Tech Talk: What Is AI?
 | Author:
Victor Sample
Vic Sample: MT43 News Treasurer |
Tech Talk: What Is AI?
Victor Sample
Last week, we discussed why Artificial Intelligence has become so important. In 2022, the company OpenAI made its AI engine, ChatGPT, available to the general public. Since then, AI has dominated technology.
It has become widely used to generate text and images. Many technology companies are laying off computer programmers, using the excuse that AI can create software faster and better than human programmers. Governments were rushing to create laws governing AI. Many “experts” are warning that AI will be the end of the human species.
In January of 2023, I was in an all-day committee meeting to define a new mission statement for one of the organizations of which I am a member. Our old mission statement was far too long and was out of date. We were trying to create a mission statement that was one or two sentences long that would effectively communicate what we are trying to do.
After hours of working on the mission statement, we felt we still weren’t really capturing what we wanted to say. One of the members suggested using ChatGPT to generate the statement. We were stunned at the result; it was amazing how ChatGPT really captured our mission in a very short statement. Ultimately, we did not use the text generated by ChatGPT, but we did use that generated statement to craft a mission statement that we felt accomplished what we wanted to say.
How did ChatGPT actually go about generating that text? In essence, AI engines are very mathematical. They are really using statistical probability to generate text. According to CoPilot (the Microsoft AI product), AI is a "Probabilistic Inference and Statistical Modeling” program.
AI engines are built around Large Language Models (LLMs), which are massive databases of statistical data derived from massive amounts of example texts. LLMs are “fed” the texts from books, essays, newspaper articles, web pages, etc., and build statistical data about words and how they are used.
When an AI engine generates text, it starts with the prompt you give it and then uses statistical probability to select the next word or phrase. It keeps doing that until it statistically determines the prompt has been answered.
The AI engine does not really understand what it said; it did not answer the prompt based on knowledge of the subject. It just generated the most statistically probable collections of words and phrases based on the massive amount of text it “learned” from.
According to CoPilot itself: “When you type a prompt, the AI doesn’t “know” the answer like a person might—it looks at all the words you’ve written, then uses patterns it has learned from reading billions of examples to guess, one word at a time, what’s most likely to come next, like a supercharged version of predictive text on your phone. Each word it adds is chosen based on probabilities, not certainty, so the response is more like a well-informed improvisation than a recitation of facts.”
Artificial Intelligence is an extremely complex mathematical process based on vector algebra. When I was a math student in college, I did a lot of vector algebra on small vectors. It was amazingly difficult and very time-consuming. AI is working with massive amount of large data vectors. It is an impressive process developed over decades of research.
AI can generate images as well as text. It is very much the same process. The AI engine produces the image based on analyzing the statistical probability of the string of pixels from the massive amount of images that it was “fed". According to CoPilot: “Imagine you’re describing a dream to a painter who’s never seen it—but who has studied millions of paintings, photos, and styles from across time. You say, "Paint me a castle in the clouds at sunset,” and instead of copying a picture, the painter begins conjuring it brushstroke by brushstroke. They don’t pull from memory—they predict, based on everything they’ve seen, what a cloud castle should look like, how the light might fall, what colors would feel right. Each pixel is like a brushstroke chosen not by certainty, but by probability—what’s most likely to match your vision, one layer at a time.”
I hope this article gives you a small understanding of what AI is actually doing. It is amazing how well it can generate well-structured, articulate answers to questions you ask, but as we will discuss in future articles, there are severe limits and problems with how AI engines generate text and images.
