.webp)
In this first module, we will cover AI’s landscape at a high level to provide you with enough contextual information to think critically about the field as well as design and develop a prototype on Wonda’s platform in the following modules.
While we will cover much ground and go deep on a few key ideas, keep in mind that the goal here is to develop an intuition on how these emerging technologies work, not to deeply understand what is going on under the hood. These technologies are changing monthly (even daily!), and having a good, robust mental model will help you think critically and strategically about new developments.
The first parts of this module are dedicated to give you an overview on the history and general principles of Generative AI and LLM.
If you want to start with the interactive part directly jump to “Hands On with AI Agent Prompting” by using the left menu. Enjoy!
In this first part, we will define our terms and get to the heart of what we mean when using the term “artificial intelligence” by viewing its history from the 1950s onwards. By the end of this module, we will be better able to answer two key questions:
Take a moment to reflect on the collage below and how many of the images relate to your everyday life:
While reflecting on the collage above, perhaps some light bulbs 💡 lit up in your mind’s eye and you made some connections to your daily routine.
The lightbulb, which tends to represent “innovation” or even the idea of an “idea” itself, is a perfect introduction to understanding the AI landscape at a foundational level.
Believe it or not, when electricity was first introduced to the public there was apprehension, confusion, and even outright fear! In fact, some people had to be convinced of the benefits of electricity.
An advertisement titled “Electricity in Industry” from the 1920s remarked:
Never before have the questions of economy and efficiency in production been of such importance as now in the industrial life of the country. This is true in the large plant as well as in the small shop. Electricity is proving the most effective agency in solving these various problems as they arise.
Sounds commonsensical to us today, but this was not always the case.
The perimeter of the advertisement is even filled with several different use-cases of how one might accomplish everyday tasks with the help of electricity, such as “Electric Dough Mixer, Eclectic Furnace, and Electric Refrigeration”:
Although it’s hard to imagine life without electricity today, its acceptance was a gradual process and people needed to be educated on not only its strengths and weaknesses, but also on all the new ways it could be used.
Andrew Ng, recently named one of Time's most 100 Influential People in the world, remarked that “AI is the new electricity”:
Just as electricity transformed almost everything 100 years ago, today I actually have a hard time thinking of an industry that I don’t think AI will transform in the next several years.
In the same way that the public took some time to fully understand electricity’s nature, potential, and possibilities in the early 20th century, AI is facing those very same challenges today.
Another useful term that Andrew and others use to describe AI is as a “general purpose technology”, much like how the steam engine, the railroad, the automobile and other information technologies affected the entire economy.
💡 TIP: As you work through the rest of this course, think about the way that AI as a “general purpose technology” can be used to “electrify” and improve a wide variety of educational outcomes.
As is probably obvious by now, the collage at the start of this Module demonstrates that AI is already a general purpose technology powering much of our 21st century lives.
But interestingly, “artificial intelligence” is not a 21st century invention! In fact, we can trace its roots as a discipline of its own all the way back to 1956 at the Dartmouth Summer Research Project on Artificial Intelligence.
Take a moment to close your eyes and clear your mind of every context in which you may have heard the term “artificial intelligence”.
Although AI is already powering much of your everyday life, the term has recently been thrown around so much and misapplied in so many ways that it's a surprise different stakeholders can have a meaningful conversation about AI at all!
In this section, we will take a quick look at the context in which the term emerged, as it will be critical in helping you develop a good intuition of how this technology works.
Take a close look at the image below of a plaque that was put up at Dartmouth University in 2006 to commemorate the 50th anniversary of a project. In particular, take note of the fact that it highlights the first time the term “artificial intelligence” was used in 1956:
You might be surprised to find that this term first originated in the 1950s – long before our god-like 21st century technologies! By today’s standards, we might even consider this time period “technologically backwards”.
Yet, in “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence”, mathematician John McCarthy organized a “2 month, 10 man study” whose goal was:
To proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be described that a machine can be made to simulate it.
Such read the first sentences of a proposal whose goals were aspirational and represented the enthusiasm of what people thought would be possible with this burgeoning new field. This initial optimism of the “2 month, 10 man study” of artificial intelligence was, however, tempered by the practical challenges and limitations of the hardware available.
For example, the most advanced computers in 1956 were probably millions of times slower than today’s smartphones.
Amazingly, however, the proposal went on to list many topics that continue to define – and challenge – the field today: natural-language processing, neural networks, machine learning, abstract concepts and reasoning, and even creativity!
While these various topics display an incredible amount of foresight on behalf of Dartmouth group, they also led to a host of conceptual divides on how to best approach solving these various topics.
The main takeaway from the workshop was that “artificial intelligence” emerged as a field of both research and practice, one that we can define formally as follows:
Artificial Intelligence: a field that includes a broad set of approaches with the goal of creating machines with intelligence, many of which evolved with and against new technologies throughout history
Instead of the simplistic way in which the term is thrown around to apply to anything and everything:
Let’s instead look at the AI field more broadly and multidimensionally:
In this latter image, we see that there are a number of different terms and relationships that holistically define the artificial intelligence field. As we continue to work through the remainder of this module, we will explore this multidimensional view of AI.
In this second part, we will explore some key figures and key domains (e.g., cognitive science and psychology) that influenced AI’s development.
Since the 1956 workshop, many researchers and practitioners quickly realized that creating intelligence was much, much harder than anticipated and it spurred on different approaches that came to define the general contours of the AI landscape.
There are many ways that different commentators tend to look at the AI landscape. Our goal here is not to cover things in technical detail, but to give you a robust and nuanced enough conceptual map of major developments.
With that in mind, we will use two lenses to view the AI landscape:
👁 Lens 1: AI Springs and AI Winters
👁 Lens 2: The Conceptual Divide between Symbolic and Subsymbolic AI
The history of artificial intelligence is often described in terms of "AI Springs" and "AI Winters”:
🌱AI Springs: periods of high interest driven by the promise of this technology
❄️AI Winters: periods of low interest when the technology has not met expectations
By following a simplified history of these seasons of hype and fall, boom and bust, we can better contextualize just how amazing these recent developments in generative AI (and beyond) are before you start to build something on your own.
AI Springs 🌱are characterized by periods of enthusiasm and progress in AI research and development, often driven by technical advancements, positive public perception, active collaboration among researchers, and ample funding.
In contrast, AI Winters ❄️are characterized by periods of skepticism and reduced funding, driven by technical limitations, unwarranted hype surrounding AI capabilities, and budget cuts. In particular, the complexities of AI and the lack of computing power to meet high expectations also contribute to the onset of Winters in AI research and development.
There are many ways to represent this timeline of AI springs and winters. Think of it less as a neatly historical view (as these "AI Springs" and "AI Winters" cannot be neatly divided into historical time periods), and more as a conceptual timeline to help build your intuition on how these seasons also follow Gartner’s hype cycle:
Now that we have a global view of the AI Seasonal Cycle over time, let’s turn to another lens through which to make sense of the AI landscape, which occurred throughout the course of the Springs and Winters.
To understand the distinctions between AI and several subfields’ approaches, we need to understand a philosophical divide that occurred early on in the AI community between Symbolic AI and Subsymbolic (or Connectionist) AI.
Melanie Mitchell, author of Artificial Intelligence: A Guide for Thinking Humans (2020) sums up the Symbolic AI view as follows:
Advocates of the symbolic approach to AI argued that to attain intelligence in computers, it would not be necessary to build programs that mimic the brain. Instead, the argument goes, general intelligence can be captured entirely by the right kind of symbol-processing program (Mitchell, 11).
Symbolic AI was originally inspired by mathematical logic as well as the way people described their conscious thought processes. In contrast, subsymbolic approaches to AI took inspiration from neuroscience and sought to capture the sometimes unconscious thought processes underlying what some have called spoken words (Mitchell,12).
So we can see that the philosophical divide was driven by other domains and led to different approaches, namely a formal mathematical approach and the other being a neuroscientific approach.
The table below summarizes the key differences:
One very important development is psychologist Frank Rosenblatt’s late-1950s brain-inspired AI program, the perceptron, which is an example of a subsymbolic approach. The perceptron is the great-grandparent of modern AI’s deep neural networks since it was inspired by the way in which neurons process information.
To a computer scientist (or, in Rosenblatt’s case, a psychologist), information processing in neurons can be simulated by a computer program – a perceptron – that has multiple numerical inputs and one output [...]
In short, a perceptron is a simple program that makes a yes-or-no (1 or 0) decision based on whether the sum of its weighted inputs meets a threshold value” (Mitchell, 13 -15).
Developments in the understanding of neuronal architecture in the brain led to attempts to represent those same processes via computation.
A biological neuron receives signals through its dendrites, processes these signals in its nucleus, and then transmits the signals as electrical impulses along its axon. These impulses travel to the axon terminals, where they are converted into chemical signals (neurotransmitters) to communicate with other neurons or target cells. This process is essential for the transmission of information within the nervous system.
Just as a biological neuron has dendrites that receive signals, a cell body that processes signals, and an axon that transmits signals to other neurons, an artificial neuron features input channels, a processing step, and a single output that can connect to multiple other artificial neurons.
Rosenblatt also proposed that much like there are networks of neurons in the brain, there should be networks of perceptrons. These ideas would serve as the foundation for neural networks:
Neural Networks: A neural network is a computational model inspired by the structure and function of biological neural networks, typically composed of interconnected artificial neurons that process and transmit information to perform various tasks.
Two Major Subfields Emerge
Thus far, AI as a specialized field branched out into some important subfields spurred on by philosophical approaches and due to the seasonal changes. We can now formally define two other major important subfields:
Machine Learning (ML): A subset of AI that empowers machines to learn from data and improve their performance without explicit programming, using algorithms and statistical techniques.
NLP (Natural Language Processing): A subfield of AI that focuses on enabling machines to understand, interpret, and generate human language.
Neural networks are a specific type of machine learning model and are widely employed in tasks such as image recognition and natural language processing.
In summary, the history of AI reflects the complexity of the field, the challenges faced, and the ongoing efforts to develop AI technologies that can meet a variety of practical and ambitious goals. The dominant approach (symbolic or subsymbolic) has shifted as researchers continue to explore the best strategies for achieving AI's potential through its many seasonal changes.
Now that we have a good understanding of the AI landscape through the lens of AI Winter and Springs as well as the two dominant approaches, we can now turn to our current season and begin to answer our key questions: why now and is this time different?
Think back to the collage of 21st century technologies, many of which are orders of magnitude more computationally powerful than what was available in the 1950s and also generate massive amounts of data.
You may already be aware that two key things simultaneously power and produce many of the technologies from that collage:
🦾 Computing Power (in terms of hardware/software developments)
📊 Big Data.
In terms of computing power, one major hardware development has been GPUs (Graphical Processing Units), chips that were originally designed for rendering graphics in video games, that have proven essential for efficient processing of large datasets of all kinds.
Big Data is often characterized by 3 Vs: the presence of vast amounts of data (volume), often of different types and origins (variety), which is being generated and collected at high speeds (velocity).
Computing power and Big Data have a relationship some have described as the Data Flywheel Effect:
Some even refer to this combination of big data and computing power as the “second resurrection of neural networks”:
The tipping point was reached recently not by fundamentally new insights, but by processing speeds that make possible larger networks, bigger datasets, and more iterations. (source)
A more fitting term for all of this is “the deep learning revolution”, which most agree began in 2012. This period marked a significant shift in the capabilities of artificial intelligence and machine learning, leading to numerous applications and advancements in various domains.
The key characteristics that drove the revolution:
These advancements have led to significant progress in solving complex tasks that were previously challenging for traditional machine learning methods, which led to a surge in research and development related to deep learning in both academia and industry.
Deep Learning: A specialized branch of ML that employs neural networks with multiple layers to process and recognize complex patterns in data, enabling tasks such as image recognition and natural language understanding.
The difference between the a simple perceptron or neural network and deep learning is summarized by Melanie Mithcell in the following way:
“Deep learning simply refers to the methods for training ‘deep neural networks’, which in turn refers to neural networks with more than one [layer] … It’s worth emphasizing this definition: the ‘deep’ in deep learning doesn’t refer to the sophistication of what is learned; it refers only to the depth in layers of the network being trained” (Mitchell, 72).
Let’s revisit our AI Landscape chart and realize that deep learning, which we now have a better understanding of, cross cuts the ML field but also the NLP field.
Recall that:
Deep Learning: A specialized branch of ML that employs neural networks with multiple layers to process and recognize complex patterns in data, enabling tasks such as image recognition and natural language understanding.
NLP (Natural Language Processing): A subfield of AI that focuses on enabling machines to understand, interpret, and generate human language. NLP is essential for ChatGPT's conversational abilities and natural language understanding.
The deep learning revolution has therefore had a profound impact on the world of Natural Language Processing (NLP), making it more powerful and versatile for a general audience. Here's just some of the ways is it has transformed the way we work with language:
In essence, the deep learning revolution in NLP has made our interactions with language more seamless and intelligent and has opened up a world of possibilities for better understanding and working with human language in various industries and applications – especially education!
They have also led to some incredible breakthroughs in LLMs, or Large Language Models such as ChatGPT.
In our final multidimensional view above, we can finally see where ChatGPT sits among the complex field of AI. We can now formally define a large language model:
LLM (Large Language Model): LLMs refers to AI models like the ones shown in the chart below (BERT-Large, GPT-2, GPT-3) which are massive language models pre-trained on vast amounts of text data. LLMs possess a broad understanding of language and can generate coherent and contextually relevant responses in conversations.
The image above will give you a sense of the different LLM models over time and how much computing power they have.
So, what’s behind the excitement around ChatGPT?
OpenAI, founded in 2015, is a research and development organization that has focused on advancing AI research and developing cutting-edge large language models, such as ChatGPT, a conversational AI model designed to engage in natural, human-like conversations.
It’s critical to note that ChatGPT is a consumer-facing product and not synonymous with AI (which by now you should understand). OpenAI’s blog post explains that chatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022.
At the same time, the large language model (LLM) it was based on called GPT-3.5 and its predecessors have been publicly available for a while.
There have been many attempts to explain how LLMs like ChatGPT works, but none is better than the Financial Time's excellent interactive article, "Generative AI exists because of the transformer", which you should absolutely read and bookmark.
Some key takeaways from this article on how it works that will be critical for you to understand are as follows:
Now that we have a strong foundational understanding of Large Language Models, let’s finally get some hands-on experience with them.
While you are probably familiar with ChatGPT, OpenAI’s consumer-facing product (a front-end interface that wraps one of Open AI’s Large Language Models), you might be less familiar with the amazing resources they have on their website.
Another way to get familiar with how ChatGPT works under the hood is to explore OpenAI’s Tokenizer Tool.
The video below walks you through an explanation and example of how you can deepen your understanding of tokens, which will aid you in prompting:
Human-Readable View of Tokens:
Machine-Readable View of Tokens:
Try it yourself! Put some text in there and see what you get. See how this model is converting text input into tokens, each which has a Token ID that can then be used in probabilistic mathematical operations.
What does a prompt engineer, a hypnotist, and a screenwriter have in common?
In real life, these professions are probably quite distinct. When it comes to steering an AI Large Language Model towards a particular goal, they all have quite a bit in common. “Prompt Engineer” is a word that has been thrown around quite a bit lately, much like “artificial intelligence”.
In this section, we want to suggest that the skills associated with “prompt engineering” are more art than science. To help us understand the artistic side of prompting, we might say that it’s better to imagine that you are “hypnotizing” the large language model or are a good screenwriter that is writing descriptive but succinct stage directions for a theatrical play using the constraints of text alone.
Before we turn to examples below, there is a concept that you need to know about when it comes to using OpenAI’s model, as well as Wonda’s AI features, which are the three roles of prompting.
System: The "System" role represents the instructions or context-setting provided by the environment or application that is using the language model. It sets the stage for the conversation and can provide high-level guidance to the model. System instructions can be used to specify the role or behavior that the AI Assistant (the large language model) should take in the conversation.
Example: "You are a helpful assistant.”
User: The "User" role represents the input or messages provided by the person or user who is interacting with the assistant. These are the queries, statements, or questions posed to the model by the user.
Please note that the “User” role is referred to as “Me” within Wonda’s Platform, which we will demonstrate in the examples below.
Example: "Can you explain the concept of a metaphor?"
Assistant: The "Assistant" role represents the responses generated by the language model, or AI. When the model assumes this role, it generates a reply to the user's input based on the provided context from the System Prompt and the conversation history.
Example: "Of course! A metaphor is a figure of speech that makes a comparison between two seemingly unrelated things to illustrate a point."
Using these three roles together helps guide the conversation and allows for more interactive and context-aware interactions with the large language model. It's a way to instruct the model on how to behave and what kind of information or responses to provide based on the given roles and context.
Let’s start by taking a look at some illustrative examples that will help us understand the interaction between these three roles. Each of these examples are adapted from OpenAI’s examples. They are all text-based.
What we want to stress here is that Wonda allows you to create immersive experiences instead of just having walls of text output to the user. Each of the examples below thus demonstrates the power of Wonda’s platform and is a low-risk way to understand and get used to the platform.
This first example is adapted from OpenAI’s Example, Marv the Sarcastic Chatbot.
By now, you may have noticed some extra prompting text in the video:
### The Emotion Tags That You Will Indicate At The Start of Every Response ###
For each message that you respond to, you must structure each response by first indicating the emotion of your response by beginning each response with an emotional tag: "joy", "anger" or "neutral". The format of this emotional tag will always be: %%{"emotion": "joy", "intensity": 1}%%
When structuring the emotional tag, you must only choose among three emotions: Joy, Anger, or Neutral and you must indicate the intensity of the emotion as a number between 0 and 1. Every response should start with this emotional tag in the specific structural format I provided you. You will never include an emotional tag anywhere except the beginning of a response.
All you need to know is that this is some boilerplate text that it allows GPT to communicate emotions to the Avatar model.“ Please do not edit or remove this text unless you want to completely remove all emotions from your AI character's avatar.
🤖 ✍️ PROMPTING TIP 1: THE PERSONA PROMPTING PATTERN
🔑 We can shape the personality of our AI’s generated output response by modifying the System Prompt using the Persona Prompting Pattern
🔑 Iteration is key! Iterate iterate iterate. Fail fast.
This second example is adapted from OpenAI’s Example, Pros and Cons Discusser.
🤖 ✍️ PROMPTING TIP 2: EXTENDING THE PERSONA PROMPTING PATTERN
🔑 We can extend the Persona Pattern technique by not only identifying a real-life person, but adding more and more context to steer our AI toward the right sort of generated content.
🔑 The more additional context we add, the better our desired output will be.
This third example is adapted from OpenAI’s Example, Socratic Tutor.
🤖 ✍️ PROMPTING TIP 3: COMBINING PROMPTING PATTERNS
🔑 We can combine different techniques together, such as Persona Patterns, Flipped Interaction Patterns, and Audience Patterns to get pretty complex with our prompting.
🔑 When we have a complex prompt, we can leverage our AI conversation history which is hidden from the user, to get more and more nuanced.
Congratulations! That’s it for Module 1!
In the next Module you will have the opportunity to develop your own experience from Scratch for your specific use case. You’ll be able to create your own 360 skybox environments and your own avatar.