Overview

[email protected]
HYML7P0S3U
ChatGPT and Generative AI

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Agenda

● What is ChatGPT?
● What are some applications of ChatGPT? What can it be used for?
● OpenAI, Microsoft and their role in ChatGPT
● The Golden Age of Generative AI - Popular Examples
● Transformers and a High-level Understanding of ChatGPT
[email protected]
HYML7P0S3U
● Prompt Engineering
● Limitations of ChatGPT
● Beyond ChatGPT

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 ChatGPT

ChatGPT sounds like something straight out of a Sci-Fi movie.

But it’s no oracle - it’s just a very smart AI model used as a chatbot

Think of it as your own personal genie with some wish-granting capabilities!

[email protected]
HYML7P0S3U State of the art Language Model with billions of parameters

ChatGPT Developed by OpenAI, a cutting-edge AI research organization

Variant of the GPT series of models from OpenAI

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Artificial Intelligence, Machine Learning and Deep Learning

A machine that mimics human intelligence

Artificial Intelligence
and behavior

A machine that automatically learns from data

Machine Learning with respect to a particular task and performance
[email protected] measure
HYML7P0S3U
A subset of Machine Learning that happens via
Deep Learning algorithms inspired by the neural networks of
the human brain
ChatGPT

A state-of-the-art Deep Learning based Chatbot that uses a

Generative AI and Transformer-based architecture

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 ChatGPT
Applications and what can you use it for?

Text Summarization Language Translation

Content
[email protected] Creation Sentiment Analysis
HYML7P0S3U

Market Research Education/Training

Code Debugging App Development

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 OpenAI & Microsoft

OpenAI is a cutting-edge AI research organization that aims to

promote and develop friendly AI in a way that benefits humanity.

OpenAI played the role of both the developer and the trainer of the
ChatGPT model, using vast amounts of text scraped from the web.
[email protected]
HYML7P0S3U

Microsoft, on the other hand, signed a partnership with OpenAI to

provide Azure supercomputing services exclusively for OpenAI to
train large language models, and become OpenAI’s preferred
partner to commercialize new technologies.

Microsoft has hence started integrating ChatGPT-like technology

into its Bing search engine, Edge browser, Azure cloud platform
and will also do so for Office products for consumers worldwide.
This file is meant for personal use by [email protected] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Breaking Down ChatGPT
Conversational Chat interface that allows the user to ask questions and interact with the
Chat

model in a to & fro manner. Conversational AI interfaces and Chatbots have been a
captivating application of AI for various business use cases.

Generative: A class of Deep Learning and AI models that are trained to generate data
G

modalities in a manner similar to humans, such as text, images, audio or video.

[email protected]
HYML7P0S3U

Pre-trained: A type of training technique used for AI models, where a model may be trained
from scratch on a large initial set of data, and this “pre-trained” model is then further
P

fine-tuned and customized for a different, specific task.

Transformer: A new class of Deep Learning models based on the

ideas of attention and self-attention. With text, Transformers
T

understand the relationship of one word with all other words in a

sentence. This file is meant for personal use by [email protected] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Generative AI
1 2

Which of these faces are real?

And which of these are fake?

They’re all fake!

These are all AI-generated

[email protected] faces,
HYML7P0S3U
not images of real people.

The key innovation behind this technology,

Generative Adversarial Networks (GANs) 3 4

were invented in 2014. Source - this-person-does-not-exist.com

The field has come even further since then.

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Generating Images from Text

“Teddy Bears mixing sparkling chemicals as mad

scientists, in a 1990s Saturday Morning Cartoon.”

Et voilà!

The state-of-the-art in Image-related Generative

[email protected]
AI, another OpenAI model called DALL-E 2
HYML7P0S3U
(2022) is capable of generating images relating
to any such input prompt.

The quality of the outputs generated by these

large models and the speed at which they’re
capable of creating them is what has fueled
the rise of Generative AI.

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 The Golden Age of Generative AI
Advancements in Hardware
1 With increased computing power available, it's possible to train ever
larger and more complex Generative AI models.

Improved Algorithms
2 Developments in Deep Learning algorithmic research, such as GANs,
Stable Diffusion and Transformer-based models, have enabled even
[email protected]
more accurate and diverse outputs.
HYML7P0S3U

Abundant Data
3 The explosive growth of the internet and the quantity & quality of data
freely available on the web have provided all the training data
needed for Generative AI.

A Multitude of Applications
4 Generative AI is being used to automate a wide range of processes
across industry verticals, such as synthetic protein creation, image
generation, programming Thiscopilots and
file is meant chatbots.
for personal use by [email protected] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Pre-Trained models
Pre-training is a technique used with models already trained on large data volumes, to “transfer” their
knowledge to more specific tasks, which may not have the luxury of large data sets for the model to
train.

[email protected] Smaller
Large Generic Corpus
HYML7P0S3U Specific
of Data Dataset for
some task

Pre-training Fine Tuning Task

Pre-trained Fine-tuned
Language Language
Model Model

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Transformers

A cutting-edge Deep Learning model primarily used for tasks

in Natural Language Processing.

Based on the idea of self-attention, which allows each word

to compute its relationship with every other word in a
sentence.
[email protected]
HYML7P0S3U

Transformers have an “Encoder” and “Decoder” portion -

the “Decoder” is responsible for generating data such as
text.

Transformers are a key building block in Large Language

Models such as the GPT series of models from OpenAI.

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 ChatGPT: A High-Level Understanding

GPT 3.5, one of OpenAI’s Large Language Models (LLMs), was trained on the entire corpus of
text data present on the internet (consisting of billions of web pages).

GPT 3.5 - the pre-trained model, was ultimately fine-tuned on a smaller Q&A dataset, to
convert questions into answers. This is the base for how ChatGPT was created.
[email protected]
HYML7P0S3U
Questions

Chain of Chain of
Question Answers
Encoders Decoders
Embeddings

Model asks
User Feedback is back propagated and used
to fine-tune the weights of the network for User
Feedback

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Prompts and Prompt Engineering
Prompt Engineering is the process of designing and crafting prompts for Conversational AI
interfaces and Generative AI models, in order to get as close as possible to the exact output
desired.

Some elements of a well-crafted prompt

[email protected]
HYML7P0S3U
“In the context of a dating show, come up with a hilarious
Context pick-up line.”

“Give me a short story about a young boy who discovers a

Task Description
mysterious object in his backyard.”

“Imagine you’re a Tech Support

Role Play Representative, and a customer calls to
say their Wi-Fi is not working. What would
This file isyour
meantresponse
for personal be?”
use by [email protected] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Prompt Engineering
Prompt Engineering is the process of designing and crafting prompts for Conversational AI
interfaces and Generative AI models, in order to get as close as possible to the exact output
desired.

Prompt
[email protected]
HYML7P0S3U

Generative
Language
Models

Generated Text

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 So what is GPT good for?
● Code Generation for known problems

● Creative Writing

● Text Summarization

● Idea Generation
[email protected]
HYML7P0S3U

In short, you want to use ChatGPT & Generative AI

for use cases where there is no right/wrong answer.

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Limitations of ChatGPT

● It is indifferent to the truth. Your prompt obliges ChatGPT

to complete the sequence.

● It doesn’t understand math. Don’t ask it to do calculus

[email protected]
HYML7P0S3U
and mathematical operations.

● It needs fine-tuning to work with internal

knowledge-bases.

● It is computationally expensive to train

This file is meant and fine-tune.
for personal use by [email protected] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Beyond ChatGPT
● Even bigger models, such as GPT-4, trained on larger chunks of text and other data
modalities like audio, images and video, which are also abundant

● Specialized versions fine-tuned on internal knowledge bases (ex: product

documentation, internal Wiki pages, medical publications etc.)

● Integration of ChatGPT-like models into an increasing number of applications across a

[email protected]
HYML7P0S3U
wide range of fields and industries

● ChatGPT rivals from other big technology firms

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Summary
A quick recap:

● ChatGPT is a large language model developed by OpenAI and it can generate

human-like text in response to a given prompt or question.

● ChatGPT has been trained on a vast amount of data, including text from the internet
and books, to improve its ability to understand and respond to natural language.

[email protected]
● Generative AI has the potential to revolutionize various fields such as language, image
HYML7P0S3U
and music generation, and is an area of growing interest and research.

● We got an overview of Language models, Transformers and Prompt Engineering.

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action.
 Happy Learning !
[email protected]
HYML7P0S3U

This file is meant for personal use by [email protected] only.

Sharing or publishing the contents in part or full is liable for legal action. 20