What Is a Large Language Model (LLM)?

The history of computing is marked by sea change moments; those times when the world seems to shift into a new possibility space almost overnight. ENIAC. The personal computer. The World Wide Web. The smartphone. And now, AI.

While the term “AI” has been applied to many new (or re-branded) services and products, the underlying technology that makes most of them feel like magic is the large language model (LLM).

What is a large language model?

A large language model is a type of computer model that analyzes a massive amount of human language to build associations between words and how they’re used.

In some ways, a large language model works similarly to the autocomplete function on your phone. It’s all about knowing which word is the most likely to come next based on context. But while your phone’s autocomplete mostly looks at the previous word in the sentence, an LLM uses multi-dimensional word associations to analyze an entire passage, weigh each of the associations for the words therein, and then use that context to determine what word should follow.

This is how LLMs turn pages upon pages of intricate documentation into an easy-to-read summary and how they can draft an email response to your colleague: one word, with a deep (though sometimes flawed) understanding of how those words relate to each other, at a time.

Here are some more in-depth concepts behind large language models:

Transformers

A transformer is a kind of computer architecture that uses neural networks to process large amounts of data in a sequence. In the case of a large language model, that data is words, but it could also be frames of a live video feed for computer vision, vocalized sounds for speech recognition, and many other types of data. (The “T” in ChatGPT stands for “transformer,” while the “GP” notes that it’s general purpose.)

Transformers are built on the concept of allowing a model to devote its “attention” to multiple parts of a data set, regardless of when those parts were introduced. This enables transformers to observe and integrate associations that other models may overlook.

Training

The LLM meaning of “training” is the data the model uses to build its understanding of language. These training materials are often a substantial portion of the publicly accessible internet, including everything from books, scholarly research, and news articles to Reddit posts.

Whether it’s a verb, an adjective, a proper noun that refers to a patented product, or a meme phrase, the LLM will see how each of these words is used within its context and use that information to build its “understanding” of what other words each one is associated with.

This training is also essential to allow general-purpose LLMs to recognize user inputs in natural language and then generate new responses based on that input.

Fine-tuning

Fine-tuning gives a pre-trained LLM AI further training for a specific purpose. For example, a company that wants to automate its customer service interactions could use ChatGPT’s API to set up a support agent, then fine-tune the agent with the contents of the company’s product documentation.

The support agent would then use the broad capabilities imparted by ChatGPT’s training to understand and respond to user requests and its fine-tuned training to recommend solutions. Fine-tuning helps distinguish models meant for general public usage and enterprise-grade AI deployments developed for particular tasks.

Opportunities and risks of large language models

Large language models can distill clear insights from large amounts of data, whether that data is spreadsheets full of statistics or a transcript of a long video call. Organizations have also begun to see the benefits of combining LLMs with fine-tuning and APIs that allow them to autonomously operate other applications in response to one-time instructions from users.

The potential productivity gains made possible by these innovations are substantial. However, the risks of using LLMs or being exposed to their products are also significant. Here are a few notable examples:

• LLM data leaks: Since most LLMs run as cloud applications hosted by their creators, any data they handle must pass outside your organization’s network and into theirs. This opens the door to potential data leaks, either within the LLM itself if the data isn’t handled properly or at connection points exploited by threat actors.
• Hallucinations: LLMs don’t always have a perfect understanding of how words (and the concepts those words represent) relate to each other. Furthermore, the reinforcement models LLMs use may prioritize giving any answer over admitting they don’t know something. This can cause an LLM to confidently state incorrect information. Since these “hallucinations” blend in with other, correct information, they may be difficult to spot — and potentially costly when they sneak through into a final product.
• Legal and ethical issues: The U.S. Copyright Office has ruled that works created solely by AI may not be copyrighted, since they aren’t considered to have human authorship. Yet, in many other respects, governments and other regulatory bodies are still far from consensus on how large language models may be used and where. This leaves organizations that integrate LLMs into their day-to-day operations potentially vulnerable to shifts in their acceptable use.
• AI-empowered threat actors: Legitimate LLMs use guardrails to deter their use for malicious purposes, but those guardrails may be subverted — and models modified for illegitimate purposes may have no guardrails at all. Threat actors may use LLMs to expand and augment their capabilities with more convincing social engineering attacks or new malware programs.

Protect your organization in the age of LLM AI

So, what is LLM security in practice? It means building a robust governance model for your AI applications, as well as everyone in your organization who works with them. Your security must reach everywhere employees are likely to use or be exposed to LLMs wherever they work, not just at their desks.

Our mobile EDR playbook will help you find strategies to protect your data without limiting the impact of potentially transformative technologies such as LLMs. We’re eager to help you protect and empower your organization as you navigate a new age of computing.