Decoding AI Jargons with Chai

Artificial intelligence is the hottest technology of this decade. Everyone wants to learn as well as know about artificial intelligence. But very few know the basic concepts that are crucial to know about artificial intelligence.

In this short article, I am going to explain 10 AI jargons which are important to know so that anyone can understand the working of Artificial Intelligence without getting overwhelmed.

1. Large Language Model (LLM)

A large language model is a type of AI that uses some algorithms which can recognise patterns, inspired by the human brain. The LLM is trained on massive datasets and is able to generate human language. LLMs can perform tasks like text generation, translation and question answering. The most popular application ChatGPT is built on a Large Language Model.

2. Generative Pretrained Transformer (GPT)

The one of the hottest jargons in the market is GPT, primarily popular because of ChatGPT.

The Wikipedia describes GPT as,

A generative pre-trained transformer (GPT) is a type of  large language model  (LLM)[1][2][3]  and a prominent framework for  generative artificial intelligence.[4][5]  It is an  artificial neural network  that is used in  natural language processing  by machines.

Many of you will get confused while reading this because of theses concepts such as generative artificial intelligence, artificial neural network. But we will understand it in a simpler language.

We can simply understand GPT as an advanced model that utilises transformer architecture to generate a human-like text. It is trained on vast amounts of data from the internet. This model is able to understand and generate coherent and contextually relevant text.

3. Transformers

A Transformer is a kind of neural network that takes a sequence of information (like words in a sentence) and turns it into another sequence (like a translated sentence). It works by understanding the meaning of each part and how the parts relate to each other.

For example, the input is “Is dog an animal?” The transformer model uses mathemtics internally and understands the relation between words such as “dog” and “animal”. After the understanding, it will generate the output like, “Yes. Dog is an animal”.

4. Encoder & Decoder

These are important components of transformers. Encoder gets an input sentence and transforms into vectors (mathematical term) and decoder generates the output sequence based on these representations while keeping the focus on relevant parts of the input.

5. Embeddings

Embeddings are numerical representations of the data submitted by the human. Inputs like text, images and audio are getting converted into mathematical representations that allow machine learning models to understand and process complex information by capturing the semantic meaning and relationships between data points.

6. Semantic Meaning

As a human, we use lots of words with more than one meaning. That makes the interpretation of words difficult to the computer. For example, when I say, “I am a huge fan of Tom Cruise”, I mean that I admire or idolise Tom Cruise. But when I say that “My fan is not working properly”, that means the fan, an electric device, is broken. In AI, the semantic meaning is the meaning of the word based on the context from the entire sentence.

7. Self Attention

Self-attention helps the model figure out how different words in a sentence are related to each other. This helps the model understand the meaning of the sentence better and make smarter decisions. It is a key component of a transformer.

8. Softmax

Imagine a neural network predicting the type of animal in an image (cat, dog, or bird). The softmax function would take the network's output (a vector of numbers representing the "strength" of each prediction) and convert it into a vector of probabilities: 0.8 (cat), 0.1 (dog), and 0.1 (bird).

9. Temperature

You should be familiar with the temperature word if you have used a GPT application apis. The temperature refers to a parameter that controls the creativity of generated response. By lowering the temperature parameter, you are restricting the model to process the input with hight probability. If you want more creative response for the input, you can increase the temperature value of a model.

10. Tokenization

It is crucial part of the text analysis. When we input the text to the model, the first task to do is to break down complex text into smaller units, known as tokens. These tokens can be words, symbols or phrases. The tokenization process simplifies the text analysis.