An AI chatbot such as ChatGPT, Claude, Copilot, or Gemini may appear to be speaking with a human when you ask it to perform something. They are able to respond to you with well-written, properly grammatized, and persuasive emails, notes, essays, or summaries of search requests.
However, you’re not interacting with a human. These chatbots aren’t as good at understanding word meanings as humans are. Rather, they serve as the interface via which we communicate with large language models, or LLMs. These underlying technologies can anticipate words, sentences, or paragraphs in the future since they have been trained to identify word patterns and commonly occurring combinations.
Several generations have passed since LLMs first evolved. GPT-4o was released by OpenAI in May, followed by GPT-4o Mini in July and OpenAI o1 in September. Google offers two variants: the 1.5 Flash and the 1.5 Pro. Anthropic has reached Claude 3.5, while Meta is now at Llama 3.
If you’re wondering what LLMs have to do with AI, this explainer is for you.
What is a language model?
A language model can be compared to a word-soothsayer.
According to Mark Riedl, an associate director of the Georgia Tech Machine Learning Center and professor in the school of interactive computing at Georgia Tech, “a language model is something that tries to predict what language looks like that humans produce.” “What makes something a language model is whether it can predict future words given previous words.”
This is the foundation for AI chatbots and text-to-autocomplete capability.
What is a large language model?
A large language model contains vast amounts of words from a wide array of sources. These models are measured in what is known as “parameters.”
What’s a parameter?
Neural networks are machine learning models that take an input and utilize mathematical calculations to produce an output. This is how LLMs use them. Parameters are the quantity of variables used in these calculations. One billion parameters or more can be found in a large language model.
“We know that they’re large when they produce a full paragraph of coherent fluid text,” Riedl stated.
Do small language models actually exist?
Yes. Tech companies like Microsoft are rolling out smaller models that are designed to operate “on device” and to not require the same computing resources as an LLM but nevertheless help users tap into the power of generative AI.
What’s under the hood of a large language model?
When Anthropic mapped the “mind” of its Claude 3.0 Sonnet large language model, it found each internal state (“what the model is ‘thinking’ before writing its response”) is made by combining features, or patterns of neuron activations. (The artificial neurons in neural networks mimic the behaviour of the neurons in our brains.)
Anthropic was able to visualize a map of Claude 3.0 Sonnet’s internal states as it produced replies by taking these neuron activations out of the system. The AI startup discovered that neuron activity patterns were associated with abstract notions such as computer code errors, gender bias in the workplace, debates about preserving secrets, cities, people, atomic elements, scientific areas, and programming syntax.
How do large language models learn?
Deep learning is a fundamental AI technique that LLMs use to learn.
“You show a lot of examples, just like when you teach a child,” remarked Momentum Worldwide’s worldwide CTO, Jason Alan Snyder.
To put it another way, you feed the LLM a library of content (also referred to as training data) that includes things like books, articles, code, and postings from social media platforms to help it learn about various contexts in which words are used as well as the finer points of language. This model processes billions of tokens, which is significantly more information than any human could possibly read in their lifetime.
Tokens help AI models break down and process text. You can think of an AI model as a reader who needs help. The model breaks down a sentence into smaller pieces, or tokens—which are equivalent to four characters in English, or about three-quarters of a word—so they can understand each piece and then the overall meaning.
The LLM is continuously improving its comprehension of language, becoming more adept at spotting patterns and forecasting future words because this prediction and correction process occurs billions of times. It can even create original text formats, translate languages, and learn concepts and facts from the data to respond to queries. However, their comprehension of word meaning is limited to statistical associations, unlike ours.
LLMs also learn to improve their responses through reinforcement learning from human feedback.
What do large language models do?
Given a series of input words, a LLM will predict the next word in a sequence.
For example, consider the phrase, “I went sailing on the deep blue…”
Most people would probably guess “sea” because sailing, deep, and blue are all words we associate with the sea. In other words, each word sets up context for what should come next.
What do large language models do really well?
LLMs are very good at figuring out the connection between words and producing text that sounds natural.
“They take an input, which can often be a set of instructions, like, ‘Do this for me’ or ‘Tell me about this’ or ‘Summarize this’ and are able to extract those patterns out of the input and produce a long string of fluid responses,” Riedl said.
What are large language models weaknesses?
First, they’re not good at telling the truth. In fact, they sometimes just make stuff up that sounds true, like when ChatGPT cited six fake court cases in a legal brief or when Bard mistakenly credited the James Webb Space Telescope with taking the first pictures of a planet outside of our solar system. Those are known as hallucinations.
“They are extremely unreliable in the sense that they confabulate and make up things a lot,” Sap said. “They’re not trained or designed by any means to spit out anything truthful.”
They also struggle with queries that are fundamentally different from anything they’ve encountered before. That’s because they’re focused on finding and responding to patterns.
And while they excel at predicting words, they’re not good at predicting the future, which
Finally, they struggle with current events because their training data typically only goes up to a certain point in time and anything that happens after that isn’t part of their knowledge base. And because they don’t have the capacity to distinguish between what is factually true and what is likely, they can confidently provide incorrect information about current events.
They also don’t interact with the world the way we do.
It is difficult for them to grasp the nuances and complexities of current events that often require an understanding of context, social dynamics and real-world consequences,” Snyder said.
How will large language models evolve?
Multimodal models, which are trained using images, video, and audio in addition to text, are already being introduced by generative AI companies such as OpenAI, Google, and Adobe.
Retrieval capabilities are also evolving beyond what the models were trained on; for example, they can now connect to search engines like Google, allowing the models to perform web searches and feed the results back into the LLM. This implies that they might respond to inquiries more quickly and comprehend them better.
“This helps our linkage models stay current and up-to-date because they can actually look at new information on the internet and bring that in,” Riedl said.
