It's probably no surprise that "artificial intelligence" (AI) could take many different meanings. The marketing hype around "AI" is quite palpable, and it's important to know what "AI" might mean depending on the context.
Oxford English Dictionary
Well, let's first just turn straight to the dictionary. The OED contains the following entry for "artificial intelligence":
“The capacity of computers or other machines to exhibit or simulate intelligent behaviour; the field of study concerned with this. In later use also: software used to perform tasks or produce output previously thought to require human intelligence, esp. by using machine learning to extrapolate from large collections of data.”
Referenced on 2024 Jan 06.
Artificial vs Natural
A distinction often made with artificial intelligence is that from natural intelligence – that is, intelligence exhibited by animals and other earthly creatures.
Defining "Artificial" and "Intelligence"
What if we were to break down "artificial" and "intelligence" independently?
- artifice → "clever or cunning devices or expedients, especially as used to trick or deceive others" (New Oxford American Dictionary 2023)
- intellect → "the faculty of reasoning and understanding objectively, especially with regard to abstract or academic matters" (New Oxford American Dictionary 2023)
These seem to imply that artificial intelligence is a kind of deception. This hints at notions like "The Imitation Game" (aka "The Turing Test," discussed later), often cited in popular culture as a test for AI. But is just being fooled by something enough to warrant the name?
As Defined in AI: A Modern Approach.
Stuart J. Russell, Peter Norvig, et al., have arguably written the book on artificial intelligence: Artificial Intelligence: A Modern Approach.
The first chapter provides a framework to unpack this question of what AI is in which the authors highlight two popular dimensions of thought. One covers a sense of intelligence along a spectrum from mimicking human performance to that of a more objective sense of rationality. The other dimension considers the internal thought processes and reasoning involved versus observed external behavior. By combining these four combinations, they present the following approaches:
- Acting humanly: The Turing test approach
- Thinking humanly: The cognitive modeling approach
- Thinking rationally: The "laws of thought" approach
- Acting rationally: The rational agent approach
I highly recommend the first two chapters. The website is here. The book's code is available here on GitHub. The detailed section outline of the 4th edition is here, and the first two chapters and sections cover the following:
- Chapter 1 Introduction ... 1
- 1.1 What Is AI? ... 1
- 1.2 The Foundations of Artificial Intelligence ... 5
- 1.3 The History of Artificial Intelligence ... 17
- 1.4 The State of the Art ... 27
- 1.5 Risks and Benefits of AI ... 31
- Chapter 2 Intelligent Agents ... 36
- 2.1 Agents and Environments ... 36
- 2.2 Good Behavior: The Concept of Rationality ... 39
- 2.3 The Nature of Environments ... 42
- 2.4 The Structure of Agents ... 47
Particularly interesting for we spatial designers are the notions of "environment" and "agent." We'll talk more later about the multiple uses of the word "agent," (which is typically used as an abstraction for things like robots), but as spatial designers, we're more interested in the design of environments, which AI tends to treat more as something to assess and respond to, rather than craft and control. This points to an interesting opportunity for computational design.
Council on Artificial Intelligence, EU AI Act
As noted later in the Ethics and Governance Module, a number of legislative acts are in the works to establish governance standards around AI and thus protections for individuals and companies. Most notable is the European Union's Artificial Intelligence Act, which has undergone a 5-year process of development and ratification in the EU.
So the section may seem bureaucratic for a creative and critical program, but it's a place at which a number of definitions have been compiled, definitions that affected the direction of the world's thinking around constraining AI.
So as part of the process, numerous interests groups were asked to contribute their points of view. One such contribution is a document entitled AI Watch: Defining Artificial Intelligence. Towards an operational definition and taxonomy of artificial intelligence., published by the Joint Research Centre (JRC), the European Commission’s science and knowledge service. This document compiled definitions of and terminology related to AI. You can download it here. (The link is towards the bottom of the page.)
Let's look at a few excerpts from this publication:
AI Domains as summarized by AI Watch
It's this list of "core" domains that's interesting at this point, further detailed in page 16:
- reasoning
- knowledge representation
- automated reasoning
- common sense reasoning
- planning
- planning and scheduling
- searching
- optimisation
- learning
- machine learning
- communication
- natural language processing
- perception
- computer vision
- audio processing
Table 3, starting on page 18, contains the list of definitions, like this.
I recommend perusing these definitions to get a sense of both the ambiguity and consistency these definitions carry.
"AI System" Definition from OECD
One definition of an "AI System" deserves a bit of attention. It's found on page 7 of the referenced document, OECD, Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449 by the Organization for Economic Co-operation and Development.
"AI System: An AI system is a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. AI systems are designed to operate with varying levels of autonomy."
This highlights a number of key concepts to consider when working with AI and evaluating the claims of various companies:
- human-defined objectives – It's interesting that objectives are human-defined, but does this mean that AI waits for our commands or tries to behave like a human?
- predictions – Thus are weather apps also AI, in a sense? Perhaps, but the general idea of predicting the future, what might happen, described statistically, is in the purview of AI. High frequency trading algorithms deployed by hedge funds and financial companies could be a form of AI. Computer vision algorithms, like facial recognition, are attempting to predict who someone might be.
- recommendations – Are product recommendations given by Netflix and Amazon the result of AI?
- decisions – Is this inherent to AI or is this more a matter of how we deploy an AI system? Does a "decision" matter if we don't allow the AI to act upon that decision itself?
- real or virtual environments - This puts AI in the realm of spatial design and virtual reality, as well as reality.
- autonomy – How much can (and should) these systems operate on their own? AI is often considered to be a technology designed to act independently of direct human control.
These are here just to get you thinking about what "artificial intelligence" might mean. Is there a definition that resonates with you?
AI Tasks and Models
More practically, let's take a preview at one of the platforms we'll be discussing in the next module. Hugging Face is an AI company based out of Brooklyn, NYC, that hosts a platform that's essentially GitHub for AI. Anyone can register and host models, use their infrastructure and SDKs, etc. More on this later, but for now, their "Tasks" page illustrates the kinds of functionality modern AI is capable of.
Note how a number of the tasks above are about translating one medium to another, like text-to-image (for generative imagery like that performed by Stable Diffusion and DALL•E) or image-to-text (or captioning performed by vision models).
Note that "tasks" are distinct from but related to "models." A task is a higher level description of the capability an AI system may provide, like detecting objects in raster images. However, a "model" is a specific implementation of a task, given a specific architecture (like a neural network or support vector machine). You may have heard of the AI model "GPT" of the famed ChatGPT, but there are many versions of GPT available. One primary task that the product ChatGPT offers is text generation (aka text completion), and the model underlying model may be GPT-4o, GPT-3.5, GPT-4-Turbo, etc. More on AI and machine learning models later.
Generative vs Non-Generative AI
Much of the hype today is due to a category of AI tasks grouped under the heading "generative artificial intelligence" or "generative AI." The name hints at the purpose, to "generate" content, particularly as a human might, like text to serve as essays or blog posts, or images given short text descriptions called "prompts."
In contrast to generative AI methods, there are non-generative methods as well, or what we might call analytical methods. This kind of AI typically is used for perceptual purposes, like identifying people from photographs of their faces, classifying photographs of real-world objects, or identifying objects and people in realtime videos. Other textual analysis include techniques like "sentiment analysis," which attempts to predict the tone and emotion in a given text, useful for companies wanting to understand how well their brand is perceived in public online forums.
This course reviews tasks and methods in both categories.
Popular Definitions
A bit of opinion… the popular sense of artificial intelligence often comes from science fiction stories. Images of autonomous robots armed with weaponry or acting as humans, but stronger and perhaps less empathetic. One could even say that we invoke "AI" as a term because we want to invoke associations with science fiction lore, to claim a machine is intelligent even though its capabilities may be more akin to a complex artifice, like the automata of the 18th Century.
In particular, the movie "Her" is at the forefront of popular conversation, particularly since OpenAI's chatbots are approaching this kind of human-like interaction. The film is ironic, however, in that it ultimately shows how human satisfaction and fulfillment in life and relationships cannot be fulfilled by the best technology. Time, as they say, will tell.
This falls under a concept of "artificial general intelligence" (AGI), also running rampant, which refers to an AI-powered system that truly "thinks" on its own or behaves like a kind of superpower. It's popular for futurists or pop-science figures to speculate on this kind of technology, with or without understanding how AI and machine learning actually work. Maybe we will reach this, but why would we want to?
Conversations of AI often focus on what could be done, rather than what should be done. Hopefully this course will dispel some of the hype and reinforce how AI as a medium should be beneficial for designing better experiences and environments for humans, if used in responsible ways.
