aises_2_1

<!-- AI Background -->

<h1 id="introduction">2.1 Introduction</h1>
<p><strong>To reduce risks from AI systems, we need to understand their
technical foundations.</strong> Like many other technologies, AI
presents benefits and dangers on both individual and societal scales. In
addition, AI poses unique risks, as it involves the creation of
autonomous systems that can intelligently pursue objectives without
human assistance. This represents a significant departure from existing
technologies, and we have yet to understand its full implications,
especially since the internal workings of AI systems are often opaque
and difficult to observe or interpret. Nevertheless, the field is
progressing at a remarkable speed, and AI technologies are being
increasingly integrated into everyday life. Understanding the technical
underpinnings of AI can inform our understanding of what risks it poses,
how they may arise, and how they can be prevented or controlled.</p>
<p><strong>Overview.</strong> This chapter mostly focuses on machine
learning (ML), the approach that powers most modern AI systems. We
provide an overview of the essential elements of ML and discuss some
specific techniques. While the term “AI” is most commonly used to refer
to these technologies and will be the default in most of this book, in
this chapter we distinguish between AI, ML, and their subfields.</p>
<p><strong>Artificial Intelligence.</strong> We will begin our
exploration by discussing AI: the overarching concept of creating
machines that perform tasks typically associated with human
intelligence. We will introduce its history, scope, and how it permeates
our daily lives, as well as its practical and conceptual origins and how
it has developed over time. Then, we will survey different “types” or
“levels” commonly used to describe AI systems, including narrow AI,
artificial general intelligence (AGI), human-level AI (HLAI),
transformative AI (TAI), and artificial superintelligence (ASI) <span
class="citation"
data-cites="bostrom2014superintelligence">[1]</span>.</p>
<p><strong>Machine Learning.</strong> Next, we will narrow our
discussion to machine learning (ML), the subfield of AI focused on
creating systems that learn from data, making predictions or decisions
without being explicitly programmed. We will present fundamental
vocabulary and concepts related to ML systems: what they are composed
of, how they are developed, and common tasks they are used to achieve.
We will survey various types of machine learning, including supervised,
unsupervised, reinforcement, and deep learning, discussing their
applications, nuances, and interrelations.</p>
<p><strong>Deep Learning.</strong> Then, we will delve into deep
learning (DL), a further subset of ML that uses neural networks with
many layers to model and understand complex patterns in datasets. We
will discuss the structure and function of deep learning models,
exploring key building blocks and principles of how they learn. We will
present a timeline of influential deep learning architectures and
highlight a few of the countless applications of these models.</p>
<p><strong>Scaling Laws.</strong> Having established a basic
understanding of AI, ML, and DL, we will then explore scaling laws.
These are equations that model the improvements in performance of DL
models when increasing their parameter count and dataset size. We will
examine how these are often power laws—equations in which one variable
increases in proportion to a power of another, such as the area of a
square—and examine a few empirically determined scaling laws in recent
AI systems.</p>
<p>
Throughout the chapter, we focus on building intuition, breaking down
technical terms and complex ideas to provide straightforward
explanations of their core principles. Each section presents fundamental
principles, shows prominent algorithms and techniques, and provides
examples of real-world applications. We aim to demystify these fields,
empowering us to grasp the concepts that underpin AI systems. By the end
of this chapter, we should have a well-rounded understanding of machine
learning, ready to delve deeper into the complexities and challenges of
AI systems, the risks they pose, and how they interact with our society.
This will provide the technical foundation we need for the following
chapters, which will explore the risks and ethical considerations that
these technologies present from a wide array of perspectives.</p>


<br>
<br>
<h3>References</h3>
<div id="refs" class="references csl-bib-body" data-entry-spacing="0"
role="list">
<div id="ref-bostrom2014superintelligence" class="csl-entry"
role="listitem">
<div class="csl-left-margin">[1] N.
Bostrom, <em>Superintelligence: Paths, dangers, strategies</em>, First
edition. <span>Oxford</span>: <span>Oxford University Press</span>,
2014.</div>
</div>
</div>