What are... large language models?

A monthly tech explainer series about the technology shaping our world today, from the Garage.

By David Rand — July 11, 2023

Why have artificial intelligence (AI) chatbots like OpenAI’s ChatGPT, Google’s Bard, and Microsoft’s Bing become so sophisticated and usable? There’s a simple explanation: Large Language Models (LLMs).

LLMs have only been around for about five years, but recently these deep learning algorithms have advanced so rapidly that you can hold natural, human-like conversations with chatbots and actually get value from them.

In fact, even though ChatGPT — the flag-bearer for this technology — only came out in November 2022, more than half of US Internet users have tried it. Granted, most folks are just experimenting. But many others are finding value in these tools for research, generating ideas, and crafting remarkably proficient essays, articles, resumes, and emails. And Wall Street has been going gaga over anything even remotely AI related, in part, because of what LLM technology makes possible.


RELATED: What is… AI? 


How it works

To understand how LLMs work, think about them like black holes. The more matter they consume, the bigger and stronger they become. LLMs have similarly insatiable appetites — for data. The more they’re fed, the better they function, which is why they are called “large” language models.

LLMs are built on neural network “transformers,” which detect how words relate to one another in order to process and understand natural language inputs and spit out human-like responses. But of course, as with any AI, you have to “train” the model to be able to access, understand, and transform data. The more learnable “parameters,” or variables, your LLM has for finding patterns in data, the more advanced its reasoning capabilities will be.

Illustration by Eric Chow

The a-ha moment

While LLMs have progressively improved since 2017, they took off like a rocket in large part because of generative AI, a type of artificial intelligence that can produce original content like text, images, and video on demand. Indeed, each version of the pre-trained transformer models underlying LLMs has added exponentially more parameter capability. For example, GPT-4, released by OpenAI in March, is thought to have around 1 trillion parameters compared to 175 billion for GPT-3.5, released in March 2022. 

There are three main types of LLMs: public models available to anyone; private models developed for use within enterprise organizations; and hybrid models that use both public and custom data. For example, there are four GPT models currently available from Open AI and Azure Open AI, and they are composed of four variants ranging from 350 million parameters on the small end to 175 billion parameters for the largest. 

What LLMs mean for everyday life

When talking about LLMs, we naturally turn to souped-up search engines like ChatGPT because, more than anything before them, they are consumerizing artificial intelligence.

Anyone can download these powerful apps for free and test them out. But they’re not perfect. They can (and do) make mistakes, such as plagiarizing content or making stuff up. But like any tool, if you use them as starting points or glorified copy editors, they can offer great time savings.

That’s today. Tomorrow, LLM-like models could very well influence other prolific digital technologies. Amazon, for example, is reportedly building a more “generalized and capable” LLM for its Alexa virtual assistant. LLMs are also  influencing the direction of existing search engines and could eventually completely change the way in which users interact with them. Opera and Microsoft Edge, for example, have announced browser features using LLMs, and others are expected to follow suit. LLMs will also likely transform many commonly used apps and tools, such as translators, spell checkers, and antivirus programs. 

What’s more, they could pave the way for a wide variety of new industry-specific applications. Google, for example, is reportedly testing an LLM-based medical chatbot called Med-PaLM 2 for answering patient questions more effectively. Heck, LLMs could even play a larger role in making humanoid robots like Ameca and Sophia smarter one day.

LLM technology is changing so quickly that they could be vastly different in just a few years. All we can say for sure is: Get ready. It’s going to be an interesting AI ride that will likely surprise us all.


Learn about how chatbots work.