What is a Large Language Model? The LLM tech powering GenAI explained (2024)

Large Language Models, also known as LLMs, are the tech that powers much of the advanced GenAI tech we’re seeing on laptops, phones and other technology in 2024 – but what is an LLM, and what exactly does it do?

In essence, LLMs are a type of artificial intelligence trained on gigabytes (if not terabytes or petabytes) of data to interpret human language and generate outputs in the form of text, audio, imagery and more – but there’s much more to it than that.

Here, we explain everything you need to know about Large Language Models and how they power popular chatbots like ChatGPT and Google Gemini.

What is a Large Language Model?

In its simplest form, a Large Language Model (also known as LLM) is a type of artificial intelligence that can recognise and generate text – though LLMs can also specialise in elements like photo generation, video generation, music creation and much more. This is essentially the underlying tech that powers Generative AI tools like ChatGPT, Google Gemini and Microsoft Copilot.

To achieve the task, the LLMs are trained on absolutely massive sets of data – hence the name – and utilise machine learning to understand what’s being asked of them, and generate something new based on that.

For context, most LLMS are trained on data found on the internet, possibly millions of gigabytes worth of text from every corner of the web, to gain as much information as possible.

However, the quality of the samples will impact how well the LLM performs its duties, so specialised LLMs may use a more curated data set. For example, a LLM trained exclusively on French language data wouldn’t be able to generate a story in English, and vice versa.

What can you use Large Language Models for?

As noted earlier, LLMs are the backbone of the Generative AI assistants we’ve seen appear over the past few years, from ChatGPT to Google Gemini and practically any other GenAI tool you can think of.

While the potential uses of GenAI and, thus, LLMs are continually expanding, the current iteration seems to focus on several key areas.

The most obvious is copywriting; LLM-powered chatbots like ChatGPT can write completely original copy based on a description that you give. This can be anything from a short children’s book to a step-by-step guide to cooking the perfect steak depending on what you ask it to do.

Similarly, LLMs are also great for answering queries about a specific product, known as knowledge base answering.

This is essentially when a company trains an LLM exclusively on its product or service, which consumers can then use to answer basic (and complex) questions without having to search the web or speak to a real person. It’s handy not only for finding out more about a product before you buy, but can also be handy for troubleshooting said product once purchased.

LLMs have also been a lifesaver for coders, generating code in a variety of coding languages based on developers’ descriptions. You won’t be able to create a new app or game entirely using ChatGPT without at least a passable knowledge of coding, but it can be a massive time saver.

Then there’s the big one; image generation. This is likely the most controversial use of LLM-powered GenAI services right now, as you can essentially get AI to create whatever you describe.

That’s all well and good until you get into the murky waters of misinformation and how easily you can create viral fake news just by using AI-powered image generators. Most popular image-based tools have limitations on the kinds of images they can generate, but particularly dedicated people can often find a way around these limitations.

What are some of the limitations of Large Language Models?

Large Language Models can do a lot of good, but it’s worth noting that there are some limitations to the tech as it stands.

The biggest issue with LLM-powered chatbots right now is hallucination. It’s a relatively new term in the world of artificial intelligence but it essentially means that the LLMs essentially create fake information when they can’t produce an accurate answer. This could be because the LLM wasn’t trained on that specific dataset, but it can sometimes just happen in regular chatbot conversations.

This is why it’s important to at least have a passing knowledge of what it is you’re getting LLMs like ChatGPT to describe or create. A seasoned coder, for example, could spot hallucinations in generated code while a newbie would take it at face value and paste it right into their project, while someone interested in phones would notice an inaccurate spec if asking an LLM about a new release.

There’s also the issue of privacy; some users may upload confidential documentation, or include confidential information in their queries, but LLMs use the inputs they receive for further training. This means that confidential information may be exposed in response to questions and queries from other users.