2024 AI Quick Learning Guide
Recently, many friends asked me how to start learning AI. We all want to catch up with AI’s trends, understand the buzzwords on the news, and accelerate daily work using AI.
You don’t need a whole computer science degree; you don’t need to be a genius. We just need some learning strategies.
The Quick Way
I saw many of my engineer friends go straight to some online classes on machine learning, starting from support vector machines or even linear algebra and statistics. While the basics are important and these resources are great to have, most people however can’t keep the stake long enough, including me.
The first secret I wanna share is, you don’t need to learn everything. We can just go to the essential part first, and learn stuff along the way. You can follow a couple of KOLs on Twitter and subscribe to some newsletters. You can follow my YouTube channel too. The fields change quickly and the tech community shares the most up-to-date news, insights, and tips.
From the online resources, you’ll see tons of technical terms but don’t worry. For each new term, just find 1–2 sentences to describe the concept, and then keep moving. If some of the terms appear repeatedly, maybe they are more important, then we can find additional resources and learn more.
And, try out as many AI products as you can. ChatGPT for sure, and Midjourney or Stable Diffusion to draw some artwork, and maybe Suno for some AI music, and so on. After trying it out, you’ll understand the power and limitations of today’s AI tools.
What’s Machine Learning?
Let’s try a simple example of machine learning.
Imagine a system that has input and output. The input could be an image, and the output is the answer yes or no to whether the image is a cat.
Now we wanna build such a system from a neural network. Neural networks can have many structures and shapes.
For now, we feed in the neural network with a lot of animal images and, for each image, tell the system if it’s a cat. These images are called training data. Then the system starts to learn and look for underlying patterns, and then it will be able to classify and pick cats.
Quick quiz: If you have played around with ChatGPT or Midjourney, what are the input and output of these machine-learning models? And a much harder question, what kind of training data do you think they used? Leave your answers in the comment!
Online Classes Recs
If you want to go deeper and learn the maths and principles behind it, go for it. It’s always cool to lay a solid foundation. I’d recommend two series of free online classes.
The first is Andrew Ng’s classes via DeepLearning AI. Depending on your background, if you are non-technical, try AI For Everyone; if you’re a startup founder trying to incorporate AI into your products, try this, Generative AI with Large Language Models.
The second is Stanford’s CS324, Large Language Models. The class was given in winter 2022. After taking it, you’ll see why ChatGPT and GenAI are an inevitable trend.
Other classes may be cool too, just be careful about your choice. The majority of machine learning classes online haven’t incorporated the trendy stuff. I’ll share what technical concepts are missing in the middle.
Trendy Concepts Explained!
First, transformers. They are a type of neural network structure. Alternatively, we have CNNs, or convoluted neural networks in the old days. They are like a sliding window. When they slide over the input, they only read and learn from the small window.
On the other hand, transformers are like scanners; they can learn the whole thing and have this attention mechanism to magically pay attention to the part that matters.
For a visual transformer in 2D, same thing. For CNN, it’s hard to tell from a small patch if this is a cat, but transformers can scan the whole image and highlight important features that tell us it’s a cat. Now we know why transformers are so powerful.
Then, people realize they can use transformers to build language models, like ChatGPT. Why are they called large language models? Because they are large! They can have billions of parameters or variables a deep neural network learns during training. In the past, machine learning models were small, of only tens of thousands of parameters, and usually trained on a specific topic for a specific task.
Now, people found that these large models can handle a variety of tasks and can be general purpose. We call it foundation models, on top of which we can build specific applications.
There are foundation models for texts, for images, for videos, for 3D, for sounds, etc. And if it takes in multiple formats, for example, both text and images, then congrats, we have a multimodal.
What’s cool is that foundation models may have a performance boost when passing a certain point, suddenly becoming super good at some tasks. We call this phenomenon emergence.
When we’re not happy about the style or traits of the output, there are several ways. We can finetune a pre-trained model with additional input data, use LoRA which is a new research method, or use prompting which is to change the input.
Running the Code
After all of these concepts, in case you wanna run the models yourself, feel free to do so. We used to find models from GitHub but now I think HuggingFace is convenient to generate some quick results, no coding skills required.
If you wanna deploy a model yourself, there are free credits from Google Cloud and AWS up to a few hundred bucks that everyone can apply to. If you go crazy like me, da-da, Nvidia 4090 is a good choice.
Final Note
This is already a ton of information. In the next videos, we’ll talk more about popular AI tools. If you have questions or topics you want me to share, please let me know. Let’s study and improve together!
