Start Here First: AI Key Terms for Beginners

If you understand these terms, you can understand most things about AI

Terminology can be one of the biggest barriers to starting with AI. Terms and concepts are often explained using circular logic and may be confusing at the start, but there is nothing to be intimidated about. Start with these few and the rest will be easier to pick up as you progress.

Algorithm: AI algorithms are, quite simply, lists of instructions. created by smart engineers using computer software and at their core algorithms are just a list of instructions. That’s it. That’s all there is to it. Just a list of instructions that computers follow in order to complete a task. To the right is an algorithm for brushing teeth. Yes, it's that simple. It's not a list of instructions a computer can follow but there are many types of algorithms. (see image below)

What's different about AI alogrithms?

AI algorithms are written using software programs so computers can perform different types of tasks. through trial and error without input from a human. That is what makes AI algorithms different from other types of algorithms.

The “trial and error” part of the AI algorithm (software/code) is considered a form of AI “learning." This is explained below under

Machine Learning.

FYI only - There are many types of alogrithms, different from those used in AI, which are designed to tackle different types of problems. Here are just a few.

Chatbot: A chatbot is an AI software program/algorithm that is designed to communicate with people through text or voice commands in a way that mimics human-to-human conversation. When you hit the "chat" or "help" button on the bottom of a website screen and type in a question, that is a chat bot you are using. Another way of thinking about it is to see a chatbot as an incredibly sophisticated Q&A software program that can provide highly detailed answers, and even anticipate what follow-up questions you may have. More about "anticipating" and "predicting" will follow.

Machine learning: When AI software is written so it can improve the accuracy of responses over time, without the help of a human, that is considered machine learning. This definition is not all-encompassing, and purists may scoff, but it is a useful working definition.

One way to improve the accuracy of responses generated by Chat GPT or Gemini - also known as "training" the AI software - is to give it the same problem to solve over and over. For example, you instruct Chat GPT or Gemini to complete the following sentence, "The quick brown _______ jumps over the lazy dog". If the response is anything other than "fox" the AI software is judged as having "failed" and the request is made repeatedly until the AI software returns the correct "fox" response. If the software does not return the correct response within a certain number of attempts the AI software coding (list of instructions) will be changed by engineers. That process continues until correct responses are returned.

Stated a little differently, when responses are generated by AI software and the responses are compared to a known correct answer ("fox"), that process is called "training a model." Machine Learning and Training a Model are different sides of the same coin. Think of Machine Learning as a broad definition what is happening and think of Training a Model as how machines are learning (getting better over time without human intervention).

An example using images instead of text: if you go to the AI image-generator site, Dall-E, and type, "show me an apple," the Dall-E AI software has already been trained to "know" what an apple looks like. When we say Dall-E "knows" something, we mean it has a database full of information about the attributes of an apple - "round/oval, red/green, baseball-sized," etc.

When it was "trained," the Dalle-E software was given a command to find pictures in different databases matching the attributes of apples, and it draws on those images as references for future images it will create later. The AI software was designed to spot patterns of similarity among all of the images it scanned and correctly identified as an apple.

More concretely, here is what happened when Dall-E was trained to create an image of an apple (not exactly, but the general idea is there):

• It scanned pictures of 500,000 apples in a database it was instructed to examine.

• It identified patterns of similarity between all 500,000 images.

• When the user asked Dall-E to "show me an apple" it constructed images that drew upon those similarities.

• After placing the images it generated onto the user's screen, it waited to see if the images met the requirements of the person who requested them.

• If it didn't meet the user's needs, the software would go back into its database of apple images, and its patterns of similarity, again to generate slightly different-looking images for the user.

• When the user selects one of the new images as "correct" or "good," the Dall-E AI software has "learned." It has one more data point and can make its responses better for the next time.

In the example above, Dall-E interpreted "show me" as an additional request for images of an apple. So it created images with people showing an apple to someone. Nicely done Dall-E.

This process is one example of Machine Learning, and it is one way to "train a model." There are a dozen different ways to train AI software. They are incredibly sophisticated and amazing feats of engineering, and they are evolving at incredible speeds. This video does a great job of explaining training models and "foundational models" in more detail.

How do you like them apples? Making an AI model better through "training": At its core, we know an AI model is a software program (also called an algorithm, which is a set of instructions). What turns a simple software program/algorithm into an AI model is the process over time of making the program/algorithm more accurate in its responses to our questions.

If we wanted to create AI software like Dall-E, capable of "generating an image of an apple" what where would we start? Well, we would engineer the software of course. But then what? We would have to train the software, which is a fancy way of saying we have to test it and see if it generates the correct results for us.

Training Data: As soon as we (AI engineers) complete the first draft of our AI software, we have to test it with questions or requests to see how well it works. Something like, "generate an image of an apple." Rather than having our AI software search the entire internet for pictures of apples, let's instead have it search only in a database of fruit pictures we have stored somewhere on our private network. Our fruit picture database is massive, but it is not the whole internet so it's manageable. It contains billions of pictures of different fruits. But only fruits, that's. So when we ask it questions or make requests, we can judge whether the software is working correctly. This database of fruit pictures is our training data.

If we ask our new AI software program to search our fruit picture database for apple images, it's possible to know when our software returns a correct or incorrect response each time we click "search for an apple." When an incorrect image is returned, our AI software program marks the image as "error" or "negative" or something like that. When it correctly finds an image of an apple, our AI software program marks the image as "correct" or "positive."

Here is the key to an AI model - over time our AI software gets better at returning the correct answers. Eventually, when we search for an apple we only receive correct responses. When the software improves - on its own - over time through trial and error, that is called "training the model."

Now imagine our AI software wasn't limited to searching for apples. Instead, if we trained the

model to distinguish between all fruits, that would be something big. And even further, if we "trained the model" to distinguish between ripe and rotten fruit, that would potentially, at some point, be very useful for farmers who are sorting real fruit coming in from their orchards. For example, set up a camera to photograph/video each piece of fruit as it comes into the barn on a conveyor belt. Using AI software, recognize the rotting fruit and sort it away from the ripe fruit. What a time and money-saving use for a simple AI software that has "learned" how to recognize different kinds of fruits.

All AI software must be trainied at some point, and the training is done in slightly different ways but you get the idea with this one apple example. An AI model relies on training data (e.g. - a fruit picture database) to recognize patterns and make predictions or decisions. Our apple example didn't talk about predictions or forecasts, but that is next.

The more data points an AI model receives - that is, the more times the software searches for apples and gets a correct/incorrect response - the more accurate it can be in its data analysis and forecasts.

One last point: I mentioned there are different ways to train an AI Model. It is helpful to know the two big concepts for training an AI software program using different approaches:

Supervised learning: This is a type of machine learning/model training where structured datasets, with inputs and labels (like the example above with a single database of fruit pictures), are used to train and develop the algorithm that is being "trained."

Unsupervised learning: This is a much more sophisticated form of training where the AI software/algorithm program is programmed to make inferences from datasets that don’t contain labels. No criteria are given to evaluate what an apple looks like, for example. These inferences are what help it to learn. There will be more on this in later blog posts. For now, understanding models and the concept of training them is great!