Welcome to introduction to generative AI.

After watching this video,

you'll be able to describe
generative AI and its evolution,
you'll also be able to explain how
generative AI differs from discriminative AI.
Artificial intelligence, or AI has been around for years,
shaping almost every sphere of
our lives and revolutionizing how we live and work.
At its core, AI can be defined as
the simulation of human intelligence by machines.
AI models learn from vast amounts of existing data.
The process of learning from data is called training.
There are two fundamental approaches to AI,
discriminative AI and generative AI.
Discriminative AI is an approach that
learns to distinguish between different classes of data.
A discriminative AI model is given a set of
training data where each data point
is labeled with its class.
The model then predicts the class of a new data point by
finding the side of the decision boundary
that the data point falls on.
Discriminative AI models use
advanced algorithms to differentiate,
classify, identify patterns,
and draw conclusions based on training data.
An example of how a discriminative
AI model works can be seen in how
email spam filters can
differentiate between spam and non-spam emails.
Discriminative AI models are
best applied to classification tasks.
They cannot, however, understand context or generate
new content based on
a contextual understanding of the training data,
and this is where generative artificial intelligence
or generative AI comes in.
Generative AI models learn to
generate new content based on the training data.
They can capture the underlining distribution of
the training data and generate novel data instances.
Generative AI starts with a prompt.
This can be text, an image,
video, or any other input that the model can process.
As an output, the model generates new content,
including text, images, audio,
video, code, and data.
Generative AI can produce output
in the same form in which the prompt is provided.
For example, text to
text or in a different form from the prompt,
such as text to image or image to video.
Here is a simple example
to understand the difference between
discriminative or traditional AI and generative AI.
Discriminative AI would be
best suited to answer questions such as,
is this image a drawing of a nest or an egg?
Generative AI would respond to prompts,
such as draw an image of a nest with three eggs in it.
While discriminative AI mimics
our analytical and predictive skills,
generative AI goes a step
further to mimic our creative skills.
As implied by this comment
from the Harvard Business Review,
AI can not only boost
our analytic and decision making abilities,
but also heighten creativity.
Generative models can take what they have learned and
create entirely new content based on that information.
Both discriminative and generative models
are created using deep learning techniques.
Deep learning involves training
artificial neural networks to
learn from vast amounts of data.
An artificial neural network is
a collection of smaller computing units called neurons,
which are modeled in a manner that is similar
to how a human brain processes information.
The creative skills of
generative AI come from generative AI models,
such as generative adversarial networks or GANs,
variational autoencoders or VAEs,
transformers, and diffusion models.
These models can be considered as
the building blocks of generative AI.
Generative AI is not a new concept,
it's roots trace back to the origins of machine learning.
In the late 1950s when
scientists proposed machine learning,
they explored using algorithms to create new data.
During the 1990s, the rise of
neural networks infused advancements in generative AI.
Further, during the early 2010s,
deep learning supported by
the availability of large datasets
and enhanced computing power
further advanced the development of generative AI.
In 2014, generative AI was transformed with
the introduction of GANs by
Ian Goodfellow and his colleagues.
GANs and other models such as
VAEs and transformers set the stage for
generative AIs growth and
the development of foundational models and tools.
Foundation models are AI models with
broad capabilities that can be adapted to
create more specialized models or
tools for specific use cases.
A specific category of
foundation models called large language models or
LLMs are trained to understand
human language and can process and generate text.
In 2018, OpenAI introduced
a transformer-based LLM called
generative pre-trained transformer or GPT.
Over the years, different LLM such as
GPT-3 and GPT-4 in the GPT series,
Google's pathways language model or PaLM,
and Metas large language model,
Meta AI or Llama,
have significantly enhanced generative AI
to generate coherent and relevant text.
There have been similar developments
and models for other use cases.
For example, stable diffusion and
DALL-E models for image generation.
The development of a variety
of generative models has led to
a growing market for generative AI tools
for diverse use cases.
For instance, you have ChatGPT
and Bard for text generation,
DALL-E 2 and Midjourney for image generation,
Synthesia for video generation,
and Copilot and AlphaCode for code generation.
The rapidly emerging models and tools have generated
a wide scope for generative
AI applications across domains.
To quote from McKenzie's report on
the economic potential of generative AI,
generative AI has the potential
to change the anatomy of work,
augmenting the capabilities of
individual workers by automating
some of their individual activities.
The report also predicts that generative AI's impact on
productivity could add trillions of dollars
in value to the global economy.
In this video, you learned that
generative AI models can generate
new content based on the data they are trained on.
Further, you learned that the creative skills of
generative AI are built from models such as GANs,
VAEs, transformers, and diffusion models.
Foundation models can be adapted to create
specialized models or tools
tailored to specific use cases.
Finally, you learned that
generative AI models and tools have
a wide scope for applications
across different domains and industries.