First step in AI and LLM.

What is AI, and What Are Its Types?

AI stands for Artificial Intelligence. It’s a big part of computer science, first created in the 1950s. AI can do many amazing things, like recognize faces, help cars drive by themselves, predict the weather, and even make music.

There are four main types of AI:

1. Artificial Intelligence (started in 1956): This is the field where people build smart computers that can think like humans.
2. Machine Learning (started in 1997): A part of AI that lets computers learn from data so they can make better choices or guesses.
3. Deep Learning (started in 2017): A type of machine learning that uses “layers” of computer networks to understand complex information.
4. Generative AI (started in 2021): This type of AI creates new pictures, sounds, or writing based on what it has learned.

What is a Large Language Model (LLM), and What Does It Do?

A Large Language Model (LLM) is a special type of AI that can understand and create text. It can recognize letters, words, and sentences and figure out how they work together to make sense.

To do this, LLMs use a method called deep learning to study tons of text. Over time, they learn to recognize patterns in words and sentences, so they can answer questions or write text without human help.

LLMs are used for lots of things, like chatbots (for example, ChatGPT), and can help people write stories, answer questions, and much more.

How Are LLMs Trained?

To train an LLM, we need a huge amount of text (billions of words) so it can learn a lot about language and how words fit together.

There are two main ways to train LLMs:

The first way is call supervised Learning: The AI learns from labeled data, which means humans help it by telling it what’s right or wrong. For example, to help the AI learn numbers, we might show it lots of pictures of numbers with labels so it knows what each one is.

One of the most famous videos about Supervised Learning applications is about Convolutional Neural Networks (CNN) that identify numbers in picture. Link video

Another way to train LLM is unsupervised learning when machine learn by themself without human instruction.

Unsupervised learning is suitable for complex processing tasks such as organizing large datasets into clusters. They are useful for identifying previously undetected patterns in data that can help identify features useful for categorizing data.

To train an LLM, there are 3 stages required to do:

1. Gather Data: The engineer gives LLM a large dataset including words, sentences, and text like posts, articles, and websites. This dataset will include billions of words. For example, Llama3 model of Meta now is training based on 15T token (over 12 billions words). Model will be training based on these natural language and try to understand the context between words.
2. Fine-tuning: Engineer will fine-tune model with data, label all the correct answer so LLM can relearn it again and again
3. Reinforcement Learning: After LLM provider correct answer, apply Reinforcement Learning (a machine learning technique that trains software to make decisions to achieve the most optimal result) to adjust correct answer and choose the best answer for LLM improve.

Props and cons of LLM

Props:

• LLMs understand human language very well, meaning they can read and make sense of words, sentences, and even complex ideas, much like a person would
• LLM application like ChatGPT can save people time by answering questions quickly and can sometimes even give better answers than people.

Cons:

• Knowledge cutoff: LLMs can’t know about events after they were last trained. For example, they might not know about very recent news.
• Hallucinations: Sometimes, the model can produce outputs that are coherent and grammatically correct but factually incorrect or nonsensical

For example, when we compare 9.11 and 9.9 which number is greater in ChatGPT we will get this answer