Large Language Model Tutorial
Step 1: Welcome
Welcome to the interactive portion of our AI Guide!
This is a brief introduction on how to use ChatGPT, a product created by OpenAI that is driven by a large language model. If some of those terms are unfamiliar, please review the glossary in Part 1: AI Starter. We are using ChatGPT as an example, but the following information applies to other large language models, such as those developed by Microsoft, Meta, Google, Anthropic, and others. This is not an endorsement of ChatGPT over the others. Future tutorials will cover image generation, code, and more.
We’ll cover essential dos and don’ts, debunk common myths, and help get you prepared to use large language models on your own. Let’s go!
Step 2: Write your first prompt
The window on the right is the AI Pedagogy Project’s ChatGPT window. It works just like the ChatGPT interface that you will find on the OpenAI website.
A large language model, the technology that drives ChatGPT, is a type of AI system that predicts the most likely words to follow in a sequence in response to text that you enter, which is called a prompt.
Let’s write your first prompt! Try typing “What are you?” “Who are you?” or “Where are you?” and press enter. Notice that ChatGPT tends to write in a friendly and agreeable tone.
|Do not enter any sensitive data such as:
Rule of thumb: if you wouldn’t post it on social media, don’t enter it into a large language model! In many cases, the company (in this instance, OpenAI) has access to the data you enter.
Later, when we cover setting up your account to use ChatGPT on your own, we’ll show you a choice you can make regarding how your data might be used.
Step 3: Try the same prompt again
Now, enter the same prompt again and press enter. For example, if you asked “What are you?” in the last step, type the exact same message and press enter.
Notice that ChatGPT’s response is different from before. Unlike a calculator, large language models usually respond in a new way even if you ask it the same question.
Large language models are not like search engines either (such as Google), which generate the same results if you search something twice in a row. Instead, large language models use “next word prediction,” like autocomplete on your smartphone. Large language models are designed to offer variety and appear convincing, rather than be entirely accurate.
Step 4: Consider when you shouldn’t use a Large Language Model
Large language models, like all tools, are better at some things than others. Since they are designed to seem accurate rather than be accurate, there are many circumstances where large language models shouldn’t be used.
Engineering problems, medical advice, or financial decisions are all situations where imprecise or inaccurate information could be dangerous. These are situations where you should consult an expert or trusted source.
If you are in need of accurate information, do not use a large language model. Large language models are not designed to provide factual information. In fact, they can confidently state falsehoods! This is known as a “hallucination.”
You should assume that any information provided by a large language model is inaccurate, unless you validate it from an external, reliable source.
Step 5: Consider when you might use a large language model
ChatGPT can be useful for experiments with language that do not demand a single (or correct) answer. For example, you might use it for brainstorming ideas, summarizing large amounts of information, or workshopping your thoughts to receive “feedback” from the model.
Try entering any of these prompts into the chat window:
- Brainstorm three ideas that explain to students how literature can impact and shape society. Include examples.
- What are some introductory terms to know in the field of cognitive science?
- Write a sonnet about Hamlet in the voice of Shakespeare.
- What is the nature of love and its complexities?
Reflect on the responses. You may be impressed by the speed and volume of writing, but be critical of the content. Are the ideas exceptionally creative?
Notice that ChatGPT produces confident and quite convincing responses. Remember that the information should be considered inaccurate until validated. A large language model is effective at predicting and generating a seemingly reasonable answer while lacking understanding of any of the concepts.
Step 6: How do large language models work?
Have you used autocomplete in your email, word processor, or smartphone? Large language models work in a similar way. The large language model underlying ChatGPT analyzes the text that you input and generates a response by selecting the next most statistically probable word. It continues until a complete response has been generated.
How does a large language model determine which word is the most statistically probable? First, engineers “train” the model on a massive data set, which entails dumping a very large corpus of text into the model for it to find connections and trends. Then the model is “fine-tuned” through a combination of automated and manual methods.
It’s worth being critical about the labor practices surrounding large language models. The manual methods for fine-tuning are often exploitative. The human feedback helps refine performance so that the large language model does not use offensive language.
One limitation of the large language models that power ChatGPT is that it may include out-of-date training data. GPT-3.5 only includes information up to January 2022, and GPT-4 was last updated in April 2023. ChatGPT has “session-length memory” in that it remembers what you have stated in any particular conversation, but it won’t be able to answer questions about recent events.
Try typing “Tell me about your training data” and press enter.
- ChatGPT from OpenAI
- Gemini from Google
- Bing Chat from Microsoft, powered by GPT-4
- Claude from Anthropic
- StabilityLM (an open-source model as described in The Verge)
- ChatSonic, JasperAI, You.com, ShortlyAI, Sudowrite, CopyAI, Ryter, StoryMachines, and ChibiAI are a few of the many writing assistant apps that draw on GPT-3/3.5/4.
- Elicit.org, an AI academic research tool powered by GPT-3
- Perplexity AI, a question-answering tool powered by GPT-3 that includes linked sources
- Poe.com–a site with access to multiple AI systems, such as OpenAI’s GPT 3.5 and 4, Anthropic’s Claude, Google’s PaLM, and Meta’s Llama
- Write with AI in Google Docs (Google Workspace Labs beta)