Welcome to the fascinating world of Large Language Models (LLMs) and the basics of prompting techniques! In this first lesson, we're going to dive into what LLMs really are. Spoiler alert! They’re essentially next-word prediction machines. This might sound simple, but there's a lot more than meets the eye. Whether you're completely new to coding or just curious about how these technological marvels work, you're in the right place. Let's embark on this exciting journey together and unwrap the mysteries of LLMs, starting with their core functionality.
Imagine you’re writing a text message or an email, and your phone suggests the next word you might want to type. That's a very basic example of what Large Language Models (LLMs) do. However, LLMs like GPT-4o, Claude 3.5 Sonnet, and LLaMA are like the superheroes of word prediction. They don't just suggest the next word in a sentence; they can generate whole paragraphs of text that make sense based on the input they receive. They do this by sequentially predicting the next word that continues the text they've already got.
At their core, LLMs analyze vast amounts of text data. Through this analysis, they learn patterns, nuances, and the structure of language. This enables them to predict what word naturally comes next in a series of words. It's like they're constantly playing a game of “fill in the blank,” but at an astonishing scale and speed.
You might wonder how LLMs are able to make these predictions. Well, it's all about training. LLMs are exposed to huge datasets containing all sorts of textbooks, articles, websites, and more. During this training phase, they learn to understand the context and flow of language. They pick up on things like grammar, style, and even the tone of the text.
When you prompt an LLM with a sentence or a question, it uses what it has learned to predict the most likely next word or words to follow. This isn't just a wild guess; it's a calculated prediction based on the patterns and rules it has observed during its training.
Let's consider some examples of one-word predictions to gain a deeper understanding. Let's say an LLM is given the prompt "better late than". It is highly likely to predict "never" as the next word.
Now, consider a prompt "a bottle of". Most LLMs will predict that "wine" is the next word. However, there is a chance an LLM will predict "water" or something else. When an LLM makes predictions, it doesn't always choose the most probable next word. Instead, it incorporates a degree of randomness to ensure that the text it generates isn't always the same for a given prompt.
During practice exercises, you will work with LibreChat – an application that allows you to chat with various LLMs. In our environment, we have a limited list of LLMs to use. The interface is straightforward: just send a message and get a response.
Note:
Entersends a messageShift + Enteradds a newline to the text message
You can also open a sidebar to access additional settings:
In the sidebar, you can see your chat history and create a new chat. Note: Cosmo will evaluate the entire chat history when checking your solution, paying close attention to the latest chat.
You can manage a chat by clicking on three dots near its name.
You can use it to delete chats where you experimented with prompts or received intermediate result, so Cosmo doesn't take them into account.
Given the probabilistic nature of LLMs, the challenge for Prompt Engineers is to guide LLMs towards highly predictable and accurate outcomes as consistently as possible.
As part of this course, you'll learn many techniques that will allow you to master the art and science of highly predictable LLM outputs. But before we go too far, let's start with a few simple practice exercises to get our gears turning.
