How do large language models work?

LLMs take a complex approach that involves multiple components.

At the foundational layer, an LLM needs to be trained on a large volume -- sometimes referred to as a corpus -- of data that is typically petabytes in size. The training can take multiple steps, usually starting with an unsupervised learning approach. In that approach, the model is trained on unstructured data and unlabeled data. The benefit of training on unlabeled data is that there is often vastly more data available. At this stage, the model begins to derive relationships between different words and concepts.

The next step for some LLMs is training and fine-tuning with a form of self-supervised learning. Here, some data labeling has occurred, assisting the model to more accurately identify different concepts.

Next, the LLM undertakes deep learning as it goes through the transformer neural network process. The transformer architecture enables the LLM to understand and recognize the relationships and connections between words and concepts using a self-attention mechanism. That mechanism is able to assign a score, commonly referred to as a weight, to a given item (called a token) in order to determine the relationship.

Once an LLM has been trained, a base exists on which the AI can be used for practical purposes. By querying the LLM with a prompt, the AI model inference can generate a response, which could be an answer to a question, newly generated text, summarized text or a sentiment analysis.

What are large language models used for?

LLMs have become increasingly popular because they have broad applicability for a range of NLP tasks, including the following:

• Text generation. The ability to generate text on any topic that the LLM has been trained on is a primary use case.
• Translation. For LLMs trained on multiple languages, the ability to translate from one language to another is a common feature.
• Content summary. Summarizing blocks or multiple pages of text is a useful function of LLMs.
• Rewriting content. Rewriting a section of text is another capability.
• Classification and categorization. An LLM is able to classify and categorize content.
• Sentiment analysis. Most LLMs can be used for sentiment analysis to help users to better understand the intent of a piece of content or a particular response.
• Conversational AI and chatbots. LLMs can enable a conversation with a user in a way that is typically more natural than older generations of AI technologies.

Among the most common uses for conversational AI is with a chatbot, which can exist in any number of different forms where a user interacts in a query-and-response model. One of the most widely used LLM-based AI chatbots is ChatGPT, which is based on OpenAI's GPT-3 model.

What are the advantages of large language models?

There are numerous advantages that LLMs provide to organizations and users:

• Extensibility and adaptability. LLMs can serve as a foundation for customized use cases. Additional training on top of an LLM can create a finely tuned model for an organization's specific needs.
• Flexibility. One LLM can be used for many different tasks and deployments across organizations, users and applications.
• Performance. Modern LLMs are typically high-performing, with the ability to generate rapid, low-latency responses.
• Accuracy. As the number of parameters and the volume of trained data grow in an LLM, the transformer model is able to deliver increasing levels of accuracy.
• Ease of training. Many LLMs are trained on unlabeled data, which helps to accelerate the training process.

What are the challenges and limitations of large language models?

While there are many advantages to using LLMs, there are also several challenges and limitations:

• Development costs. To run, LLMs generally require large quantities of expensive graphics processing unit hardware and massive data sets.
• Operational costs. After the training and development period, the cost of operating an LLM for the host organization can be very high.
• Bias. A risk with any AI trained on unlabeled data is bias, as it's not always clear that known bias has been removed.
• Explainability. The ability to explain how an LLM was able to generate a specific result is not easy or obvious for users.
• Hallucination. AI hallucination occurs when an LLM provides an inaccurate response that is not based on trained data.
• Complexity. With billions of parameters, modern LLMs are exceptionally complicated technologies that can be particularly complex to troubleshoot.
• Glitch tokens. Maliciously designed prompts that cause an LLM to malfunction, known as glitch tokens, are part of an emerging trend since 2022.

What are different types of large language models?

There is an evolving set of terms to describe the different types of large language models. Among the common types are the following:

• Zero-shot model. This is a large, generalized model trained on a generic corpus of data that is able to give a fairly accurate result for general use cases, without the need for additional training. GPT-3 is often considered a zero-shot model.
• Fine-tuned or domain-specific models. Additional training on top of a zero-shot model like GPT-3 can lead to a fine-tuned, domain-specific model. One example is OpenAI Codex, a domain-specific LLM for programming based on GPT-3.
• Language representation model. One example of a language representation model is Bidirectional Encoder Representations from Transformers (BERT), which makes use of deep learning and transformers well suited for NLP.
• Multimodal model. Originally LLMs were specifically tuned just for text, but with the multimodal approach it is possible to handle both text and images. An example of this is GPT-4.

The future of large language models

The future of LLM is still being written by the humans who are developing the technology, though there could be a future in which the LLMs write themselves, too. The next generation of LLMs will not likely be artificial general intelligence or sentient in any sense of the word, but they will continuously improve and get "smarter."

LLMs will continue to be trained on ever larger sets of data, and that data will increasingly be better filtered for accuracy and potential bias. It's also likely that LLMs of the future will do a better job than the current generation when it comes to providing attribution and better explanations for how a given result was generated.

Enabling more accurate information for domain-specific knowledge is another possible future direction for LLMs. There is also a class of LLMs based on the concept known as knowledge retrieval -- including Google's REALM (Retrieval-Augmented Language Model) -- that will enable training and inference on a very specific corpus of data, much like how a user today can specifically search content on a single site.

There is also ongoing work to optimize the overall size and training time required for LLMs, including Meta's LLaMA (Large Language Model Meta AI), which is smaller than GPT-3, though its backers claim it can be more accurate.

What's likely is that the future of LLMs will remain bright as the technology continues to evolve in ways that help improve human productivity.