Artificial Intelligence Fundamentals PDF

Artificial Intelligence Fundamentals.. Get Started with Artificial Intelligence.. Artificial intelligence (AI) has been a dream of many storytellers and sci-fi fans for years. But for a long time most people hadn’t given AI much serious thought because it was always something that might happen far...

Artificial Intelligence Fundamentals.. Get Started with Artificial Intelligence.. Artificial intelligence (AI) has been a dream of many storytellers and sci-fi fans for years. But for a long time most people hadn’t given AI much serious thought because it was always something that might happen far into the future. Well, researchers and computer scientists haven’t been waiting for tomorrow to arrive, they’ve been working hard to make the dream of AI into a reality. In fact, some have said we’ve already entered the Age of AI. It’s unclear just how deeply AI will become part of our daily lives. But what is certain is that for us to have meaningful conversations about AI, we need a shared vocabulary and a solid foundation of core concepts to build upon. As it stands, if you ask 10 people to define artificial intelligence, you’re likely to get 10 different answers. In this badge we try to reach an agreed-upon definition by exploring AI’s current capabilities. We also investigate how computer scientists create the AI systems that perform such incredible feats. The Difficulty of Defining AI.. The first step in defining AI is to recognize that our current notion of AI might be a little distorted. A steady diet of science fiction books and movies where AI is seen as a nefarious entity bent on conquering the world hasn’t helped. Science fiction isn’t the only thing that’s complicated our view of AI. Generally speaking, we humans tend to think quite highly of ourselves; the benchmark by which everything else is measured. So when we speak of artificial intelligence, we can’t help but to compare it to our own intelligence. The problem is that humans aren’t the only intelligent beings out there. Animals, from crows to octopuses, use tools and problem solving to perform complex tasks. Even slime molds can solve mazes if given enough time. And as we’ve begun to appreciate the huge spectrum of intelligence in the animal kingdom, we’ve also started to recognize the great diversity in our own human intelligence. Maybe you’ve met someone who’s fantastic at public speaking but can’t do math to save their life. Or someone who can always tell when you’re feeling a little anxious, but would trip over a soccer ball at the first opportunity. The point is that our intelligence is expressed in many, specialized forms. We need to think of artificial intelligence in the same way. There are specific kinds of AI that are good at specific kinds of tasks. So let’s bring some definition to what we mean by artificial intelligence by taking a close look at what AI can do today. Main Types of AI Capabilities.. Right now there’s no singular AI that’s good at everything. That idea, known as general AI, is still far into the future. Instead, over the years we’ve developed several specialized AI systems that are designed to perform specific tasks. The kinds of tasks they do generally fall into one of a few broader categories. Numeric Predictions.. Have you looked at a weather forecast recently? Predicting rain or shine helps you decide if you should grab an umbrella. Although we’ve made weather predictions for thousands of years, AI can do it better than any previous method. A good prediction can help you answer all sorts of questions. Is this customer likely to renew their subscription? Are you at risk for a medical condition? Will there be high demand on the power grid this evening? Often AI predictions take the form of a value between 0 (not going to happen) to 1 (totally going to happen). Numeric predictions include more than just percent values, they can predict any numeric value, such as dollars. Maybe your business wants to predict next quarter’s sales, or figure out the optimal pricing for your latest service: Widget+. And as a consumer you’re probably already affected by these kinds of numeric predictions, even more than you realize. Just imagine a trip overseas: the airline tickets, hotel room, ridesharing, and travelers insurance are all likely to be priced by AI to perfectly balance supply and demand. Classifications.. Is a hot dog a sandwich? This question has led to countless hours of friendly philosophical debate about how we categorize things. But in the real world, the stakes can be much higher. Is this plant edible or poisonous? Is that email legitimate or a phishing attempt? Classification is often the first step in taking some kind of action, making it an incredibly valuable skill. So it isn’t surprising that computer scientists have worked hard to create AI that’s good at classifying data. Identifying plants and phishing emails is only the tip of the iceberg. Financial institutions need to flag fraudulent transactions. Medical professionals must diagnose illnesses. Social media platforms want to identify toxic comments. All of these are examples of classification problems. AI can effectively make the first pass at classifying, and then the professionals can take it from there. Often, AI classifiers can do the job just as well, or better, than humans. That said, each classifier is only good at one, narrow task. So the AI that’s great at detecting phishing emails would be lousy at identifying pictures of actual fish. Robotic Navigation.. Some AIs excel at navigating a changing environment, and that might mean actual navigation in the case of autonomous (hands-free) driving. AI-powered cars are already quite capable of keeping centered in a lane and following at a safe distance on the highway. They adapt to curves in the road, gusts of wind from semi trucks, and sudden stops due to traffic. AI that can adapt to changing environmental conditions have all sorts of real-world applications. For example, businesses need to produce and deliver products to their customers every day. Lots of market conditions play a role in how quickly that gets done: materials availability, manufacturing capacity, existing inventory, transportation costs, even real-time traffic. AI can optimize the supply chain even while conditions are changing. And let’s not forget robots! Even the modest robot floor sweeper can avoid stairs and chairs. On a bigger scale, assembly lines are being fitted with robots that become faster and more efficient over time. Those same robots can adjust for changes to the production method without costly reprogramming. And researchers are creating rescue robots that can traverse disaster areas, such as a collapsed building. A robot-caterpillar that can squeeze through tiny cracks could deliver aid and hope to those trapped inside. Language Processing.. On November 30, 2022, Merriam-Webster’s word of the day was quiddity. Those who learned that word got a little better at what might be the most important skill of all: communication. On that same day, the world was introduced to ChatGPT, an artificial intelligence that demonstrated its own communication skills. It could write long responses to questions about almost any topic. And the responses seemed like they were written by a human. ChatGPT is one of the most capable AIs built to interpret everyday language and act on it in some meaningful way. This is known in the industry as natural language processing, or just NLP. Natural language processing relies on an understanding of how words are used together, and that lets AI extract the intention behind the words. For example, you might want to translate a document from English to German. Or maybe you want a short summary of a long, scientific paper. AI can do that too. NLP is a huge part of generative AI, a subcategory of AI that takes words and turns them into unique images, sounds, and of course other words. Generative AI is such a disruptive technology that we’ve devoted a whole badge to Generative AI Basics. Check it out when you’re done here. In Summary.. Artificial intelligence can be thought of as the ability for a computer to perform skills typically associated with human intuition, inference, and reasoning. At this time, AI skills are very specialized, and fall into some broad categories like numeric predictions and language processing. Now that you have a sense of what AI is (and isn’t), you’re ready to explore how computer scientists and researchers create AI. Turn Data into Models.. What AI can do may seem like magic. And like magic, it’s natural to want a peek behind the curtain to see how it’s all done. What you’ll find is that computer scientists and researchers are using lots of data, math, and processing power in place of mirrors and misdirection. Learning how AI actually works will help you use it to its fullest potential, while avoiding pitfalls due to its limitations. The Shift from Crafting to Training.. For decades, programmers have written code that takes an input, processes it using a set of rules, and returns an output. For example, here’s how to find the average from a set of numbers. Input: 5, 8, 2, 9. Process: Add the values [5 + 8 + 2 + 9] then divide by the number of inputs. Output: 6. This simple set of rules for turning an input into an output is an example of an algorithm. Algorithms have been written to perform some pretty sophisticated tasks. But some tasks have so many rules (and exceptions) that it’s impossible to capture them all in a hand-crafted algorithm. Swimming is a good example of a task that is hard to encapsulate as a set of rules. You might get some advice before jumping in the pool, but you only really figure out what works once you’re trying to keep your head above water. Some things are learned best by experience. What if we could train a computer in the same way? Not by tossing it into a pool, but by letting it figure out what works to succeed at a task? But just like learning to swim is very different from learning to speak a foreign language, the kind of training depends on the task. Let’s check out a few of the ways AI is trained. Experience Required. Imagine that every time you went to the store to pick up milk, you tracked details of the trip in a spreadsheet. It’s a little weird, but go with it. You set up the following columns. 1. Is it the weekend? 2. Time of day 3. Is it raining or not? 4. Distance to store 5. Total minutes of trip. After several trips you start getting a feel for how conditions affect how long it’ll take. Like, rain makes the drive longer, but it also means fewer people are shopping. Your brain makes connections between the inputs (weekend [W], time [T], raining [R], distance [D]) and the output (minutes [M]).. But how can we get a computer to notice trends in the data so it can estimate too? One way is the guess-and-check method. Here’s how you do it. Step 1: Assign all of your inputs a “weight.” This is a number that represents how strongly an input should affect the output. It’s OK to start with the same weight for everything. Step 2: Use the weights with your existing data (and some clever math we won’t get into here) to estimate the minutes for a milk run. We can compare the estimate to the historic data. It’ll be way off, but that’s OK. Step 3: Let the computer guess a new weight for each input, making some a little more important than others. For example, the time of day might be more important than whether or not it’s raining. Step 4: Rerun the calculations to check if the new weights result in a better estimate. If so, it means the weights are a better fit, and changing in the right direction. Step 5: Repeat steps 3 and 4, letting the computer tweak weights until its estimates aren’t getting any better. At this point the computer has settled on weights for each input. If you think of weight as how strongly an input is connected to the output, you can make a diagram that uses line-thickness to represent the weight of a connection. For this example it looks like the time of day has the strongest connection, but apparently rain doesn’t make much of a difference. This process of guess-and-check has created a model of our milk runs. And like a model boat, we can take it to the pool to see if it floats, so to speak. That means testing it in the real world. So for your next several milk runs, before you leave, have the model estimate how long it’ll take. If it’s right enough times in a row, you can confidently let it do the estimating for every future trip. Use the Right Data for the Right Job.. This is a very simple example of using training to make an AI model, but it touches on some important ideas. First, it’s an example of machine learning (ML), which is the process of using large amounts of data to train a model to make predictions, instead of handcrafting an algorithm. Second, not all data is the same. In our milk run example, the spreadsheet is what we would call structured data. It is well organized, with labels on every column so you know the significance of every cell. In contrast, unstructured data would be something like a news article, or an unlabeled image file. The kind of data that you have available will affect what kind of training you can do. Third, the structured data from our spreadsheet lets computers do supervised learning. It’s considered supervised because we can make sure every piece of input data has a matching, expected output that we can verify. Conversely, unstructured data is used for unsupervised learning, which is when AI tries to find connections in the data without really knowing what it’s looking for. Letting the computer figure out a single weight for each input is just one kind of training regimen. But often interconnected systems are more complicated than what 1-to-1 weighting can represent. Thankfully, as you learn in the next unit, there are other ways to train! Understand the Need for Neural Networks.. The Need for Neural Networks.. No conversation about AI is complete without mentioning neural networks. Neural networks are important tools for training AI models, so it’s good to have some idea of what they are. But before we get into the details, let’s first discuss why we need neural networks in the first place. In the previous unit, you learned that we can train an AI model by letting it guess-and-check the importance-weight of each input. But the milk run example was actually overly simplified. Our model would give us pretty rough estimates. To understand why, let’s consider two scenarios. 1. It’s raining on a Tuesday evening. You’d rather not get wet, so you (and many others like you) decide shopping can wait until tomorrow. In this scenario, rain is a significant factor. 2. It’s raining on Saturday afternoon. For many people, this is the only time of week when they can go shopping. So the store will be busy, rain or shine. In this scenario, rain doesn’t make much difference. The problem is that our original model can only assign one weight to rain, but we know it’s more complicated than that. There is a solution, though, and it starts by representing the two scenarios in two separate graphs. Again, line thickness shows importance. In the first, “weekend” and “time” are weak, while “rain” is strong. For the second, “weekend” is strong, while “time” and “rain” is weak. We know these two scenarios are significant because we’re smart and have experience buying milk. But a computer just starting to learn about milk runs doesn’t know anything yet! It has to consider many scenarios: weekend-evening-rain, weekday-morning-shine, and so forth. Instead of two graphs, eight might better represent the kinds of scenarios you encounter. That’s a lot of very similar graphs. Since the three inputs always represent “weekend,” “time,” and “rain,” you can overlap them. If you move the outputs so they aren’t touching, you get a combined graph that looks like this. The importance of each scenario depends on the specific inputs. But knowing the importance is only half the battle. Each scenario needs to affect the final estimate in its own way. For example, the weekend-afternoon-shine milk run should take much longer. So let’s give it an adjustment of +5. When we do the math to calculate an estimate, it results in a larger number. While we’re at it, let’s give the weekday-morning-rain scenario an adjustment of -4 since we know milk runs are shortest at that time. Each scenario gets its own adjustment, which is what we call a bias. In this case, bias is a good thing because it helps us get a more accurate estimate. Let’s redraw our graph to include the bias of each scenario. So what do we do with these eight scenarios and their biases? Using some more clever math, we can combine them into a final estimate. Some scenarios should contribute more than others, so you guessed it, we need more weights! We can update our graph to show how the scenarios connect to the final estimate with different strengths. This is our new model. More connections will hopefully mean better estimates. This web of connections, guided by weights and biases, is an example of a neural network. We call it that because the connections, forged by experience (data), resemble how the neurons in a brain are connected. And while scenario is a good beginner word for describing a unique combination of factors, we should really use the word node for that concept. It’s what AI experts use, so moving forward we’ll use it too. Adding Complexity to Neural Networks.. Our new milk run model is a pretty basic example of a neural network. In practice, they can get quite complex. Let’s explore some of the ways researchers set up neural networks to get better results for specific tasks. First, you might be wondering why we chose eight nodes to stand between our inputs and output. There’s actually some flexibility in that number. We know that having no nodes at all will give us rough estimates. In the same way, having too few might not capture all of the nuance of the system we’re trying to model. But having too many nodes is a problem, too. We don’t want to make the computer do more calculations than necessary. So there’s a sweet spot for the number of nodes where we get good results for the least effort. Choosing the right number is part of designing a good neural network. There’s something else we can do to make artificial neural networks more like our own, organic ones. It has to do with how our minds often leap from idea to idea to find connections between two things that are not obviously related. Some of the most brilliant insights are the result of several leaps. So, what if we could make a neural network that could make more leaps, too? We can! We do it by adding more nodes as layers, connecting each node to its neighbor. Training AI by adding extra layers to find hidden meaning in data is what’s called deep learning. Thanks to an abundance of computing power, many neural networks are designed to have multiple layers. Again, the best number of layers is a balance between the number of calculations required and the quality of results they produce. More Than Mental Math, It’s Neural Network Math.. So, about those calculations. Up to this point we’ve glossed over the math part of training neural networks. That’s for a few reasons. First, the math can get really complicated, really fast. For example, here’s a snippet of a research paper about neural networks. Yeah, it’s intense! Second, the exact math is going to depend on what kind of task you’re training the neural network to do. Third, each new research paper updates the math as we learn what works better at training different models. So, designing a neural network involves choosing the number of nodes, layers, and the appropriate math for the task it’s training for. With the model architecture ready, you have to let the computer use all that fancy math to do its guess-and-check routine. Eventually it’ll figure out the best weights and biases to give good estimates. And this brings us to something that’s a little unsettling about artificial neural networks. Imagine a skilled talent scout who’s looking for the next great baseball player. “I’ll know them when I see them,” they might say. They can’t explain how they’ll know, they just will. In the same way, our neural network can’t explain why certain factors are important. Sure, we can look at the values attached to each weight and bias, but the relevance of a number that’s the result of a connection of a connection of a connection will be lost to us. So just like how the mind of the talent scout is a black box, so is our neural network. Because we don’t observe the layers between the input and output, they’re referred to as hidden layers.. To summarize. Neural networks are a mix of nodes, layers, weights, biases, and a bunch of math. Together they mimic our own organic neural networks. Each neural network is carefully tuned for a specific task. Maybe it’s great at predicting rain, maybe it categorizes plants, or maybe it keeps your car centered in the lane on the highway. Whatever the task, neural networks are a big part of what makes AI seem magical. And now you know a little bit about how the trick is done. Generative AI Basics.. Explore the Capabilities of Generative AI.. You may have noticed a lot of discussion about artificial intelligence (AI) as of late—it’s almost overwhelming. But why the huge spike in interest? AI isn’t exactly new; a lot of businesses and institutions have used AI in some capacity for years. The sudden attention to AI was arguably caused by something called ChatGPT, an AI-powered chatbot that can do what no others could. ChatGPT can respond to plain-language questions or requests, and those responses seem like they were written by a human. And because it was released to the public, people could experience firsthand what it was like to have a conversation with a computer. It was surprising. It was eerie. It was evocative. So of course people started paying attention! An AI that can hold a natural, human-like conversation is clearly different from what we’ve seen in the past. As you learn in the Artificial Intelligence Fundamentals badge, there are a lot of specific tasks that AI models are trained to perform. For example, an AI model can be trained to use market data to predict the optimal selling price for a three-bedroom home. That’s impressive, but that model produces “just” a number. In contrast, some AI models can produce an incredible variety of text, images, and sounds that we’ve never read, seen, or heard before. This kind of AI is known as generative AI. It holds within it massive potential for change, both in and out of the workplace. In this badge you learn what kinds of tasks generative AI models are trained to do, and some of the technology behind the training. This badge also explores how businesses are coalescing around specialties in the generative AI ecosystem. Finally, we end by discussing some of the concerns that businesses have about generative AI. Possibilities of Language Models.. Generative AI might seem like this hot new thing, but in reality researchers have been training generative AI models for decades. Some have even made the news in the past few years. Maybe you remember articles from 2018 when a company named Nvidia unveiled an AI model that could produce random photorealistic images of human faces. The pictures were surprisingly convincing. And although they weren’t perfect, they were definitely a conversation-starter. Generative AI was slowly beginning to enter the public consciousness. As researchers worked on AI that could make specific kinds of images, others were focused on AI related to language. They were training AI models to perform all sorts of tasks that involved interpreting text. For example, you might want to categorize reviews of one of your products as positive, negative, or neutral. That’s a task that requires an understanding of how words are combined in everyday use, and it’s a great example of what experts call natural language processing (NLP). Because there are so many ways to “process” language, NLP describes a broad category of AI. (For more on NLP, see Natural Language Processing Basics.) Some AIs that perform NLP are trained on huge amounts of data, which in this case means samples of text written by real people. The internet, with its billion web pages, is a great source of sample data. Because these AI models are trained on such massive amounts of data, they’re known as large language models (LLMs). LLMs capture, in incredible detail, the language rules humans take years to learn. These large language models make it possible to do some incredibly advanced language-related tasks. Summarization. If you’re given a sentence and you understand how all the words come together to make a point, you can probably rewrite the sentence to express the same idea. Since AI models know the rules for syntax, and they’ve learned which words can be swapped for others, they can make a remix too. Taking a whole paragraph and condensing into one or two sentences is just another kind of remix. This kind of AI-assisted summarization can be very helpful in the real world. It can create meeting notes from an hour-long recording. Or write an abstract of a scientific paper. It’s the ultimate elevator-pitch generator. Translation. LLMs are like a collection of rules for how a language structures words into ideas. Each language has its own rules. In English, we typically put adjectives before nouns, but in French it’s usually the other way around. AI translators are trained to learn both sets of rules. So when it’s time for a sentence remix, AI can use a second set of rules to express the same idea. Voilà, you have yourself a great translation. And programming languages are languages too. They have their own set of rules, so AI can translate a loose set of instructions into actual code. A personal pocket programmer can open a lot of doors for a lot of people. Error correction. Even the most experienced writers make the occasional grammatical or spelling mistake. Now, AI will detect (and sometimes auto-correct) anything amiss. Also, patching up errors is important when simply listening to someone speak. You might miss a word or two because you’re in a noisy environment, but you use context to fill in the gap. AI can do this too, making speech-to-text tasks like closed captioning even more accurate. Question answering. This is the task that launched generative AI into the limelight. AIs such as ChatGPT are capable of interpreting the intention of a question or request. Then, it can generate a large amount of text based on the request. For example, you could ask it for a one-sentence summary of the three most popular works of William Shakespeare, and you’d get: "Romeo and Juliet" - A tragic tale of two young lovers from feuding families whose love ultimately leads to their untimely deaths. "Hamlet" - The story of a prince haunted by his father’s ghost, grappling with revenge and the existential questions of life and death. "Macbeth" - A chilling drama of ambition and moral decline as a nobleman, driven by his wife’s ambition, succumbs to a bloody path of murder to seize the throne. Then, you could continue the conversation by asking for more information about Hamlet, as though you were talking with your Language Arts teacher. This kind of interaction is a great example of getting just-in-time information with a simple request. Guided image generation. LLMs can be used in tandem with image generation models so that you can describe the image you want, and AI will attempt to make it for you. Here’s an example of asking for “a 2D line art drawing of Juliet standing in the window of an old castle.” Because there are so many descriptions and images of Romeo & Juliet on the internet, the AI generator didn’t need any further information to make a guess at an appropriate image. Related to guided image generation, some AI models can add new content into existing images. For example, you could extend the borders of a picture, allowing the AI to draw in what is likely to appear based on the context of the original picture. Text-to-speech. Similar to how AI can convert a string of words into a picture, there are AI models that can convert text to speech. Some models can analyze audio samples of a person speaking. It learns that person’s unique speech patterns, and can reproduce them when converting text to new audio. To the casual listener, it’s hard to tell the difference. These are just a few examples of how LLMs are used to create new text, images, and sounds. Almost any task that relies on an understanding of how language works can be augmented by an AI. It’s an incredibly powerful tool that you can use for both work and play. Impressive Predictions.. Now that you have an idea of what generative AI is capable of, it’s important to make something very clear. The text that a generative AI generates is really just another form of prediction. But instead of predicting the value of a home, it predicts a sequence of words that are likely to have meaning and relevance to the reader. The predictions are impressive, to be sure, but they are not a sign that the computer is “thinking.” It doesn’t have an opinion about the topic you ask about, nor does it have intentions or desires of its own. If it ever sounds like it has an opinion, that’s because it’s making the best prediction of what you expect as a response. For example, asking someone “Do you prefer coffee or tea?” elicits a certain kind of expected response. A well-trained model can predict a response, even if it doesn’t make sense for a computer to want any kind of drink. Understand the Technology Ecosystem of Generative AI.. Supercharging Generative AI Training. Generative AI has gained a lot of capabilities in what seems like a very short amount of time. The incredibly fast pace of improvement is largely due to three big factors. The first is the availability of huge amounts of training data. As mentioned in the previous unit, the more than a billion web pages on the internet are a great source of writing samples. But data is only good if you have a way to use it. That’s where the second big change comes in: better training. As you learn in Artificial Intelligence Fundamentals, researchers design neural networks that use sophisticated math to train AI models. The architecture of neural networks is a field of study that’s constantly progressing. In 2017, researchers at Google published a game-changing paper about training large language models. They proposed a new AI architecture called a transformer. As you can imagine, the details of the research are pretty complicated. But to simplify (greatly), the new architecture was capable of identifying important relationships between words, no matter how far away they appear within a block of text. It could retain that connection even after processing lots of words. The new transformer architecture brings us to the third major factor in the rapid advancement of generative AI: computational power. It takes a lot of processing power to do the math behind AI model training. Historically, AI models are designed in a way that requires a sequence of calculations, run one after the other. The transformer architecture is different—it relies on many separate, concurrent calculations. So, one computer processor can do the first calculation while a different processor does the second at the same time. That’s called parallel computing, and it greatly reduces the time it takes to train a transformer. On top of that, in recent years processors that can perform parallel computing have become much more powerful and abundant. These three factors of data, architecture, and computing have converged for just the right conditions to train very capable large language models. One of the biggest LLMs is the GPT language model, which stands for generative pre-trained transformer. In other words, a model that’s already been trained that can be used to generate text-related content. Emerging Ecosystem.. Right now, there are already hundreds of sites on the internet where you can go to get hands-on with generative AI. When you visit one of those sites, you’re at the tip of a technology iceberg. And that technology can come from a lot of different sources. Let’s investigate the tech stack that makes it possible to bring awesome generative AI experiences to the masses. At the bottom of the iceberg, we start with the compute hardware providers. Training an LLM can take a staggering amount of computational power, even if you're training a transformer. It also takes computing power to process requests to actually use the model after it’s been trained. Technically you can train AI models on any computing hardware, but processors that excel at parallel computing are ideal. Today the biggest name in AI compute is Nvidia. Next are the cloud platforms that allow developers to tap into the compute hardware in a cloud deployment model. Devs can rent the appropriate amount of time for a specific project, and the platforms can efficiently distribute requests for computing time across a connected system. Google, Amazon, Microsoft, and Oracle are the main tech providers in this space. AI models, including LLMs are the next layer. These models are carefully crafted using research techniques and trained using a combination of public and privately curated data. Developers can connect to LLMs through an application programming interface (API), so they can harness the full power of NLP in their own applications. The trained and accessible AI model is commonly referred to as a foundational model. Because these models are accessed through an API, developers can easily switch from one foundational model to another as needed. A few examples of foundational models are GPT4, Claude, Stable Diffusion, and LLaMA. The next layer is infrastructure optimization, which is all about providing tools and services that make for more efficient and higher-quality model training. For example, a service might offer perfectly curated data sets to train on. Another might provide analytics to test the accuracy of generated content. It’s also at this point where foundational models can be fine-tuned with specialized, proprietary data to better meet the needs of a particular company. This is a busy space in the AI ecosystem, with many companies offering a variety of optimization services. Finally, we find ourselves back at the tip of the iceberg: the applications. Developers of all kinds can tap into optimization services and foundational models for their apps. Already we see LLM-powered standalone tools, as well as plugins for mainstream applications. This thriving ecosystem of technology companies has grown at an incredible rate just over the past few years. Some companies will specialize in one particular segment. For example, one in the foundational model space may want to focus on training new and better performing models to differentiate themselves. Other companies will create solutions that span multiple layers of the tech stack, creating their own proprietary LLM to use for their application. Many businesses are just starting to get a handle on what AI can do for them. Given the unprecedented demand for AI technology, there’s a huge amount of opportunity for businesses to make their mark at several levels of the AI tech stack. Common Concerns About Generative AI.. Generative AI is going to lead to many changes in how we interact with computers. With any disruptive technology, it’s important to understand its limitations and causes for concern. Here are a few of the main concerns with generative AI. Hallucinations. Remember that generative AI is really another form of prediction, and sometimes predictions are wrong. Predictions from generative AI that diverge from an expected response, grounded in facts, are known as hallucinations. They happen for a few reasons, like if the training data was incomplete or biased, or if the model was not designed well. So with any AI generated text, take the time to verify the content is factually correct. Data security. Businesses can share proprietary data at two points in the generative AI lifecycle. First, when fine-tuning a foundational model. Second, when actually using the model to process a request with sensitive data. Companies that offer AI services must demonstrate that trust is paramount and that data will always be protected. Plagiarism. LLMs and AI models for image generation are typically trained on publicly available data. There’s the possibility that the model will learn a style and replicate that style. Businesses developing foundational models must take steps to add variation into the generated content. Also, they may need to curate the training data to remove samples at the request of content creators. User spoofing. It’s easier than ever to create a believable online profile, complete with an AI generated picture. Fake users like this can interact with real users (and other fake users), in a very realistic way. That makes it hard for businesses to identify bot networks that promote their own bot content. Sustainability. The computing power required to train AI models is immense, and the processors doing the math require a lot of actual power to run. As models get bigger, so do their carbon footprints. Fortunately, once a model is trained it takes relatively little power to process requests. And, renewable energy is expanding almost as fast as AI adoption! In Summary Generative AI is capable of assisting businesses and individuals alike with all sorts of language-based tasks. The convergence of lots of data, clever AI architecture, and huge amounts of computing power has supercharged generative AI development and the growth of the AI ecosystem.