have a much longer history than most people realize. In fact, most of us have already been using the approaches they are based on for years in our existing technology. LLMs are a particular type of language model, which is a mathematical representation of language based on probabilities. If you've ever used predictive text) on a mobile phone or asked a smart speaker a question, then you have almost certainly already used a language model. But what do they actually do and what does it take to make one? Language models are designed to estimate how likely it would be to see a particular sequence of words. This is where probabilities come in. For example, a good language model for English would assign a high probability to a well formed sentence like "the old black cat slept soundly" and a low probability to a random sequence of words such as "library a or the quantum some." Most language models can also reverse this process to generate plausible-looking text. The predictive text in your smartphone uses language models to anticipate how you might want to complete text as you are typing. The earliest method for creating language models was described in 1951 by Claude Shannon, a researcher working for IBM. His approach was based on sequences of words known as n-grams -- say, "old black" or "cat slept soundly." The probability of n-grams occurring within text was estimated by looking for examples in existing documents. These mathematical probabilities were then combined to calculate the overall probability of longer sequences of words, such as complete sentences. Estimating probabilities for n-grams becomes much more difficult as the n-gram gets longer, so it is much harder to estimate accurate probabilities for 4-grams (sequences of four words) than for bi-grams (sequences of two words). Consequently, early language models of this type were often based on short n-grams. However, this meant that they often struggled to represent the connection between words that occurred far apart. This could result in the start and end of a sentence not matching up when the language model was used to generate a sentence. To avoid this problem, researchers created language models based on neural networks -- AI systems that are modeled on the way the human brain works. These language models are able to represent connections between words that may not be close together. Neural networks rely on large numbers of numerical values (known as parameters) to help understand these connections between words. These parameters must be set correctly in order for the model to work well. The neural network learns the appropriate values for these parameters by looking at large numbers of example documents, in a similar way that n-gram probabilities are learned by n-gram language models. During this "training" process, the neural network looks through the training documents and learns to predict the next word based on the ones that have come before. These models work well but have some disadvantages. Although in theory, the neural network is able to represent connections between words that occur far apart, in practice, more importance is placed on those that are closer. More importantly, words in the training documents have to be processed in sequence to learn appropriate values for the network's parameters. This limits how quickly the network can be trained. The dawn of transformers A new type of neural network, called a transformer, was introduced in 2017 and avoided these problems by processing all of the words in the input at the same time. This allowed them to be trained in parallel, meaning that the calculations required can be spread across multiple computers to be carried out at the same time. A side effect of this change is that it allowed transformers to be trained on vastly more documents than was possible for previous approaches, producing larger language models. Transformers also learn from examples of text but can be trained to solve a wider range of problems than only predicting the next word. One is a kind of "fill in the blanks" problem where some words in the training text have been removed. The goal here is to guess which words are missing. Another problem is where the transformer is given a pair of sentences and asked to decide whether the second should follow the first. Training on problems like these has made transformers more flexible and powerful than previous language models. The use of transformers has allowed the development of modern large language models. They are in part referred to as large because they are trained using vastly more text examples than previous models. Some of these AI models are trained on over a trillion words. It would take an adult reading at average speed more than 7,600 years to read that much. These models are also based on very large neural networks, some with more than 100 billion parameters. In the last few years, an extra component has been added to large language models that allows users to interact with them using prompts. These prompts can be questions or instructions. This has enabled the development of generative AI systems such as ChatGPT, Google's Gemini and Meta's Llama. Models learn to respond to the prompts using a process called reinforcement learning, which is similar to the way computers are taught to play games like chess. Humans provide the language model with prompts, and the humans' feedback on the replies produced by the AI model is used by the model's learning algorithm to guide further output. Generating all these questions and rating the replies requires a lot of human input, which can be expensive to obtain. One way of reducing this cost is to create examples using a language model in order to simulate human-AI interaction. This AI-generated feedback is then used to train the system. Creating a large language model is still an expensive undertaking, though. The cost of training some recent models has been estimated to run into hundreds of millions of dollars. There is also an environmental cost, with the carbon dioxide emissions associated with creating LLMs estimated to be equivalent to multiple transatlantic flights. These are things that we will need to find solutions to amid an AI revolution that, for now, shows no sign of slowing down.