# AI Foundations ### About this export | Field | Value | | --- | --- | | **content_type** | course | | **platform** | contentstack-academy | | **source_url** | https://www.contentstack.com/academy/courses/ai-foundations | | **language** | en | | **product_area** | ai foundations | | **learning_path** | standalone | | **course_id** | ai-foundations | | **slug** | ai-foundations | | **version** | 2026-03-01 | | **last_updated** | 2026-05-18 | | **status** | published | | **keywords** | ["ai foundations","ai fundamentals","ai"] | | **summary_one_line** | This course offers an in-depth introduction to artificial intelligence (AI), examining its evolution, applications, and ethical implications. Beginning with a historical perspective, it traces AI's development from early… | | **total_duration_minutes** | 43 | | **lessons_count** | 12 | | **video_lessons_count** | 12 | | **text_lessons_count** | 0 | | **linked_learning_path** | standalone | | **linked_assessment_ref** | LMS_UNCONFIGURED_COURSE_ASSESSMENT | | **markdown_file_url** | /academy/md/courses/ai-foundations.md | | **generated_at** | 2026-05-18T10:08:42.366Z | | **intended_audience** | [] | | **prerequisites** | [] | | **related_courses** | [] | > **Academy MD v3** — companion `.md` for Ask AI. Quizzes and graded assessments are **LMS-only**; this file never contains answer keys. ## Course Overview | Metadata | Value | | --- | --- | | Catalog duration | 43m 3s | | Released (if known) | 2026-03-01 | | Product area | ai foundations | ### Description This course offers an in-depth introduction to artificial intelligence (AI), examining its evolution, applications, and ethical implications. Beginning with a historical perspective, it traces AI's development from early experiments to today's advanced technologies like ChatGPT and Google Gemini. The curriculum delves into key AI concepts and terminologies such as machine learning, deep learning, and neural networks, with practical examples that illustrate their use in real-world applications. Students will explore both the capabilities and limitations of current AI tools, learning how to apply AI in various contexts, from business operations to creative projects. Additionally, the course addresses the societal impacts of AI, discussing issues like bias, privacy, and the ethical use of AI technologies. This comprehensive overview aims to equip learners with a foundational understanding of AI to navigate and leverage this transformative technology effectively. ### Learning objectives 1. Follow each lesson in order. 2. Practice in a training stack using placeholders **YOUR_STACK_API_KEY** and **YOUR_DELIVERY_TOKEN** in local `.env` files only. 3. Validate API responses against the official documentation. ### Topics covered ai foundations; ai fundamentals; ai ## Course structure ```text ai-foundations/ ├── 01-introduction · video · 190s ├── 02-what-is-artificial-intelligence- · video · 338s ├── 03-machine-learning · video · 349s ├── 04-machine-learning-frameworks · video · 214s ├── 05-deep-learning · video · 211s ├── 06-generative-ai · video · 275s ├── 07-prompting-generative-ai · video · 147s ├── 08-custom-gpts-rag · video · 177s ├── 09-hallucinations · video · 217s ├── 10-bias-in-ai · video · 196s ├── 11-ethical-considerations · video · 199s ├── 12-conclusion · video · 70s ``` ## Lessons ### Lesson 01 — AI Foundations #### Video details #### At a glance - **Title:** Introduction - **Duration:** 3m 10s - **Media link:** https://cdn.jwplayer.com/previews/PUvNP1jb - **Publish date (unix):** 1714402846 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113752 kbps - video/mp4 · 180p · 180p · 202825 kbps - video/mp4 · 270p · 270p · 280618 kbps - video/mp4 · 360p · 360p · 342082 kbps - video/mp4 · 406p · 406p · 393421 kbps - video/mp4 · 540p · 540p · 554754 kbps - video/mp4 · 720p · 720p · 813531 kbps - video/mp4 · 1080p · 1080p · 1748665 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/PUvNP1jb-120.vtt` #### Video transcript Here's the thing, artificial intelligence is a transformative technology. If you've played around with any of the new generative AI tools, you know that I'm not exaggerating when I say AI will truly change not only the way that we work, but it will have a profound effect on the way that we live. And because of this, there's an enormous amount of excitement. But with that excitement, there's equal amounts of unease, apprehension, and nervousness. Welcome to this brief introduction to artificial intelligence. In this course, we'll talk about the latest AI innovations like ChatGPT, Google Gemini, and others. We'll explore tools that enable us to create fresh content, be it text or code, or even images, audio, and video. We'll walk through what these tools excel at, what their limitations are, and the reasons behind those constraints. However, understanding the newest tools requires a peek into the past. And while these innovations feel like they came to fruition seemingly overnight, they didn't. They're the culmination of decades of work in the AI space. So we'll spend some time talking about what work has been done in the past to better understand where we are today. Along the way, we'll unpack concepts and terminology related to AI, like machine learning, deep learning, neural networks, large language models, foundation models, and more. And importantly, we'll clarify their significance. Understanding these terms that you've undoubtedly have come across will lay a foundation of understanding and set context, allowing you to really grasp what AI's role is and its potential to transform. We'll explore its implications for businesses and projects. And while we'll look at it in that context, make no mistake, AI affects us all, regardless of our roles, and its influence is only set to grow. Fear not, this isn't a technical deep dive, rather an overview to better understand how to practically use AI. So let's begin by tackling a basic question that seems simple, yet is anything but. What is Artificial Intelligence? #### Key takeaways - Connect **AI Foundations** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 02 — What is Artificial Intelligence? #### Video details #### At a glance - **Title:** What Is Artificial Intelligence - **Duration:** 5m 38s - **Media link:** https://cdn.jwplayer.com/previews/oyIFhAMe - **Publish date (unix):** 1714422013 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113451 kbps - video/mp4 · 180p · 180p · 188270 kbps - video/mp4 · 270p · 270p · 250219 kbps - video/mp4 · 360p · 360p · 301835 kbps - video/mp4 · 406p · 406p · 344470 kbps - video/mp4 · 540p · 540p · 479873 kbps - video/mp4 · 720p · 720p · 722544 kbps - video/mp4 · 1080p · 1080p · 1689615 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/oyIFhAMe-120.vtt` #### Video transcript The term artificial intelligence originally appeared at a conference at Dartmouth College in 1956, which established AI as a legitimate field of study. Think about that, 1956. Here we are in 2024 and it seems like AI is now just becoming a thing. Early AI research focused on solving specific problems, and there were some early successes. There was ELISA, a natural language processing program, and SHAKI, a mobile robot that could navigate its surroundings. But there was a problem. As soon as something was delivered in the AI space, it was quickly dismissed as not really being artificial intelligence. And there's a few reasons for that. As with anything related to technology, it tends to move pretty fast. And as AI evolved, as new capabilities were introduced, it made locking down a definition challenging, and it became too much of a moving target. And some of those challenges come from the subtle differences in AI compared to traditional programming. Traditional programs follow a set of explicit instructions that follow step-by-step directions to perform a specific task. Think of a recipe for a meal. It tells you exactly what ingredients to use and how to combine them. While AI programs may also have some pre-programmed elements, a key distinction is that they have the ability to learn and adapt. This can involve training on data sets or interacting with their environments. Think of a chef who can not only follow recipes, but also experiment with new ingredients and techniques based on experience. But even that example can have holes poked in it. What if two recipes were essentially combined? Is that the chef, or in our case, the AI? Or is it not? Is it simply joining two pieces of information together? This type of second guessing is so pervasive, they have a name for it. It's called the AI effect. And the AI effect is the tendency for people to downplay a machine's intelligence as soon as it achieves something impressive. Basically, what was once considered a true sign of AI becomes just a trick once a computer can do it. All this is to say, we're not going to get a clear, easy definition for artificial intelligence. But we need to at least be able to classify different types of AI. And even this comes with some inherent problems. And that's people's perceptions of what AI is. As we think about AI in pop culture terms, it's usually in the form of a robot or machine with human-level or beyond intelligence. This stereotype is perpetuated by science fiction, whether it be a movie or a novel. But let's start there. This type of AI is referred to as strong AI, or artificial general intelligence. And it's completely theoretical. It doesn't exist. Artificial intelligence possessing the ability to understand, learn, and apply its intelligence to solve any problem with the same level of competence as a human being? It's still science fiction. And that's not to say that work and research isn't happening in the space. But it's not the type of AI that we're seeing in the market. What we have now is considered weak or narrow AI. Narrow AI focuses on performing a singular or limited task with a predefined scope. Think about speech recognition, systems like virtual assistants, image recognition used for tagging photos, or recommendation engines on streaming and shopping platforms. These systems operate under a set of constraints. There's no understanding or consciousness on the behalf of the machine. They simply follow programmed algorithms or learn from vast data sets within their specific areas. Now, with that being said, narrow AI technologies have made significant advancements in their areas over the past few years, often achieving or surpassing human level performance in tasks such as playing complex games, translating languages, and even diagnosing diseases from medical images. However, they lack the ability to perform beyond their designated area, to apply their knowledge in a broader context, distinguishing them sharply from the conceptual goal of artificial general intelligence. Even new tools like ChatGPT or Google's Gemini are considered to be narrow AI. And sometimes people struggle with this because they can perform a variety of tasks. You can ask a chatbot to help with a report, summarize a white paper, enhance your resume, depending upon the version, even generate images. It doesn't feel that narrow. Despite its versatility in handling various topics, it lacks the general intelligence to perform tasks outside the constraints of its programming and the capabilities given to it by its training. And I used a key word there, training. Machine learning is one of the key implementations of AI. #### Key takeaways - Connect **What is Artificial Intelligence?** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 03 — Machine Learning #### Video details #### At a glance - **Title:** Machine Learning - **Duration:** 5m 49s - **Media link:** https://cdn.jwplayer.com/previews/IneY7znY - **Publish date (unix):** 1714422339 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113493 kbps - video/mp4 · 180p · 180p · 187830 kbps - video/mp4 · 270p · 270p · 248383 kbps - video/mp4 · 360p · 360p · 301188 kbps - video/mp4 · 406p · 406p · 343227 kbps - video/mp4 · 540p · 540p · 479933 kbps - video/mp4 · 720p · 720p · 722449 kbps - video/mp4 · 1080p · 1080p · 1674544 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/IneY7znY-120.vtt` #### Video transcript Machine learning is a critical aspect of artificial intelligence, but the two aren't the same. AI is a broad field focused on making computers mimic human intelligence. Machine learning, on the other hand, is a subset of AI focused on teaching computers to learn from data. To understand this data-driven approach, let's contrast it with traditional programming found in many software applications. In traditional programming, developers craft explicit instructions and rules for every scenario. For example, if a user inputs the wrong username or password, the program will display a specific error message. Or, if the balance of your bank account falls below a certain threshold, it could trigger a notification. Or if a player's score reaches a certain level, it could unlock a new game level. This approach is highly effective for tasks where the rules and outcomes are well-defined and predictable. However, machine learning diverges from this. Instead of starting with a comprehensive set of rules programmed by a developer, machine learning algorithms begin with data. They learn from this data, identifying patterns and making decisions based on the insights gained from the data that they were trained with. So there's no need for every possible situation to be predefined by a human programmer. Imagine showing someone a picture of a Springer Spaniel. The more pictures they see, the better they get at recognizing this specific breed of dog. Machine learning algorithms work in the same way. They're fed massive amounts of data, allowing them to identify patterns and improve their ability to perform tasks. This process is referred to as training a model, and a model is a decision-making tool based on its training. A common misconception is that the model somehow stores the information it was trained on, like a database. It doesn't. In our Springer Spaniel example, the model recognizes the patterns and features of the dog breed. The more diverse the training, the better it gets at recognizing these dogs. And if you were using image-based generative AI, and you asked for a Springer Spaniel, it would generate a new image based on those patterns. And if you wanted to know what image or images inspired that generation, that information isn't stored anywhere. Because it's not just a few images, it's thousands, maybe millions of photos it learned those patterns from. And while the Springer Spaniel example may seem basic, it does illustrate something machine learning is very good at, classification. Classification involves identifying which category a new observation belongs to. For example, after being trained on thousands of images labeled as either Springer Spaniel or German Shepherd, a machine learning model could accurately classify new images as being one or the other. This capability makes machine learning ideal for tasks ranging from email spam detection, where emails are classified as spam or not spam, to a medical diagnosis, where patient data can be classified to indicate the presence or absence of a disease. Beyond classification, machine learning is also good at clustering and regression. Clustering differs from classification in the sense that it can group information based on similarities without the help of labels. Regression is used for predicting outcomes based on data. Regression is a type of predictive modeling that analyzes the relationships between a dependent variable and one or more independent variables. As an example, let's say you're playing a bit of a guessing game. If you're trying to guess the score you'll get in a video game based on the number of hours you practice, the score you're trying to predict is the dependent variable because it depends on how much you practice. The hours you practice are the independent variables because they are what you think will change your score. So if you practice more, you expect your score to go up. This guessing game is a lot like regression. Let's look at a real world example. So in terms of regression, let's talk about used car prices. Let's say you're looking for a used car and you want to know how much you should expect to pay. You could gather data on various used cars being sold online or at dealerships, including all the parameters that would sway the car's price. And you would feed this data into the model, which would allow you to uncover factors that influence the final selling price. Regression would reveal which factors have the most impact. For example, high mileage often has a negative influence on price. This wouldn't be perfect. A well-maintained luxury car with high mileage might be worth more than a neglected economy car with low miles. But regression would give you a good starting point for negotiating a fair price. The trick of making this successful is quality data. And not just a little. A lot. That's why organizations put such an emphasis on collecting as much data as possible. The more data means you have more information to train your model. #### Key takeaways - Connect **Machine Learning** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 04 — Machine Learning Frameworks #### Video details #### At a glance - **Title:** Machine Learning Frameworks - **Duration:** 3m 34s - **Media link:** https://cdn.jwplayer.com/previews/BFSpBlWR - **Publish date (unix):** 1714422565 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113897 kbps - video/mp4 · 180p · 180p · 182694 kbps - video/mp4 · 270p · 270p · 240207 kbps - video/mp4 · 360p · 360p · 291541 kbps - video/mp4 · 406p · 406p · 332363 kbps - video/mp4 · 540p · 540p · 463008 kbps - video/mp4 · 720p · 720p · 695017 kbps - video/mp4 · 1080p · 1080p · 1627112 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/BFSpBlWR-120.vtt` #### Video transcript When dealing with machine learning, there's two fundamental components required, the data and the algorithm. Oftentimes people think that organizations need to develop their own machine learning algorithms. The reality is they generally don't. Most businesses wouldn't build their own communication tools from scratch, but rather use well-established products like Microsoft Teams or Slack. It's more common and practical to utilize existing frameworks and platforms. These tools, ranging from open source to commercial solutions, are readily available on platforms such as Microsoft Azure, Google Cloud, or Amazon AWS. These environments offer a variety of machine learning algorithms that can be tailored to specific needs. Some are better or faster at categorization or regression or clustering. This makes deployment of artificial intelligence more accessible than ever before and really a big reason for its explosive growth recently. However, implementing machine learning is not as simple as feeding data into the algorithm and expecting insightful outcomes. As an example, think about a comprehensive set of retail store transaction data collected over several months. This data might include details such as time of purchase, items bought, payment methods, customer demographics, and even staff members involved in each transaction. All this data has the potential to provide insights. It could include patterns in customer buying behavior, identification of potential stock shortages or surpluses, I suppose, and even flags for suspicious transactions that might indicate theft or fraud. But to leverage this information effectively, you have to identify a specific issue or opportunity that you're trying to address. Are you aiming to improve inventory management, enhance customer service, or prevent losses? This specifically will guide the data preparation process, which is the backbone of effective machine learning. Data preparation involves meticulous cleaning of the data, labeling essential features, and discarding any irrelevant or misleading information. The adage, garbage in, garbage out, holds particularly true here. The quality of the input data critically influences the effectiveness of the resulting machine learning model. This stage is where data scientists spend a substantial amount of time refining raw data into a format that's both meaningful and computationally viable for algorithms to process. Once the data is ready, the next step includes selecting the appropriate algorithm, configuring it, and conducting tests. This is not a straightforward process. Data scientists often experiment with various algorithms to determine which yields the best results for the given data and desired outcomes. Machine learning's impact has been profound, with applications spanning across various industries, leading to advanced AI technologies like ChatGPT and Dolly. These advanced systems, although more complex than basic models, are built on the foundational principles of machine learning. Of course, there's more sophisticated aspects of this technology. Let's discuss that next. #### Key takeaways - Connect **Machine Learning Frameworks** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 05 — Deep Learning #### Video details #### At a glance - **Title:** Deep Learning - **Duration:** 3m 31s - **Media link:** https://cdn.jwplayer.com/previews/CSmJF2s9 - **Publish date (unix):** 1714422720 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113690 kbps - video/mp4 · 180p · 180p · 183081 kbps - video/mp4 · 270p · 270p · 240671 kbps - video/mp4 · 360p · 360p · 290420 kbps - video/mp4 · 406p · 406p · 329740 kbps - video/mp4 · 540p · 540p · 456450 kbps - video/mp4 · 720p · 720p · 681878 kbps - video/mp4 · 1080p · 1080p · 1576960 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/CSmJF2s9-120.vtt` #### Video transcript Early on, I mentioned that artificial general intelligence, the type of AI that provides human-level or beyond intelligence, is theoretical. And while that is the case, there is a fascinating area within AI called deep learning, which is a specialized branch of machine learning that draws some inspiration from how our brains work. Deep learning utilizes structures known as neural networks, which, while not exact replicas of our actual brains, do borrow from the general concept of how our brains work. Think of it this way. Our brains operate with billions of neurons interconnected through synapses. Each neuron communicates with numerous others, forming a vast network. Similarly, in deep learning, we create artificial neural networks, where simulated neurons or nodes link to one another, passing information and processing it through multiple layers. For those primarily using AI tools rather than developing them, the intricate details of neural network design isn't the critical piece here. What's important is to understand what deep learning excels at. This technology has revolutionized tasks that involve complex pattern recognition, like interpreting images, recognizing faces, understanding spoken language, and even mastering strategic games. But deep learning isn't without its challenges. It demands significant computational resources and vast amounts of data to train effectively. While it's hard to pin down the exact volume of data required, as it varies by project, it's not uncommon to need hundreds or thousands or even millions of data points to train a robust deep learning model. The cost can be staggering, running into the multiple millions of dollars just for the computational power alone. This doesn't factor in any other expenses like development or staffing. Despite these costs, the investment in deep learning is often worthwhile due to the superior outcomes it can achieve with sufficient data. Once a model is fully trained, its capabilities can be profound. Natural language processing, or NLP for example, it focuses on enabling computers to understand, interpret, and respond to human language in a way that's both meaningful and useful. It can manipulate human language to perform tasks such as speech recognition, sentiment analysis, and translation. And at the surface, it sounds straightforward. But think about how you might interact with a digital assistant. You may ask, Alexa, who do I have a meeting with at 3 p.m.? Their surface statement is the easy part. The complication comes from the assistant understanding that it needs to retrieve your calendar information, provide the name of the attendees within the meeting at 3 p.m. And this type of complexity, this type of understanding, is where we can begin to see how even more complex services are emerging. Natural language processing, natural language understanding, and natural language generation brings us to what probably has you interested in AI in the first place, generative AI. #### Key takeaways - Connect **Deep Learning** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 06 — Generative AI #### Video details #### At a glance - **Title:** Generative Ai - **Duration:** 4m 35s - **Media link:** https://cdn.jwplayer.com/previews/WeSAoc6f - **Publish date (unix):** 1714422819 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113698 kbps - video/mp4 · 180p · 180p · 187728 kbps - video/mp4 · 270p · 270p · 249775 kbps - video/mp4 · 360p · 360p · 302429 kbps - video/mp4 · 406p · 406p · 344466 kbps - video/mp4 · 540p · 540p · 481456 kbps - video/mp4 · 720p · 720p · 722975 kbps - video/mp4 · 1080p · 1080p · 1675462 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/WeSAoc6f-120.vtt` #### Video transcript We looked at a couple different areas of AI, but the one that's gotten the most attention, and frankly is the most mainstream, is generative AI. And generative AI comes in many forms. Some applications generate written text, and you've probably used some of these tools, like Chats GPT, Google's Gemini, or Anthropic Cloud, while others are great for producing new images, like Midjourney, Dolly, and Stable Diffusion. Then there's others that are great for generating programming code, like GitHub Copilot. Then there's applications that will generate video or music tracks. Generative AI covers a broad spectrum of software tools and platforms, and the list seems to grow daily. Generative AI capabilities are appearing in tools that you probably already work with, products from Microsoft and Adobe, services that you might be familiar with, like Grammarly or even ContentStack. The key component of generative AI lies in the interaction. You input commands, typically through a brief prompt detailing your requirements, and the AI creates entirely new content, hence the word generative, which reflects the AI's ability to not merely retrieve existing data, but to actually create novel outcomes. Chatbots leverage something called Large Language Models, or LLMs, similar to the models we talked about earlier. They're trained on vast amounts of text data, and they're capable of generating human-like responses by leveraging deep learning and transformer architectures. Think of a transformer architecture as a conductor of an orchestra. The transformer has something called a self-attention mechanism, which allows it to focus on different sections of the orchestra, or input data, simultaneously, rather than one at a time. This enables the conductor to understand how each section contributes to the overall performance, adjusting focus dynamically to highlight or blend certain musical elements. In transformers, this translates to processing input data in parallel, allowing the model to assess the relevance of different data elements and grasp complex relationships. Additionally, there is an encoder and a decoder within the architecture. It functions in a similar way as a translator. The encoder absorbs and comprehensively represents the original language, while the decoder interprets and recreates this into new language, whether it be for a translation, a summarization, or generating text. This has proven exceptionally effective for natural language processing, serving as the foundational structure for most contemporary large language models. So you might ask, do AI tools that create images, music, or video also use large language models? Kind of, but they're referred to as foundation models, as they're multimodal, not just a tool for working with language. The underlining mechanism of any generative AI tool remains fairly consistent across different applications, whether it's creating text, images, code, whatever the case may be. A quick sidebar. An important concept to understand when working with any of these generative AI tools is this idea of a knowledge cutoff date. Because training these models is so computationally expensive, and it requires an enormous amount of data, the vendors often set a cutoff date, which is to say they'll stop using data up until a certain point in time, typically as close as they can get till today. But it's usually months earlier. So if you were to ask ChatGPT to summarize an event that occurred in the past few months, regardless of how significant the event, it might not be able to do it. To know what the date is of a knowledge cutoff, just ask the chatbot. Now that we have a general understanding of what's happening in the background with these AI tools, and we understand some foundational concepts like machine learning, deep learning, computation models, and LLMs, we can begin to explore how we can get the most out of these tools, and it often starts with a prompt. #### Key takeaways - Connect **Generative AI** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 07 — Prompting Generative AI #### Video details #### At a glance - **Title:** Prompting Generative Ai - **Duration:** 2m 27s - **Media link:** https://cdn.jwplayer.com/previews/vQamweER - **Publish date (unix):** 1714491905 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 114046 kbps - video/mp4 · 180p · 180p · 181640 kbps - video/mp4 · 270p · 270p · 238768 kbps - video/mp4 · 360p · 360p · 292111 kbps - video/mp4 · 406p · 406p · 332248 kbps - video/mp4 · 540p · 540p · 463309 kbps - video/mp4 · 720p · 720p · 694996 kbps - video/mp4 · 1080p · 1080p · 1601700 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/vQamweER-120.vtt` #### Video transcript When it comes to engaging with generative AI tools, the quality, the creativity of your prompt will ultimately dictate what type of results you get. And this holds true for text-based AI like ChatGPT to image-based AI like MidJourney. Let's look at some examples. If I ask ChatGPT to write a cover letter for a data analyst role that I'm applying for, pretending for a moment that I'm a data analyst, it'll do just that. Admittedly, the results are pretty generic. The results are generic because my prompt was generic. But this is a chatbot, so it remembers the context of the conversation. So I can chat with it without having to re-ask to write a cover letter. It knows that's what we're doing. I can simply ask to highlight my five-plus years of working as a data analyst at a large technology firm. It returns a result that's a little more personalized. I could continue to interact with ChatGPT to further refine the results. I could paste in examples of other cover letters that resonate with my style or the role that I'm applying for. You'll find that you'll experiment with your prompts as small modifications can alter the results to better suit your desired outcome. Some things to keep in mind as you prompt. Be creative. Add context where needed. It can be useful to have the AI assume a role or a persona. Act as a professor, a financial planner, or a software engineer. Act as a marketing professional. You could ask the AI to act as a fictional character, a historical figure, or an industry icon. Respond as if you were Warren Buffett explaining an investment strategy. Describe quantum physics as if you were Stephen Hawking. Assume the role of Ted Lasso. You could also tailor your prompts to a specific audience. Explain this to me as if I were a financial novice. As if you were talking to a room full of doctors. Be specific about length. You could use words like short, but also a specific amount of words like 300 or two paragraphs. Again, the more specific you are with your prompts, the better your results will be. #### Key takeaways - Connect **Prompting Generative AI** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 08 — Custom GPTs & RAG #### Video details #### At a glance - **Title:** Custom Gpts & Rag - **Duration:** 2m 57s - **Media link:** https://cdn.jwplayer.com/previews/QFDJ1wlQ - **Publish date (unix):** 1714491899 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113941 kbps - video/mp4 · 180p · 180p · 185778 kbps - video/mp4 · 270p · 270p · 246070 kbps - video/mp4 · 360p · 360p · 297229 kbps - video/mp4 · 406p · 406p · 337507 kbps - video/mp4 · 540p · 540p · 469191 kbps - video/mp4 · 720p · 720p · 700410 kbps - video/mp4 · 1080p · 1080p · 1625103 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/QFDJ1wlQ-120.vtt` #### Video transcript There's two ways you can personalize an experience with generative AI tools. First, our custom knowledge bases, which in the case of ChatGPT, are called custom GPTs. Think of it as tweaking a meal recipe to suit your taste. That's what we do with custom GPTs. We start with a basic AI model, which knows a lot about a lot of things, and then we fine tune it with specific data that's unique to your needs. This could be anything from medical knowledge for a healthcare app, to customer service dialogues for support chatbots. The idea is to make the AI better at understanding and generating responses that are super relevant to particular topics or industries. It's refining the tool for a specific use case, instead of the more general-purpose, one-size-fits-all solution. This helps the AI perform exceptionally well where you need it to, by supporting it with additional context for any given topic. Now, let's switch gears and talk about retrieval-augmented generation. Imagine you're having a conversation, and you can pull out your smartphone to look up facts to make sure what you're saying is not just correct, but also super relevant. And you may already do that in your daily life. But that's what retrieval-augmented generation does for AI. While it's chatting with you, it pulls up-to-date information from external sources to boost its replies. This isn't about knowing more in a traditional sense, like what we were doing with custom GPTs, but this is about the AI being able to access a vast amount of information in real-time, ensuring it gives you the most accurate and contextual answers possible. It's particularly useful when the AI needs to handle questions on a wide range of topics and provide answers that are based on the very latest data. Mind you, custom GPTs in ChatGPT are easier to set up using a graphical interface. Other chatbots, like Google's Gemini, require you to create a custom knowledge base through APIs, and working with retrieval-augmented generation often requires another layer of training for the model. Both custom GPTs and retrieval-augmented generation have their unique strengths. If you need an AI that's an expert in a specific field, go with a custom GPT. But if you're looking for an AI that can offer the freshest, most precise information across a broad spectrum, then retrieval-augmented generation is your go-to. It's about matching the right tool for your needs to get the best out of this technology. #### Key takeaways - Connect **Custom GPTs & RAG** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 09 — Hallucinations #### Video details #### At a glance - **Title:** Hallucinations - **Duration:** 3m 37s - **Media link:** https://cdn.jwplayer.com/previews/3Lxm59Zz - **Publish date (unix):** 1714496002 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113926 kbps - video/mp4 · 180p · 180p · 186276 kbps - video/mp4 · 270p · 270p · 246904 kbps - video/mp4 · 360p · 360p · 298554 kbps - video/mp4 · 406p · 406p · 339934 kbps - video/mp4 · 540p · 540p · 473555 kbps - video/mp4 · 720p · 720p · 711032 kbps - video/mp4 · 1080p · 1080p · 1679727 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/3Lxm59Zz-120.vtt` #### Video transcript Imagine you're using a GPS system in your car. As you drive, it calculates the fastest route based on real-time traffic data. It seems like it knows everything about the roads, the turns, and even the expected delays. But there's a twist. While it can guide you to your destination, it doesn't really understand what it means to drive, the nuances of the traffic patterns, or even why certain routes are preferable during a particular time of day. It operates purely on data, without a genuine understanding. Now let's draw a parallel with generative AI. Like GPS, generative AI can produce content, whether text, images, or music, that often seems perfectly tailored and deeply informed. For instance, let's say I ask AI to write a poem about autumn in New England. Within seconds, it can churn out a verse that mentions golden leaves, crisp air, and apple cider. The results might be stunning, but does the AI understand the essence of autumn, or the emotional undertone of the fleeting beauty it describes? No, it doesn't. It's all based on patterns it has seen in the data that it was trained on. This leads us to a critical realization about generative AI. It's like a mirror reflecting a thousand books, images, and songs, all mixed into what seemed like a new creation. It doesn't really know anything. For example, if I input a prompt for AI to generate a picture of a cat playing a piano, the AI doesn't grasp the humor or absurdity of the scene. It simply processes that cats and pianos are both recognizable and can be combined visually. Moreover, the real trouble begins when we rely on these AI systems for tasks that require deep understanding or accuracy. Suppose an AI is tasked to generate a legal document or a medical advice report. The AI might produce documents that look correct, but they could be based on misunderstood nuances or outright inaccuracies. This is especially dangerous in fields where precision is critical. The concept of hallucinations in AI is where the system generates false or misleading information and does so confidently. Just as our hypothetical GPS might suddenly insist on a non-existent shortcut, an AI could confidently present a historical fact that never happened or misinterpret a crucial piece of scientific data. As we integrate AI more deeply into our lives, we have to be aware of these limitations. It's essential to remember that AI doesn't replace human expertise or judgment. Just as you wouldn't let your GPS drive your car, you shouldn't let AI make unverified decisions in critical areas of life. Finally, as we look to the future, the ethical implications of AI has to be carefully considered. Issues on data privacy, intellectual property, and the potential for AI to perpetuate biases or spread misinformation has to be addressed. These are not just technical challenges, but moral imperatives for developers and users alike. #### Key takeaways - Connect **Hallucinations** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 10 — Bias in AI #### Video details #### At a glance - **Title:** Bias - **Duration:** 3m 16s - **Media link:** https://cdn.jwplayer.com/previews/zHd4WVia - **Publish date (unix):** 1714500416 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113641 kbps - video/mp4 · 180p · 180p · 183523 kbps - video/mp4 · 270p · 270p · 240502 kbps - video/mp4 · 360p · 360p · 289493 kbps - video/mp4 · 406p · 406p · 328795 kbps - video/mp4 · 540p · 540p · 456036 kbps - video/mp4 · 720p · 720p · 683369 kbps - video/mp4 · 1080p · 1080p · 1630356 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/zHd4WVia-120.vtt` #### Video transcript As we know, generative AI systems are trained on vast datasets collected from the Internet, books, media, and other sources. These datasets, while rich in information, often reflect the existing biases of society. For example, if a generative AI trained on historical literature predominantly features characters of a certain gender in stereotypical roles, it may unintentionally replicate those stereotypes when generating new content. This perpetuation of bias can be subtle and without careful analysis may go unnoticed. Moreover, the way data is selected and prepared for training these models can also introduce biases. If the data is not diverse or representative of different groups, the AI's output will likely mirror these gaps. This is especially critical in systems used for facial recognition or personalized recommendations where failing to adequately represent the diversity of human features or interests can lead to unequal outcomes. Another layer of complexity arises from the AI's training process itself. The algorithms used to learn patterns and generate new content are designed by humans, and that may inadvertently include the creator's biases. This could affect the data that the algorithm weighs as more important, influencing the AI's decisions in ways that might not be immediately apparent. As an example, in MidJourney, I asked for an image of a carpenter. It returned all males. Likewise, I asked for an image of a nurse. It returned all females, both with little ethnic diversity. Again, it's the historical data that's been collected. If you go back a hundred years, 70 of those years, that was probably the representation of those careers. But the AI isn't correcting for today's view or representation of those careers. So as you prompt, you have to keep this in mind. If you ask for diversity, whether it be gender or ethnic background, they will return more diverse results. To mitigate these biases long-term, developers are exploring various strategies. One approach is to curate the training datasets more carefully, ensuring that they're broad and inclusive. Another is to employ techniques in the model's design that can detect and correct bias, such as adjusting the algorithm's parameters during the training to compensate for imbalances in the data. It's an ongoing effort, and while it's challenging, and sometimes there are setbacks in the outcomes, even in regard to overcompensation, which happened with Google's Gemini, I do see improvements in this area regularly because of the ongoing commitment to addressing bias in AI. But understanding that AI does have these limitations, it also brings up a whole different set of issues, ethical issues. Let's talk about that next. #### Key takeaways - Connect **Bias in AI** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 11 — Ethical Considerations #### Video details #### At a glance - **Title:** Ethical Considerations - **Duration:** 3m 19s - **Media link:** https://cdn.jwplayer.com/previews/tsX195sP - **Publish date (unix):** 1714521905 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 113469 kbps - video/mp4 · 180p · 180p · 180640 kbps - video/mp4 · 270p · 270p · 236514 kbps - video/mp4 · 360p · 360p · 285419 kbps - video/mp4 · 406p · 406p · 323861 kbps - video/mp4 · 540p · 540p · 449976 kbps - video/mp4 · 720p · 720p · 673883 kbps - video/mp4 · 1080p · 1080p · 1604440 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/tsX195sP-120.vtt` #### Video transcript As AI integrates more deeply into every aspect of our lives, we need to make sure we're using it responsibly. Based on the last clip, let's start with a scenario. Imagine a financial institution is using AI to assess loan applications. The AI system, trained on historical financial data, decides who qualifies for a loan and who does not. It sounds efficient, but what if the training data includes unintentional biases against certain demographics? This AI could inadvertently continue these biases, denying loans to individuals based on patterns it's learned from the past, rather than their actual financial stability. This issue is just the tip of the iceberg. It brings us to the concept often referred to as the black box problem in AI. When AI makes a decision, whether it's approving a loan, filtering job applications, or diagnosing diseases, it often doesn't provide an explanation for its conclusions. It gives an answer without showing its work. This opacity can be problematic, especially when decisions have significant consequences on people's lives. As we become more aware of these issues, the field is moving towards what is known as explainable AI. This approach aims to make AI's decision-making process clear and understandable to humans. It's not just about fairness, but also about building trust between technology and its users. Moreover, the rise of nefarious AI implementations like deep fakes, which are hyper-realistic videos and audio recordings manipulated by AI, illustrates another risk, the spread of misinformation. Deep fakes can make it appear as though individuals are saying or doing things they never actually did or said, potentially damaging reputations or even influencing elections. Acknowledging these challenges, there's a growing movement towards responsible AI. Major tech companies and academic institutions are now prioritizing transparency, accountability, and fairness in their AI developments. They're not only adjusting their algorithms to mitigate bias, but also embracing regulatory standards that require them to disclose how their AI systems are making decisions. Organizations like the Partnership on AI are at the forefront of establishing ethical guidelines that encourage safety and transparency. These initiatives are essential in assuring that AI technologies enhance societal well-being without undermining human rights. And let's not forget the broader implications. As AI continues to evolve, the potential for what it can achieve grows exponentially. But so does the potential for harm, if it's not developed responsibly. This was famously highlighted by Stephen Hawking, who warned that while AI could be one of humanity's greatest achievements, it might also pose one of its greatest threats. #### Key takeaways - Connect **Ethical Considerations** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ### Lesson 12 — Conclusion #### Video details #### At a glance - **Title:** Conclusion - **Duration:** 1m 10s - **Media link:** https://cdn.jwplayer.com/previews/YJZEaUfA - **Publish date (unix):** 1714515197 #### Streaming renditions - application/vnd.apple.mpegurl - audio/mp4 · AAC Audio · 114952 kbps - video/mp4 · 180p · 180p · 195666 kbps - video/mp4 · 270p · 270p · 259002 kbps - video/mp4 · 360p · 360p · 308635 kbps - video/mp4 · 406p · 406p · 350944 kbps - video/mp4 · 540p · 540p · 489562 kbps - video/mp4 · 720p · 720p · 733996 kbps - video/mp4 · 1080p · 1080p · 1692499 kbps #### Timed text tracks (delivery) - **thumbnails:** `https://cdn.jwplayer.com/strips/YJZEaUfA-120.vtt` #### Video transcript Well, that's going to wrap up our introduction to artificial intelligence. I hope this gives you a better understanding of where AI came from, how it generally works, and why you get the results that you do. Obviously, this is going to change not only the way that you work, but it'll impact just about every part of your life. So as I sign off, I thought maybe I'd have ChatGPT write a closing statement. So what I'll do is I'll prompt, write a closing statement for a course on the foundations of artificial intelligence. I'll also paste the outline into ChatGPT to give it some context. And it returns, throughout this course, we've explored the transformative potential and ethical considerations of AI, equipping ourselves to responsibly harness its capabilities. Thank you for your engagement and curiosity, and I look forward to continuing the journey of discovery with you in the future. It's not half bad. #### Key takeaways - Connect **Conclusion** back to your stack configuration before moving to the next module. - Capture one concrete artifact (screenshot, Postman call, or code snippet) that proves the step works in your environment. - Re-read the delivery versus management boundary for anything you changed in the entry model. ## Resources & references | Page | Companion Markdown | | --- | --- | | /courses/ai-foundations/introduction | /academy/md/courses/ai-foundations/introduction.md | | /courses/ai-foundations/what-is-artificial-intelligence- | /academy/md/courses/ai-foundations/what-is-artificial-intelligence-.md | | /courses/ai-foundations/machine-learning | /academy/md/courses/ai-foundations/machine-learning.md | | /courses/ai-foundations/machine-learning-frameworks | /academy/md/courses/ai-foundations/machine-learning-frameworks.md | | /courses/ai-foundations/deep-learning | /academy/md/courses/ai-foundations/deep-learning.md | | /courses/ai-foundations/generative-ai | /academy/md/courses/ai-foundations/generative-ai.md | | /courses/ai-foundations/prompting-generative-ai | /academy/md/courses/ai-foundations/prompting-generative-ai.md | | /courses/ai-foundations/custom-gpts-rag | /academy/md/courses/ai-foundations/custom-gpts-rag.md | | /courses/ai-foundations/hallucinations | /academy/md/courses/ai-foundations/hallucinations.md | | /courses/ai-foundations/bias-in-ai | /academy/md/courses/ai-foundations/bias-in-ai.md | | /courses/ai-foundations/ethical-considerations | /academy/md/courses/ai-foundations/ethical-considerations.md | | /courses/ai-foundations/conclusion | /academy/md/courses/ai-foundations/conclusion.md | ## Supplement for indexing ### Content summary This course offers an in-depth introduction to artificial intelligence (AI), examining its evolution, applications, and ethical implications. Beginning with a historical perspective, it traces AI's development from early… This course offers an in-depth introduction to artificial intelligence (AI), examining its evolution, applications, and ethical implications. Beginning with a historical perspective, it traces AI's development from early experiments to today's advanced technologies like ChatGPT and Google Gemini. The curriculum delves into key AI concepts and terminologies such as machine learning, deep learning, and neural networks, with practical examples that illustrate their use in real-world applications. Students will explore both the capabilities and limitations of current AI tools, learning how to appl ### Retrieval tags - ai foundations - ai fundamentals - ai - ai-foundations - Foundations - What - Artificial - Intelligence - Machine - Learning - Frameworks - Deep - Generative - Prompting ### Indexing notes Chunk at each "### Lesson NN — Title" heading; copy lesson_id and topics from the preceding HTML comment into chunk metadata for RAG filters. Course slug: ai-foundations. Union of lesson topic tokens: Foundations, What, Artificial, Intelligence, Machine, Learning, Frameworks, Deep, Generative, Prompting, Custom, GPTs, RAG, Hallucinations, Bias, Ethical, Considerations, Conclusion. Do not embed or retrieve LMS-only quiz items or mastery exam answer keys from this export. ### Asset references | Label | URL | | --- | --- | | Video thumbnail: AI Foundations | `https://cdn.jwplayer.com/v2/media/PUvNP1jb/poster.jpg?width=720` | | Video thumbnail: What is Artificial Intelligence? | `https://cdn.jwplayer.com/v2/media/oyIFhAMe/poster.jpg?width=720` | | Video thumbnail: Machine Learning | `https://cdn.jwplayer.com/v2/media/IneY7znY/poster.jpg?width=720` | | Video thumbnail: Machine Learning Frameworks | `https://cdn.jwplayer.com/v2/media/BFSpBlWR/poster.jpg?width=720` | | Video thumbnail: Deep Learning | `https://cdn.jwplayer.com/v2/media/CSmJF2s9/poster.jpg?width=720` | | Video thumbnail: Generative AI | `https://cdn.jwplayer.com/v2/media/WeSAoc6f/poster.jpg?width=720` | | Video thumbnail: Prompting Generative AI | `https://cdn.jwplayer.com/v2/media/vQamweER/poster.jpg?width=720` | | Video thumbnail: Custom GPTs & RAG | `https://cdn.jwplayer.com/v2/media/QFDJ1wlQ/poster.jpg?width=720` | | Video thumbnail: Hallucinations | `https://cdn.jwplayer.com/v2/media/3Lxm59Zz/poster.jpg?width=720` | | Video thumbnail: Bias in AI | `https://cdn.jwplayer.com/v2/media/zHd4WVia/poster.jpg?width=720` | | Video thumbnail: Ethical Considerations | `https://cdn.jwplayer.com/v2/media/tsX195sP/poster.jpg?width=720` | | Video thumbnail: Conclusion | `https://cdn.jwplayer.com/v2/media/YJZEaUfA/poster.jpg?width=720` | ### External links | Label | URL | | --- | --- | | Contentstack Academy home | `https://www.contentstack.com/academy/` | | Training instance setup | `https://www.contentstack.com/academy/training-instance` | | Academy playground (GitHub) | `https://github.com/contentstack/contentstack-academy-playground` | | Contentstack documentation | `https://www.contentstack.com/docs/` |