If you’re not a tech person but still want to learn some basic AI knowledge, you have to watch this video, because I’ve condensed the content of Google’s 4-hour AI literacy course into this 10-minute video. Actually, I didn’t have high expectations for this course at first, because I thought it mainly talked about conceptual things, and our channel doesn’t talk about empty talk. Based on my understanding of Google, this course will most likely be removed in an hour. And I actually found that the underlying conceptual knowledge taught in this course enabled me to better use tools like ChatGPT and Google Bard, and also helped me clear up a series of misconceptions about AI machine learning and large language models that I was not aware of. So let’s start with the broadest question: What is artificial intelligence?

I can only admit with embarrassment that I just realized I didn’t know the answer before. Artificial intelligence, like physics, is an entire field of study. Machine learning is a subfield of artificial intelligence, just like thermodynamics is a subfield of physics. Going down one level, we have deep learning, a subfield of machine learning, and deep learning models can be further categorized into discriminative models, generative models, and large language models, or LLMs.

The technology behind applications such as ChatGPT and Google Bard, which we are all familiar with, also belongs to deep learning and overlaps with LLM and generative AI.

Could you please tell me in the comments if you knew this before? Now that we have a basic understanding of the field and how different disciplines relate to each other, let’s look at some key points at each disciplinary level. In general, machine learning is a process that uses input data to train a model.

The trained model can then make predictions based on new data. For example, if you train a model using Nike sales data , the model can then predict the sales performance of new Adidas shoes based on Adidas sales data.

The two most common types of machine learning models are supervised learning and unsupervised learning. The key difference between the two is that supervised learning uses labeled data, while unsupervised learning uses unlabeled data. In this supervised learning example, this scatter plot shows the relationship between the total bill and the tip amount for a single restaurant . And this data is labeled. Blue dots represent customer-pickup orders, and yellow dots represent merchant-delivered orders.

So using a supervised learning model like this, we can predict how much tip we’ll receive next time based on the total bill and whether the order is for pickup or delivery .

For unsupervised learning, we look to see if the raw data can be automatically grouped. For example, this graph shows the relationship between the number of years an employee has worked at a company and their salary income. We can see that the salary/year ratio of this group of employees is higher than that of the group below. We also know that this data is unlabeled.

If this data were labeled, we would see gender, years of experience, department, and so on. We can now use this unsupervised learning model to solve problems such as whether this new employee is growing rapidly.

If they are in the group on the left, the answer is yes. If it is on the right, then no. Pro tip, there is one more significant difference between the two models.

After a supervised learning model makes a prediction, it compares the prediction to the data used to train the model. If there is a difference between the two, it will try to narrow the difference, which an unsupervised learning model will not do. Oh, and this video isn’t sponsored, but you can support me by paying for a subscription to my Google Software Tips articles. You can visit my website at jeffsu.org/productivity-ping for more details.

Now that we have a basic understanding of machine learning, we can start to understand deep learning. Deep learning is actually a type of machine learning that uses artificial neural networks to learn.

Do n’t worry, all you need to know is that artificial neural networks are inspired by the human brain, and they look something like this: layers of nodes and neurons. The more nodes and neurons there are, the more powerful the model will be. Thanks to these neural networks, we can perform semi-supervised learning, which is training deep learning models using a small amount of labeled data and a large amount of unlabeled data .

For example, a bank might use a deep learning model to detect fraud. The bank will take some time to label 5% of its transaction data as fraudulent and non-fraudulent transactions. Since banks do not have enough time and resources to label all the data, the remaining 95% of transaction data is unlabeled data. The secret of this model is that it uses this 5% of labeled data to learn the basic concepts involved in a task, which are good and which are bad, okay. The model will apply what it has learned to the remaining 95% of unlabeled data, using this entire new dataset to make predictions about future transactions.

That’s cool, but it’s not over yet. There are two types of models in deep learning, namely discriminative models and generative models. Discriminative models learn the relationships between the labels of the data, and they can only classify these data points as fraudulent or non-fraudulent. For example, now you have a bunch of images, or data points. You now label them as dogs or cats.

The discriminative model will learn the label of a cat or a dog. If you submit a picture of a dog, it will predict the label of this data point: a dog. Now we finally get to generative AI. Unlike discriminative models, generative AI models learn the patterns of data. After receiving some input, such as a text instruction, they will generate new content based on the data patterns they have just learned .

Let’s use the animal example again. These images or data points are not labeled as cats or dogs.

The generative model will look for patterns, and see that these data points all have two ears, four legs, a tail, love to eat dog food, and bark. When you ask it to generate a thing called a dog, the generative model will generate a brand new picture based on the patterns it has learned before. There is a simple way to easily tell whether a model is generative AI.

If the output is a number, a category (like spam or not spam), or a probability, it’s not generative AI. It is generative AI only when it generates natural language (text or speech), images, or sounds . Generative AI generates entirely new examples similar to the data used to train it.

Next, let’s talk about different types of generative AI models. Most of us are familiar with text conversion models such as ChatGPT or Google Bard.

Other common model categories include text-to-image models such as Midjourney, DALL·E, and stable diffusion. They can not only generate images, but also edit them. Text-to-video models, surprise, they can generate and edit videos. For example, imagen video, CogVideo and the very creatively named make a video. Text-to-3D models can be used to create game assets.

A lesser-known example is OpenAI’s shape-e model. Finally, text-to-task models are trained to perform a specific task. For example, if you type @Gmail please summarize my unread emails, Bard will read through your inbox and summarize the contents of your unread emails. Next, let’s talk about large language models. Don’t forget that LLMs are also a subset of deep learning, and although there is some overlap, LLMs and generative models are not the same thing.

There is an important difference between the two, that is, large language models are usually pre-trained with a large amount of data and then fine-tuned according to specific purposes. What does this mean? Suppose you have a pet dog.

You can train it in advance to learn some basic commands, such as sit, come, lie down, and stay still. After learning it, it becomes a good dog and a generalist.

But if this good dog is to become a police dog, a guide dog, or a hunting dog, it needs more specific training to be fine-tuned into that particular specialist dog. The same is true for large language models. They are first trained to develop basic language processing skills, such as classifying text, answering questions, summarizing documents, and generating text. Then, smaller industry datasets are used to fine-tune these large language models to become industry experts to solve specific industry problems, such as retail, finance, healthcare, and entertainment. In the real world, this might mean that large language models pre-trained by major tech companies are fine-tuned by hospitals using their own medical data to improve the diagnostic accuracy of X-rays or other tests .

It’s a win-win situation because large companies spend billions of dollars to build a general model, or a large language model, and then sell these models to smaller institutions like retailers, banks, hospitals, etc. Although they have professional industry data to fine-tune the model, they do not have enough resources to build their own large language model.

Pro tip, if you want to take this free course, you can go to cloudskillsboost.google/course_templates/536. If you want to take notes, you can right-click on the video and copy the video URL at the current time, and you can quickly find that video clip.

This course has five modules, and you ‘ll get a small badge for completing each module.

Since this course is quite theoretical, you should definitely check out my video on how to master AI prompting skills. See you in the next video, and in the meantime, have a great day!.