1.3: How do chatbots come up with text?

How do chatbots come up with text when we ask them a question or give them a prompt?

Here’s a set of jargon-free explanations of increasing length and specificity:

• AI uses math/statistics
• AI uses math with words / statistics with words
• AI uses statistical prediction with words
• AI uses statistical prediction of next words based on lots of text
• AI uses statistical prediction of next words based on patterns in a large chunk of the text on the Internet.

To oversimplify a bit, chatbots take the words you give them and assign numbers to them. Then they feed those numbers into a complex formula developed automatically during their training.

1. The human gives a chatbot some words (the prompt).
2. The chatbot assigns numbers to the words.
3. The chatbot feeds those numbers into a complicated formula it came up with during training.
4. The formula spits out some numbers that correspond to predicted next words.

How do the chatbots get trained? This is a time, money, energy, and data-intensive process that involves processing a huge amount of text to come up with a mathematical formula that encapsulates patterns in that text. Here are the steps in the training:

1. Start out with a random guess as to what formula would get us useful next words.
2. Try this formula on the first part of some text you already have.
3. See how well the formula predicts the rest of the text you have.
4. When it’s not right, adjust the formula so it’s better at predicting the text you already have.
5. Keep doing that until it’s good at predicting the text you have.
6. Use that adjusted formula to predict text with new prompts.

So you could say chatbots are answering the question, “Given the patterns in all the training text, what word is mathematically likely to come next?” You could paraphrase that as “Based on much of the Internet, what would a human say next?” Chatbots answer these questions over and over to come up with a series of words and serve it to us.

Then there’s another layer of training where either humans or AI or both rate chatbot performance. The ratings are used to adjust the chatbot formulas to make them more likely to give higher-rated answers.

Yet another layer comes when you give a chatbot extra information to focus on. You might upload an image, a document, or a spreadsheet that you want it to consider in addition to your instructions. Or the chatbot might be allowed to do searches on the Internet or other data and take what it finds into account when it gives an answer.

What powers chatbots is still statistical word prediction, but that capability will continue to be revised and extended as software products combine them with other tools.

Please take my explanations with a grain of salt; they are approximations of what is really going on in these systems. Really chatbots don’t predict whole next words but rather chunks of words called tokens. Would you like to learn more? Want to read about large language models, (LLMs), Natural Language Processing (NLP), neural nets, tokens, weights, transformers, attention, constitutional AI, reinforcement learning from human feedback (RLHF), and retrieval-augmented generation (RAG)? Don’t be intimidated! You can find explanations at many levels of difficulty and specificity. A few popular ones are listed below.

• An Introduction to Large Language Models, a video by former OpenAI engineer Andrej Karpathy
• What Is ChatGPT Doing … and Why Does It Work? By Stephen Wolfram
• Language Models: A Guide for the Perplexed by Sofia Serrano, Zander Brumbaugh, and Noah A. Smith
• A Very Gentle Introduction to Large Language Models without the Hype by Mark Reidl, April 13, 2023
• How AI Works: An entirely non-technical explanation of LLMs By Nir Zicherman
• How AI chatbots like ChatGPT or Bard work – visual explainer by Seán Clarke, Dan Milmo and Garry Blight, The Guardian
• Talking about Large Language Models by Murray Shanahan