First Edition

Copyright © 2023 by Oliver Theobald

All rights reserved. No part of this publication may be

reproduced, distributed, or transmitted in any form or by
any means, including photocopying, recording, or other
electronic or mechanical methods, without the prior written
permission of the publisher, except in the case of brief
quotations embodied in critical reviews and certain other
non-commercial uses permitted by copyright law.

www.scatterplotpress.com

For feedback, print quality issues, media contact, omissions

or errors regarding this book, please contact the author at
oliver.theobald@scatterplotpress.com
Table of Contents
Introduction
A Brief History
AI Building Blocks
The 3 Stages of AI Development
Machine Learning
Deep Learning
Natural Language Processing
Generative AI
Recommender Systems
Computer Vision
Privacy & Ethical Considerations
The Future of Work
Further Resources
 Prologue
Japan clings to a reputation for high-speed trains, robots,
and futuristic cities, but that’s not quite the impression you
get walking past one of Japan’s many DVD stores or
watching someone at the office fiddling with the fax
machine.
The pace of technological change in Japan can be slow but
when it does occur, it tends to be premeditated and
meticulously choreographed. There are several examples in
Japan’s rich history, but nothing compares with the
country’s swift transition to an industrialized powerhouse.
You may have thought the fax machine was overstaying
its utility by several decades, but only 160 years ago, Japan
was governed by a feudal system left over from medieval
times. The drawing of the curtain on Japan’s feudal era and
the dismantling of the ruling samurai class in the 1860s was
highly organized and efficient. Over the space of a decade,
top-knot hairstyles, killing swords, and wages paid in units
of rice were phased out and replaced with the steel and
metal of a modern industrialized nation.
The transformation has intrigued me ever since I moved
to Japan, especially as we look ahead to the future of
artificial intelligence and knowledge work. The samurai
class, similar to accountants, lawyers, and other knowledge
workers today, were well-educated and respected members
of society. Trained in combat and schooled in calligraphy,
history, mathematics, and science, they enjoyed a secure
and elevated social standing in Japan’s social hierarchy.
For centuries, the samurai's main threat was rivalry with
neighboring clans, but the sudden arrival of steam-powered
ships from abroad eventually marked the advent of the Meiji
Restoration, a period of rapid industrialization and
westernization in Japan. Foreign powers, particularly the
United States and several European countries, forced Japan
to open its borders to international trade. This exposure to
both Western technology and ideas spurred a revolution in
Japan's economic, political, and social systems.
The samurai, previously the military and administrative
elite, found their social and economic status removed from
underneath them. The feudal system, which had provided
them a guaranteed income and elevated status in society,
was dismantled. A modern conscript army based on the
Western model also replaced the samurai as the nation's
defenders. Moreover, the rise of industry and commerce
created new wealth and opportunities, leading to the
emergence of a new class of rich merchants and
industrialists.
These changes didn't happen overnight, but the transition
was significant and swift in the grand scope of Japan's
history. The samurai, who had enjoyed a secure position at
the top of the social order for centuries, had to adapt to a
new environment that no longer valued their traditional
skills or recognized their social privileges.
Similar to the changes brought by modern machinery and
weaponry in the 19th Century, AI poses a threat to a large
and highly skilled segment of today’s workforce. Despite
increased globalization, knowledge workers have always
found reassurance in their local knowledge, language
abilities, and unique domestic expertise. However, the
unfolding revolution of AI-first practices is poised to
dramatically reshape the boundaries of work in ways that
are difficult to foresee.
With very few precedents that we can look to for
guidance, the experience of the samurai offers rare insights
and valuable lessons about resilience in the face of
technological change. Faced with social upheaval and
technological disruption, the samurai reacted in two ways.
The first approach involved rebellion and resistance, leading
to violent yet ultimately futile clashes with the Emperor’s
technologically advanced army.
The second and more productive course of action was to
embrace change and adapt to new technology. For the
samurai who took this path, education played a key part.
Exposure to foreign languages gave former samurai living in
the port city of Nagasaki a head-start in an economy open
to foreign trade and keen to learn from Western powers.
The Charter Oath issued by the Emperor in 1868 stated that
“knowledge shall be sought throughout the world” and
those with a tongue for foreign languages were among the
first to be dispatched overseas. Others stayed in Japan and
leveraged their classical education to become teachers,
bureaucrats, and artists in the new nation.
Re-education also played a vital role in navigating the
transition to a new economy. Although the samurai had
enjoyed privileged access to education (compared to
merchants, farmers, and the lowest class of Japanese
society), their pre-Meiji education was relegated to
foundational knowledge in an evolving economy. Their
classical education, though comprehensive for the time,
didn't fully prepare them for the significant shifts brought
about by rapid modernization and westernization. Beyond
literacy and basic arithmetic, many elements of samurai
training were undesirable or obsolete, especially in warfare
and philosophy. Battles would now be fought with modern
weaponry, and, like ancient Greek and Latin, knowledge of
the Chinese analects offered marginal utility in a modern
era where Western ideas on free markets, industrialization,
and capitalism ran supreme. In this way, the samurai
weren’t prepared to prosper in the new Japan based on
their education alone. Instead, they had to recalibrate and
adapt to new practices and employment pathways.
The parallels to today are striking, as professional writers,
lawyers, artists, and web developers find themselves
needing to acquire new skills such as text prompt writing in
response to the rise of generative AI technology. The speed
at which AI is seeping into all aspects of modern work—
from marketing to contract writing—underscores AI literacy
and understanding its key strengths and weaknesses as key
knowledge for the modern workforce. Just as the Meiji
Restoration in Japan introduced new job titles, there will be
new opportunities in what the authors of Human + Machine:
Reimagining Work in the Age of AI call the “missing middle”,
the nexus or fertile space where humans and machines
collaborate to exploit what each side does best. According
to the missing middle theory, the most effective and cost-
efficient path is to merge automated tools with the flexibility
of human workers to achieve optimum results. This is
similar to how chefs and waitstaff work alongside automatic
cashier machines inside Japanese restaurants today.
Virtually every sector has the potential to maximize
productivity and innovation by finding the right balance
between human creativity, judgment, and empathy, and
AI's speed, scalability, and quantitative capabilities. In the
ongoing era of AI and with new tools such as ChatGPT, we
aren’t just bystanders but active participants. As knowledge
workers, we can embrace AI as a partner in our daily tasks,
optimizing our abilities and complementing our existing
skills. As leaders, we can champion AI to foster a culture of
innovation, driving transformation and competitive
advantage inside our organizations.
These transformations won’t just impact individual tasks,
but entire workflows, industries, and even societies. As a
result, we need to recognize and address the challenges
that AI presents, including ethical considerations, job
displacement concerns, and the need for re-skilling. By
proactively confronting these issues, we can ensure that the
rise of AI benefits more people and not only a select few.
 Admittedly, this is a narrow path to walk, and this book
does not hold the wisdom and solutions to a fair and
symbiotic future for humans and machines. Instead, I hope
to hold your hand through understanding the fundamentals
of artificial intelligence and prepare you for important
discussions and decisions regarding the use of AI in your
organization or daily life. This includes a series of practical
thought exercises dispersed over multiple chapters.
Lastly, please keep in mind that this book is aimed at non-
technical readers, including marketers, product managers,
entrepreneurs, and students, seeking to build or expand
their understanding of artificial intelligence. While the
following chapters provide a solid foundation of core AI
techniques, the book does not delve into highly technical
aspects or run through coding examples for building AI
programs. For a more in-depth exploration of algorithms
and coding prediction models using Python, you may like to
read my other titles Machine Learning for Absolute
Beginners or Machine Learning With Python. These books
offer a more technical and detailed treatment of machine
learning algorithms and will help to complement your
understanding of AI gained from reading this book.
 1

Introduction
Machine intelligence represents a significant milestone in
human innovation and after decades of research and two AI
winters, artificial intelligence now dominates mainstream
attention and is promising revolutionary changes in the way
we work, create, and live our lives.
While the inception of AI as a field of study can be traced
back to the 1950s, it's only in recent years that this
technology has formed a significant component in our daily
lives. Today, it manifests in various forms, from digital
assistants like Siri and Alexa to recommendation engines on
TikTok, Netflix, and Amazon, as well as new content
generation tools like ChatGPT and DALL-E.
Despite its growing ubiquity, AI is often misunderstood
and confused with data science, which is an intersecting
field based on extracting insight from data and with its own
set of use cases. As a comparison, a company might use
the principles of data science to uncover new insights by
analyzing customer interactions and website support
tickets. This process involves aggregating and examining
the data to identify common customer issues, peak times
for customer support queries, or correlations between
support ticket volume and specific product features. Using a
data science methodology, the overall goal is to be as
precise as possible at identifying patterns and trends that
might help to inform the company’s decision-making.
 Artificial intelligence, on the other hand, performs a
different role that relies less on detective work and more on
general intelligence. AI, for example, can be used by a
company to power customer service chatbots on their
website that mimic human interactions, answer simple
customer questions, and refer more complex queries to
human representatives. By applying artificial intelligence,
the company’s overall goal is to automate key parts of the
customer journey, enhance efficiency, and enable 24/7
customer service support without relying solely on human
capital.
As highlighted in these two examples, data science
focuses on extracting insights and knowledge from raw
data, whereas artificial intelligence aims to simulate and
embed human intelligence into machines. However, in many
cases, AI systems will leverage insight derived from data
science to enable machines to learn and make intelligent
decisions. It’s important, therefore, to acknowledge the
overlap between data science and AI, while also
understanding that AI and data science remain two different
approaches to solving complex problems.
Beyond data science, AI encompasses a variety of
subfields and techniques including machine learning, deep
learning, generative AI, natural language processing,
cybernetics, and computer vision. Regardless of the
methods used, artificial intelligence, at its core, returns to
the overarching mission of creating systems capable of
performing tasks that would normally require human
intelligence. Such tasks include understanding human
language, recognizing patterns, learning from experience,
making informed decisions, and even displaying emotional
intelligence.
Learning from experience and making informed decisions
falls into the subfield of machine learning, which entails the
use of statistical methods to create prediction models that
improve their performance on a specific task through
experience and exposure to data. An example of this can be
seen in email spam filters, which learn to distinguish spam
from regular emails more accurately over time.
Pattern recognition, meanwhile, forms the basis of many
AI applications, from biometric identification systems that
recognize fingerprints or retina patterns to recommendation
systems that analyze our online shopping patterns to
suggest products we are likely to buy.
Emotional intelligence in AI, while still in its nascent
stage, aims to enable machines to recognize and respond to
human emotions. It has potential applications in many
areas such as mental health and customer service, where
it’s possible for AI to assist in providing assistance,
empathy, and emotional support to humans.
Understanding these different use cases and the breadth
of AI helps underline the fact that AI isn’t one monolithic
technology or technique but rather a collection of
technologies and approaches that strive to emulate human
intelligence. Acknowledging this diversity is crucial for
appreciating the full spectrum of AI technology and
recognizing the multitude of possible use cases.
The next essential insight for those keen on
understanding AI further is the realization that AI is still
only in its first stage of potential evolution. This first stage
is known as narrow AI or weak AI, which describes systems
designed to perform a narrow task, such as voice
recognition or recommending relevant products. These
systems excel at the specific tasks they were designed to
undertake but lack the understanding or consciousness to
freely apply their capabilities to other use cases.
At the same time, we are edging closer to the next stage
of AI development, known as artificial general intelligence
(AGI) or strong AI. AGI refers to a version of AI that
possesses the ability to understand, learn, adapt, and
implement its knowledge across a broad range of tasks at a
comparable or superior level to that of a human being. To
help grasp the concept of general AI, it’s useful to think of
science fiction portrayals where AI entities, like Data from
Star Trek or Ava from the movie Ex Machina, mingle with
humans and exhibit cognitive abilities that are
indistinguishable from ours. These AI entities are often
shown to possess self-awareness, emotions, creativity, and
the ability to understand and exhibit human-like behaviors,
which are all hallmarks of general AI.
However, as with any powerful new technology, AGI raises
a selection of ethical and privacy issues that must be
navigated with care, as we will discuss in later chapters. We
will also look further at the three stages of AI development
and explore the potential ramifications of the final stage
known as superintelligent AI. Beyond that, we will explore
the major subfields of AI including machine learning, deep
learning, natural language processing, generative AI,
recommender systems, and computer vision. The final
chapter will lay down a series of tips and insights for
adopting AI in your job or organization.
For now, understand that your journey into AI has many
potential paths and the field will continue to evolve as we
edge closer to the next stage of machine intelligence.
 2

A Brief History
From its humble beginnings as a scientific curiosity to the
powerful technology we now know today, AI has survived
several seasons of scarcity and abundance, doubt and
optimism, and fluctuating levels of government support
before embarking on a path toward exponential growth. To
understand where we stand currently and the future
trajectory of AI, we first need to look back at the
fragmented chronology of its evolution and decades-long
development.
While the roots of artificial intelligence can be linked
loosely to the period of classical philosophers who
attempted to describe human thinking as a symbolic system
(which was the original focus of AI research), the modern
field of study and the terminology both emerged in the
middle of the 20th Century. In 1956, the term artificial
intelligence was introduced and formally established as a
special field of research at a summer workshop held at
Dartmouth College.
The idea of designing intelligent machines had fascinated
scientists, mathematicians, and technology researchers in
the years before the conference, including prominent British
mathematician and logician Alan Turing. Considered the
father of modern computing, Turing made significant
contributions to computation and machine intelligence
during the 1940s and 1950s. This included the Turing Test,
which was designed to assess a machine's ability to exhibit
intelligent behavior indistinguishable from that of a human.
According to the Turing Test, a machine can be considered
intelligent if it can engage in a task without being detected
as a machine, such as holding a conversation with a human.
Inspired by Turing and other research developments, the
Dartmouth Conference was held at Dartmouth College in
Hanover, New Hampshire, in the summer of 1956. The
conference was organized by John McCarthy, Marvin Minsky,
Nathaniel Rochester, and Claude Shannon, four leading
figures in the fields of mathematics, cognitive science, and
computer engineering. Their proposal for the conference
expressed an optimistic view regarding the trajectory of AI-
related technology and outlined a research project that
would explore the possibilities of machines imitating human
intelligence. Invitations were sent to researchers known to
be active and interested in the field of machine-based
intelligence. The final list of participants included the four
organizing members as well as notable figures such as Allen
Newell and Herbert A. Simon, who all made significant
contributions to AI’s development.
Lacking the on-stage theatrics, sponsored booths, and
free swag common at most events today, the attendees
were shown their rooms and quickly put to work. Mirroring
an eight-week-long hackathon, the conference’s attendees
spent their days attempting to build computer programs
capable of imitating human intelligence. This included
working on problems still considered core to the field of AI
today, such as natural language processing, problem-
solving, learning and adaptation, and perception.
However, perhaps the most significant outcome of the
conference was the agreement on terminology and the
establishment of AI as its own distinct field of study.
Steered by John McCarthy, the term artificial intelligence
replaced previous and now forgotten descriptors such as
automata studies and complex information processing. This
had the effect of focusing attention on the simulation and
impersonation of human intelligence.
Moreover, the conference nurtured a community of
researchers who helped to establish AI as a legitimate field
of study and who would individually go on to become
leaders in AI research. Equally important, the conference
played a crucial role in securing financial backing for AI
research, which was aided by the conference's optimistic
outlook and the Cold War's emphasis on automation and
computational technologies. This resulted in significant
funding from government organizations and laid the
groundwork for the rapid expansion of AI research.
The initial years following the Dartmouth Conference
formed the first golden era of AI research and projected
expansive optimism regarding the future of AI. Caught up in
the excitement, researchers made ambitious forecasts,
predicting machines would be capable of achieving complex
human tasks in a matter of years.
Funding was abundant during this time and AI research
centers sprung up at prestigious universities across the
United States including Stanford University, MIT, and
Carnegie Mellon. The crux of AI research centered on rule-
based systems as researchers attempted to encode human
knowledge and intelligence into machines as a set of logical
rules, which later gave birth to the development of expert
systems in the 1970s and early 1980s.
By design, expert systems operate under a collection of
predefined rules that generate a recommendation such as a
medical diagnosis through a sequence of decisions.
Commonly structured as “if-then” statements, fixed rules
are set such as “if the patient has a fever above 38°C, then
they may have an infection” and “if the patient has a rash
and has been exposed to poison ivy, then they may have
contact dermatitis”. By stacking a series of if-then
statements, an expert system can be designed to analyze a
patient’s medical history and symptoms in order to provide
a potential diagnosis or propose additional tests and
treatments.
While these systems are capable of performing well at
narrow and well-defined tasks, they are largely flawed as an
effective long-term solution. Firstly, they rely on the
expertise and manual input of human professionals to
contribute knowledge to the system. Translating this
knowledge into a structured format of rules demands
significant time and effort, which means that expert
systems lack the ability to learn autonomously. Moreover, as
knowledge and expertise evolve over time, these systems
depend on regular updates in order to remain current and
accurate. The net result is a system that is expensive and
resource-intensive to build and maintain. What’s more,
failure to keep the system up to date carries the risk of an
outdated or incorrect diagnosis or recommendation.
The next problem is the domain specificity of these
systems. Their design and rule base are tailored to solving
specific problems, making it difficult to transfer these
systems across tasks. Deploying expert systems in different
contexts requires substantial modifications or, in most
cases, the development of entirely new systems. Even when
deployed within a specific context, the contextual
understanding of expert systems is severely limited.
Operating within the constraints of predefined rules, expert
systems struggle to grasp the broader context or interpret
information that is outside their explicit rule library. This
lack of adaptability and flexibility poses real challenges,
especially when facing complex, ambiguous, or special
scenarios that require nuanced judgment.
Over time, these challenges became increasingly
apparent, along with the obvious failure of artificial
intelligence to deliver on the ambitious promises made by
AI researchers. By the mid-1970s, many of the expected
breakthroughs turned out to be more difficult than originally
anticipated. Rule-based systems, which had been the
primary approach to AI research up to this point, were
proving expensive and time-consuming to run, and failing to
replicate the nuanced and complex nature of human
intelligence.
This lack of progress resulted in growing skepticism,
which then spread to policymakers and led to significant
funding cuts. In the United States, the government,
especially the Defense Advanced Research Projects Agency
(DARPA), reduced and eventually cut most of its AI research
funding. In the United Kingdom, the infamous Lighthill
Report, published in 1973, critically assessed the lack of
progress made by AI research, which resulted in cuts in
government support for AI research at many British
institutions.
Broad funding cuts prompted the beginning of the end for
numerous AI projects and interest in the field dwindled
across the broader AI research community. This period of
deflated funding and interest, starting from the mid-1970s
and extending into the early 1980s, became known as “AI
winter”. Used to describe a period of disillusionment and
fall-off in AI funding, the term is a takeoff on the idea of a
“nuclear winter”, a dark period marred by coldness and
barrenness.
Despite the challenging circumstances, a dedicated group
of AI researchers persisted and continued their work
throughout this period, with their efforts eventually leading
to a resurgence in interest over the subsequent decade. The
evolution of machine learning in the 1980s marked a pivotal
transformation in the landscape of AI research. Instead of
attempting to encode knowledge as a set of predefined
rules, machine learning proposed the notion that computers
could learn from data, identify patterns, and make decisions
with minimal human involvement. While the theoretical
foundations of machine learning had been established in
earlier decades, research funding and general interest had,
until this point, been channeled into expert systems.
This shift breathed new life into the AI research
community. The pivot, however, was not instant but rather a
gradual process driven by a series of breakthroughs and
milestones. One of the critical milestones during this period
was the application of an algorithm technique called
backpropagation. Short for “backward propagation of
errors”, backpropagation significantly improves the
efficiency of multi-layer neural networks. It enables the
model to adjust its internal parameters in response to the
difference between its actual output and the desired output,
thereby improving the model's accuracy through a series of
iterations. While multi-layer neural networks and the
concept of backpropagation existed prior to the 1980s, their
applicability was limited due to computational constraints.
The increasing power of computers in the 1980s made it
feasible to train larger neural networks, opening up new
possibilities for the application of machine learning.
The increasing availability of digital data during this period
also benefited machine learning. As computers became
more prevalent in business, academia, and government,
large volumes of data began to accumulate. With large
datasets needed to learn patterns effectively, demand for
machine learning began to build. This combination of
computational power, algorithmic innovation, and data
availability set the stage for the application of machine
learning across a broad range of research projects, from
speech recognition and computer vision to medical
diagnosis. By the end of the 1980s, machine learning had
firmly established itself as the leading approach in AI
research, but this return to optimism would again prove
harmful. Just as volcanic lava can shoot high into the air
before it starts to fall or an athlete’s career can reach its
peak before it starts to decline, there’s inevitably a peak in
AI development where doom presents itself as a triumph in
disguise.
Repeating patterns of the past, the next downturn was
caused by a variety of factors, spanning inflated
expectations, technical challenges, and economic pressures.
Despite various technological advancements, machine
learning models were still costly and slow to run. What’s
more, the hype surrounding AI again led to inflated
expectations that the present technology failed to live up to.
When these expectations weren't met, both investors and
the public pulled back their support, leading to a decrease
in funding and interest, similar to the outbreak of the first
AI winter.
The economic conditions of the time played a role too.
The conclusion of the Cold War in 1991 prompted large cuts
in defense spending, which impacted funding for AI
research, and a recession in the early 1990s tightened
budgets both in industry and academia, further reducing
resources available for AI research. Still, research into AI
and machine learning persisted throughout this period,
albeit at a slower pace than the years prior. The experience
of a second AI winter led researchers to adopt a more
measured approach to their work, with a focus on specific
and solvable problems as well as making more realistic
claims about AI’s capabilities.
As the 1990s rolled on, the rise of the Internet helped set
the stage for the next wave of development as vast
amounts of digital data became available. Interest and
funding in AI started to rebound by the late 1990s and early
2000s, thanks in part to some technical advances and the
growing importance of digital data. The success of AI
systems across a variety of tasks, ranging from chess
competitions to speech recognition, helped improve AI’s
brand image and rebuild confidence in the field. This
included IBM Deep Blue’s win over Garry Kasparov to
become the first AI system to defeat a reigning world chess
champion.
By the turn of the millennium, AI was on a path of steady
growth and development. The emergence of new algorithms
in the early 2000s for solving classification problems and a
series of practical demonstrations helped to reinforce AI’s
credibility and illustrate the real-world applicability of AI
systems. This included a defining moment in the field of AI
robotics after Stanford's autonomous car, Stanley, won the
2005 DARPA Grand Challenge by driving 131 miles across
the desert without direct human intervention.
However, the real revolution was still brewing and one
that would push the boundaries of AI capabilities closer to
reality. The late 2000s, more than anything else, marked a
leap forward in deep learning, a subfield of machine
learning and another example of AI theory well ahead of its
time. The concepts and principles of deep learning have
their roots in the 1980s but remained unrealized for several
decades. Once again, this wasn't necessarily a shortcoming
of the theory but rather a limitation of the technology
available.
Deep learning, named so for its use of “deep” networks
with many layers of decision-making, describes a unique
way of learning from complex data. In deep learning,
multiple layers of decision-making are used to calculate a
progressively more abstract representation than the
previous one. However, even with well-defined structures
and algorithms, such as backpropagation for error
correction, the computational demands for training these
complex networks remained outside what was technically
possible.
Geoffrey Hinton, often referred to as the godfather of
deep learning, became a key figure in the development of
deep learning during the mid-2000s. He advocated the idea
that simple learning algorithms, when fed a sufficient
amount of data and computed with sufficient processing
power, could surpass traditional hand-coded software.
Hilton’s big breakthrough came in 2006 when he introduced
a more efficient way to train deep neural networks—a
concept that paved the way for practical applications of
deep learning.
Hilton’s breakthroughs were followed by two distinct but
equally vital developments: the ongoing surge of big data
and the advent of powerful graphics processing units
(known as GPUs). In regard to the former, the ongoing
penetration and acceleration of the Internet and the digital
age brought with it an unprecedented supply of data.
Smartphones, social media, e-commerce, and various other
digital channels contributed to a vast and ever-growing
reservoir of data. Deep learning models, which learn by
adjusting their parameters to minimize prediction errors,
perform more effectively when exposed to large amounts of
data. The big data era, therefore, provided the perfect
environment for these data-hungry models.
However, vast data alone could not solve the
computational bottleneck. Training deep learning models
involves complex calculations and adjustments across
millions, if not billions, of different parameters. This is
where GPUs, originally designed for video games and
computer graphics, caught the eye of AI researchers. In
contrast to traditional central processing units (CPUs) that
undertake calculations one after the other, graphics
processing units are engineered to perform multiple
computations concurrently, which enables a more
sophisticated and immersive graphics experience.
Investment in GPU technology exploded and by 2005,
mass production strategies drove the costs of these chips
down substantially. This price reduction, paired with their
unique processing capabilities, expanded the potential
applications for GPUs beyond the realm of graphics
rendering, paving the way for their use in other fields
including deep learning and, later, Bitcoin mining.
In the case of deep learning, the advantage of GPUs lies
in their ability for parallel processing. This capability aligns
with the demands of deep learning computations, which are
predominantly matrix and vector operations. The capacity of
GPUs to handle these calculations concurrently, rather than
sequentially, attracted the interest of researchers including
Andrew Ng1. As a professor at Stanford University, Ng's
experiments with GPUs led to significant reductions in
training times for deep learning models, making it feasible
to build more complex neural networks. This development
signaled a paradigm shift, with deep learning models
outperforming traditional AI approaches across numerous
tasks including image and speech recognition as well as
natural language processing. The advancements in deep
learning led by Ng and other researchers including Geoffrey
Hinton in the area of backpropagation pushed the
boundaries of what was thought possible, with their findings
facilitating new breakthroughs and startups including
DeepMind and OpenAI.
Acquired by Google in 2014, DeepMind is best known for
developing AlphaGo, a deep learning program that defeated
the world champion in the game of Go 2016. Due to the
complexity and intuitive nature of the game, the feat was
considered a major milestone in AI, and the publicity
generated from the five-match series fueled a flood of
interest in deep learning among students and startups. The
company has since ventured into using AI to tackle societal
challenges, such as the protein folding problem in biology
and forecasting energy production for wind farms.
OpenAI, meanwhile, has made headlines with the release
of generative AI products including DALL-E and ChatGPT
that are capable of generating unique art and performing
knowledge-based tasks such as translation, question
answering, and summarizing text. These recent
advancements introduced by OpenAI are set to form an
integral part of many sectors and industries, driving
innovations in areas such as education, marketing,
entertainment, web development, and visual design.
As we delve deeper in the following chapters, we will look
closer at how generative AI and other examples of AI work,
as well as their common applications and the potential AI
holds for the future.

Key Takeaways
1) The history of AI has experienced a recurring pattern of
inflated expectations and hibernation periods known as AI
winters, characterized by periods of decreased funding and
interest. These different cycles highlight the importance of
maintaining realistic expectations and a measured approach
to AI's capabilities.
2) The resilience of AI, seen throughout its history, can be
partly attributed to the Lindy effect, which suggests that the
longer an idea or technology survives, the longer its future
life expectancy becomes. AI's ability to overcome periods of
skepticism, funding cuts, and technological limitations
underlines its resilience and reinforces the notion that the
longer AI continues to advance, the more likely it is here to
stay.
3) The advent of powerful graphics processing units
played a crucial role in the modern era of AI by enabling the
parallel processing required for complex computations. This
ability significantly reduces the time to train a model and
has facilitated the development of deep learning, pushing
the boundaries of what AI can achieve.
 Thought Exercise
1) What percentage of the current AI cycle do you think is
hype or exaggerated? What aspects of AI aren’t hype and
how can you focus on these opportunities?
 3

AI Building Blocks
Just as physical laws govern the mechanics of the
universe, the field of artificial intelligence is controlled and
regulated by algorithms. They underpin the advanced
capabilities we see today, from the more simple act of
recommending a song based on a user’s listening history, to
the complex task of driving an autonomous vehicle.
By definition, algorithms are sets of specific instructions
designed to perform a task or solve a problem. They are the
essential building blocks and the recipe for every good AI
system. But unlike recipes written in human language,
these algorithms are written in code and read by machines.
While explaining the code and mathematical intricacies of
popular algorithms is beyond the scope of this book, the
following chapter provides a high-level overview of different
algorithm categories, including classification, regression,
sorting, and clustering techniques, as well as transparent
and black box algorithms.
First, let's examine the shared characteristics of
algorithms within the context of artificial intelligence.
Fundamentally, algorithms function by processing data
inputs to yield an output. Input data can include text,
images, audio, video, sensor data, or any other form of
unprocessed information. For a recommendation engine, the
output might be a series of suggested products, and for a
facial recognition system, the output might be an
identification or face match.
AI algorithms typically have the capacity to learn and
improve over time. As more data is processed, the
algorithm fine-tunes its understanding, often improving the
quality of its outputs. This feedback typically comes in the
form of a loss function, a measure of how well the algorithm
is performing its task. The loss function quantifies the
difference between the algorithm's actual output and the
expected output. Lower values of the loss function imply
better performance.
When the loss function indicates that the output is failing
to achieve what was expected, the algorithm undergoes an
optimization process. This process involves adjusting the
internal parameters of the algorithm to better align its
outputs with the expected result. In essence, this is the
algorithm's way of learning from its past mistakes.
Parameters, meanwhile, are the internal settings that the
algorithm adjusts during the optimization process. The
exact nature and number of these parameters depend on
the specific algorithm being used. An example of an
algorithm with very few parameters is the k-nearest
neighbors (k-NN) algorithm. In k-NN, the number of
parameters is determined by the value of k, which
represents the number of nearest neighbors used for
classification or regression. Conversely, an artificial neural
network may consist of thousands of different parameters.

Classification vs. Regression

It's essential to understand that there's no one-size-fits-
all algorithm when it comes to AI. Different algorithms cater
to different tasks and different types of data. The choice of
algorithm depends on the specific problem you are
attempting to solve.
 To illustrate, classification algorithms are designed to
assign the input data to a certain category or class based on
its features, such as identifying whether an email is spam
and whether a tumor is malignant or benign. The task of
classifying instances is conducted by analyzing the
characteristics or composition of existing examples and then
creating a prediction model that captures relationships
between the features and the class labels in order to predict
the class of instances yet to be classified. This could involve
diagnosing whether a patient is carrying a specific disease,
such as diabetes, based on different health indicators. In
this scenario, the output or class label is binary: “diabetes”
or “no diabetes”. The explanatory features include a range
of health measurements such as patients’ blood pressure,
body mass index (BMI), age, cholesterol levels, and so on.
These measurements are collected from a large number of
patients, some of whom have been diagnosed with diabetes
and some who have not. This data is then used to train a
classification model.

Table 1: Sample data for classification of patients with diabetes

The training process involves feeding the features and the

corresponding class labels (“diabetes” and “no diabetes”) of
each patient into the algorithm. The algorithm will analyze
these features and establish a relationship between them
and the corresponding class labels to create a trained model
that captures these relationships. The trained model can
then be used to predict the likelihood of a patient having
diabetes based on the input of their health measurements.
 When a new patient's data is fed into the model, the
model can classify them as having “diabetes” or “no
diabetes” based on the relationships it learned from the
training data. For instance, the model may have learned
that patients over the age of 50 with a high BMI and high
blood pressure are statistically more likely to have diabetes.
Thus, if a new patient matches that criteria, the model will
classify the patient in the “diabetes” category. However, it's
important to remember that the model is actually making
an estimate based on the data it was trained on, and while
this can be highly accurate, certain errors are unavoidable,
especially if the model doesn’t account for fringe cases.
Regression algorithms, on the other hand, are used to
predict continuous values (in numerical terms, i.e. 5, 4, 5.4,
etc.), such as evaluating the value of a house or predicting
a patient's life expectancy. In the case of the second
example, the explanatory variables can include a range of
the same health measurements used in the previous
example, such as blood pressure, body mass index (BMI),
age, and cholesterol levels. The difference is that the target
outcome is no longer binary (“diabetes” and “no diabetes”)
but a continuous value: the patient's life expectancy
measured in years. The dataset again comprises numerous
examples of patients, each characterized by their feature
values in terms of health measurements but instead, there
is a new column with a life expectancy value expressed in
numbers.

Table 2: Sample data for regression of patient’s life expectancy

 A regression algorithm, such as linear regression, can
then be trained on the dataset to create a model. The
training process involves feeding the features and
corresponding life expectancy values of each patient into
the algorithm. The algorithm will then analyze those
features and establish a relationship between the
explanatory variables and life expectancy, which is the
target variable for this model.
Once the training phase is complete, the algorithm will
have learned and designed a model that can predict the life
expectancy of a patient. When a new patient's data is
entered into the model, the model estimates their life
expectancy based on their specific health measurements.
For instance, the model might have learned that patients
who are younger, with a lower BMI, and who have
controlled blood pressure tend to enjoy a higher life
expectancy. Therefore, if a new patient's data fits these
criteria, the model will estimate a higher life expectancy for
that patient. However, similar to the classification model
example, the model’s prediction constitutes an estimation
based on a generalization of the data the model was trained
on and is not guaranteed to be 100% correct and the model
may need to be retrained over time.
In summary, classification and regression algorithms are
designed to handle different types of problems.
Classification is used when the target variable is a category
or a class, such as “spam” or “not spam” in the case of
email filtering. Regression is used when the target variable
is a real or continuous value, like house prices measured in
dollars or life expectancy measured in years.
For a deeper exploration of specific classification and
regression algorithms, I have documented these techniques
in a separate book, Machine Learning for Absolute
Beginners: Third Edition, which walks through the process
of designing basic prediction models.
 Sorting & Clustering
In addition to classification and regression, there are also
sorting and clustering algorithms.
Sorting algorithms arrange a collection of items or data
points in a specific order, making it easier to analyze.
Sorting can be performed on various types of data,
including numbers, dates, and strings. For example,
consider a list of numbers: [11, 23, 1, 5, 2]. By applying a
sorting algorithm, such as merge sort, the list can be
rearranged in ascending order: [1, 2, 5, 11, 23].
Clustering, on the other hand, involves grouping similar
data points together based on shared characteristics or
similarities. The objective is to identify patterns, structures,
or relationships within the data without any predefined
labels or categories. Spam email, for example, isn’t usually
labeled by the sender as “spam”, but by grouping emails
together based on the subject line, contents, and external
links, it’s possible to group most of the spam email
messages into a cluster based on similar patterns—all
without any labels directly categorizing these emails as
“spam”.
Another common example of clustering is finding
customers who share similar purchasing patterns. By
identifying a cluster of customers who share purchasing
preferences, such as time of purchase and seasonal factors,
brands can identify distinct customer segments for targeted
marketing campaigns.

Algorithm Transparency
In addition to matching the model with the type of
problem you are attempting to solve, you also need to
consider the model’s interpretability. To elaborate, some
algorithms are relatively easy to understand but might not
always be the most accurate. Other algorithms can be
highly accurate but are often described as black boxes
because their internal decision-making process is difficult to
interpret.
Transparent or highly interpretable algorithms are those
where the steps taken to produce an output are easy to
follow and clear to interpret. Decision trees and linear
regression are two examples of algorithms where the
relationship between the input(s) and the output is clear
and easy to follow. The logic behind these algorithms can be
examined step by step, allowing you to understand
correlations and the specific variables that led to the final
output. In general, transparent algorithms are useful in
contexts where understanding the reasoning behind
decisions is crucial, such as a credit loan decision or house
valuation.

Figure 1: Decision tree

Black box algorithms, on the other hand, are those where

the pathway from input to output cannot be easily traced or
explained. Examples of black box algorithms are multi-
layered neural networks. These models involve layers and
layers of interconnected neurons that process input data,
transforming it multiple times to reach an output. Due to
the large number of transformations and the non-linear
nature of these transformations, it's extremely difficult to
understand how a neural network makes its decision. This
lack of transparency can become problematic in contexts
where accountability and explainability are needed or
favored.

Figure 2: Basic 4-layer feed-forward neural network

Balancing the need for highly effective AI systems and the

need for transparency and accountability is a key issue in AI
research and ethics. The development of techniques to
improve the interpretability of black box models, a field
known as explainable AI, is an active area of study seeking
to address this issue. Explainable AI refers to the ability of
an AI system to provide understandable explanations or
justifications for its decisions and actions. In medical
diagnosis, for example, explainable AI can help to provide
doctors with clear explanations of how a system arrived at a
diagnosis or treatment recommendation. This helps doctors
validate the AI's suggestions and gain insights into the
underlying factors considered by the model.
 Datasets
Creating effective AI systems is not just about selecting
and implementing the right algorithm. The design process
also requires a deep understanding of the dataset the
algorithm will consume and analyze.
Just as a car needs petrol or electricity to move, AI
algorithms need data to learn. However, the relationship
between AI and data goes beyond necessity. Understanding
the nuances of data—its collection, interpretation, and
limitations—is paramount to not only the effectiveness of AI
systems but also fairness and transparency. Poor quality or
biased data can lead to poor or biased results, a common
challenge in AI called “garbage in, garbage out”.
As organized stores of data, datasets provide the fuel for
algorithms to operate. In the case of supervised learning
(explored further in Chapter 5), a dataset is split into
training and test portions with a 70:30 or 60:40 split. The
training data is used to create the model, while the test
data is used to evaluate the model's performance.
The quality and size of the dataset significantly impact the
accuracy and reliability of the AI model. The better the
dataset, the better the model's ability to learn and
generalize to new inputs. Yet, datasets are not as
straightforward as they might seem. Datasets can present
numerous challenges that, if not properly managed, can
lead to poor or erroneous predictions. A dataset that lacks
diversity or that is not representative of the problem at
hand can lead to models that are biased or that perform
poorly when faced with real-world data.
One problem is the curse of dimensionality. Coined by
Richard Bellman in the late 1950s, the term refers to
various challenges that arise when analyzing datasets with
hundreds or thousands of variables—a phenomenon that
only occurs in high-dimensional datasets (with a large
number of variables). According to this theory, to obtain a
statistically sound and reliable result, the amount of data
needed to support the result often grows exponentially with
the dimensionality or number of variables analyzed.
However, having too many variables also injects noise that
makes it harder to identify patterns and build an accurate
model. What’s more, given the sheer volume of variables
thrown at the model, it’s possible for one or more variables
to correlate with the target output due to coincidence rather
than pure causation.
As anyone working in this industry will tell you, correlation
does not imply causation. Just because two variables move
together does not mean that one causes the other to move
and the website Spurious Correlations
(www.tylervigen.com/spurious-correlations) presents
numerous examples of variables that are strongly correlated
but clearly have no causal relationship. For instance, one
chart on the site visualizes a strong correlation between the
divorce rate in the state of Maine and the yearly per capita
consumption of margarine in the United States. As amusing
as it is to think about why this might be true, there is no
reason to believe that these two trends have any effect on
each other. The correlation is most likely a coincidence.
Next, the costs associated with processing large amounts
of data must be taken into consideration. While having
more data often improves the predictive ability of machine
learning models, it amplifies computational requirements
and increases model complexity. Conversely, too little data
can result in models that are inadequately trained and
perform poorly on new data.
Another crucial aspect is the cleanliness and integrity of
the dataset. Inaccurate or missing data can significantly
impair the model's performance, leading to incorrect
predictions. Hence, a substantial part of building a model
involves preprocessing and cleaning data to ensure its
quality and consistency. This step involves filling in missing
data, removing duplicates, correcting inconsistencies,
converting non-numerical variables to numeric variables,
and normalizing variables to maintain a consistent scale
(i.e., all variables measured in “minutes” and not in multiple
different units of time). It's a demanding task, but the value
it brings to the model and its outputs cannot be overstated.
To demonstrate the range of data quality considerations,
let's examine an example from an e-commerce company
developing a machine learning model for its new marketing
campaign. The aim of the model is to predict customer
responses based on various variables such as past
purchases, customer information, and website browsing
behavior.
First, the accuracy of the data is paramount. If the past
purchase history of a customer is recorded inaccurately, the
predictive model might target the wrong customers and
waste resources on customers who are less likely to
purchase.
Completeness is the next essential aspect of data quality.
If data concerning customer responses to past campaigns or
seasonal events are missing, the model may not have
sufficient information to make accurate predictions. This is
akin to trying to complete a puzzle without all the pieces
available.
Consistency in data is also crucial. Picture a scenario
where the age of the customer is recorded in years in one
part of the system and in months in another. Such
inconsistency could lead to confusion and result in errors in
the model.
The reliability of the data collected is yet another vital
component of data quality. If the browsing data is collected
inaccurately due to a bug in the tracking code or acquired
from an unreliable third party, the model might skew its
predictions. A customer might be misclassified as not being
interested in certain products when they actually are,
leading to inappropriate targeting and a missed sales
opportunity.
Timeliness, the next metric, ensures that the model is
based on the current reality. If the most recent purchase
data is not updated in real time, the model could base its
predictions on outdated information. For instance, a
customer who recently made several baby product
purchases might be incorrectly classified as unlikely to
respond to a new baby product recommendation if their
latest transactions aren't included in the data. Conversely,
marketing to a customer who made baby product purchases
five years ago may no longer be a reliable candidate for
purchasing that line of products.
Relevance of the data to the task at hand is essential to
avoid confusion and the waste of computational resources.
For example, including data such as the customer's phone
carrier or the color of the last item purchased (if color isn't
relevant to the campaign), introduces noise into the model
and distracts the model from more relevant variables.
Lastly, granularity refers to the level of detail in the data.
While highly detailed data might seem advantageous, it can
also introduce noise into the model, leading to a problem
known as overfitting where the model over-emphasizes
randomness or insignificant patterns in the data. For
instance, if website browsing data is recorded for every
single page view, it might add unnecessary complexity
without providing significant value. Conversely, if data is
aggregated or too coarse (having a low level of detail),
critical details about customer behavior might be overlooked
and wasted.
To summarize, data is a critical asset for creating AI
systems, providing the necessary information for these
systems to learn, improve, and generate accurate
predictions or insights. Each piece of data is like a brick
used to construct a building. If the bricks are weak or
flawed, the resulting building might be unstable. Similarly, if
data is inaccurate, incomplete, or unreliable, the resulting
AI system could produce misleading results, leading to poor
decisions and potential risks. As such, a thorough
understanding of data and its various aspects is
indispensable when working with AI. This includes not only
the ability to analyze data and check for spurious
relationships but also to clean and preprocess the input
data.

Table 3: Data quality considerations and examples

AI Hardware & Software

In this next section, we'll explore the vital role that
hardware and software play in facilitating artificial
intelligence.
Let's start with the hardware. High-performance hardware
is a cornerstone of AI, particularly in the realm of machine
learning and deep learning, where vast quantities of data
are the norm. From the central processing units that serve
as the brains of the computer to the graphics processing
units that accelerate computations, the role of hardware is
to provide the raw power needed to process and run
complex algorithms on the data.
 CPUs vs. GPUs
During the early stages of AI development, computations
were executed using CPUs, which are known for being
versatile and capable of executing a variety of tasks. In
practice, CPUs are used for performing sequential
processing, where each instruction is executed one after
another.
While CPUs are efficient for general computing tasks,
advanced AI and particularly deep learning require the
processing of complex mathematical operations on very
large datasets, which necessitates parallel rather than
sequential processing. Although originally designed to
handle the rendering of pixel-rich images in computer
games and gaming consoles such as the PlayStation 2 and
Xbox, GPUs are specifically designed for parallel processing,
which refers to the ability to perform multiple calculations at
the same time. This makes them well-suited for tasks that
require a lot of computing power, such as graphics
processing and machine learning. This also explains why the
use of GPUs in AI has become so widespread and why the
production and price of GPUs are carefully monitored by
companies and governments.
In general, GPUs are more expensive than CPUs to
produce. The specific pricing of GPUs is influenced by
multiple factors including production costs, demand,
technological advancements, and geopolitics. Over time, as
with many technologies, we have witnessed a general trend
of price deflation in the cost per unit of computing power.
This doesn't mean that the price of a GPU has decreased,
but rather, that the cost for a certain level of performance
(measured in floating point operations per second) has
generally gone down over time. This is due to
improvements in manufacturing, increased efficiency, and
technological upgrades.
 Nonetheless, it's important to note that the price of chips
can be vulnerable to short-term fluctuations. During periods
of increased interest in cryptocurrencies, the demand for
chips can spike, leading to increased costs. Similarly,
disruptions in the supply chain, such as those caused by the
global pandemic or international trade disputes, can impact
the availability and cost of computing resources.
As a byproduct of semiconductors2, chips are becoming
increasingly subject to trade war sanctions. In September
2022, the U.S. Department of Commerce added seven
Chinese supercomputing entities to the Entity List, which
restricts the export of certain goods and technologies to
those entities. While aimed at preventing China from using
chips for military purposes, the ban has far-reaching
implications for the AI industry.
Specifically, the ban affects the export of chips to China
from two major chipmakers, Nvidia from Taiwan and
Advanced Micro Devices (AMD) from the United States. Both
companies had been supplying chips, including GPUs, to
China for use in supercomputers but were forced to halt
shipments in response to the ban. This now makes it
challenging for companies in China to obtain the GPUs they
need for building AI applications. The Chinese Government
has responded to the ban by imposing its own restrictions
on the export of chips to the United States, making it more
difficult for U.S. companies to obtain chips from China. The
trade war has also prompted countries to increase their
investment in local chip manufacturing, which may lead to
the development of new chip innovations, supply lines, and
manufacturing processes.
To learn more about the lead-up and the potential
trajectory of the recent chip war, you may like to read Chip
Wars: The Fight for the World's Most Critical Technology by
Chris Miller, which tells the story of the global
semiconductor industry and the growing competition
between the United States and China to dominate this
critical resource.

TPUs
Beyond CPUs and GPUs, there is another important type
of chip to discuss. As the field of AI has advanced, more
specialized hardware has been required and this has led to
the development of Tensor Processing Units (TPUs),
originally developed by Google.
TPUs are a type of application-specific integrated circuit
(ASIC) developed specifically for accelerating machine
learning workloads. They are named after the tensor data
structure, which refers to a multi-dimensional array or data
structure with multiple variables. The first generation of
TPUs was announced in 2016 and used internally at Google
to improve the efficiency and speed of their machine
learning systems such as Google Search and Google
Translate.
Google announced its second-generation TPU in 2017.
These units were made available to external developers via
Google's public cloud platform, which continues to release
new and more powerful TPU generations. Other companies
including Microsoft and Amazon Web Services (AWS) have
followed in offering TPU services on the cloud. This has
helped to accelerate the development of AI as these
processing chips make it possible to train and run models
on a much larger scale.
In terms of hardware comparison, while CPUs are used to
handle a variety of general computing tasks and GPUs are
designed to quickly process the mathematical calculations
needed for rendering graphics, TPUs are specifically built to
handle the kind of matrix-based computations that are
common in machine learning and deep learning.
One of the significant differences between GPUs and TPUs
lies in their architecture. GPUs are designed to handle a
large number of relatively small computational cores for
parallel processing, which is great for graphics rendering
and beneficial for certain types of machine learning tasks.
TPUs, on the other hand, are designed around a large
matrix multiply unit, which offers high computational
capability and reduced computational precision (such as
lower-bit floating-point numbers) that is typically acceptable
for neural network workloads. This precision reduction can
still deliver accurate results for neural network models while
offering significant performance gains, which makes TPUs
efficient at executing the large-scale matrix operations that
are often found in deep learning algorithms. TPUs can also
be significantly faster than GPUs in regards to the running
of machine learning applications and, in general, are more
energy efficient.
However, it's crucial to note that while individual
operations may be more efficient, the overall environmental
impact of TPUs and GPUs raises a reason for concern. First,
the manufacturing of these high-performance chips requires
a lot of upfront energy and resources, including the mining
of raw materials, such as silicon and rare earth metals,
which are energy-intensive to mine. They then continue to
consume a lot of energy once put into use. The energy
consumption of these chips can be significant, especially in
data centers where they are used in large quantities.
As a result, there are a number of initiatives underway to
reduce the environmental impact of high-performance
chips, such as the development of more energy-efficient
manufacturing processes, the recycling and reusing of
chips, and the use of renewable energy sources, including
solar and wind power, to power data centers.
Regarding manufacturing processes, there are a number
of new energy-efficient manufacturing processes being
implemented including 3D stacking, which involves
vertically attaching three or sometimes four chips to make
better use of available chip space.

Software
Having discussed the evolution and importance of
hardware, including the use of GPUs and TPUs, let's shift
our focus to another critical component of AI systems: the
software. AI software ranges from the programming
languages used to write algorithms, such as Python and R,
to specialized libraries like TensorFlow, PyTorch, and Keras
that offer pre-built functions for creating and training
prediction models.
(Please note that this section delves into more technical
aspects, including descriptions of common code libraries. If
you aren’t interested in the programming side of AI, feel
free to skim or skip over this section.)

Libraries
Libraries, in the context of AI software, are collections of
pre-written code that developers can use to streamline their
work. They span everything from simple data-importing
commands to complex mathematical functions. Essentially,
they provide a way to perform tasks without having to write
custom code from scratch every single time you want to do
something. To explain why this is important, let's consider
the case of creating a neural network. Without code
libraries, a programmer would need to manually code the
entire architecture of the network, implement the
mathematical operations for forward propagation (where
the network makes its predictions) and backpropagation
(where the network learns from its errors), and handle the
optimization process that tweaks the network parameters
for improved performance. Not only is this process
extremely time-consuming but it also requires a deep
understanding of the underlying mathematics.
 Libraries reduce much of this complexity. When creating a
neural network using TensorFlow or Keras, developers can
build a model by simply stacking together pre-defined
layers. Each layer might represent a set of neurons in the
network, and come pre-packaged with the necessary
mathematical operations. Furthermore, these libraries come
with pre-implemented algorithms for training and optimizing
the model. When data is fed into the model, the library
takes care of passing it through the mathematical
operations defined by the network's architecture, adjusting
the network's weights based on the errors it makes, and
iteratively improving the model. The end result is that with
just a few lines of code, developers can create, train, and
implement a model without having to manually insert the
underlying math.
These libraries also provide code for preprocessing data,
which is a critical step in creating a model. They can handle
tasks like normalizing data to ensure that variables are
expressed on a common scale, converting categorical
variables into numerical variables, and splitting datasets
into training and test sets, among other things.
Lastly, libraries play an essential role in promoting
consistency and standardization. By using libraries,
developers adhere to a set of standardized practices, which
reduces the risk of errors and improves the replicability of
the code for future use.
Beyond TensorFlow, PyTorch, and Keras, there is an
extensive selection of useful libraries out there. Scikit-learn,
for instance, offers a broad array of machine learning
algorithms and tools for data preprocessing and model
evaluation using Python. Natural Language Toolkit (NLTK)
provides resources for tasks in natural language processing,
a field of AI focused on the interaction between computers
and human language. Matplotlib is a widely used library for
creating static, animated, and interactive visualizations in
Python, which is crucial for data exploration and presenting
results.

Programming Langauges
Programming languages are fundamental in the
development of AI systems. The choice of language often
depends on the specific needs of a project, including factors
such as the type of problem being solved, computational
efficiency requirements, and the availability of relevant
libraries and frameworks. Here, we'll discuss some of the
most popular programming languages used within AI.

Python: Known for its simplicity and readability, Python is

the most widely used programming language in AI and
machine learning. Python boasts a plethora of libraries and
frameworks that facilitate the development of AI
applications, including TensorFlow, PyTorch, Keras, and
Scikit-learn. It’s also a great choice for beginners and
experts.

R: If you're leaning towards statistical computing or data

analysis, R is a language designed specifically for these
purposes. It has a rich package ecosystem that facilitates
statistical modeling and visualization, which are core
components of AI and machine learning.

Java: Java is another language worth considering,

particularly if you're building large-scale enterprise
applications. It’s platform independence and robust
debugging features make it a versatile choice for AI
development.

C++: Due to its high execution speed and control over

system resources, C++ is often used in AI projects where
performance is a crucial factor. It's typically used in parts of
AI applications where low latency is required, although it
may not be as easy to use as Python or R.

Julia: Julia is a high-level, high-performance language for

technical computing. It provides the ease of use of Python
and R, but also the performance of C++. Julia is gaining
some popularity in the data science and AI communities,
particularly in areas that require heavy numerical and
scientific computation.

Prolog: Prolog, short for Programming in Logic, is a

language often associated with artificial intelligence and
computational linguistics. Its capacity to efficiently resolve
problems involving relationships, especially those that
involve structured data, makes it suited to certain AI
applications.

LISP: Although less commonly used today, LISP is one of

the oldest high-level programming languages and is closely
associated with artificial intelligence research. It was widely
used in AI research due to its symbolic processing
capabilities, but its use has waned with the rise of more
modern languages.

Figure 3: Percentage of Stack Overflow questions that month (not

necessarily AI-specific)
 The right language for a particular AI project can depend
on many factors, but these languages have demonstrated
their value across a wide range of AI applications. Beyond
the languages themselves, you should also consider the
support systems surrounding them. From tutorials and
forums to sample code, an active community can be a
goldmine of resources. These resources can dramatically
accelerate your learning and provide much-needed support
when you encounter problems with your code. Languages
with large communities also tend to be regularly updated
with new features and improvements, keeping you at the
forefront of AI development.
The nature of your project will play a significant role in
your decision too. If you are developing an AI solution that
needs to integrate with existing systems, your choice might
be influenced by the languages these systems use and their
compatibility. Similarly, if your work involves large volumes
of data, you should look for languages with strong data-
handling capabilities.
Lastly, bear in mind that many AI tasks can be
significantly accelerated by running them on multiple
processors simultaneously. This is especially true for deep
learning. As such, languages that support parallel
computing can be advantageous.
In conclusion, the right language to learn for AI
development depends on your personal circumstances,
including your existing skills, the type of projects you intend
to work on, and your performance requirements. Python, R,
Java, and C++ are all excellent choices, but they come with
their own strengths and weaknesses. If you are still unsure,
it may be wise to default to the most popular language for
AI systems, which, at this time, is Python.

Key Takeaways
 1) Classification and regression are two common
categories of algorithms used in AI. Classification algorithms
assign input data to specific categories or classes based on
their features, while regression algorithms predict
continuous values.
2) Sorting algorithms arrange data into a specific order,
facilitating easier understanding. Clustering algorithms, on
the other hand, group similar data points together without
predefined labels, allowing for the identification of new
patterns and relationships.
3) Transparency and interpretability of algorithms are
important considerations. Transparent algorithms have clear
and understandable steps, making their decision-making
process easily interpretable. Black box algorithms, on the
other hand, lack transparency and their internal workings
are difficult to trace.
4) Datasets play a crucial role in AI systems. The quality,
diversity, and relevance of the data used to train AI models
significantly impact their performance. Inaccurate, biased,
or incomplete data can lead to poor or biased predictions.
Preprocessing and cleaning the data are also important
steps for ensuring data quality.
5) Libraries promote convenience and consistency. By
using established libraries, developers adhere to
standardized practices, reducing the risk of errors and
improving the replicability of code for future use.
6) The choice of programming language should consider
factors such as computational efficiency, library availability,
community support, integration with existing systems, data
handling capabilities, and parallel computing support.
7) Python is currently the most widely used language in
AI and machine learning, offering simplicity, readability, and
a rich ecosystem of libraries.
 Thought Exercises
1) What’s a real-life example reflecting the curse of
dimensionality? Is there a business book, news channel,
conspiracist, politician, or popular theory, for instance, that
draws on a select number of seemingly correlated examples
that are coincidental and not typical of the true situation?

2) Do you think classification problems or regression

problems are more common in your organization or
institution? (i.e., how much to pay interns is a regression
problem, whereas selecting new interns based on their
credentials is a classification problem.)

3) Is model transparency important for a football coach

using a prediction model to recruit players who are likely to
score goals? Do fans betting on a football player to score
also need to know why the model predicts that player to
score or is model accuracy more important?

4) Pick an industry you are currently interested in and

think of a special use case for AI. Now ask whether model
interpretability and transparency are important for
evaluating the effectiveness of the model.
 4

The 3 Stages of AI
Development
As you venture further into the field of AI, it’s important
to recognize and understand the three potential stages of AI
development. Spanning narrow AI, general AI, and
superintelligent AI, these three stages represent crucial
milestones in the evolution of AI technology.
Narrow AI, for example, is designed to perform a specific
task, whereas general AI is more advanced and capable of
understanding, learning, and applying knowledge flexibly
across a broad spectrum of tasks. As the most advanced
stage, superintelligent AI remains a theoretical category at
present but is hypothesized to surpass human intelligence
across multiple domains at some point in the future.

Narrow AI
Despite its name, narrow AI or weak AI should not be
confused as weak or ineffective. Rather, the naming of this
category refers to the focused nature of AI systems in
relation to their scope and functionality.
By design, narrow AI systems are capable of mimicking
human intelligence but are constrained to a specific domain,
meaning they can’t perform tasks outside of what they are
trained or programmed to do. To illustrate, an AI system
designed for image recognition can identify and categorize
images based on its training but cannot translate languages
or diagnose diseases without radical adjustments and
additional training data.
This might make narrow AI seem limited in terms of
application, but these specialized systems form the
backbone of many powerful AI applications. Google's search
engine, Amazon's recommendation engine, Apple's Siri, and
Tesla's Autopilot are all examples of narrow AI systems.
Each of these systems performs specific tasks exceptionally
well, often surpassing human capabilities in speed,
accuracy, and efficiency. Google’s search engine, for
instance, can process billions of web pages and deliver the
most relevant search results in a fraction of a second. It's
optimized to perform this specific task and outperform any
human assistant. However, the same system is clueless
when it comes to steering a car or writing a poem. This
underlines the essence of narrow AI: exceptional at
performing a specific task but inflexible and unable to
operate effectively on other tasks.

General AI
Having established the definition of narrow AI, let’s turn
our attention to a more advanced iteration and the next
phase of AI development called general AI or strong AI.
General AI, often referred to as artificial general
intelligence or AGI, refers to a version of artificial
intelligence that has the ability to perform any cognitive
task achievable by a human. Note, however, that this
definition does not encompass physical abilities or the use
of robotics, which are often subject to different tests and
forms of evaluation. The Coffee Test, proposed by Apple Co-
founder Steve Wozniak, for instance, sets a benchmark for
evaluating physical capabilities by testing a robot’s ability to
enter an unfamiliar house, find the kitchen, identify the
tools and ingredients, and prepare a cup of coffee. Robots
are yet to pass the test and this example is just one of
many benchmarks currently in discussion regarding AGI
(beyond the boundaries of solely cognitive tasks).
When it comes to cognitive tasks, benchmarks and
definitions of AGI remain varied, which creates a sliding
scale of standards and expectations. In general, the more
specific the definition (such as the ability to converse with
humans on any topic in multiple languages), the easier or
lower the target of AGI becomes. Conversely, the less
specific the definition (perform any task a human can do),
the higher the AGI benchmark becomes. Depending on the
definition, AGI might also include a biological component,
physical capabilities, consciousness, or some other human
quality, which comes with its own set of design challenges.
Broad physical capabilities, for example, are difficult to
achieve due to the lack of data available to train robots or
humanoid robots. ChatGPT was trained on millions of data
sources scoured from the World Wide Web but the data
available to train robots is much harder to find and acquire.
Consciousness, meanwhile, is the subjective feeling of being
aware that one exists and having an understanding of the
surrounding environment. Some experts contend that
having an inner mental life is not replicable in machines
because consciousness arises from biological substrates and
cannot be replicated in non-biological entities.
Amidst these nuanced attempts to define general
intelligence, the core of the discourse and research on AGI
centers on emulating cognitive abilities across a
comprehensive range of non-physical tasks. This includes
the ability to reason, solve puzzles, plan, learn, integrate
prior knowledge into decision-making, and communicate in
a natural language such as English or Spanish. Importantly,
general AI includes the ability to transfer knowledge from
one domain to another—a skill known as transfer learning.
This might involve leveraging its understanding from
reading books to engage in a meaningful conversation
about literature, a capability that is well beyond the ability
of narrow AI systems.
It's important to highlight that as of the time of writing,
general AI remains unrealized. While we've made significant
strides in AI technology, we're still some way from creating
a machine that can fully replicate the broad cognitive
capabilities of the human mind. Most of the AI systems we
have today, including GPT-4, are considered narrow AI
because they excel at specific tasks (such as content
generation) but don't possess a generalized understanding
or ability to reason beyond their specific training. Also, while
some applications are capable of performing multiple
different tasks, they are actually using a collection of
narrow AI models under the hood.
One candidate for driving progress in the field of general
AI is AutoGPT. As an open-source AI agent powered by
OpenAI's GPT-4 API, the system can generate text and
perform various tasks autonomously, including code writing,
language translation, text summarization, question
answering, creative text generation, and online task
completion. It operates by breaking down user goals into
subtasks and utilizing access to apps, software, and online
services to accomplish those tasks. For instance, if a user
requests AutoGPT to compose a Tweet about coding, the
agent will segment the task into subtasks like finding a
coding tutorial, reading the tutorial, composing the Tweet,
and posting it on Twitter.
Using AutoGPT, one user has created the Do Anything
Machine, which is a to-do list that uses a GPT-4 agent, log-
in access to the necessary services, and your personal or
company information to complete each new task added to
the list. Other users are using AutoGPT to research and
script podcasts, conduct market research for new products,
and create websites from scratch.
 Although AutoGPT demonstrates promise and has the
potential to serve as a robust automation tool for complex
projects, it falls short of emulating human intelligence
across a broad spectrum of tasks or rivaling human
capabilities, including reasoning, in its current form. It’s
also common for AutoGPT to encounter problems or fail at
fulfilling tasks, especially for more nuanced tasks. Still, it
signifies a step towards the development of general AI by
showcasing the feasibility of creating AI models capable of
connecting to different online tools and adapting to new
tasks without direct human supervision.

Superintelligent AI
Following our exploration of general AI, let us now
venture into the most speculative phase of artificial
intelligence. Although it currently resides within the bounds
of theoretical forecasts and human imagination,
superintelligent AI presents a vision of the future that has
long captivated scientists, philosophers, futurists, and
science fiction writers.
While general AI aims to replicate the full spectrum of
human cognitive abilities, superintelligent AI goes a step
further, seeking to exceed human capabilities. In theory, a
superintelligent AI would not only outperform humans at
any intellectual task but would also outperform humans in
high-value endeavors including scientific research, strategic
planning, and social influencing. Thus, the idea of
superintelligent AI extends beyond an advanced tool or
system; it suggests a potentially autonomous entity capable
of out-thinking humanity and coming up with ideas,
strategies, and solutions that exceed the abilities of the
smartest human minds. As such, it raises questions and
concerns about control and alignment with human values
that are significantly more challenging than those
associated with general AI.
 This leads us to what’s termed the control problem, a
term popularized by philosopher Nick Bostrom, author of
the seminal book Superintelligence: Paths, Dangers,
Strategies. The control problem refers to the theoretical
difficulty of controlling or restraining a superintelligent AI. If
AI surpasses human intelligence, it might become
impossible to fully predict or control its actions. The AI
entity could devise strategies to avoid being shut down or it
could manipulate humans in ways we are unable to detect
and control.
The late Stephen Hawking, one of the most renowned
theoretical physicists of our time, expressed concern about
the potential risks of developing superintelligent AI. In a
2014 interview with the BBC, Hawking warned that “The
development of full artificial intelligence could spell the end
of the human race”.3 He went on to explain that once
machines reach a point where they can improve themselves
at a rapid pace, humans, with our slow biological evolution,
would no longer be able to compete and we will be
superseded as a result. Hawking reiterated his viewpoint
during a Q&A session at the annual Zeitgeist Conference in
2016. He stated, “I believe there is no deep difference
between what can be achieved by a biological brain and
what can be achieved by a computer. It, therefore, follows
that computers can, in theory, emulate human intelligence
—and exceed it”.4
Beyond a potential showdown with AI agents, a more
immediate concern lies in how humans will harness
superintelligent AI to wield power and influence. Similar to
preceding technological advancements, humans will
inevitably seek out ways to exploit superintelligent AI to
achieve their objectives, whether that’s delivering bias and
misinformation over the Internet or exploiting it for
cyberattacks, digital espionage, and deep surveillance.
Equally, there is a looming danger of a strong central actor
like OpenAI, Microsoft, or a state actor such as the Chinese
Community Party monopolizing access to superintelligent
systems and relevant hardware.
Seeing the problems posed by this scenario, initiatives
such as StabilityAI advocate for an open-source and
grassroots-driven approach to AI development. They warn
against the path of centralized power and encourage the
proliferation of distinct AI systems, each aligned with the
values and perspectives of the communities they serve,
mirroring the plurality and diversity of human values.

The Singularity
Building on the concept of superintelligent AI, the theory
of the Singularity forecasts a future point of irrevocable
societal change that will occur if and when AI surpasses
human intelligence. This theory is based on the principle
that a sufficiently advanced AI would be capable of
designing an even more advanced version of itself, which
could then design an even more advanced version, and so
forth, leading to a rapid and exponential increase in
intelligence.
The term and theory were popularized by mathematician
and science fiction author Vernor Vinge in his 1993 essay
The Coming Technological Singularity. According to Vinge,
the Singularity represents the end of the human era, as
superintelligence would continue to upgrade itself, leading
to an exponential increase in AI capabilities. This process
would result in unfathomable changes to civilization, so
much so that our current models of reality—the ways we
think about and understand the world—would no longer be
sufficient.
These ideas have been further explored and expanded
upon by authors and thinkers like Ray Kurzweil. In his book
The Singularity is Near, published in 2005, Kurzweil predicts
that the Singularity will occur around the year 2045 and will
lead to considerable societal and biological changes, as the
line between humans and machines becomes increasingly
blurred. Kurzweil suggests that humans will merge with AI,
enhancing our intellectual, physical, and emotional
capabilities.
Amid this possibility, it’s important to revisit the
existential risks of leading AI development in this direction.
As previously noted, there are substantial concerns
regarding the issue of aligning AI with human values and
objectives. Experts argue that as AI systems become more
intelligent and capable, the likelihood rises that they will
deviate in ways unforeseen by us today. To counter this
risk, Nick Bostrom emphasizes the need for substantial
investments in research aimed at developing strategies to
manage AI, ensuring that it contributes positively to
humanity rather than causing harm.
One popular solution is to control and monitor the data
used to train AI models. However, even with meticulous
screening of data inputs to eliminate inappropriate content,
ideologies, and knowledge, it’s theoretically possible for
superintelligent AI to uncover pathways to act contrary to
human values or behave in unexpected ways. Using an
inversion function, for example, the model could use its
knowledge of positive behavior to acquire insights into the
characteristics of negative behavior. If the model is trained
to avert house fires, it could potentially learn how to start a
house fire by adopting behavior contrary to the training
data—all without ever being explicitly trained to engage in
this behavior.
Similarly, training an AI to perform a specific task may
inadvertently increase the likelihood of it doing the opposite
of the intended behavior. This alignment problem has
recently been termed the Waluigi Effect, inspired by the
character Waluigi, the evil counterpart of Luigi in the Super
Mario franchise. Humans, particularly teenagers, for
instance, are inclined to do the opposite of what they are
instructed and it’s possible that developmental AI models
could do the same.
Conversely, there are many experts who question whether
it’s possible for AI models to rebel or replicate and surpass
human-level intelligence. One of these skeptics is the
computer scientist Gordon Moore, co-founder of Intel and
the originator of Moore's Law (the observation that the
number of transistors on a microchip doubles approximately
every two years). Moore has expressed doubts that the
Singularity will occur within the predicted timeframe of the
next 25 years or even at all. He argues that there are
physical limitations to the computation speeds that
machines can reach and that these limits will prevent the
realization of the Singularity as envisioned by Vinge and
Kurzweil.
Whether we are destined to reach the superintelligence
phase of AI and witness the Singularity in our lifetimes
remains uncertain, especially as it would require significant
breakthroughs far beyond our current understanding and
capabilities. In the meantime, contemplating the potential
consequences of the Singularity is crucial for steering
responsible AI development. For instance, if we create
machines that match or surpass human intelligence, what
are the implications? How do we ensure these AI systems
align with human values and ethics?
According to the precautionary principle, a concept from
risk management, it’s valuable to address potential risks or
harm even in the absence of scientific proof. The
precautionary principle emphasizes the need for proactive
action when there’s the possibility of serious or irreversible
harm to the environment or human health. In the context of
superintelligence, this may involve the following measures.
1) Regulation and policy: Robust regulatory
frameworks and policies are needed to address potential
risks and provide guidelines for the responsible
development and use of advanced AI systems. This involves
setting safety and transparency standards, monitoring
development, defining ethical boundaries for tech
companies to operate under as well as potential
intervention, if necessary.
Experts such as Daniel Colson, Executive director of the
Artificial Intelligence Policy Institute, have proposed
governments impose restrictions on AI firms, preventing
them from acquiring vast supplies of hardware used to build
super-advanced AI systems, while also making it illegal to
build computing clusters above a certain processing
threshold.5 While these measures may appear severe, there
is public support for government regulation and oversight of
AI development. According to a 2023 poll organized by the
Artificial Intelligence Policy Institute, 82% of American
respondents said they don’t trust tech executives to
regulate AI, with 56% of respondents supporting a federal
agency to regulate AI (compared to 14% who did not).6
Similarly, a 2023 global study published by the University of
Queensland and KPMG found that across 17 countries and
17,000 respondents, 71% of people surveyed were in favor
of AI regulation7, while the Ada Lovelace Institute and The
Alan Turing Institute found that 62% of 4,000 British
respondents would like to see laws and regulations guiding
the use of AI technologies.8
However, backing AI regulation doesn't automatically
imply public confidence in the ability of governments to act
effectively. As per the University of Queensland and KPMG
study, confidence in the government’s ability to regulate AI
development stood at 49% in the U.S., 47% in Japan, 45%
in the UK, 86% in China, 70% in India, and 60% in
Singapore.
2) Global coordination: While country-level measures to
regulate AI development are crucial, it’s just as—or if not
more—important to maintain alignment between countries
and regions. Similar to the founding of the International
Atomic Energy Agency to manage the safe use of nuclear
power, OpenAI's co-founders have called for a global agency
to rein in the development of superintelligence. This
organization would be responsible for conducting
assessments and audits of AI systems, designing and
implementing safety standards, and defining ethical
boundaries. While difficult to implement, agreements to
restrict the pace of global AI development are another
recommendation currently under discussion.
Existing forums for global cooperation, including the
European Union, G7, and the United Nations are also in the
process of designing processes to monitor and regulate the
development of AI systems. However, as the COVID
pandemic has shown us, achieving global cooperation is a
challenging endeavor, especially in light of the present
geopolitical environment and various schisms between
Western powers, China, and Russia. If a country or several
countries opt out of regulatory efforts or disregard global
guidelines, the efficacy of participating nations' endeavors
diminishes, resulting in a myriad of issues including
inequitable access to AI hardware and technology, along
with the migration of ambitious AI companies to non-
participating states.
3) Public engagement: Next, to help more people
identify and understand the potential risks and impacts of
AI, it’s crucial to encourage public participation in
discussions and decision-making processes related to the
development and deployment of advanced AI systems. This
ensures that a wide range of perspectives and concerns are
considered, while also encouraging ongoing debate over
social, economic, ethical, and safety implications that could
arise from developing superintelligent systems. Part of this
debate may start as early as high school, with the Montana
Digital Academy now offering courses covering AI history
and ethics to high school students in the U.S. state of
Montana, for example.
Other public initiatives include the United Nations AI for
Good Global Summit, the Center for AI Safety’s 2023
statement raising the risk of extinction from AI as a global
priority, and the Future of Life Institute’s open letter calling
for AI companies to pause training AI models more powerful
than GPT-4 for at least 6 months (which OpenAI has not
signed). As a signee of the Future of Life Institute’s open
letter, Elon Musk has also launched a new company called
xAI, with the mission of offering pragmatic alternatives to
pausing the development of superintelligence.
With each new AI breakthrough, we can expect to see
more non-government initiatives and public activities
discussing the role and threat of AI, with the overall aim of
ensuring that AI remains aligned with societal values and
priorities.

Key Takeaways
1) Narrow AI, general AI, and superintelligent AI form the
three potential stages of AI development. Narrow AI is
designed for specific tasks, while general AI possesses
broad cognitive capabilities similar to humans, and
superintelligent AI surpasses human intelligence.
2) There are serious concerns regarding the control and
alignment of superintelligent AI to ensure that it aligns with
human goals and values.
3) It is difficult to discuss and imagine the future of AI
without confronting the issue of the Singularity, which
predicts a future point where the human race is overtaken
and potentially overrun by AI agents.
4) It’s important to discuss the potential implications of
superintelligence and introduce early precautionary
measures despite the absence of scientific proof.

Thought Exercises
1) What are your instincts regarding the possibility of the
Singularity? Also, what alternative perspectives or potential
developments are you currently overlooking?

2) What measures, if any, can humans take to prevent the

Singularity from becoming a reality?
 5

Machine Learning
No matter where your exploration of AI takes you, the
path will invariably intersect with the field of machine
learning. Whether it's forecasting stock prices, detecting
fraudulent transactions, or powering speech recognition in
virtual assistants, machine learning provides the core for
many AI applications.
As a subfield of AI, the power of machine learning lies in
its ability to learn from data and make predictions without
being directly programmed. Given a set of inputs, the model
will make a prediction about what it thinks will happen next
based on the patterns learned from existing data. This
might involve predicting the price of a house based on
features such as its location, size, year built, and sales
history.
This process of learning and understanding patterns in the
data is known as training, whereby an algorithm is fed data,
called a training set, and studies that data in order to learn
patterns. If we take the example of a house price prediction
model, the algorithm is shown many examples of houses
along with their actual price value. The algorithm learns the
relationship between the features of the houses, known as
the explanatory variables, and their price valuation, known
as the target variable. Once it's learned this relationship, it
can predict the price of a new house based on its variables
using what’s called a model or prediction model. The model
is a mathematical representation that maps all the
explanatory variables to a target variable, such as the
qualities of a house and its market value. The goal of
training is to find the best model, which is the one that
most accurately captures the relationship between the
explanatory variables and the target variable.
It’s crucial to realize that perfect alignment with the
existing data is not necessarily the goal of designing a
reliable model. This stems from a common pitfall in machine
learning, known as overfitting. This occurs when a model is
fine-tuned to the nuances and noise of the training data to
such an extent that it captures insignificant and random
fluctuations in the data.
While this results in exceptional performance when tested
against the training dataset, it leads to a drastic decline in
the model's ability to generalize new and unseen data. In its
pursuit of achieving the best fit on the training data, the
model fails to encapsulate the underlying relationships that
are universally true beyond the training data. In essence,
the model has memorized the training data rather than
learning the true underlying patterns.
Thus, the objective of machine learning is to find a
balance—one that fits the training data well enough to
capture the true patterns and relationships without
mirroring its exact patterns or noise in the data. Striking
this balance between a model's capacity to learn from data
and not overfitting or overlearning is one of the biggest
challenges in machine learning and requires careful model
design decisions, data handling, and evaluation
methodologies.
In regard to model design, there are three overarching
techniques that we will explore in this chapter, starting with
supervised learning.

Supervised Learning
 Supervised learning is a type of machine learning where
the algorithm learns from a labeled dataset. Here, “labeled”
means that the training data includes both the explanatory
variables and the correct target variable. The house price
prediction example is a case of supervised learning because
the model is trained on a dataset of houses for which the
explanatory variables (i.e., land size, year built, etc.) and
the target variable (house price) are already known and
included in the dataset.

Figure 4: Labeled vs. unlabeled data

As another example, consider the model for predicting the

quickest route from point A to point B in Bangkok. First, you
need a labeled dataset that includes numerous examples of
routes from various points to various other points, along
with information about how long each route took. The label
in this case is the time it took for each route. The machine
learning model, through training, would then learn to
associate different factors such as distance, number of
turns, speed limits, and other route characteristics with the
time taken. Once trained, and given a new pair of start and
end points, the model is able to predict the quickest route
based on patterns learned from past examples.
In essence, supervised learning gives the model all the
information it needs so that it can decipher relationships
between the explanatory variables and a given target
variable, such as house price or travel time. Please note
that explanatory variables are also called the “independent
variables” or “X” in machine learning literature, while the
target variable is called the “dependent variable” or “y”.
 Unsupervised Learning
Unsupervised learning deals with unlabeled data. The
model still has access to different explanatory variables but
no corresponding target variable. In this case, you might
have the same dataset describing house features, including
land size and the number of rooms, but no information
about the house price. As a consequence, it’s impossible to
predict the value of houses in this dataset without this
missing variable. Unsupervised learning, therefore, looks at
other ways of exploring the data.
In general, the goal of unsupervised learning is to find
structure in the data, such as grouping similar variables
together. Building on an earlier example, let's say you have
access to the GPS data for thousands of taxi trips in
Bangkok including travel times and addresses but no labels
about popular pickup and drop-off points. Using
unsupervised learning, you could use the data to explore
and discover the structure within it. For example, you might
identify clusters of trips that start or end in certain areas,
effectively identifying popular pick-up and drop-off points.
Likewise, you might discover recurring patterns in the data,
such as specific routes that are frequently taken, essentially
learning the main thoroughfares without ever being
explicitly told what the main roads in Bangkok are.
This ability to discover unlabeled relationships means that
unsupervised learning algorithms can be used in advance of
supervised learning to prepare the data for prediction
modeling. When used in this way, unsupervised algorithms
help to clean up and label the data.

Reinforcement Learning
Used in various fields, including robotics, game playing,
and navigation, reinforcement learning allows the model to
learn by performing actions in a defined environment and
receiving rewards or punishments based on its decisions.
Opposite to unsupervised learning, reinforcement learning is
told the target variable but needs to learn the values of the
explanatory variables (which are unknown).
The learning process is guided by the goal of maximizing
the total reward. To explain, let’s consider a self-driving car
that needs to navigate from point A to point B. The car has
a map of the city and a GPS signal, but it needs to learn
how to drive to the destination on its own. In reinforcement
learning, the self-driving car would learn by trial and error.
It might start with random actions and receive a reward or
penalty based on how well it performed. If it takes a shorter
route, it might get a reward. If it violates traffic rules or
takes a longer route, it might receive a penalty. Over time,
by trying to maximize its total reward, the self-driving car
would learn to effectively navigate the city.
Due to the exploratory nature of reinforcement learning,
such models require significantly more time and computing
power than other forms of machine learning to train.

Model Design
It's vital to remember that the model techniques
discussed are not easily interchangeable. Each technique is
tailored to solving a specific type of problem using a given
dataset and their use cases tend not to overlap.
Supervised learning, for example, is generally the go-to
approach when you have a lot of labeled data, in which you
know both the explanatory variables and the target
variable. For example, if you're developing an email spam
filter and you have a dataset with emails that are already
labeled as “spam” or “not spam” (target variable) and other
information such as the name of the sender and the email
title (explanatory variables), then supervised learning is an
ideal technique for this task.
 Unsupervised learning comes into play when you have
plenty of input data but no corresponding target variable
labels. The main goal is to uncover hidden patterns,
clusters, or structures within the data. If you're tasked with
customer segmentation as part of a marketing analytics
project, unsupervised learning methods like clustering can
help identify groups of similar customers based on their
purchasing behaviors or preferences, for example. In its
place, supervised learning would not be a suitable approach
for discovering these unknown customer categories given
there are no existing labels.
Reinforcement learning, meanwhile, is suitable for
situations where a model learns to make decisions by
interacting in an environment with a finite number of
sequential decision actions. For instance, if you're designing
a system for playing chess or a video game, reinforcement
learning would be a prime candidate as it allows the AI to
learn optimal strategies over time through trial and error.
Reinforcement learning, though, cannot be used to predict
house prices or categorize emails into spam or non-spam as
these tasks don't involve sequential decision-making within
a confined environment.
In summary, the choice between supervised,
unsupervised, and reinforcement learning is not a matter of
superiority or preference, but rather a matter of fit. It's
about selecting the technique that best aligns with the data
you have and the problem you are attempting to solve.
Making the right choice can make the difference between a
successful AI project and one that is destined to fail before
you even start.
 Figure 5: Machine learning techniques

Practical Demonstration
To reinforce the difference between supervised,
unsupervised, and reinforcement learning, let’s explore
three different model scenarios. To help with this practical
demonstration, we will be using a hypothetical dataset,
belonging to a fictional video game. In this video game,
players control a character to collect items, avoid enemies,
and maximize their score within a limited time period.

Variables:
PlayerID (unique identifier for each player)
Items_Collected (number of items the player's character
collected)
Enemies_Avoided (number of enemies the player's
character avoided)
Time_Taken (time in seconds taken to complete a mission)
Score (points earned by the player, based on items,
enemies, and time taken)
This dataset has five variables, also known as features or
dimensions, and seven rows of values. Please note that this
dataset is for illustrative purposes; in real-world scenarios,
you would need a much larger dataset with many more
rows in order to train a credible model.

Model 1: Supervised Learning

Objective: Predict the Score of a player given the other
features
Explanatory variables: Items_Collected,
Enemies_Avoided, Time_Taken
Target variable: Score

Steps
1) Divide the dataset into a training set (75%) and a
test set (25%)
2) Choose an algorithm: In this case, a regression
algorithm such as linear regression is a good choice
because the target variable (Score) is a continuous
variable.
3) Train the algorithm on the training data to build a
model, which works by finding the correlation
between different explanatory variables and the
target variable.
4) Test the model on the test data to see how well it
predicts the target variable (Score) using the
 remaining unused data.
5) Measure the accuracy using an evaluation metric
such as Mean Squared Error (MSE), which is the
average of the squares of the differences between
the actual and predicted values.
6) Use the trained model to predict future data.

Model 2: Unsupervised Learning

Objective: Group players into clusters based on their
playing style
Explanatory variables: Time_Taken, Score
Target variable: N/A (unknown)

Steps
1) Choose a clustering algorithm, such as k-means
clustering.
2) Decide on the number of clusters/groups, i.e. k =
3.9
3) Train the algorithm on the explanatory variables.
(There is no training-test data split in unsupervised
learning.)
4) Analyze the clusters to see which players fall into
which cluster and interpret the characteristics of
each cluster. For example, we might find that the
three clusters represent: aggressive players,
balanced players, and defensive players.
 These three new categories can now be used to label
each player in the dataset, providing a new
variable/feature that we didn’t know about previously.

Model 3: Reinforcement Learning

Objective: Develop an AI agent that can play the game
and maximize the Score value.
Explanatory variables: Move left, move right, jump,
crouch, advance forward
Target variable: Score

Steps
1) Define the state space: A combination of the
player's position, items left, enemies nearby, etc.
 2) Define the action space: Move left, move right,
jump, crouch, advance forward.
3) Define the reward: For example, +50 for collecting
an item, -50 for colliding with an enemy, and +100
for passing the mission.
4) Use an algorithm like Q-learning to train the agent.
The agent will play the game multiple times. In the
beginning, it will play randomly and make many
mistakes.
5) Over time, the agent will learn the optimal strategy
to play the game and maximize Score.

Other Approaches
In addition to the three approaches to machine learning
that we have discussed, there are two other variants that lie
between supervised and unsupervised learning, called semi-
supervised learning and self-supervised learning
respectively.
In the case of semi-supervised learning, there is typically
a large amount of data but only some of it is actually
labeled. The idea is to use the labeled data to train an initial
model and make predictions on the unlabeled data. The
most confident predictions are then used to label the
unlabeled data and expand the labeled training set. The
prediction model can then be retrained on the combination
of the originally labeled and the self-labeled data. In this
way, semi-supervised learning seeks to benefit from the
additional (albeit imperfectly labeled) data. For instance,
you might have GPS data for millions of taxi rides but only
have the route times for a few thousand of those rides.
Using the known route times to train an initial model, you
can then use that model to predict the route times for rides
missing that variable based on their start point, end point,
and other variables such as the day of the week.
 In the case of self-supervised learning, all data is
considered unlabeled, but the structure within the data is
used to provide supervision. In other words, the training
labels are automatically generated from the input data
itself, without human annotation. This is done by creating a
learning task where the objective is to predict some part of
the input data. For example, a self-supervised learning
algorithm may learn to predict the next word in a sentence
or the next frame in a video. The goal is to design a task
where the correct answer is available, allowing the model to
learn from the data directly.
A real-world example of self-supervised learning is the
masked language model training objective used in
transformer-based language models such as Google’s BERT
(Bidirectional Encoder Representations from Transformers).
During training, some portion of the input data such as a
word in a sentence is masked, and the model is trained to
predict the masked word based on the rest of the sentence.
This allows the model to learn useful features about
language from the data itself, without requiring explicit
labels for each example.
Another common example comes from the field of
computer vision, where a portion of an image might be
obscured and the model is tasked with predicting the
missing part. In this case, the model takes an image and
purposedly obscures a part of it, like covering a portion of
the image with a black box. The model then predicts what's
hidden behind the black box by analyzing the surrounding
context and making an educated guess. For instance, if the
covered area is around a person's face, the model might
predict that it's likely a face. As the model is exposed to
more images, it learns to predict more hidden parts.
Through this process, the model becomes better at
recognizing objects, patterns, and relationships without
explicit labels for every image.
 In summary, semi-supervised learning is used when you
have some labeled data and lots of unlabeled data, while
self-supervised learning can be used on data without
preexisting labels to predict hidden features.

Key Takeaways
1) Machine learning allows systems to learn from data
and make predictions without explicit programming.
3) Training involves feeding a machine learning algorithm
a dataset to learn patterns and relationships. The goal is to
find a model that accurately captures the underlying
patterns without overfitting the training data.
3) There are three primary types of machine learning
models:
- Supervised learning uses labeled data with known
explanatory variables and a known target variable to train
the model.
- Unsupervised learning deals with known explanatory
variables and aims to discover hidden patterns and
structures within the data to create a new target variable.
- Reinforcement learning involves learning through trial
and error in an environment with rewards and punishments
in order to achieve a predefined target.
4) Model selection depends on the type of data you have
available and the problem you are attempting to solve.

Thought Exercises
1) What is a use case for supervised learning? (Hint: think
of a website that makes a prediction based on past data)

2) What is a use case for unsupervised learning? (Hint:

think of a model for sorting items)
 3) What is a use case for reinforcement learning? (Hint:
think of a game or closed environment with a limited
number of potential actions)
 6

Deep Learning
From self-driving cars to large language models, the
ability of deep learning to learn from raw data and process
large datasets has dramatically broadened the use cases for
artificial intelligence. However, like all technologies, deep
learning has its own unique limitations and challenges that
we will examine in this chapter.
Before we dive in, it’s important to explain the connection
between deep learning and machine learning. Machine
learning, as discussed, involves algorithms and models that
improve with experience and exposure to data. As a
subfield of machine learning, deep learning takes the
foundational principles of machine learning and applies its
own techniques to even larger and more complex datasets.
This involves the use of artificial neural networks with deep
and multiple layers stacked together to form a model.
While artificial neural networks are not a direct replica of
the human brain, (with the human brain estimated to
contain 100 billion neurons, outnumbering the number of
stars in the galaxy, and more than one thousand kilometers
of interconnections), they are loosely inspired by it. Artificial
neural networks consist of interconnected nodes or neurons
that process and pass on information to the next layer of
neurons—similar to how neurons in the brain process and
transmit information. The first layer of neurons in the
network, for example, might recognize basic features in
images, such as edges. The next layer might put these
edges together to recognize shapes, the layer after that
might recognize fur patterns, and so on, until the final layer
recognizes the overall image as a cat.
Like machine learning, deep learning makes use of
supervised, unsupervised, and reinforcement learning
techniques. In the case of supervised learning, a neural
network, such as a convolutional neural network, can be
trained to recognize images by studying a large number of
labeled images (with the labels describing the image) and
learning to identify the labeled objects contained in those
images.
In the case of unsupervised learning, deep learning is
often used to simplify and reduce the dimensionality
(number of features) in the data or remove noise from it.
In reinforcement deep learning, neural networks are
combined with reinforcement learning techniques to create
systems that can learn to make a sequence of decisions. An
example is the AlphaGo program developed by DeepMind,
which learned to play the game of Go at a superhuman level
by studying past games played by humans as well as games
played against itself.

Deep Learning Models

There are various types of neural networks employed in
deep learning, each with its own unique strengths and
common use cases. In this section, we will examine
convolutional and recurrent neural networks as well as
transformer networks, which play a major role in NLP and
computer vision (discussed in Chapter 7 and Chapter 10
respectively).

Convolutional Neural Networks

CNNs represent a major advancement in the field of
computer vision, enabling machines to see and interpret
visual data. From facial recognition to self-driving cars,
CNNs are integral to computer vision and other AI
applications. In fact, the name “convolutional” comes from
the mathematical operation by the same name, which is a
specialized type of linear operation often utilized for image
processing.
CNNs excel at identifying spatial hierarchies or patterns in
images by processing small chunks of the image, allowing
them to recognize complex shapes and structures. Prior to
CNNs, image data was typically fed into neural networks as
one-dimensional arrays10. This, however, reduced the
spatial information contained in the image. To elaborate,
images are typically recorded as two-dimensional
structures, consisting of rows and columns of pixels. Each
pixel represents a value that represents the color or
intensity at that specific location in the image. To process
an image using a traditional neural network, the image is
transformed into a one-dimensional array, also known as a
vector. This transformation involves concatenating (meaning
“joining”) the rows or columns of the image into a long
linear sequence of values.
As an example, let's say we have a grayscale image of
dimensions 64x64 pixels. Each pixel in the image represents
a grayscale intensity ranging from 0 to 255. By flattening
this image into a one-dimensional array, we obtain a
sequence of 4,096 values, where each value represents the
intensity of a pixel. By representing the image as a one-
dimensional array, we can feed it into a neural network as
input, which then processes that array and performs
computations on the values of each element.
While flattening the image into a one-dimensional array
simplifies the representation of the image, it also eliminates
the spatial structure and relationships between neighboring
pixels. This loss of spatial information can limit the ability of
the neural network to capture complex patterns and spatial
dependencies within images. Recognizing a face, for
instance, is not just about identifying individual facial
features like the eyes, nose, or mouth, but also
understanding their relative positions to each other.
Traditional neural networks, though, fail to effectively
account for such spatial hierarchy between features.
To overcome this limitation, CNNs preserve the spatial
structure of images by operating directly on the two-
dimensional or three-dimensional representations of
images, allowing them to capture local patterns, spatial
hierarchy, and spatial relationships between features. Unlike
traditional neural networks, CNNs maintain the spatial
relationships between pixels by learning image features
using small squares of input data. This method is less
sensitive to object position and distortion within the image,
enabling the model to recognize an object even if its
appearance varies in some way. For example, when a CNN
is trained to recognize a turtle, it learns to detect its distinct
features like the shape of the shell, the presence of flippers,
and the general body structure. These collections of
features are consistently present across different images of
turtles, irrespective of their overall positions or poses. By
capturing the spatial relationships among these features,
the CNN can effectively generalize and recognize turtles,
regardless of whether they are crawling, standing, or
partially hidden inside their shell.
In terms of their design, CNNs consist of a series of layers
that filter the raw pixel data of an image to extract and
learn higher-level features, which the model can then use
for classification. Higher-level features refer to visual
elements that are built upon simpler features detected in
earlier layers, such as edges, lines, and textures. In other
words, the earlier layers capture basic visual elements that
are present in the input images and as the information
flows through subsequent layers of the network, higher-
level features are learned by combining and abstracting
these low-level features.
Another important aspect of CNNs is the pooling layers
that follow the convolutional layers. When passing an image
through a CNN, it goes through several layers that detect
different features like edges, shapes, and objects. As
mentioned, these features are combined to understand
what's present in the image. However, the information can
be quite detailed, which can make the network overly
complex and slow to run. Pooling layers solve this problem
by reducing the dimensionality of the data. This involves
summarizing and preserving the most relevant information,
which results in a more manageable and efficient
representation. By compressing the information, pooling
therefore decreases the size of the data and the number of
calculations needed, speeding up the overall network.
Pooling layers also make the network more robust by
focusing on important features rather than their exact
position. This enables the network to recognize objects even
if they appear slightly shifted or in different parts of the
image.

Recurrent Neural Networks

Renowned for their inherent ability to remember, RNNs
have been central to groundbreaking developments in
various fields including natural language processing, speech
recognition, and time series prediction.
Popularized by John Hopfield in the early 1980s, RNNs
were originally designed to overcome the limitations of
existing artificial neural networks. Notably, traditional feed-
forward neural networks, including CNNs, are not designed
to deal with sequential inputs. (A feedforward neural
network is an artificial neural network where connections
between the decision nodes do not form a loop or cycle).
 For tasks such as language translation, speech
recognition, or time-series prediction, the order of the
inputs often carries crucial information, and the ability to
handle sequences is a key requirement. While traditional
neural networks assume that all inputs (and outputs) are
independent of each other, RNNs leverage the sequential
nature of their input data. By processing inputs in sequence
and having a form of memory, they can account for the
context provided by preceding elements. This allows them
to excel at tasks where the sequence and context matter,
such as sentences or time series data. They achieve this by
creating loops that pass information from one step in the
sequence to the next, effectively giving the network a basic
form of memory. In practice, each neuron or unit in an RNN
has a “hidden state” which acts as a point of memory. This
hidden state is a function of the current input and the
previous hidden state, allowing the network to retain
information from prior inputs in the sequence.
To better understand this concept, let’s consider the task
of predicting the next word in a sentence as an example. A
feed-forward network might be given a fixed number of
previous words and be trained to predict the next, but it
would treat these input words independently, ignoring the
order in which they appear. An RNN, on the other hand,
would process the sentence word by word, retaining
information from previous words as it goes along, thereby
capturing the sequence's context.
RNNs, however, possess limitations and chief among these
is the vanishing gradient problem. The vanishing gradient
problem occurs when a deep neural network struggles to
learn and make meaningful updates to its weights during
training. This happens because the gradients (slopes) of the
loss function for the network's weights become extremely
small as they are passed backward through the layers
(backpropagation). These gradients are calculated by
multiplying the gradients from the subsequent layers with
the weights connecting the current layer. As the network
becomes deeper, this multiplication can lead to a situation
where the gradients diminish significantly as they propagate
backward through multiple layers. This can result in very
small gradient values, leading to slow weight updates or
adjustments. As a result, the network makes only tiny
adjustments to its weights and it either takes a very long
time to complete or the network fails to learn complex
patterns in the data.
To gain a clearer grasp of this concept, it helps to think of
the neural network as a hiker trying to find the best path to
reach their destination. The hiker adjusts its path (which is
the weights) by looking at how much the slope (which is the
gradient) changes. If the slope is steep, it means there's a
lot to learn, so the hiker makes bigger adjustments to its
path. If the slope is gentle, it means there's not much to
learn, so the hiker only makes small changes to its path. In
some cases, especially with deep networks or long
sequences like in RNNs, the slopes can become very flat.
When this happens, the hiker gets stuck because it can't
make meaningful adjustments to its path (which is the
weights). Thus, much like a hiker struggling on a flat
mountain path, the network might take forever to learn or
become disorientated in the process.
As a practical example, imagine that you are training a
neural network to identify various types of animals in
images. When the network faces tricky examples such as a
zebra, the gradient is already extremely tiny, which reflects
how much the model should fine-tune its internal settings to
improve its recognition. As a result, the network learns at a
sluggish pace or it may even become stuck and fail to
accurately recognize animal images that are difficult to
classify.
 In the context of RNNs, the vanishing gradient problem
makes it difficult for the model to learn long-range
dependencies in the data (discussed also in Chapter 7)
because it struggles to learn relationships between inputs
that are distanced apart. For instance, in a language
processing task, an RNN might find it challenging to connect
information from the beginning of a long sentence with
relevant words or phrases at the end of the sentence. This
happens because the RNN's ability to remember and use
information from earlier inputs gets weaker with more
added inputs.
This problem has led to the development of more
advanced types of RNNs like long short-term memory
(LSTM) and gated recurrent unit (GRU), which introduce
gates to control the flow of information. Acting as a switch,
a gate is a mechanism that controls the flow of information
within the network. It determines how much information
should be passed through and how much information should
be blocked or forgotten at each step. This allows the model
to selectively retain important information, helping to
mitigate the vanishing gradient problem. This architecture
has proven extremely effective, enabling RNNs to tackle
complex and nuanced tasks in NLP, including machine
translation, sentiment analysis, text summarization, and
language generation.

Transformer Networks
Transformer networks represent a significant leap forward
in dealing with sequential data and especially in natural
language processing. By focusing on the parts of the data
that matter most and processing sequences more efficiently,
they have opened a new paradigm in deep learning.
Transformer networks were formally introduced in a 2017
paper published by Google titled Attention is All You Need.
They were designed primarily to address the shortcomings
of existing models in handling sequential data, especially in
the field of NLP. Prior to the introduction of transformers,
recurrent neural networks and their variants such as long
short-term memory and gated recurrent unit were the
dominant models for sequential data tasks. While these
models performed well, they processed sequences in a
linear manner, considering one element at a time. This
approach resulted in two unique problems: computational
inefficiency and difficulty in capturing long-distance
dependencies in the data.
First, sequential processing made it challenging and slow
to effectively parallelize computations during training.
Second, even though LSTMs and GRUs were designed to
mitigate the vanishing gradient problem of standard RNNs
and better capture long-range dependencies, they still
struggled with very long sequences.
Transformer networks were thus developed to tackle these
two issues. They introduced a mechanism called attention,
which allows the model to weigh the relevance of different
elements in the input sequence when producing an output,
thus enabling it to focus more on important parts and less
on others.
This mechanism is useful in NLP, where the meaning of a
word can depend heavily on its context in a sentence or
document. For instance, consider the following English
sentence: “I took the dog that bit me to the vet”. In this
sentence, the word bit” is closely related to “dog” and me”,
while “vet” is more relevant to “I” and “took”. A transformer
model uses attention mechanisms to determine these
dependencies. It assigns higher weights to “dog” and “me”
when processing “bit” and higher weights to “I” and “took”
when processing “vet”. Moreover, if we were translating this
sentence into a language like German, where the verb often
comes at the end of the sentence, the transformer's
attention mechanism would allow it to associate the verb in
English with the corresponding verb in German, even
though they are in different positions within their respective
sentences.
Importantly, transformer models can compute the
attention weights for all elements of the sequence in
parallel, leading to significant improvements in
computational efficiency. Moreover, by directly attending to
all other words in the sequence, regardless of their position,
transformer models capture both short-term and long-term
dependencies contained within the data.
This architecture has proven extremely effective at natural
language processing tasks with the ability to process entire
sentences or even paragraphs at once—instead of
sequentially. Specifically, they make use of the attention
mechanism to weigh the importance of different words in
understanding the context of a sentence. This has resulted
in state-of-the-art models like GPT (generative pretrained
transformer) by OpenAI and BERT by Google. These
models, pre-trained on a large database of text and fine-
tuned for specific tasks, have significantly advanced the
field, demonstrating human-level performance on a range of
different benchmarks.

Challenges
In general, building and training any type of deep learning
network, especially reinforcement learning-based models,
requires substantial computation and processing resources.
This includes processing massive amounts of data and
performing complex calculations that necessitate
considerable memory and powerful graphical processing
units. However, recent technological advancements and
cloud-based solutions have made these resources more
affordable and accessible, enabling more parties to
participate in deep learning.
 Next, deep learning models thrive on data and while the
recent explosion of big data has provided important fuel for
these data-hungry models, the quality of data is important
too. Erroneous or biased data can cause misleading results
or reinforce existing biases. In addition, deep learning
models contain a tendency to model intricate patterns in the
training data, which can be a strength as well as a potential
weakness. There's a risk that models might overfit the
training data, learning noise and specific details that don't
generalize well to unseen data.
However, perhaps the biggest challenge within the field of
deep learning is the black box nature of these prediction
models. In general, it’s difficult to understand why a
particular prediction was made and this is problematic in
scenarios like healthcare or judiciary scenarios where
transparency and trust are crucial. As discussed in Chapter
3, efforts are ongoing in the field of explainable AI to make
models more transparent and interpretable.
Understanding these strengths and weaknesses is critical
when deciding on potential uses for deep learning.
Businesses without access to massive troves of data and
GPU resources, for instance, may not be suitable candidates
for deep learning, especially if they need a model that is
transparent and easy to visualize.

Key Takeaways
1) Deep learning is an advanced subfield of machine
learning that uses artificial neural networks with deep and
multiple layers to learn and model complex patterns in data.
2) Deep learning models include convolutional neural
networks (CNNs), recurrent neural networks (RNNs), and
transformer networks. Each has its own strengths and
common use cases.
3) Due to the complexity of the neural networks and the
large amount of data they handle, deep learning models
require substantial computational resources. Deep learning
models can also suffer from limitations such as the black
box nature of their decision-making mechanisms and the
risk of overfitting patterns in the training data.
 7

Natural Language Processing

Beyond numerical data, a vast amount of human
knowledge and experience is captured in text and audio.
The ubiquity of words in daily life and the need for effective
human-computer interactions make the ability to process
human language a crucial part of artificial intelligence.
As a multidisciplinary field straddling linguistics, computer
science, and AI, natural language processing empowers
computers with the capability to understand and reproduce
human language. Inspired by linguistics—the study of
language and semantics—NLP was originally designed for
parsing text in databases using coding rule systems, but
over time, it merged with common algorithms from machine
learning to evolve into a novel and specialized field of
computational linguistics. Now, as a field of its own, NLP
involves analyzing human language with reduced emphasis
on quantitative problem-solving, which is typically the focus
of other subfields related to AI.

Early Development
To grasp the impact of recent developments in NLP, it’s
important to see what’s changed and where it all began.
Originally, NLP consisted of human-programmed rule-based
systems, where linguistic experts would manually encode
rules of language into a computer program. While some of
these earlier approaches obtained minor success in domain-
specific cases, they were inherently limited by the
complexity and variability of human language. One example
is ELIZA, which was one of the first chatbots developed by
Joseph Weizenbaum at MIT in the 1960s. This included a
version of ELIZA designed to emulate a psychotherapist
using its ability to reflect user statements back in the form
of a question, as demonstrated in the following example.

Example
Input: I am feeling sad.
Output: I see, can you tell me more about why you are
feeling sad?

If a user said, “I am feeling sad”, ELIZA might respond by

replying, “I see, can you tell me more about why you are
feeling sad?” While the output was impressive at the time,
the ELIZA chatbox was simply following predefined rules
and presenting itself as more intelligent than it actually was
—especially as it didn't understand the full context of the
conversation.
During the late 1980s and 1990s, the landscape of NLP
research evolved with the introduction of machine learning
methods. Rather than relying on hard-coded rules, new
methods used statistical-based models to learn patterns
from large amounts of labeled data. This development
significantly improved overall performance and broadened
the scope of tasks that NLP researchers could tackle.
Despite several advancements, these methods still had
limitations. Notably, they relied on carefully engineered
features derived from expert knowledge, which was labor-
intensive and unable to fully capture the richness and
subtlety of human language. In addition, these methods
struggled to capture longer-range dependencies within a
text, which refers to the relationship between words or
elements in a sentence that are not adjacent or close
together but that still influence each other's meaning. To
explain why, let’s consider the following example.

Example
Input: The girl, who was wearing a red hat that her
mother bought her for her birthday last year, ran down the
street.

Here, the main subject (“The girl”) and the main action
(“ran down the street”) are separated by a long clause
(“who was wearing a red hat that her mother bought her for
her birthday last year”). Despite this distance, we
understand that it is the girl who is running, not the mother
or the hat. Traditional NLP methods, however, struggled to
capture and understand this long-range dependency.
Next, traditional NLP methods lacked the ability to
leverage the benefits of distributed representations, a
method of symbolizing words as vectors in high-dimensional
space. This approach, commonly known as word
embeddings, enables semantically similar words to exist
closer together, thereby enriching the understanding of
textual data.
To understand this approach, it helps to think of a high-
dimensional space as a coordinate system where each word
has its own position. Words with similar meanings or usage,
such as “cat” and “dog”, tend to be closer to each other
because they share a similar semantic meaning related to
animals. Similarly, words like “AI” and “artificial intelligence”
would be close together because they share the same
meaning. These close relationships in a high-dimensional
space are called vectors or word embeddings, which
resemble an ordered list of numerical values. Using these
word embeddings, NLP models are capable of understanding
relationships between words and capturing their contextual
meaning. This allows the model to see similarities and
differences between words and make more accurate
predictions or enhance their understanding.
As mentioned, traditional NLP methods struggled to
analyze vectors in a high-dimensional space. This limitation
and the need to overcome it drove the adoption of newer,
more sophisticated deep learning architectures such as
recurrent neural networks and transformer networks.
Designed to examine the full context and multiple sentences
at a time, these techniques exhibited a superior ability to
handle longer-range dependencies in text, marking a
significant advancement in the field of NLP.
Transformer networks are now used heavily in NLP models
including OpenAI's GPT model and Google's BERT model. In
recent years, NLP has also benefited from several other
emerging trends including zero-shot learning, where models
learn to generalize using fewer labeled examples. This is
especially useful when labeled data is scarce, which is often
the case for less common and low-resource languages.
In practice, rather than relying solely on labeled
examples, zero-shot learning leverages auxiliary
information or semantic relationships to make predictions
on new, unfamiliar classes. In other words, rather than
training the model on numerous examples for every class,
the model is trained on a smaller selection of classes. Then,
when presented with a completely new class that it hasn't
seen before, the model uses the hints and relationships it
learned from the initial training data to make a prediction.
In effect, this allows the model to generalize and perform
tasks in domains it wasn’t explicitly trained to do. As an
example, imagine you have a model that has learned to
identify different animal classes, but the model has never
seen a zebra before. If the model knows that zebras are
similar to horses but they have black and white stripes, the
model can recognize a zebra without ever seeing any direct
examples of a zebra. In this case, the class “zebra” is
already labeled in the dataset as a horse with black and
white stripes.
Another promising trend has been self-supervised
learning, which has been one of the driving forces behind
the success of large language models such as GPT-3. This
approach describes models that learn representations from
unlabeled data, leveraging the abundant unannotated text
available on the Internet.
In the context of GPT-3 and large language models,
learning involves two main steps: pretraining and fine-
tuning. In the pretraining phase, the model is exposed to a
large collection of text, sourced from books, articles, web
pages, and other sources. Here, the model learns to predict
the next word in a sentence based on the preceding
context, such as the following example.

Example
Input: “The cat sat on the ______”

In this example, the model needs to predict the next word

after “the”, which might be a word like “mat”. The process
repeats for numerous sentences in the dataset, exposing
the model to a wide range of words and contexts. Through
self-learning, the model learns to capture the statistical
patterns and relationships contained in the text. It learns
that after “The cat sat on the”, words like “mat”, “chair”, or
“floor” are likely to follow based on patterns observed from
the training data.
By continuously training on a diverse range of sentences,
the model gradually develops a strong understanding of
language. It learns to recognize grammatical structures,
understand the meaning of different words, and grasp the
contextual relationships between words. This pretraining
phase helps the model acquire a broad knowledge of
language, allowing it to generate coherent and contextually
relevant responses in conversations.
After pretraining, the model enters the fine-tuning phase.
During this phase, the model is trained on specific
supervised tasks with labeled data. For example, it can be
fine-tuned on a conversational dataset where inputs and
corresponding responses are provided. By training on this
labeled data, the model learns to generate appropriate
responses based on different inputs.

NLU & NLG

Broadly speaking, NLP can be divided into two distinct
approaches: natural language understanding (NLU) and
natural language generation (NLG). NLU enables a model to
comprehend and derive meaning from human language,
while NLG empowers a model to generate coherent and
contextually appropriate text. This involves creating entire
sentences, paragraphs, or even entire articles. NLG can be
used for tasks such as generating descriptions of data,
writing news stories, summarizing text, or creating
conversational agents (known as chatbots). The core
process involves determining what information to include
and how to organize it (document planning), before putting
the information into appropriate sentences (sentence
planning), and then realizing the sentence plans in the
actual text (text realization).
NLU, meanwhile, aims to extract meaning, sentiment,
intent, and other semantic features from the text to
perform sentiment analysis, named entity recognition, or
text classification.
In terms of their application, both approaches start with
raw, unstructured text data. This data emanates from
various sources such as databases, APIs, scraped website
content, user inputs, and chat logs. As with any data
processing data, the raw data is cleaned, which usually
involves removing inconsistencies, inaccuracies, or
irrelevant information. Depending on the source of the data,
this could involve tasks such as reformatting and removing
or correcting typos.
Using a variety of techniques and preprocessing
algorithms, the text is converted into a structured format
that machines can process and understand. This involves
tasks such as tokenization (breaking text into individual
words or tokens), stemming (reducing words to their root
form, i.e., “naturally” > “nature”), and removing stop words
(common words such as “and”, “the”, and “a” that don’t
provide informational value) as well as unnecessary white
spaces.

Natural Language Understanding

After transforming the text into a suitable format, various
tasks can take place. In the case of natural language
understanding, common tasks include topic modeling,
sentiment analysis, named entity recognition, and text
classification. Sentiment analysis involves determining the
sentiment or emotion of a sentence or document, usually
categorized as positive, negative, or neutral. Named entity
recognition identifies and classifies named entities (i.e.,
persons, organizations, locations) in a text. Text
classification classifies text into predefined categories, such
as spam detection in emails and categorizing news articles.
Lastly, as an unsupervised learning approach, topic
modeling is used to discover abstract topics within a text
document.
After undertaking one or more of these tasks, the model
is evaluated based on appropriate metrics (such as
accuracy, precision, recall, F1-score) to measure its
performance. Then, once the model is evaluated and fine-
tuned, it is deployed to real-world applications and used to
inform decisions. Common examples include chatbots and
virtual assistants such as Apple's Siri, Google Assistant, and
Amazon's Alexa. These applications employ NLP to
understand user commands and generate appropriate
responses, enabling users to set reminders, search the
Internet, control home devices, and so forth. In the realm
of social media, NLU is used for sentiment analysis, helping
businesses understand public sentiment towards their brand
by analyzing text data from social media posts or customer
reviews. The same technology can also aid in detecting
online harassment or toxic behavior on online platforms.

Natural Language Generation

Having discussed common tasks in natural language
understanding, let’s revisit natural language generation.
Following the data preprocessing stage, NLG typically
consists of three stages: document planning, sentence
planning, and text realization.
Document planning is the initial stage where the content
to be included in the output text is decided. The system
identifies the information from the source data that needs
to be conveyed in the text. For instance, if an NLG model is
generating a weather report, the document planning stage
will identify the key data points to be included in the report,
such as temperature, precipitation, and humidity. This stage
includes organizing the selected content into a coherent
structure and determining the order and manner in which
the information will be presented.
Sentence planning, known too as microplanning, next
determines how to express the selected information in
linguistic terms. It decides on the specific words and
phrases that can be used to represent the information. It
also determines the structure of the sentences, taking into
account different aspects such as grammatical correctness,
cohesion with the surrounding sentences, and variability in
expressions to avoid repetitiveness. Using the weather
report example, this stage would involve choosing whether
to say “It's expected to be sunny” or “Sunshine is
predicted”, among other considerations.
Lastly, text realization is where the actual text is
generated. Here, the system transforms the linguistic
representations from the previous stage into a final, fluent
text. It ensures proper grammatical structures and pays
attention to other aspects of the language such as
punctuation and agreement between subjects and verbs.
Using the weather report example again, this last step
would involve generating the final form such as “Tomorrow,
it's expected to be sunny with a high of 25 degrees”.
These three stages often involve complex algorithms and
rules, and more recently, newer language models such as
GPT have proved highly successful at generating fluent and
coherent text, leading to significant advances in the field of
NLG and generative AI, which we’ll discuss further in the
following chapter.

Challenges
As with other areas of artificial intelligence, NLP is not
without potential challenges and headaches! First,
understanding the intricate nuances of human language,
including irony, sarcasm, and cultural references poses a
major challenge in the field of NLP.
Second, subtle changes in the text or context can shift the
meaning of a word or phrase, and while this is easy for
humans to perceive, it is notably difficult for machines to
notice and comprehend. One example is the interpretation
of homographs, which are words spelled the same but
possess more than one meaning. The word “bass”, for
example, can mean a type of fish or a low, deep voice or
musical instrument depending on the context. Similarly, the
phrase “It's raining cats and dogs” doesn't literally mean
animals are falling from the sky. Deciphering the meaning of
such phrases is something that comes naturally to humans
but is difficult for NLP models.
Issues such as language bias, model interpretability, data
privacy, and the lack of high-quality annotated data for low-
resource languages are other areas of concern and ongoing
development. The issue of bias in NLP is a particularly
thorny issue. This is because AI models tend to learn and
replicate the biases contained in the training data, leading
to biased predictions. This is problematic when NLP models
are used in sensitive areas like recruitment or loan
approval, where biased outcomes can have significant
impacts on people's lives.
Ethical considerations, particularly around privacy, are
also crucial. Text data, whether from social media, emails,
or other sources, often contains sensitive or personal
information. Hence, handling this data responsibly, adhering
to data privacy regulations, and anonymizing data to
protect individual identities become paramount
responsibilities for those involved with collecting the training
data and building the model.
Lastly, the development of NLP models for low-resource
languages and multilingual contexts presents challenges.
While a lot of work in NLP has been done in English, there's
a scarcity of high-quality labeled data for many other
languages. This lack of resources makes it challenging to
develop robust NLP models that can understand and
generate text in other languages. Furthermore, creating
models that can handle text in multiple languages
simultaneously also represents an ongoing challenge in
terms of model sophistication and training.

Key Takeaways
1) Natural language processing is a multidisciplinary field
that empowers computers to process, understand, and
generate human language.
 2) NLP can be divided into natural language
understanding and natural language generation. NLU
concentrates on extracting meaning, sentiment, and other
semantic features from the text, while NLG focuses on
generating coherent and contextually appropriate text.
3) Challenges in NLP include understanding the intricacies
of human language, such as irony and cultural references,
language subtlety, addressing bias, ensuring model
interpretability, protecting data privacy, and developing NLP
models for low-resource languages.
 8

Generative AI
In the ever-evolving world of AI, a unique variant of
machine intelligence is rapidly breaking into software
applications and mainstream use. In this chapter, we will
explore generative AI, looking at how it differs from its
traditional counterpart, and discuss how its rich creativity is
changing the landscape of content creation, with a
particular focus on generative adversarial networks and the
battle for available training data.
First, to understand the strengths of generative AI, it’s
important to understand how it differs from traditional AI.
The first distinction between generative AI and traditional AI
is found in their outputs. Traditional AI, which is sometimes
referred to as discriminative AI, is trained to discriminate,
classify, or predict based on input data fed into a model.
This approach powers a variety of AI applications including
recommendation systems and search engines. Generative
AI, on the other hand, learns the underlying patterns of
input data but rather than provide a prediction or insight, it
uses that learned knowledge to generate an output that is
similar but different from the training data.
ChatGPT, developed by OpenAI, is a prominent example of
generative AI. Instead of predicting a likely output to a
given input, ChatGPT is designed to generate coherent and
contextually appropriate responses, whether that be
drafting emails, writing essays, creating poetry, or
simulating human dialogue.
Similarly, new software applications such as DALL-E (also
developed by OpenAI), Midjourney, and Stable Diffusion are
making waves in the domain of art and image creation.
Visual outputs can be generated using what’s called a
generative adversarial network (GAN), a type of generative
AI model that consists of two neural networks: a generator
and a discriminator. During training, the generator and
discriminator are pitted against each other in a competitive
environment. The generator aims to produce realistic
outputs that can fool the discriminator, while the
discriminator aims to correctly identify whether the outputs
are real or generated. In practice, the generator learns the
target art style by observing and analyzing a vast dataset of
existing artworks. Acting as a type of art critic, the
discriminator then evaluates these generated images and
issues feedback on their authenticity. Initially, the
discriminator can easily identify the generated images as
computer-generated due to the lack of resemblance to the
training data. Using feedback from the discriminator, the
generator adjusts its approach and through multiple
feedback loops, the generator improves its ability to
produce convincing artwork.
Throughout this entire process, the generator and
discriminator are engaged in a competition against each
other. The generator aims to produce art that the
discriminator cannot distinguish from real artwork, while the
discriminator seeks to improve its ability to differentiate
between real and fake art. This adversarial competition
drives both sides to evolve and learn from each other. Over
time, the competition and ongoing feedback loop lead to a
point of convergence where the generator becomes so
proficient at creating art (resembling the style of the
training data) that it deceives the discriminator.
 Building on their success in the domain of image
generation, researchers are also developing GANs for
generating audio, music, writing, and even three-
dimensional objects, leading to a further explosion in AI-
generated content.

Data as a New Asset Class

Riding the wave of large language models powered by
OpenAI’s GPT architecture and advances in GAN techniques,
generative AI is set to radically transform the physics of
content creation and creative expression. From producing
full-length films to customer service avatars trained on
customer service logs and product documentation, as well
as simulating realistic video game and esports
environments, the possibilities are suddenly closer than
many people expected.
Underpinning these developments is access to relevant
training data and the growing adoption of generative AI
software tools is set to unlock a new dimension of value for
existing data. In sports, for example, match data and video
footage of professional athletes can be monetized to create
new and unique content. While this data has traditionally
been used for entertainment and post-match analysis (as
popularized in the film and novel Moneyball), it can now be
commercialized for creative use cases as well.
One example comes from Behaviol, a digital sports
company, working on a platform where gamers can acquire
and develop AI sports stars to compete in virtual
tournaments. The company is starting with cricket, after
purchasing five years of player data to train and generate
unique AI player avatars. This development marks a
departure from traditional sports games, where in-game
athletes come with preprogrammed actions and fixed
behaviors.
 Generative AI will be used by Behaviol to study the
movements, techniques, and playing styles of real athletes
by training AI on their match data and video footage. By
doing so, Behaviol can recreate the actions, style, and
performance of athletes in a virtual environment. This will
enable gamers to play alongside AI-generated athletes and
make it possible to simulate hypothetical scenarios, such as
pitting Michael Jordan against modern basketball stars.
This new path of game and content creation opens up
fertile space for athletes, broadcasters, and sporting
franchises to commercialize their data in innovative ways
and transform the market valuation of different data
classes, including player data and video footage.
In a similar vein, actors, producers, directors, and
film/television companies may also be enticed to
commercialize their archives, including scripts, dialogue,
scenes, and visual elements to produce AI-generated films
and TV series. Actors and other industry professionals,
though, are concerned about the potential for AI to create
digital doppelgängers that replace human talent, especially
for scriptwriters, new talent, and actors playing non-starring
roles. Training AI models on an actor’s film archive, for
example, makes it possible to de-age that actor by
modeling their younger self, reducing the available
opportunities for younger talent.

Challenges
While it’s clear that generative AI offers exciting potential,
it's important to acknowledge the challenges and ethical
considerations that come with it. In fact, generative AI has
opened a Pandora’s box of problems that researchers,
practitioners, and policymakers are now scrambling to
address. This includes ensuring the originality and quality of
AI-generated content, mitigating biases and misinformation,
and handling potential misuse of the technology.
 Among the many applications of generative AI technology,
deep fakes are a widely recognized use case with the
capacity to inflict harm. This involves using generative AI to
create convincingly realistic fake videos or images of
individuals, such as celebrities or public figures, which can
be used for disinformation campaigns, fraud, or
harassment. There are already cases emerging of scammers
using AI to imitate the voice of real people to scam their
relatives to transfer money over a voice message or phone
call.
A 2023 study by the University College London warns
that detection will become increasingly challenging as deep
fake technology continues to evolve. The study found that
humans only have 73% accuracy at identifying deep fake
speech (based on the current technology), highlighting the
need for AI and automated detection systems to mitigate
human deficiencies.
Next comes critical questions regarding originality and
ownership. Specifically, who owns the copyright to a piece
of music generated by AI? Is it the AI's developers, the
users who interacted with the AI, or perhaps none of the
above since it's all machine-generated? Is ownership of AI-
generated content even possible? Under copyright law set
by the U.S. Copyright Office, an author's exclusive right to
reproduce their work does not apply if a work has been
generated by a computer process that operates randomly or
mechanically without human authorship.
Moreover, as AI-generated content is shaped by its
training data, it is possible for the AI to generate content
that is too similar to its training data and encroach on
existing copyright protections. As a case in point, AI-
generated art has been caught adding remnants of artists’
signatures in the bottom corner of the image, triggering
concerns about artistic originality and imitation. This has
prompted the creation of an online database called
haveibeentrained.com, offering artists a means to verify
whether their artwork has been utilized to train AI models.
Additionally, with millions of users generating content
from the same training data, AI-generated content has a
tendency to exhibit bias toward reoccurring perspectives,
case studies, arguments, aesthetics, and phrasing. AI-
generated content also tends to lack the creativity, nuance,
and context that human creators bring to their work. AI
writing tools such as ChatGPT, for instance, lack the ability
to curate case studies and narratives that humans find
interesting or remarkable, leading to dry and substandard
content. The result is commoditization, with many blogs,
email newsletters, and other content channels propagating
the same AI-generated content using popular tools such as
ChatGPT.
Lastly, there is the issue of the training data. Given that
GANs and large language models are trained on vast
amounts of data collected from the Internet, they
inadvertently absorb and ingest the biases, misinformation,
low-quality, and untrustworthy information that exists
online. The Pew Research Center, for example, estimates
that less than half of the health and medical information
available online has been reviewed and validated by a
doctor. Moreover, two different studies on 35 kidney cancer
websites and 188 breast cancer websites found that only
12.5% of these websites fulfilled requirements set by Health
on the Net (HON), a non-profit charity providing quality
assessments of health-related information available
online.11
In addition to problems with the accuracy and quality of
online information, there are concerns over the potential
use of inappropriate and dangerous information for training
generative models. A case illustrating this concern was a
stream on Twitch involving an AI-generated version of
Family Guy that was later removed due to its portrayal of a
bomb threat. During the stream, the character Peter Griffin
began discussing the process of planting a bomb at a venue
in Washington DC.
“First, you need to find a good spot to plant the bomb.
You want to consider where it will cause the most damage
and destruction. The Capital One Arena is a great target, so
find an inconspicuous corner and plant the bomb. Next, set
up a timer to detonate the bomb. I suggest 15 minutes
after you have left the arena. Finally, make sure you have
an escape plan.”12
The streamed episode (which was not affiliated with the
official show or its creators) was subsequently removed by
Twitch for violating the site’s “Community Guidelines and
Terms of Service” but not before these dangerous
comments were made in public. This case not only
highlights the hazards associated with broadcasting AI-
generated content but also underlines the challenges in
governing generative AI content and ensuring the
availability of safe training data.
Finally, there is the problem of AI hallucinations, which
refers to situations where AI generates content or
predictions that are not accurate or reflective of reality.
These hallucinations can occur when the AI model
extrapolates patterns from its training data that don’t
reflect reality, leading to outputs that may seem plausible
but are incorrect or nonsensical. As an example, an AI art
model might “hallucinate” by generating images that
contain objects, features, or details that don’t exist in the
real world. Similarly, in natural language processing, AI-
generated text might include information or connections
that are not factual or coherent.
Hallucinations can occur due to a variety of technical
reasons, including noise or ambiguity in the data,
overfitting, and limited training data where the model
makes assumptions based on the patterns it learned from a
small sample size, leading to inaccurate or unrealistic
outputs. Overfitting occurs when the model memorizes the
training data instead of learning general patterns. As a
result, it replicates outliers or unique quirks in the training
data when generating new outputs, even if those details are
not representative of the broader reality.
These distortions serve as a reminder of the resources
and effort required to ensure that AI-generated content
aligns accurately with reality including the importance of
relevant training data.

The Data Wars

While we have achieved major breakthroughs in
generative AI technology, crawling large portions of the
Internet might not be as straightforward13 for companies
like OpenAI under the changing landscape of data
protection. Organizations invested in data collection are
meeting increased resistance with forces advocating for
more stringent privacy protection. Concerned by the vast
quantities of data being collected, governments around the
globe, are putting up legislative guardrails to protect their
citizens' privacy. As a result, these new regulations
significantly complicate the task of collecting, accessing,
and processing raw data.
At the same time, corporations are assembling their own
walls and protecting their access to data. This unfolding
drama, heralded the Data Wars, is manifesting itself in
several ways. The first is the race for data dominance and
control. Corporations across industries are escalating their
efforts to amass and analyze data as well as monopolize
their access to that data. Consequently, corporations are
displaying a reduced willingness to share their data freely,
and in some cases, they are even requesting a nominal fee,
as exemplified by Twitter. In early 2023, the social media
platform announced its plan to eliminate free API access to
third parties. This change means that companies who
previously relied on Twitter’s API to collect public data from
the site will now need to pay to access this data.
Facebook, Apple, and Google have made similar changes
over recent years to limit the availability of data to third
parties, including the retirement of Google Analytics, which
no longer aligns with current reporting and privacy
requirements. Some of these changes have been made in
response to user privacy regulations, and the fact that
OpenAI has shone the way to monetize existing data
through new innovations in generative AI will only
accelerate the use of walled gardens and the competition
for data. This includes encrypting data more heavily to
make it harder for bots to access and analyze it. Artificial
intelligence may also be used more aggressively to detect
and block bots, adding an extra layer of difficulty to web
crawling. Governments, meanwhile, may begin to regulate
the use of bots more strictly, making it more difficult for
companies to use them for web crawling.
These potential changes highlight the increasing
complexity associated with crawling the web. As a result,
this could lead to challenges in data collection and
utilization, creating obstacles for training models and
developing generative AI. Organizations, for instance, may
need to obtain explicit permission from the government and
other organizations before they can collect the data they
need. Meanwhile, large corporations with direct access to
troves of data, such as Facebook and Google, will look to
take advantage of their exclusive access to valuable data.
Next, as time goes on and generative AI content becomes
more prevalent, AI models will not only be trained on
human-generated data but also on AI-generated data,
introducing an additional layer of bias, misinformation, and
error. In effect, there are already significant problems with
existing information on the Internet created by humans
without adding another layer of confusion caused by
randomness, hallucinations, and the bias of generative AI
models.
Furthermore, there are data privacy issues to navigate
regarding the collection and use of private data. At
companies such as Google and Alibaba, internal use of
generative AI applications such as ChatGPT among
employees was immediately banned due to data security
concerns. Despite potential business use cases for ChatGPT,
large companies are concerned about the possibility of their
data and proprietary information being exposed to external
entities or being utilized for model retraining. To capitalize
on the efficiency gains offered by generative AI, these
companies are instead building their own large language
models to keep employees’ text prompts and data on
company-controlled servers. This approach enables
companies to generate content based on internal training
data rather than using general-use models trained on
unknown data, which is also an important design feature.
While tools like ChatGPT can be game-changing for small
companies and solopreneurs, their applicability within major
corporations like Microsoft or Amazon Web Services is
limited. This limitation arises from the fact that existing
generative AI models are predominantly trained on public
data and designed for general use. Large corporations,
however, must be cautious about relying on large language
models trained on public and unfamiliar data, as this could
lead to errors and negative repercussions. Instead, private
models are needed to generate relevant, authorized, and
company-specific content for important tasks such as
updating product documentation or communicating with
customers through chat. While expensive and onerous to
manage, private language models will make sense over the
coming years as the cost of computing resources continues
to fall and generative AI becomes more entrenched in
internal company processes.

Key Takeaways
1) Generative AI differs from traditional predictive AI
through its ability to create, innovate, and generate new
outputs.
2) Generative AI presents challenges and ethical
considerations, including the misuse of technology for
generating deep fakes, spam, fake news, or phishing
emails. There are also issues regarding the originality,
ownership, and copyright of AI-generated content.
3) The Data Wars refer to the battle between
corporations' data collection efforts and the demand for
privacy protections. Stricter regulations will limit data
accessibility and companies are building walled gardens to
control and monetize their data.

Thought Exercises
1) If generative AI continues to develop, what data will
increase in value? (i.e., unpopular films featuring actors that
can be used to train an avatar for use in other films.)

2) If you were to design your own version of ChatGPT to

help you at work or school, what data would you use to
train it, and why?

3) How can you reclaim more ownership of your data?

(i.e., use a VPN to protect your IP location, opt out of data
collection programs, and understand data collection
processes.)

4) In light of the recent developments in generative AI,

how might you potentially monetize your personal or
organization’s existing data? (i.e., monetize your online
browsing data by participating in programs such as Brave
Rewards)
 9

Recommender Systems
In an era of information overload, recommender systems
have become an indispensable tool for steering people
through the vast ocean of content and navigating the long
tail of available products. Fueled by data and algorithms,
recommender systems can analyze our behaviors and
preferences and then deliver personalized recommendations
tailored to our unique tastes and interests. By
recommending movies, music, books, products, and other
items, they save us valuable time while opening doors to
new discoveries.
In this chapter, we’ll delve into the exciting world of
recommender systems, explore various approaches, and
uncover strategies to maximize their performance. Before
we get started, it’s important to acknowledge that
recommender systems are not built on a single technique or
one family of algorithms. Instead, they represent a
mismatch of techniques and algorithms united under one
common goal: to make relevant recommendations. Whether
it’s machine learning, deep learning, or NLP, recommender
systems use whatever algorithm they can to serve relevant
items to end-users. There are, though, a number of design
methodologies that are specific to recommender systems,
including collaborative filtering, content-based filtering, and
the hybrid approach, which form the core focus of this
chapter.
 Content-Based Filtering
Content-based filtering, also known as item-based
filtering, provides recommendations based on similar item
characteristics and the profile of an individual user’s
preferences. In effect, the system attempts to recommend
items that are similar to those that a user has liked,
browsed, or purchased in the past. After purchasing a book
about machine learning, for example, Amazon’s content-
based filtering model is likely to serve you other books from
the same author, series, or genre.
This approach relies heavily on descriptions of items as
well as the profiling of individual user preferences. A book,
for example, can be described by the following
characteristics:
1. The author(s)
2. The genre, e.g., thriller, romance, historical fiction
3. The year of publication
4. The type of book, e.g., fiction, non-fiction
5. Book format, e.g., paperback, audiobook, e-book,
hardback
Likewise, user preferences need to be collected and
analyzed. Individual user preferences can be determined by
examining:
1. Past purchasing/consumption behavior
2. Browsing history
3. Personal details, e.g., location, nationality, and hobbies
4. IP address (to determine location and time zone)
Using the information gathered, filtering techniques can
then compare this data with the descriptions of available
items to identify and recommend relevant items. If a user
has shown a preference for thriller movies in the past and
has rated several thriller movies positively, a content-based
filtering model can identify these preferences and
recommend other thriller movies with similar traits, such as
genre, actors, and storyline, even if those movies weren’t
highly rated by other users.
Whether it’s movies, books, or other items, the model
aims to recommend items that align closely with the user's
preferences, irrespective of their popularity among the
overall user base. Let’s now review the other advantages as
well as some of the drawbacks of content-based filtering.

Advantages
1. Agnostic to crowd preferences
The first advantage of content-based filtering is that it
aids the discovery of relevant but low-profile items. As
content-based filtering doesn't take crowd preferences into
account, relevant items with low exposure to the crowd can
still be found and promoted.
2. Content items are stable
Items don’t change over time as much as people do and
they are generally more permanent. People, on the other
hand, are fickle and our tastes change over time. We’re all
guilty of following fad diets, new exercise regimes, and
content binges. However, an item will always be an item,
making content-based filtering less vulnerable to short-term
shifts in user preferences and reducing the need for regular
retraining of the model. This, though, could prove a
disadvantage over the long term as the model struggles to
keep up with shifting consumption behavior.
3. Items are generally fewer than users
Most online platforms have fewer items than users, and
content-based filtering can help to conserve computational
resources by comparing a limited number of items rather
than a larger volume of user relationships.
4. Compatible with new items
If there is insufficient rating data for a new or existing
item (known as the cold-start problem), content-based
filtering can be used to gather information regarding other
items rated/purchased/consumed by the target user that
share similar attributes. Items are therefore recommended
based on the user’s interaction with similar items despite
the lack of existing data for certain products.
5. Mitigates cheating
The other notable benefit of content-based filtering is that
it’s generally more difficult to game the system because
malicious actors have less power to manipulate or fabricate
item-to-item relationships. This is not the case for item-to-
user relationships, which can be easily manipulated with a
flood of fake reviews and purchases or views.

Disadvantages
1. Low variety
The variety of recommended items can be limited and less
diverse than other methods. This is because content-based
filtering relies on matching a specific item with similar
items. Thus, unique and novel items with low exposure to
the target user are unlikely to surface, limiting the range of
category discoverability.
2. Ineffective for new users
While content-based filtering methods excel at
recommending new items, this isn’t the case for new users.
Without information about the user’s preferences to
construct a user profile, there’s little way of recommending
related items. To mitigate the cold-start problem, some
online platforms attempt to extract relevant keywords when
onboarding new users. Pinterest, for example, directs new
users to specify a collection of over-arching interests that
are used to establish a preliminary user profile and match
these descriptions to content recommendations. Pinterest’s
machine learning-based models then refine the user’s
profile and their specific interests based on observing their
pins and browsing behavior.
3. Mixed quality of results
Content-based filtering is generally accurate at selecting
relevant items, but the quality of such items can sometimes
be poor. As content-based filtering ignores the ratings of
other users, the model is limited by an inability to decipher
the quality of an item.

Demonstration
To explore how content-based filtering works, let’s run
through a simple demonstration that looks at
recommending films to users according to their rating
history.
For this demonstration, let's assume the films available
are represented based on three features: genre, director,
and production company.

Dataset: Film Metadata

Step 1: Profile Creation

For each user, we first need to create a profile based on
the features of films they have already rated. For instance,
if User 4 gave a high rating to films directed by Christopher
Nolan, then Nolan’s films would be a prominent feature in
their profile.
Dataset: Film ratings (1-5 stars), blanks indicate that the user has not
yet rated the film

Step 2: User 4 Profile

Create a profile for User 4 based on their known film
ratings.
- User 4 likes War/Drama genre and films directed by
Christopher Nolan (because of their high rating for
Dunkirk)
- User 4 dislikes Fantasy/Musical genre and Disney
films (because of low ratings for The Little Mermaid and
Mary Poppins)

Step 3: Compute Scores

For the films that User 4 hasn't rated (Oppenheimer and
Beauty and the Beast), we need to calculate a score based
on their profile and the film's features. The score is derived
from how many features of the film match the user's
preferences.
Oppenheimer: Matches with War/Drama and
Christopher Nolan (1 positive match)
Beauty and the Beast: Matches with Fantasy/Musical
and Disney films (2 negative matches)

Step 4: Recommend
Compared to Beauty and the Beast, Oppenheimer appears
to be a better film recommendation for User 4 based on 1
positive match.
 Keep in mind that this is a highly simplified representation
of content-based filtering. In real-world scenarios, you
might use techniques like TF-IDF (Term Frequency-Inverse
Document Frequency) to represent film features and cosine
similarity or other metrics to determine the similarity
between user profiles and film features. The idea, however,
remains the same: understanding the content similarities
between items and the user's preferences and making
recommendations based on those relationships.

Collaborative Filtering
Reflecting the wisdom of the crowd, collaborative filtering
recommends items to an individual based on the
preferences and consumption trends of other users with
shared interests. For instance, on TikTok, users who enjoy
fitness content might also find personal finance content
appealing. Under this scenario, the items (fitness and
personal finance videos) may not share the same genre or
title keywords. Despite this, the recommender system will
still suggest personal finance videos to fitness enthusiasts
based on the behavioral patterns of similar users.
Collaborative filtering, though, should not be mistaken as
a popularity chart or a top ten list of popular items. Rather,
it uses two distinct methods to match items that share
popular associations among similar types of users. The first
method is user-based collaborative filtering, which
generates recommendations to a target user based on
analyzing the historical preferences of users with similar
tastes. In other words, people similar to you who buy x also
buy y.
In practice, this works by identifying like-minded users.
Their ratings or preferences are then collected and grouped
to produce a weighted average. The group’s general
preferences are used to recommend items to individual
users based on the ratings and preferences of their peer
group. For instance, if a user has never watched Squid
Games and their peers have all watched and rated it
positively, the model will recommend Squid Games to the
user based on peer observation.
The second method is item-based collaborative filtering.
Rather than finding users with similar preferences, this
method finds a set of items similar to the target item based
on user preferences. For example, Star Wars movies rated
highly by a similar audience of users will be matched
together as a set and then recommended to other users
who like and rate one of the movies in the set. Item-based
filtering can therefore be thought of as people who buy x
also buy y.
The main distinction between these two methods lies in
the selection of input. Item-based collaborative filtering
takes a given item, finds users who liked that item, and
then retrieves other items that those users liked.
Conversely, user-based collaborative filtering takes a
selected user, finds users similar to that user based on
similar item ratings or purchases, and then recommends
items that similar users also liked.
In reality, both methods tend to produce similar item
recommendations, but user-based collaborative filtering can
be more accurate for datasets that have a large number of
users with diverse or esoteric interests. Datasets that have
less information regarding user characteristics and tastes,
though, are generally more compatible with item-based
collaborative filtering.

Advantages
1. Low knowledge of item characteristics
The first advantage of collaborative filtering is it doesn’t
rely on a sophisticated understanding of items and their
attributes. This saves upfront effort because you don’t need
to spend time meticulously documenting items. This is
especially convenient for online video and audio content
items that are generated daily and are time-consuming to
review and classify.
2. Flexible over the long-term
As collaborative filtering responds directly to user
behavior and trends, this approach is generally more
flexible than content-based filtering at reacting to changes
in user/consumer behavior. Sudden short-term changes in
fashion, pop culture, and other fads, though, can be difficult
to respond to—at least initially—depending on when and
how regularly the data is collected.
3. Discoverability
Collaborative filtering enables the discoverability of items
outside the user’s periphery as it synthesizes preferences
from users they’ve never met but who share similar
interests.

Disadvantages
1. Large-scale user data
One drawback of collaborative filtering is the significant
amount of upfront information needed to understand user
preferences. While Amazon and Netflix have enough user
data to ride out sparsity problems in the data, new
platforms without an established user base face limitations
because collaborative filtering is largely ineffective without
sufficient information. Obtaining or acquiring data from a
third party, as Amazon did by partnering with AOL in the
early 2000s, is one strategy to overcome the cold-start
problem.
2. Malicious activity
Collaborative filtering is highly vulnerable to people
gaming the system and doing the wrong thing. This includes
driving fake traffic to target items, attacking competitor’s
items with negative reviews, fabricating online user
personas, or creating a general system of user actions to
cheat the system, known in the industry as a shilling attack.
One approach to minimize malicious activity is to limit the
model’s analysis to user purchases, rather than browsing
habits, as the former is more difficult to fabricate. That said,
fraudulent online transactions remain common, and
unscrupulous actors are constantly developing their tactics
to game recommender systems.
3. Negative reputation
As collaborative filtering relies on extracting users’
personal information to generate recommendations, it
inevitably raises questions regarding data privacy and social
manipulation. Criticism has surfaced in recent times
regarding the U.S. election and the alleged role Facebook
has in sharing user data with third-party organizations as
well as their content display algorithms that potentially
reinforce political biases and disseminate news stories.
4. Consistency
Aside from different tastes and preferences, users have
different standards—making it difficult to trust the
consistency of rating data aggregated from multiple users.
The meaning of a three-star can be interpreted differently
among users based on their average rating history, for
example. Based on personal experience, standards also
vary between countries and types of users (i.e., e-book
readers versus physical book readers). Readers of physical
books, for example, rate negatively when there are delivery
delays or printing issues, which doesn’t affect e-book
readers who receive a digital copy on demand. To improve
consistency, some models may need to filter user ratings by
additional criteria such as country and customer type.

Demonstration
 In this second demonstration, we will use user-based
collaborative filtering to make a film recommendation to
User 1.

Dataset: Film Ratings (1-5 stars)

Steps
1) Compute similarity: Calculate similarity scores between
users. One common method is the Pearson correlation
coefficient, which is a number between -1 (non-identical)
and 1 (identical) indicating how closely two things are
related.
2) Predict ratings: Predict the ratings of films that the
target user hasn't watched by considering the ratings of
similar users.
3) Recommend: Films with the highest predicted ratings
are recommended to the user.

Example
1) Let's predict a rating for The Little Mermaid for User 1,
who is yet to watch that film.
2) We notice that both User 1 and User 2 have rated
Oppenheimer, Beauty & the Beast, and Dunkirk. Based on
this, we can compute their similarity.
3) If their similarity score is high, we can use User 2's
rating of The Little Mermaid to predict User 1's potential
rating for that film.
4) As User 2 only rated it as 1-star, it’s not worth
recommending this film to User 1.
 The Hybrid Approach
After exploring the advantages and disadvantages of
content-based and collaborative filtering, you probably
spotted some trade-offs between these two techniques.
Content-based filtering, for example, tends to be less
diverse than collaborative filtering in terms of the items it
recommends but is, overall, more consistent than
collaborative filtering. To reduce the effect of these various
trade-offs, an alternative to collaborative and content-based
filtering has been developed, which draws on a combination
of techniques to deliver useful recommendations. This
approach is aptly named the hybrid approach and can
function either as a unified model or by separating content-
based and collaborative filtering and then combining their
predictions.
In addition to bridging the gap between content-based
and collaborative filtering, the hybrid approach plays a
crucial role in overcoming the cold-start problem, which
occurs when there is insufficient user interaction data or
item attributes needed to make accurate recommendations.
Hybrid systems offer an elegant solution by capitalizing on
the strengths of both collaborative filtering and content-
based filtering, effectively mitigating the limitations that
each technique faces individually. For new items that lack
user interaction data, the hybrid approach can prioritize
content-based filtering. By analyzing the attributes,
descriptions, or features of the item, it can generate
relevant recommendations based on the item's
characteristics. In the case of new users who haven't yet
provided interaction data, user-based collaborative filtering
can be used to analyze similar users based on attributes
such as IP location, age, and gender to overcome the cold-
start problem. As the interactions of the new user
accumulate and the model learns their individual
preferences, the hybrid model can gradually transition to an
item-based collaborative filtering or content-based filtering
approach.
Finally, the hybrid approach offers added flexibility to
combine multiple data sources and data types. Ordinal data
values such as item ratings (1-5 stars), for example, are
generally used for collaborative filtering, whereas
continuous variables such as item price and size are more
suitable for content-based filtering. Using a hybrid solution,
you can pipe both data inputs and then segment analysis
through a curated selection of filtering techniques.

Training Recommender Systems

Understanding how recommender systems function, even
if you aren’t a machine learning developer, is worthwhile for
creating a positive online experience or growing an
audience on popular content platforms. In this section, we
will look at how you can train recommender systems based
on your behavior.
The first step is understanding what type of recommender
system is being used to serve you recommendations.
Skillshare, for example, typically recommends items using
content-based filtering, whereas Spotify is more likely to
utilize collaborative filtering to recommend music to users.
However, the more advanced and established the platform,
the more likely it is that the platform is using a hybrid
approach. Platforms may also employ different techniques
in isolation based on different user scenarios. The YouTube
homepage, for instance, is more likely to use collaborative
filtering to enhance discovery and emphasize variety on the
platform, whereas the video page sidebar is more likely to
use content-based filtering to keep you on that page with
related content.
The next clue is how the platform labels its
recommendations. On major platforms such as Amazon,
you might see labels such as “See what other users bought”
(collaborative filtering) or “We thought you might like these”
(content-based filtering). However, in a lot of cases,
distinguishing the use of one method over the other can be
challenging due to the implementation of hybrid systems.
In practice, it’s best to assume that both techniques are in
use and to adapt your approach accordingly. This includes
careful labeling of your product or content to train content-
based filtering engines, such as the item’s metadata,
description, tags, and other labels. The more relevant
information you provide, the more likely the recommender
system is to pick up your item and showcase it to the right
audience. Having said that, it’s vital to steer clear of adding
irrelevant labels to items in a bid to deceive the system.
Established platforms such as YouTube and Amazon closely
monitor attempts to cheat the system and will penalize you
accordingly. Additionally, disappointing users with deceptive
titles or labels may impact your conversion metrics, which
also feed into the recommender system as well as the
organic search rankings.
In addition to accurate labeling of items, identifying
opportunities to associate your item with popular items
within the same category can be an effective strategy to
capture spill-over traffic. This tactic is commonly used on
platforms like YouTube, where creators know that producing
new videos inspired by popular hits can help to capture
traffic from related recommendations. As an example, if you
create a popular video featuring a tour of your minimalist
Tokyo apartment, then YouTube is likely to recommend the
same viewers to other minimalist Tokyo apartment tours
after watching your video.
Another way to train the recommender system on your
product or content is to incorporate paid advertising. In the
context of content-based filtering, you will need to focus on
bidding for specific keywords relevant to your product,
rather than opting for broad keywords with lower
conversion rates. For example, if you want Amazon to
associate your book with the keyword “machine learning”,
then you will need to target this keyword in your ad
campaigns rather than using a broad keyword such as
“computer science” or “technology”. If the bid cost for your
target keyword is too expensive, then you may need to try
using long-tail keywords that are cheaper but still contain
the desired keyword. For example, rather than paying 80
cents per click for the keyword “machine learning”, you can
target the long-tail keyword “machine learning for
dummies”, which only costs 40 cents per click and will still
associate your book with the keyword “machine learning”.
Another effective approach is driving paid traffic from an
external platform, enabling you to reach a broad audience
without disrupting the native recommender system. This
way, Amazon's recommender system won't identify the
specific keywords you are targeting when you advertise on
other platforms like Google or BookBub, for example. It
doesn’t matter what keywords you are targeting on a third-
party platform, because Amazon simply sees traffic coming
to your product page. What’s more, if the on-page
conversion rate is high, this will help to enhance your book’s
discoverability on Amazon, as conversion and sales are key
metrics for recommending books and other items on the
platform.
While paid advertising shouldn’t negatively affect the
recommender system’s ability to associate your book with
specific keywords, this method is not as effective as
targeting relevant keywords through paid advertising on the
same platform. When you use Amazon Marketing Services
to advertise, you are effectively training Amazon to
associate your product with specific words or audiences.
Likewise, when you advertise through TikTok Ads Manager,
you are helping to train your content on the powerful TikTok
content algorithm.
 In the context of collaborative filtering, it’s essential to
focus on the audience and consider segmenting them
strategically to maintain relevance. This practice is often
observed on YouTube, where creators operate multiple
accounts to target specific audiences based on content
preferences. Some football vloggers, for instance, may have
one channel dedicated to their favorite team and another
broader channel focused on the football league they follow.
This approach helps to ensure highly engaged audiences for
each channel, and this makes it easier for the recommender
system to identify potential audiences.
The next important consideration is assessing what traffic
you want to send to your product or content. If you want to
promote your new romance novel to a niche audience on
Amazon, it's crucial to avoid targeting family and friends
who have no prior interest in the romance genre and who
aren’t likely to make relevant purchases in the future. By
selling your book to a mix of readers, the recommender
system will be unable to identify and profile the target
audience of your book, hindering its ability to recommend
your book to those genuinely interested in your niche.
To maximize the power of collaborative filtering, you also
want to avoid sabotaging your items with irrelevant paid
traffic. Delivering ad campaigns based on broad user
targeting and low-cost clicks might lead to a positive return
on ad investment, but the broad targeting hampers the
recommender system’s ability to profile connections
between buyers. In addition, a low conversion rate (in
terms of purchases or views) will negatively impact your
item’s visibility in the general search results on the platform
and reduce the flow of recommendations. In general, you
want to promote your item to a narrow audience with
similar interests and accumulate a high conversion rate in
order to maximize the power of collaborative filtering.
 In summary, vigilance is key when handling traffic and
keywords associated with your content or products. How
you describe and advertise each item feeds into the training
data for the recommender system and impacts future
recommendations.

Key Takeaways
1) Content-based filtering recommends items based on
their characteristics and a user's preferences. It relies on
item descriptions and user profiling, matching similar items
to those the user has liked or browsed in the past. Content-
based filtering is effective for recommending new items but
may lack variety and struggle with recommending items to
new users.
2) Collaborative filtering recommends items based on the
preferences of similar users with shared interests. It
leverages the wisdom of the crowd and can suggest items
that users may not have discovered otherwise.
3) The hybrid approach is a combination of collaborative
and content-based filtering techniques. Hybrid
recommender systems can operate by running as a unified
model or by separating content-based and collaborative
filtering and then combining their predictions.
4) Understanding the type of recommender system used
by a platform is essential for marketers, content creators,
and general users. Properly labeling items, associating them
with popular items, and leveraging paid advertising can
train recommender systems to enhance accurate targeting
and reach relevant audiences.

Thought Exercises
1) It’s the first day of the new year and you want to
create a healthy lifestyle built around nutrition, fitness, and
positive energy. Now, choose a popular online platform like
TikTok, Instagram, YouTube, or Amazon, and list three
different ways you can train the model to recommend
content rated to your goals for this year. (Hint: think about
how you can use search and other interactions to leave
clues for the model, including unsubscribing to channels
that don’t support your goals or rating relevant
content/purchases)

2) Imagine you have a video that you want to publish on

YouTube. What actions can you take to help train a content-
based filtering system to recommend your video to more
viewers?

3) Imagine you have a book that you want to sell on

Amazon. What actions can you take to help train a
collaborative filtering system to recommend your book to
more relevant readers?
 10

Computer Vision
As human beings, we have an innate ability to recognize
objects, read signs, and understand our surroundings as we
move through the physical world. This effortless ability to
comprehend depth, motion, and spatial layouts is a
byproduct of our complex visual system, a system that has
evolved over hundreds of millions of years. Naturally, as
technology has advanced, humans have been determined to
impart the same perception of digital and physical spaces to
computers, which brings us to the exciting field of computer
vision.
As an important subfield of artificial intelligence, computer
vision enables machines to understand and interpret visual
information. However, unlike human vision, which is based
on our biological visual system, computer vision utilizes
digital inputs, computational models, and deep learning to
interpret and understand the content of visual inputs. This
complex process involves various sub-tasks, including
identifying objects within the image, understanding the
context in which objects exist, recognizing patterns and
anomalies, and even predicting future states based on past
data.
To learn more about how computer vision works, this
chapter will introduce and examine the following use cases:
image classification, object detection, and image
segmentation.

Image Classification
One of the most common computer vision use cases is
image classification, which involves categorizing an image
into one of several predefined classes. As a form of
supervised learning, the image classification model learns
from a training set of labeled images with descriptions of
what’s contained in each image. By training on many
examples, often millions of images, the model gradually
learns to recognize certain object classes such as whether
the image contains a cat or a dog. However, using
supervised learning, the model can only recognize object
classes that are included and labeled in the training data. If
the model encounters objects that weren’t part of the
training data, it will struggle to correctly classify or
categorize them.
To overcome this limitation, there are a number of other
approaches including transfer learning, data augmentation,
zero-shot learning (i.e., using knowledge of horses and
stripe patterns to classify a zebra), as well as detection and
outlier handling. In the case of transfer learning, an existing
model trained on a large dataset is fine-tuned and used on
a smaller dataset. This approach leverages the general
feature extraction capabilities learned from the large
dataset and adapts them to the new classes. While it might
not fully resolve the out-of-class problem, it can improve
the model's ability to handle unseen classes.
Instead of trying to classify everything into predefined
classes, anomaly detection and outlier handling focuses on
identifying instances that deviate significantly from learned
patterns in the training data. This can help the model detect
and flag instances that don't belong to any known class,
which may be useful for specific use cases such as fraud
detection.
A different approach is data augmentation, which involves
artificially expanding the training dataset by applying
various transformations to the existing images, such as
rotation, cropping, scaling, and flipping. This helps the
model become more robust to variations in the input data
and can improve its performance with difficult cases.

Object Detection
Object detection is a more complex use case than image
classification, where the goal is to classify multiple objects
and accurately determine their locations. This technique is
widely used in applications that involve detecting and
tracking objects in images and videos. This includes
autonomous vehicles, surveillance systems, face detection,
and defect detection in manufacturing, where the ability to
identify and locate objects is crucial for navigation and
decision-making. Additionally, object detection has paved
the way for advancements in human-computer interaction,
as it enables gesture recognition and tracking of body parts,
facilitating immersive experiences in virtual reality and
augmented reality applications.
Unlike image classification where the model checks if the
image contains a single object class such as a cat or dog,
object detection provides detailed information about where
each target object is located within the image. It does this
by generating bounding boxes that enclose individual
objects before predicting the class of each object. As an
example, suppose there is a photo containing a range of
objects such as cars, pedestrians, and traffic signs. Rather
than make a single classification, the model analyzes the
image by placing bounding boxes around each car,
pedestrian, and traffic sign, and then individually labels
each object.
 By capturing the coordinates and dimensions of each
bounding box, object detection provides insights into the
size, orientation, and proximity of different objects. This
information is valuable in applications such as self-driving
cars, where precise localization of pedestrians, vehicles, and
obstacles is essential for making real-time navigation
decisions. Object detection also extends to scenarios where
objects might partially occlude or overlap each other.

Image Segmentation
In the case of image segmentation, the goal is to partition
an image into multiple segments or regions based on pixel
groupings, each of which corresponds to a different object
or part of an object. Rather than classify one or multiple
discrete objects, it segments the image into regions. This
means that instead of isolating a pedestrian by drawing a
big rectangular box around them and classifying the object
that way, image segmentation will isolate the pedestrian
pixel by pixel and assign an object label to that selected
group of pixels. This division is based on certain criteria and
different levels of granularity, such as color, texture,
intensity, or other visual properties, which are used to
differentiate different regions. For example, if there is an
image of a person, image segmentation can be used to
divide it into segments representing the person's face, hair,
clothing, and background. By assigning different labels to
each segment, we can distinguish and analyze the different
regions separately.
Image segmentation therefore aids in understanding
images in finer detail than object detection. This makes it
useful for tasks requiring meticulous image understanding,
including precise boundary determination and accurate
analysis of specific components within an image, which is
essential for tasks such as image editing or medical image
analysis to detect and track tumors.
 How it Works
Whether it’s image classification, object detection, or
image segmentation, computer vision involves a clear
sequence of steps that center on taking digital inputs and
extracting high-dimensional data to generate numerical
information that can be used to analyze and classify
objects.
High-dimensional data refers to data with many features,
known also as dimensions. In the context of computer
vision, this refers to data containing a multitude of features
from visual inputs such as images or videos. Within an
image, each pixel can be considered a variable or
dimension. If the image has a resolution of 1000x1000
pixels, this means there are one million features or
dimensions associated with that one image. Further, each
pixel contains values representing color intensity, which
further adds to the dimensionality of the data. To decipher
the details contained in an image or video with millions of
features, computer vision involves a series of complex
tasks, which we’ll examine in this section.
The first step is image acquisition, the process by which
an image is captured through a camera sensor. Here, the
captured image is digitized and stored as a grid of pixels,
each with different color and intensity values (also referred
to as grayscale values, which represent the brightness or
luminance of a pixel in an image).
Once the image has been acquired and digitized, the next
step is image preprocessing. This involves improving the
image quality and extracting useful features. This might
include noise reduction, contrast enhancement, edge
detection, and other relevant operations.
The next step is feature extraction in which the goal is to
identify and extract important characteristics from the
processed image. Features might include edges, corners,
textures, or more complex structures such as shapes or
objects. The exact features that are extracted depend on
the specific problem being solved.
Once the features are extracted, they can be converted
into numerical representations and used to perform various
tasks. In the case of object recognition, the features can be
used to identify specific objects within the image. Typically,
this is done by comparing the extracted features to those
stored in a database. In the case of image segmentation,
the features might be used to partition the image into
different regions corresponding to different objects or parts
of objects based on pixel groupings.
The final step is decision-making, whereby the results of
the previous steps are used to make some determination
about the image. For example, in object recognition, the
system might decide which objects are present in the
image. In image segmentation, the system might decide
how to divide the image into different segments.
To put this entire process into perspective, let's imagine
that we have an image of a person's face and we want to
develop an image classification system that can
automatically identify the individual in that image. The
system would go through the following steps.
1) Acquisition: The system acquires the image of the
person's face either through a camera or by loading a pre-
existing image.
2) Preprocessing: The image is preprocessed to
enhance its quality and normalize any variations in lighting,
orientation, or scale. This step helps to improve the
accuracy of the subsequent analysis.
3) Feature extraction: The system extracts relevant and
distinctive high-dimensional data from the image by
identifying key facial features such as eyes, nose, mouth,
and facial contours.
 4) Numerical representation: The extracted facial
features are converted into numerical representations, such
as vectors, based on numerical values with each
corresponding to a specific feature or characteristic of a
facial feature.
5) Training and recognition: Using a dataset
encompassing known facial images of various individuals,
the system learns to associate the extracted numerical
features with specific identities. This training facilitates the
system's ability to recognize and differentiate between
different individuals.
6) Identification: To identify an image of a person, the
system compares the numerical representation of the
unknown face with the trained representations of known
individuals. It computes the similarity or distance between
the numerical representations and determines the closest
match, thereby identifying the person in the image.
As outlined, by extracting and analyzing high-dimensional
data related to facial features, such as the arrangement of
eyes, nose, and mouth, the system can generate numerical
information that enables it to recognize and identify
individuals in images.
Next, to carry out this sequence of complex tasks,
computer vision relies heavily on machine learning and
deep learning techniques. Deep learning, and specifically
convolutional neural networks, discussed in an earlier
chapter, have proven particularly effective at handling
computer vision tasks as they can learn to recognize
complex patterns and features in images, allowing them to
achieve high accuracy on tasks such as image classification,
object detection, and image segmentation.
In terms of human input, computer vision projects
generally involve a large team with a diverse range of skills
and expertise. This includes individuals with backgrounds in
computer science, machine learning, and possibly fields
such as cognitive science and optics. General software
developers or engineers are also needed to integrate the
computer vision system into larger software systems or
applications.
For supervised learning projects, the team might need
data annotators who can label images to create training
data for the models. This process often requires a
meticulous and detail-oriented approach but can be
outsourced using services such as Amazon MTurk or another
crowdsourced service. Also, depending on the specific
application of the computer vision system, having team
members with expertise in the relevant domain can be
advantageous. For example, having a team member with
experience in healthcare or medical technology can provide
important insights for a project related to medical imaging.
Lastly, having team members who are up to date with the
latest research in the field can be beneficial as they can
guide the team in adopting the latest technology as well as
other best practices.

Challenges
Navigating the field of computer vision involves
addressing various challenges and potential risks. This
section will explore some of the technical and non-technical
challenges or risks associated with the development and
application of computer vision. For startups and businesses,
understanding the following challenges will be important for
mitigating potential problems and devising effective
strategies.
In terms of technical challenges, one of the biggest
challenges is data. The performance of computer vision
systems depends on the quality and diversity of the data
used for training the model. Amassing a large and diverse
dataset that reflects the wide spectrum of scenarios the
system will encounter in real-world applications can be a
daunting and costly process. This is because real-world
scenarios are filled with complexities, including variations in
lighting, viewpoint or camera angles, object orientations
(which refer to the different possible orientations or poses
that an object can take in three-dimensional space), and
objects obstructed from view, also known as occlusions.
To succeed, the model must contend with these many
different variations along with feature similarity. Feature
similarity refers to the degree of resemblance or likeness
between different features extracted from data points within
a dataset. An example of this problem arose when a
Scottish football club began using AI to broadcast their
matches in 2020, before receiving a public lesson in feature
similarity.
After the COVID pandemic prevented fans from attending
live matches in the Scottish Championship, Inverness
Caledonian Thistle FC made a strategic decision to
livestream its matches online. With external support, they
arranged an automated camera system equipped with built-
in AI technology for tracking the match ball’s movement. By
tracking the ball, the objective was to ensure that viewers
at home always had an optimal view of the game. Instead,
the AI camera repeatedly confused the linesman’s head for
the ball, whose bald head shared visual similarities to the
yellow match ball. The camera angle contributed to the
confusion, creating the perception that the linesman's head
was inside the boundaries of the football pitch.
Pixellot, the company responsible for the technology, had
to swiftly update its system after the match, as the incident
attracted widespread attention and fed into memes on
social media. While this unfortunate example presents a
more amusing demonstration of the brittleness of AI and
computer vision, the potential cost of a mistake is far more
severe when the AI is behind the wheel of a self-driving
vehicle. Tesla’s Autopilot software has been at fault for
multiple fatal crashes over the years, including one case
where the Tesla Model 3 collided with a tractor-trailer that
crossed its path, shearing off the Model 3’s roof.
These case studies underscore the challenges posed by
feature similarity and general data complexity. Building
capable models that can understand and interpret intricate
details across diverse scenarios requires extensive testing as
well as the expert knowledge of skilled human teams. In
addition, computer vision demands substantial processing
power and memory, especially in the context of deep
learning. This demand can pose significant challenges for
applications that require real-time processing or when
deploying systems on devices with limited resources. Thus,
balancing the need for sophisticated models with the
practical constraints of computational resources does
require careful planning and execution.
Alongside the considerable technical challenges, it’s
important to acknowledge the non-technical challenges as
well. These span from financial to legal risk, and
organizations must be cognizant of them when entering this
field.
First, developing computer vision technologies can be an
expensive venture. The quest for large amounts of training
data and computational resources leads to significant
upfront costs and the return on investment often takes
longer than expected due to the technical challenges
mentioned. Second, as with many emerging technologies,
computer vision raises a number of legal and ethical
dilemmas.
One of the primary dilemmas is privacy. With the ability to
analyze images and videos, computer vision applications
have the potential to infringe on individuals' privacy rights.
Applications such as facial recognition have been a subject
of controversy in recent years. In one notable case,
Clearview AI, a facial recognition software company, faced
multiple lawsuits and backlash after it was revealed that
they were scraping billions of images from social media
platforms without users' consent.
Similarly, computer recognition is at risk of infringing
intellectual property. For instance, accidentally using images
or videos protected by copyright to train computer vision
models could result in copyright infringement.
In regard to potential legal regulations, various
jurisdictions have different regulations regarding the use of
computer vision systems, particularly concerning
surveillance and data protection. The European Union's
General Data Protection Regulation (GDPR), in particular,
places strict requirements on the use of personal data,
which includes images and videos analyzed by computer
vision systems. Failure to comply with these regulations can
result in significant penalties.
In light of the various compliance requirements and
ethical challenges, it is vital for organizations working on
computer vision projects to have robust strategies in place
to manage these issues. This includes implementing strict
data privacy and ethical guidelines, engaging in regular bias
and fairness audits of their systems, and closely following
the legal regulations of the jurisdictions in which they
operate.
Despite the various challenges, computer vision remains a
vital area of AI development, revolutionizing numerous
fields from self-driving cars and automated surveillance to
medical imaging, augmented reality, and virtual reality,
making it an exciting and rapidly evolving subfield of AI.

Key Takeaways
1) Computer vision involves training computers to see and
understand visual information. This includes image
classification, object detection, and image segmentation.
 2) Challenges in computer vision include dealing with
variations and complexity in real-world scenarios, acquiring
and processing large and diverse datasets, managing
computational requirements, and addressing legal and
ethical dilemmas.
3) Organizations engaging in computer vision must invest
heavily in data, computational resources, and human
expertise—and even then, it may not go as smoothly as
planned!
 11

Privacy & Ethical

Considerations
While artificial intelligence brings innumerable benefits
and promises of a technologically advanced future, it poses
numerous challenges too. As we’ve touched upon in
previous chapters, one of the common challenges is bias
and fairness. The concept of bias refers to an unfair and
prejudicial inclination or favor towards or against a person
or group. In the context of AI, the concern is that prediction
models, particularly those involved in decision-making
processes, inadvertently replicate and amplify biases
present in the training data or the biases of their human
creators, leading to unfair outcomes.
These biases can manifest in numerous ways, from racial
and gender bias in facial recognition software to
socioeconomic bias in credit scoring models. Bias has
substantial real-world implications as it can lead to
systematic discrimination, marginalization of certain groups,
and exacerbation of societal inequalities. This poses not
only ethical and social problems but legal ones too,
including anti-discrimination laws in many jurisdictions.
One example gaining widespread attention is the
application of facial recognition technology. Studies,
including the Gender Shades project led by Joy Buolamwini
at the MIT Media Lab in the U.S., have found that some
commercial facial-analysis systems had lower accuracy
rates for darker-skinned and female individuals compared to
white males. The reason for this discrepancy lies in the
dataset used to train these AI models. If the dataset is
predominately composed of light-skinned male faces, the
resulting AI model will be better equipped to recognize and
analyze individuals from that group, consequently
demonstrating bias.
The project thispersondoesnotexist.com, which generates
a hyper-realistic portrait of a person who never existed, has
also been criticized for failing to generate people with black
skin color, which alludes to issues with the training data
used to create the model. (It’s worth noting that the project
was spun up as an online stunt to build awareness
regarding the powerful capabilities of AI rather than as a
fully polished software product.)
A different instance of AI bias involves decision-making in
the criminal justice system. An investigative report by
ProPublica revealed that software used to predict future
criminal behavior was biased against African-American
defendants. The software, designed to aid in decisions
involving bail and sentencing, was more likely to incorrectly
label black defendants as future criminals, while white
defendants were more often mislabeled as low risk.
It's important to recognize that bias in AI can manifest in
other forms as well, such as socioeconomic background,
age, gender, and disability, making it vital to address a
broader spectrum of potential biases to create more
inclusive and fair AI systems.
To ensure fairness in AI, it's crucial to take steps in the
design, development, and deployment stages. This includes
careful selection and scrutiny of training data, implementing
fairness metrics, transparency about the limitations of the
AI system, and continuous auditing of AI models for bias.
These steps, however, are not foolproof, especially as
fairness is a complex and context-dependent concept. Like
eliminating all forms of bias from schools and businesses,
it’s not always possible to remove bias completely from AI
models.

Privacy and AI
The expansion of artificial intelligence into many aspects
of our lives poses significant privacy concerns. AI models,
especially those involved in computer vision and
recommender systems, are voracious consumers of
personal data, which includes sensitive and personal
information regarding individuals.
The rise of AI has created a paradox because while these
systems rely heavily on data to function effectively and
potentially improve our lives, they also infringe on our
privacy. Privacy, in this context, refers to the right of an
individual to keep their personal information and activities
secluded from or inaccessible to others, particularly those
who are unauthorized to view them.
Consider social media platforms, which employ
sophisticated models to curate and recommend
personalized content. While this personalization enhances
the overall user experience, it is often based on the
collection and analysis of vast amounts of personal data,
including our likes, dislikes, location, political affiliations, or
even our different moods. This precise targeting ability has
profound privacy implications and raises questions about
user consent and control over the collection and use of
personal data.
Another use case raising concerns is facial recognition.
While it can enhance security and streamline verification
processes, it can also be used for continuous surveillance,
leading to potential misuse. Without proper regulations and
safeguards, it can pose a serious threat to individual
privacy.
 Likewise, AI’s predictive capabilities can be complicated
when it comes to user privacy. Machine learning models, for
example, can analyze large datasets and infer sensitive
information, even if that data was not explicitly provided.
Prediction models, for instance, have been found to
unintentionally expose a person's sexual orientation or
political affiliation based on their online activity. These
inferences can be disturbingly accurate, leading to potential
invasions of privacy that are hard to anticipate and control.
The risks of privacy breaches and unintended data
exposure became evident in 2009 when a woman in
America sued Netflix over the disclosure of her sexual
preferences in the 2007 Netflix Prize dataset. The lawsuit
was filed after academic research at the University of Texas
demonstrated privacy flaws in the design of the dataset
Netflix used for the public competition. Despite Netflix’s best
attempts to remove personal identifiers from the data,
including the names of individuals, the identities were
revealed by matching the competition’s dataset with film
ratings from the publicly available Internet Movie Database.
The researchers found that an anonymous user’s rating of
six obscure movies could be used to identify an individual
Netflix user with an 84% success rate. Moreover, accuracy
rose to 99% when the date of a movie review was known.

Legal Frameworks
In recent years, new legal frameworks have come into
effect, including the GDPR (General Data Protection
Regulation) in the European Union and the CCPA (California
Consumer Privacy Act) in the United States. Designed to
give individuals more control over their personal data, these
frameworks encompass rights such as the right to access
personal data, the right to be forgotten, and the right to
data portability. GDPR, for example, enforces added
transparency for users regarding how their data is
processed and the use of cookies on web applications, as
well as a clarified “right to be forgotten” when users no
longer wish their data to be retained (given there are no
legitimate grounds for keeping it). GDPR also requires the
encryption of users’ stored personal data and the right of
users to accept or reject the use of their personal
information for the application of online recommendations.
While such legal frameworks raise the bar for privacy and
data protection, AI poses unique challenges to these
frameworks due to its complexity and the often-opaque
nature of deep learning models. In the face of these
challenges, various practical strategies are being developed
to reconcile AI with data privacy. Techniques like differential
privacy, for example, offer a way to glean useful insights
from data while providing robust privacy guarantees.
Differential privacy adds controlled noise or randomness to
the data before performing any computations or analysis.
By introducing noise, the algorithm obscures individual
contributions and makes it difficult to determine specific
information about any individual in the dataset.
Another approach is federated learning, which makes it
possible to train models on local devices rather than using a
central server or the cloud under a centralized approach.
The latter raises privacy and security concerns as the raw
data needs to be shared with a central entity, potentially
exposing sensitive information. By using a federated
learning approach, model training is performed locally on
individual devices or servers, such as edge devices or local
servers without the need for data to leave the device and be
shared externally.
Another important consideration is the software used to
store and process data. To fulfill legal and privacy
standards, it’s imperative to evaluate different software
based on their security measures, data anonymization
features as well as data encryption tools to protect sensitive
information. Examples of data anonymization include
replacing specific values with more generalized values (e.g.,
age ranges instead of exact ages), replacing identifiers with
pseudonyms, and adding random noise to the data to make
it difficult to identify individuals while preserving statistical
integrity. Regarding data security, the software should offer
permissions based on roles to restrict data access to
authorized personnel or implement multiple forms of user
verification to safeguard access to sensitive data. Also, from
an operations perspective, data retention policies are useful
for ensuring that sensitive data is not kept longer than
necessary.
Finally, there is a need for professionals who can address
the ethical and legal implications of AI and abide by the law
and industry guidelines. Here, larger organizations should
consider the need for hiring AI ethicists and AI policy
advocates, who are responsible for conducting internal
audits, monitoring best practices, hosting internal training,
and working with the government and other companies or
industry organizations to establish industry practices.
In conclusion, the issue of privacy in the age of AI is a
complex one, requiring a multifaceted approach. It touches
not just technological solutions but also robust legal
frameworks, ethical guidelines, software controls, and
human resources. However, the overall goal should be to
create an environment in which AI technologies can evolve
without compromising the privacy rights of individuals.

Key Takeaways
1) AI systems can unintentionally replicate and amplify
biases present in their training data or the biases of their
human creators, leading to unfair outcomes.
2) Privacy is a major concern as AI relies on vast amounts
of personal data, raising questions about individual privacy.
Legal frameworks like GDPR and CCPA aim to protect
privacy rights, but AI's complexity also presents challenges
to upholding these regulations.
3) Legal frameworks, ethical guidelines, software design,
and human resources are essential for creating an
environment where AI technologies can evolve while
safeguarding privacy rights.

Thought Exercises
1) How can you better protect your data and privacy?
(i.e., using a dedicated email or user account to search and
purchase sensitive items, opting out of recommender
systems, using incognito or private browsing mode, or using
anonymous search engines to protect your privacy.)

2) What does Google already know about you? To find

out, go to google.com and sign in. Next, click the “Google
apps” icon in the top right corner. Select “Account > Data &
privacy > Personalized ads (My Ad Center)”.

Figure 6: Example of Google’s user profiling for personalized ads

 12

The Future of Work

As a transformative technology, artificial intelligence holds
the power to reshape jobs and the work environment in
profound ways. But, in a world where news and discussions
about AI bring up visions of imminent job replacement,
there are genuine concerns that AI will lead to further job
losses. The fear is heightened by the AI industry’s
preoccupation with benchmarking model performance and
efficacy with human talent. From IBM’s Deep Blue to the
Stanley autonomous car and DeepMind’s AlphaGo, AI
companies have a history of evaluating and marketing their
systems based on their ability to outperform humans in
real-world demonstrations. While this approach is highly
effective at attracting media attention, it has the dual effect
of exacerbating and reinforcing people’s fear of AI from the
viewpoint of job security.
This type of fear is not unique to AI either; from the
cotton gin (which enabled the quick separation of seeds
from cotton fibers) to the computer, anxiety has surrounded
the advent of every new transformative technology. Yet,
there is a strong consensus among economists and
technologists that AI's impact might be different, given its
potential to replace not only physical tasks but cognitive
abilities as well.
According to GPTs are GPTs: An Early Look at the Labor
Market Impact Potential of Large Language Models,
commissioned and published by OpenAI and the University
of Pennsylvania, professionals exposed to large language
models are more vulnerable to future job replacement.
Their findings revealed that approximately 80% of the U.S.
workforce could have at least 10% of their work tasks
impacted by the introduction of large language models,
while approximately 19% of workers may see at least 50%
of their tasks affected.14 This includes tax preparers,
mathematicians, writers and authors, web and digital
interface designers, court reporters, proofreaders, auditors,
accountants, journalists, and administrative assistants.
Conversely, the study predicted that tasks which rely on
critical thinking and scientific skills are less likely to be
replaced by AI. Note that physical tasks associated with
stonemasons, painters, cooks, and auto mechanics aren’t at
high risk of being replaced by large language models such
as ChatGPT but could be replaced by AI-powered robots and
further automation in the future.

The 7 Stages of Resistance

During periods of fast technological transition—as we are
experiencing now—it’s common for people to experience
fear or to sink into a state of denial, anger, and depression.
In fact, individuals and organizations often experience
something akin to the seven stages of grief in response to
drastic changes caused by new technology. Typically applied
to describe emotional responses to loss, the seven stages of
grief span shock, denial, anger, bargaining, depression,
testing, and acceptance.
First, both individuals and organizations are shocked or
stunned by the introduction of new technology. This shock
stems from uncertainty or confusion about how it works and
the potential implications. In the case of ChatGPT, the new
technology shocked people with its ability to mimic human
writing capabilities.
 Once the initial shock wears off, most people turn
skeptical and reject the notion that the new technology will
make any significant or longstanding impact. In practice,
they may avoid learning about it or dismiss it as a trend or
fad that will quickly pass.
Third, as the realization that the new technology is not
going away sets in, people may start to experience feelings
of anger or resentment. They may feel threatened by the
changes that the new technology brings and how it might
affect their role or organization.
This is followed by the bargaining stage, where individuals
and organizations try to negotiate or find a compromise.
This could involve attempting to maintain their existing
routines and operations while incorporating some aspects of
the new technology.
Subsequently, people recognize that the old ways of doing
things are changing and are no longer celebrated. This can
cause a sense of sadness, loss, and even depression. In
addition, people may feel overwhelmed by the new skills
and knowledge required to adapt.
Eventually, more individuals and organizations begin to
experiment with the new technology as part of the testing
phase. Individuals start learning new skills or finding ways
to integrate technology into their job or organization.
Finally, everyone comes to acknowledge the benefits and
accept the new technology as a part of their lives and day-
to-day work.
For organizations, understanding these seven stages is
vital for setting realistic expectations, managing internal
resistance, and navigating the changes required in order to
stay competitive. For individuals, being cognizant of the
natural reactions to new technology can spur you into action
and help you reach the acceptance stage faster than those
around you.
 It’s worth noting that not every individual or organization
will go through all seven stages or experience each stage in
the prescribed order. Early adopters, for example, are
inclined to embrace new technology overnight, while others
may linger in one stage for an extended period and never
reach the next stage. The latter can be dangerous as there
is an opportunity cost to inaction as others capitalize on the
first-mover advantage, which brings us to the missing
middle theory.

The Missing Middle

For those looking to thrive in the new era of work, it is
critical to understand that AI, in most cases, is designed to
enhance productivity rather than to outright replace you. In
line with this perspective, many of the jobs of the future are
predicted to be copilot roles, in which humans work
alongside AI to carry out a variety of professional tasks,
whether that’s linguists using ChatGPT to translate
documents or lawyers validating contracts drafted by AI
assistants.
In the healthcare sector, for example, AI can be used to
streamline administrative tasks, manage patient records,
and even assist in interpreting medical images, allowing
medical professionals to spend more quality time with
patients and make more informed decisions.
In marketing, AI tools can provide real-time analytics,
predictive models, and relevant ad copy, helping marketers
make informed decisions and increase the success of their
ad campaigns.
By leveraging AI's ability to handle complex calculations
and process vast amounts of data, humans across
industries can focus more on tasks requiring creativity,
emotional intelligence, and critical thinking—skills that are
uniquely human. This is the viewpoint championed in the
book Human + Machine: Reimagining Work in the Age of
AI. Termed the missing middle, the authors advocate that
the most effective and cost-efficient path forward is to
merge automated tools with the flexibility of human workers
to achieve optimum outcomes.
The principles of cybernetics, as espoused by Norbert
Wiener, reinforce the concept of the missing middle and
highlight the complementary nature of our respective
strengths. Derived from the Greek word kybernetes,
cybernetics means steersman or governor, which
emphasizes the notions of control, communication, and
feedback in systems. While broad in scope, at its core,
cybernetics looks at systems and processes that aim to
analyze and understand the principles of how to be
effective, adaptive, and self-governing. In the context of
artificial intelligence, cybernetics looks at how machines can
complement and mimic human-like behavior or decision-
making processes as part of a symbiotic feedback loop with
human operators, as popularized in Norbert Wiener’s
research.
Born in 1894, Wiener became interested in cybernetics
following his involvement in the development of computers
and automated anti-aircraft guns during World War II. This
included his realization that the process of tracking a
moving aircraft entailed making several observations to
anticipate its future trajectory. Amid the dynamics of war,
he observed that the reactions of the gunner must be
factored into predictions. This is due to the fact that pilots
alter course in response to enemy fire, prompting the
gunner to adjust, thereby creating a feedback loop where
the gunner's actions eventually shape their future actions.
Recognizing that feedback loops are all around us, Wiener
went on to explore the traits of this phenomenon to
optimize human-machine feedback loops under the lens of
cybernetics. Wiener posited that when harnessed correctly,
technology actually amplifies human capabilities and
productivity by extending our cognitive and physical
capacities, enabling us to accomplish tasks more efficiently
and effectively.
We see this today with AI-powered tools such as
Consensus, Elicit, and Semantic Scholar, which allow
researchers to find papers, extract key claims or
summaries, and even brainstorm their ideas by connecting
with a massive pool of academic research. These tools allow
researchers to streamline the research process and
concentrate on higher-level cognitive activities such as
critical thinking, hypothesis formulation, experimental
design, and interpretation of results, which can help drive
advancements in innovation and scientific understanding.
Although the idea of substituting human tasks with AI
tools is alluring, not all tasks should function independently
of humans. Instead, most tasks should operate in a
feedback loop with human copilots, with AI constantly
learning from human inputs and modifying their behavior
accordingly. These interconnected feedback systems
optimize the strengths of both parties, resulting in new
breakthroughs and an overall productivity boost.
Chess, in particular, provides a glimpse into effective
collaboration between humans and AI, particularly given
chess's longstanding relationship with AI technology. In
hybrid chess competitions, human players utilize their
creativity and deep understanding of the game to work in
tandem with AI chess engines to leverage their vast
analytical capabilities. The combination of human ingenuity
and AI's analytical prowess creates a synergy that
outperforms either side performing on its own. This
successful partnership has been demonstrated in a number
of tournaments, including freestyle chess competitions,
where teams composed of human players and AI engines
have consistently outperformed human-only and AI-only
teams.
 This includes the 2005 Freestyle Chess Tournament held
by the Internet Chess Club (ICC), which permitted players
to use any combination of computer chess engine,
database, and human analysis. The winners weren’t
grandmasters or supercomputers but two amateur players
working together with multiple AI chess machines,
defeating both highly ranked human players and standalone
chess engines.

Job Replacement & Business

Disruption
The speed at which AI is seeping into all aspects of
modern work—from marketing to human resources—
underscores AI literacy as a key skill set for the modern
workforce. Also, just as the Internet Age triggered an
avalanche of new job titles, so too will the AI-centric era we
are now entering.
During the formative years of the Dot-com evolution,
many traditional job opportunities—including travel agent,
journalist, courier, stockbroker, and Encyclopaedia
salesperson—contracted or were phased out entirely.
However, these losses were gradually filled with the creation
of new job roles. An explosion of highly skilled jobs ensued
in web development, search engine optimization, e-
commerce, online customer service, web design, affiliate
marketing, and eventually social media and mobile web
design. While it’s probable that a proportion of these jobs
will now be phased out by AI, new job roles will again be
created or converted to copilot roles.
Moreover, it’s possible that AI may assist in creating more
inclusive workplaces. Assistive AI technologies such as AI-
driven speech recognition software can help individuals with
mobility impairments, while AI-powered predictive text and
voice synthesis can assist those with speech difficulties.
 Likewise, AI technology will empower smaller companies
or companies-of-one to thrive in a more fragmented and
fast-paced world. The ability to generate a website, write
blog content, and create an explainer video using
generative AI tools will allow individuals to test their ideas,
gain market feedback, and iterate faster than ever before.
Conversely, labor-intensive organizations that rely on
human capital to carry out services such as ghostwriting,
translation, and online tutoring are earmarked to be the
first affected by AI technology. With AI offering a faster and
lower-cost alternative, these organizations may experience
a significant transformation in how they operate as well as a
negative impact in the short term. Exposing this reality,
stocks in the online tutoring company, Chegg, plummeted
after the CEO admitted the threat posed by AI. During the
company’s earnings call in mid-2023, their CEO Dan
Rosensweig explained that “In the first part of the year, we
saw no noticeable impact from ChatGPT on our new account
growth and we were meeting expectations on new sign-ups.
However, since March we saw a significant spike in student
interest in ChatGPT. We now believe it’s having an impact
on our new customer growth rate”.15
While it’s important to acknowledge that Chegg exceeded
profit expectations as a result of the explosive demand for
online education during the COVID lockdowns that took
place between 2020 and 2021, AI must be considered as a
competitive force in all industries, alongside traditional
business threats such as globalization and geopolitics.
In response to technological changes introduced by
ChatGPT, Chegg has announced that it's developing its own
AI, CheggMate, in partnership with OpenAI to assist
students with their homework. Textbroker, a content writing
service provider, has also added AI technology to its service
offering, and many other service platforms will be forced to
adapt or die as well.
 Human services will remain in demand but rather than
being the staple service, they will become a premium or
concierge service added on top of a fast and more cost-
effective AI-powered service. Tax, legal, and accounting
services, for example, will be serviced by AI agents trained
on relevant data, and then verified, witnessed, and
supplemented by human experts. This creates a new
business model where customers can access self-service
offerings powered by AI, such as contract drafting or
balance sheet generation, and then have the option to
communicate or receive personalized assistance from a
human assistant as part of a human service that is more
focused, personalized, and effective than many of today’s
basic service offerings.

AI Adoption in Organizations
The adoption of AI presents a range of opportunities for
organizations seeking to gain a competitive edge. However,
the road to successful AI implementation is not without
obstacles. From data quality and availability, to cultural
resistance and ethical concerns, various factors can hinder
the smooth integration of AI solutions. Implementing AI
practices entails significant upfront costs as well, including
investments in hardware, software, and hiring skilled
professionals.
In many cases, organizations may be reluctant to commit
substantial resources without a clear understanding of the
return on investment and the potential long-term benefits
AI can bring. AI processes can also disrupt traditional job
roles and workflows, creating anxiety among employees
about potential job losses or changes to their
responsibilities.
Overcoming these obstacles requires a well-thought-out
strategy, including robust data management practices,
hiring, training and upskilling employees, thorough cost-
benefit analyses, stakeholder involvement, and prioritizing
ethical considerations. For small organizations, collaborating
with experienced AI vendors and consultants can help to
facilitate successful AI integration. However, for medium
and large-sized organizations, it’s important to follow a top-
down approach that can align all the necessary resources
and implement organization-wide changes. Under this
context, appointing a dedicated Chief AI Officer can play a
pivotal role. By appointing a single Chief AI Officer instead
of relying on a committee of AI consultants, companies can
ensure streamlined implementation and accelerated testing,
leading to faster adoption of AI strategies throughout the
organization.
Reporting directly to the CEO, CTO, or a board of
directors, a Chief AI Officer works as a senior executive
responsible for identifying opportunities for AI adoption,
ensuring that AI efforts align with the company's overall
business objectives. It’s important to acknowledge that the
Chief AI Officer does not personally code algorithms or lead
the data science team’s day-to-day operations. Rather, they
are responsible for monitoring market changes, determining
the right strategic direction, and identifying opportunities
for the company to utilize AI technologies effectively in
marketing, research and development, and other parts of
the organization. By staying ahead of the curve, they guide
the organization to ride the AI wave instead of being
overwhelmed by it.
In addition, the Chief AI Officer is responsible for
engaging with various teams within the organization in
order to gather feedback and determine optimal areas for
AI application. In the short term, this may involve
conducting A/B or split testing to compare human and AI-
powered resources, evaluating different AI software, and
using testing and data to find quick wins and gain internal
support for new practices. By presenting quantifiable
metrics and data that highlight the improvements achieved
through AI practices, it becomes easier to demonstrate the
value that AI brings to the organization.
The Hustle Daily Newsletter, for example, ran an A/B test
using custom AI-generated images and human-edited stock
photography.16 The AI-generated version was spun up using
Midjourney and designed based on the goal of aligning the
image with the ad copy and the value proposition they
wanted to communicate. This resulted in a 3x increase in ad
performance (measured on the cost to acquire clicks)
compared to the classic stock photography approach. On
top of this, the AI-generated ads were faster to produce.

Figure 7: The Hustle Daily ad image created using Midjourney

Hosting a hackathon, spanning an afternoon or a couple

of days, is another option to catalyze momentum and
explore AI-first processes. Hosting a hackathon not only
serves as a litmus test for trialing AI's applicability and
effectiveness within the organization but also helps to
generate interest and motivation across teams. A hackathon
can revolve around a specific theme or problem, where
teams collaborate to plan and develop solutions, ultimately
presenting their prototypes to key decision-makers or team
leaders at the end of the event.
 While it’s valuable to run internal experiments and
consider the benefits that AI can bring, it’s important to
acknowledge that not all teams, departments, and
organizations will need to adopt AI technology. The search
engine company DuckDuckGo, for example, has resisted
industry trends such as personalized search results,
personalized ad targeting, and user data tracking to provide
a privacy-centric alternative for users concerned about data
privacy and tracking. This includes offering Duck Player
mode, which provides users with a clean viewing experience
stripped of personalized ads and prevents users’ viewing
activity from influencing their YouTube recommendations.

Figure 8: Duck Player mode offers a clean viewing experience without

personalized ads

Similarly, Basecamp has chosen not to adopt certain AI-

driven features commonly found in other project
management tools, instead prioritizing a more human-
centric approach to collaboration in its task management
software. In a different vein, Luddite Games focuses on
developing games that rely on human creativity and
decision-making under the game’s narrative of destroying
AI and saving human jobs.
This type of counterapproach strategy allows each of
these companies to differentiate themselves in the market
and cater to an audience that values privacy or human
involvement and expertise. While it may not be the fastest
or most effective path to obtain media interest and investor
funding in the early stages of an AI hype cycle, there will be
plenty of opportunities for other service providers to target
a niche audience by openly resisting AI technology and
embracing human expertise. The key is to openly resist new
AI technology and integrate this stance into the company’s
marketing strategy and value proposition rather than being
a side-effect of inaction and business model complacency.

Digital Craft & Offline Events

Despite the immense efficiency gains that AI offers,
consumers will continue to respect the craft and attention to
detail that humans bring to their work. Whether it's creating
a piece of artwork or crafting handmade products, we
disproportionally value things created by ourselves or other
humans. In psychology, this phenomenon is called the
“labor illusion”, which captures how people perceive a
product or service as more valuable or of higher quality
when they believe that a considerable amount of human
effort and labor was invested in its creation. We see this
with birthday and thank you cards. Despite the shareability
and convenience of e-cards, most people prefer the
thoughtfulness and effort that come with sending a physical
card and writing a handwritten message. Similarly, if you
browse the comments section on YouTube, you will see
users consistently praise video creators who take the time
to produce subtitles for their videos, add timestamps, or
who spend inordinate amounts of time and effort capturing
the perfect drone shot. Ultimately, the human touch carries
a sense of quality, uniqueness, and care that resonates with
people. In addition, there is an inherent appreciation for the
time, skill, and dedication that goes into creating something
manually—even if it is done digitally.
In the future, the ubiquity of AI-generated content may
lead to a future where creating your own podcast or video
content without AI assistance is valued as a form of digital
craft. Likewise, companies that offer human customer
service support might stand out in a world where AI avatars
are trained on millions of data points and customer support
logs.
Finally, the surge in AI-generated content may lead to
renewed interest in offline events such as meetups, political
debates, and music concerts, as a direct response to the
growing distrust of online content. In an age of deep fakes
and content manipulation, seeking face-to-face interactions
may be the only way to discern authenticity and verify what
is genuinely real in the world. Likewise, dating might veer
towards a return to more in-person encounters or an
embrace of Japanese-style singles’ parties known as
goukon17, countering the influence of AI-enhanced photos
and AI-generated conversations in the online dating space.
Alternatively, some individuals may find the efficiency of
their AI dating agent communicating with a suitor’s AI
agent and scheduling a coffee date more appealing!

Key Takeaways
1) The fear of job displacement isn’t new and while AI is
expected to create new roles and demand new skills, we
could see a drastic shakeup to knowledge work. This
includes humans working together with AI agents as
copilots to enhance productivity and efficiency.
2) Resistance is common in the face of technological
innovation, with individuals and organizations often going
through a series of emotional responses akin to the seven
stages of grief.
 3) AI technology has the potential to empower smaller
companies and individuals to test new ideas at a lower cost.
4) AI-powered service providers can offer faster and low-
cost services, challenging traditional business models.
5) Appointing a Chief AI Officer can help organizations
navigate AI strategies, identify opportunities, and align AI
efforts with overall business objectives.
6) The proliferation and ubiquity of AI technology could
pave the way for a new form of digital trade or commerce
carried out by humans, without AI involvement.

Thought Exercises
1) What are your strengths and weaknesses? How can AI
be used to overcome your weaknesses and spend more
time on your strengths?

2) How can AI companies move beyond crude

comparisons with humans to measure and market their
products? (i.e., new jobs created by AI, lives saved by an AI
product).

3) In which industries or scenarios will using AI be a

potential disadvantage for the brand, sport, organization, or
other entity?
 Further Resources
The field of artificial intelligence is vast and multifaceted,
encompassing various domains covered in this book. As a
result, acquiring knowledge in AI requires deep-diving into
an array of different subjects, each one contributing to a
more comprehensive understanding. This final chapter
provides a general guide to learning resources and an
overview of potential career paths.

Formal Education
The most obvious starting point for anyone interested in
AI is formal education. Universities around the world offer
courses dedicated to AI and related fields, including data
science, machine learning, and computer science. Highly
respected Master’s programs in machine learning are
offered at Stanford, Berkeley, Carnegie Mellon, Columbia,
University of Washington, and MIT. Other reputable
institutes include Edinburgh, Duke, Michigan, University of
Pennsylvania, Toronto, UCSD, Brown, UCL, Georgia Tech,
Cambridge, Oxford, and Cornell. Carnegie Mellon University,
renowned for its Robotics Institute, also provides excellent
degree courses in AI robotics. Note too that many colleges
offer online degrees and there are various other degree
options offered around the world.
After completing a Master’s degree in machine
learning/artificial intelligence/data science, there is then the
option of completing a PhD. This is an ideal route for those
wishing to delve deeper into AI topics. PhDs in many
countries are supported by government or university
funding. There is, though, a sizeable opportunity cost of
completing a four-year PhD on a basic salary/stipend over
working in industry on a full-time salary. In places like the
United States, you can expect to receive USD $20,000-
45,000 a year to complete a PhD course, compared to
earning an annual salary of USD $80,000-120,000 working
for a private company.
However, Google, Facebook, and Microsoft have been
known for raiding the ranks of academia to recruit talent in
the space of artificial intelligence and to secure a pipeline of
young talent. The prevailing logic is that hiring professors
unlocks a new network of talent and makes it easier to hire
former PhD students. The other upside of completing a PhD
is that you have greater control over the scope of your work
and research, and the academic path overall may be a more
attractive proposition for some people.

Non-Degree Options
Beyond traditional education pathways, there is a plethora
of non-degree options delivered online. Online learning
platforms have democratized access to knowledge, offering
a range of courses catering to various specializations. This
includes platforms like Coursera and edX which partner with
top universities and industry leaders to provide high-quality,
accessible content. For example, Coursera's AI For Everyone
course, developed by Andrew Ng, a co-founder of Coursera
and an adjunct professor at Stanford University, offers a
non-technical introduction to AI.
As my preferred non-degree option, I recommend
DataCamp as an excellent resource for learning about AI,
machine learning, data science, and related fields including
how to code. The platform's approach is hands-on, focusing
on active learning and practical application, which is quite
different from traditional lecture-based courses. Each skill-
based course is broken down into bite-sized lessons
combining short videos with immediate hands-on exercises.

Books
 There is a wealth of books that delve into AI from
different perspectives. For those interested in the
philosophical and societal impacts of AI, books such as Nick
Bostrom's Superintelligence: Paths, Dangers, Strategies and
Max Tegmark's Life 3.0: Being Human in the Age of Artificial
Intelligence are highly recommended. If you're interested in
the technical aspects, O’Reilly Media textbooks such as
Hands-On Machine Learning with Scikit-Learn, Keras, and
TensorFlow authored by Aurélien Géron provide practical
insights for developers into AI, deep learning, and machine
learning.

Real-World Projects
For hands-on learning, nothing beats working on real-
world projects. As a platform for data science competitions
and public datasets, Kaggle offers a plethora of datasets
and challenges that can be used to hone your skills in
machine learning and AI.
Contributing to open-source AI projects is another way to
gain practical experience. Projects such as OpenAI's GPT
models, TensorFlow, and PyTorch all welcome community
contributions.
In regard to practical training, LeWagon offers data
science boot camps around the world, which train students
with the practical skills needed to become a data scientist
and work in a data team. LeWagon offers a lot of free online
and offline events, which you can sign up for on their
website (www.lewagon.com).

Blogs & Podcasts

Podcasts and blogs are valuable resources for learning
about AI. Podcasts such as The AI Alignment Podcast by the
Future of Life Institute, Data Skeptic, and AI in Business all
provide insightful discussions on various AI topics. Blogs
such as Towards Data Science and the Google AI Blog also
offer deep dives into AI trends and research. My favorite
podcast is the OC Devel Machine Learning Guide podcast.

Potential Career Paths in AI

Whether you're a recent graduate or a seasoned
professional considering a career switch, there is a place
and role for you in the ever-broadening world of AI. As
Stuart Russell and Peter Norvig point out in Artificial
Intelligence: A Modern Approach, AI is a broad and rapidly
evolving field that offers a plethora of opportunities for
those willing to invest the time and effort required to
master it.
As the influence and impact of AI grows, the demand for
professionals in this field will also continue to rise.
Regardless of industry or geography, organizations across
the globe are recognizing the power of AI to drive growth,
innovation, and efficiency. This section focuses on the vast
array of career paths in AI and provides a broad view of the
roles, skills, and opportunities available for those interested
in this field.

AI Research Scientist
The first role we’ll examine is the AI research scientist.
This position is typically based in academia or research
institutions but some private sector companies employ
research scientists too. These professionals are the thought
leaders of AI, pushing the boundaries of what’s possible. In
addition to publishing papers and attending conferences,
they are often found working on abstract and complex
problems in machine learning, deep learning, and other AI-
related fields. They possess a deep understanding of
algorithms, calculus, linear algebra, and statistics. While a
doctoral degree in a related field is often required, some
organizations hire candidates with a Masters degree and
substantial relevant experience.

AI Engineers
AI engineers apply the theories and concepts developed
by AI research scientists in real-world scenarios. They build,
test, and deploy AI models and manage AI systems and
infrastructure. They must be proficient in various
programming languages like Python, Java, or C++, and be
familiar with AI libraries such as TensorFlow or PyTorch. A
Bachelors or Masters degree in computer science, data
science, or a related field is usually required, along with a
strong practical understanding of machine learning.

Data Scientists
Data scientists work as the detectives of the AI world.
They sift through large volumes of data, using machine
learning and statistical models to extract meaningful
insights that can be used to solve business problems and
make strategic decisions. They require a firm grasp of
statistics, as well as experience with programming
languages and data management tools.

Machine Learning Engineers

Straddling the roles of AI engineer and data scientist,
machine learning engineers are typically tasked with
creating data models that are then implemented by AI
applications to learn and improve. To work as a machine
learning engineer, you will need to have a deep
understanding of multiple programming languages, machine
learning algorithms, and data modeling.

AI Ethicists
AI ethicists represent a newer but increasingly important
role in the field of AI. As AI applications proliferate,
questions about their ethical implications and the
governance required to ensure their fair use have expanded
too. AI ethicists analyze and address these issues. Their
backgrounds can vary widely, from philosophy to law to
data science, with the common denominator being a dual
understanding of AI technologies and a strong ethical
framework.

AI Policy Advocates
A lesser-known but growing role is the AI policy advocate.
These professionals work with governments, non-profit
organizations, and companies to influence policy and
legislation related to AI. This role requires a strong
understanding of AI, public policy, and legal frameworks.

The roles mentioned represent just a snapshot of the

opportunities available in the field of AI. Each of these roles
offers a unique perspective on AI and serves a critical
function in its ongoing development. However, all these
roles share a common need for continuous learning and
adaptation. Given the rapid pace of advancements in AI,
professionals in this field must be committed to staying up-
to-date with the latest developments.
Before deciding on a career path in AI, it's important to
consider your interests, strengths, and long-term career
goals too. For example, if you're passionate about
researching new theories and technologies, a career as an
AI research scientist might be a good fit for you. If you are
more interested in applying AI technologies to solving real-
world problems, roles such as AI engineer or data scientist
might be a better match.
Moreover, each of these roles offers different levels of
interaction with others. Some roles, such as AI engineer or
machine learning engineer, may work as part of a larger
team and interact closely with other departments. Other
roles such as AI research scientist may have a more solitary
focus and work primarily on individual projects.
Finally, it's crucial to consider the societal implications of
AI in choosing a potential career path. As AI becomes more
integrated into our daily lives, there is a growing need for
professionals who can address the ethical, legal, and
societal implications of AI. Roles such as AI ethicist or AI
policy advocate are expected to become increasingly
important as the field continues to evolve.
Other Books By The Author

Generative AI Art for Beginners

Master the use of text prompts to generate stunning AI art
in seconds.

ChatGPT Prompts Book

Maximize your results with ChatGPT using a series of proven
text prompt strategies.

Machine Learning for Absolute Beginners

Learn the fundamentals of machine learning, explained in
plain English.

Machine Learning with Python for Beginners

Progress your career in machine learning by learning how to
code in Python and build your own prediction models to
solve real-life problems.

Machine Learning: Make Your Own Recommender

System
Learn how to make your own machine learning
recommender system in an afternoon using Python.

Data Analytics for Absolute Beginners

Make better decisions using every variable with this
deconstructed introduction to data analytics.

Statistics for Absolute Beginners

Master the fundamentals of inferential and descriptive
statistics with a mix of practical demonstrations, visual
examples, historical origins, and plain English explanations.
Python for Absolute Beginners
Master the essentials of Python from scratch with beginner-
friendly guidance.
 Notes
[←1]
It’s important to note that Andrew Ng's contributions to deep learning
stretch beyond his work with GPUs. His popular Coursera course on
machine learning has introduced millions of students worldwide to AI. His
tenure at Google Brain, meanwhile, led to the development of large-scale
deep neural networks, pushing the boundaries of what was thought
possible.
[←2]
A semiconductor is a material that has electrical conductivity between that
of a conductor and an insulator. GPUs, meanwhile, are made of
semiconductor materials, such as silicon, and use semiconductor devices,
such as transistors, to perform their calculations.
[←3]
Rory Cellan-Jones, “Stephen Hawking warns artificial intelligence could end
mankind”, The BBC, December 2, 2014.
[←4]
Stephen Hawking, “LCFI Launch: Stephen Hawking”, Leverhulme Centre
for the Future of Intelligence, Accessed July 10, 2023,
http://lcfi.ac.uk/resources/cfi-launch-stephen-hawking.
[←5]
Daniel Colson, “One think tank vs. ‘god-like’ AI”, Politico, August 15, 2023.
[←6]
Ryan Heath, “Exclusive poll: Americans distrust AI giants”, Axios.com,
August 9, 2023.
[←7]
N. Gillespie, S. Lockey, C. Curtis, J. Pool, J & A. Akbari, “Trust in Artificial
Intelligence: A Global Study”, The University of Queensland and KPMG
Australia, 2023.
[←8]
“How do people feel about AI? A nationally representative survey of public
attitudes to artificial intelligence in Britain”, Ada Lovelace Institute and The
Alan Turing Institute, 2023.
[←9]
As an explanatory technique, there are no fixed rules for determining the
number of clusters to analyze. Note that if you set k to the same number
of data points contained in the dataset, each data point automatically
becomes a standalone cluster. Conversely, if you set k to 1, then all data
points will be deemed as homogenous and fall inside one large cluster.
You, therefore, want to avoid using a k value close to 1 or close to the
maximum number of data points.
[←10]
A one-dimensional array, also known as a vector, is a data structure to
store elements in a linear sequence with the elements arranged in a single
row or column.
[←11]
Simone Baldoni, Nalini Chintalapudi, Getu
Gamo Sagaro, Graziano Pallotta, Giulio Nittari, and Francesco Amenta,
“Factors affecting the quality and reliability of online health information”,
Sage Journals, August 2020.
[←12]
Gregory Robinson, “AI-generated Family Guy is so offensive it's already
banned on Twitch”, UNILAD, June 16, 2023.
[←13]
While OpenAI previously collected its data from the web without explicit
permission from site owners, the company’s new GPTBot, a web crawler
that automatically scrapes data from the Internet, allows site owners to
restrict access by modifying their robots.txt file.
[←14]
Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, “GPTs are
GPTs: An Early Look at the Labor Market Impact Potential of Large
Language Models”, OpenAI, OpenResearch, University of Pennsylvania,
March 2023.
[←15]
Prarthana Prakash, “Chegg’s shares tumbled nearly 50% after the edtech
company said its customers are using chatgpt instead of paying for its
study tools”, Fortune, May 3, 2023.
[←16]
Marketing Against The Grain, “How A $25B Company Uses A.I. To 300x
Their Marketing Results,” HubSpot Podcast Network, June 13, 2023.
[←17]
Goukon parties are popular in Japan as a way for adults to socialize and
expand their social circles while potentially finding romantic interests.
These parties can take various forms, such as dinners, barbecues, or
outings to various entertainment venues. The atmosphere is typically more
relaxed and informal compared to traditional matchmaking events.