How Does AI Actually Work? 🤖 | AI, Machine Learning & Generative AI Explained
Publicado no IA Dots ·
AGIAI SignalANIASIidioma-enInglêsyt-LOtWiKQ6710
Vídeo em Inglês, publicado pelo canal UpskillNexus.
Video transcript
Transcript from available captions; it may contain errors.
Hi everyone. Welcome to Upskill Nexus channel, and today we're going to understand the foundations of AI. So, we're going to understand what exactly is AI, how does AI function, what are the different subtypes of AI, how is AI classified, what are the technical aspects, and as a practitioner, uh what are the applications of AI? Just a quick check, this is part of our AI foundation series where we're trying to unpack and make AI easier and simpler for people to understand. So, don't forget to subscribe to the channel, like this video so that you keep getting our videos in the suggestions, and start getting more and more information about the AI ecosystem. Now, let's get on to it. Okay, let's start with understanding the definitions of what exactly is artificial intelligence. Now, if you think about artificial intelligence, right? In terms of the core definition, it basically revolves around the science of making intelligent machines. It tries to simulate human cognitive functions. So, for example, these are systems that can perceive, learn, or act how the humans performs particular functions. So, a simple example could be how language is used to communicate between two people. If you talk to, let's say, your Alexa or to your Siri, they basically respond to you in the same manner that a human would actually respond to you. Now, if you look at artificial intelligence, right? You can break the word into two parts. So, you've got artificial and then you've got intelligence. Artificial essentially refers to human-made, so something that's not natural. Intelligence, on the other hand, here refers to pattern prediction and not human understanding. Human understanding is a process that's extremely complex, because first we take note of the senses around us, then we perceive the world through these senses, and develop some kind of models about the world in our head. Still human understanding is something that psychologists are trying to unpack better, which is why whenever we talk about AI, we talk about intelligence in terms of pattern prediction, which is the more technical aspect, right? Also a quick note here, AI often lacks biological consciousness. We have not reached that part. We will cover this when I talk about the different intelligence levels that AI actually understands or how researchers right now are defining the AI intelligence levels. If you think about AI currently, right? So in 2026, AI can be narrowly I would say by trifurcated into three different parts. So you've got artificial narrow intelligence, you've got artificial general intelligence, and then you've got artificial super intelligence. And all of them kind of uh are stacked like a pyramid with artificial super intelligence being on top, the second level being artificial general intelligence, and the lowest level being artificial narrow intelligence. Artificial narrow intelligence is basically intelligence that is specialized for specific tasks. So for example, if I were to open my phone, I would use Face ID to open my phone. If I were to create a system that sorts out my emails and detects spams, it's for a specific task, that is spam detection. These kind of systems are already prevalent today. That is what we use with um AI chatbots like ChatGPT, that's how we work with Claude, Claude code. All of these are artificial narrow intelligence products. Then the second rung with artificial general intelligence. These are basically human-level versatile intelligence where um they are theoretical ideas and this is something that most big AI companies like Anthropic, Open AI, Google DeepMind, all of them are investing money to reach to this future goal, which is to reach the level of human-level versatile intelligence. Humans can perceive the world and can create ideas about the world, which are very different from how, let's say, machines perceive the world, right? Machines are only trained to recognize patterns as of now. But when you add capabilities like reasoning, planning, autonomous learning, all of these are capabilities right now that are only present inside a human, but not inside any kind of artificial intelligence system. So this is something that most conglom- conglomerates are working towards, and this is something that also companies want to reach first because the person who's going to reach that level first will always have an advantage over others. And the third level is artificial superintelligence. So this is the point where after artificial general intelligence, where both of them, that is, humans and AI is at the same level, when AI surpasses human intellect, right? This is again theoretical and only hypothetical, something that was mentioned in the AI 2027 paper, something that's not uh you know, that understood as of now. But basically, when we talk about emotional intelligence, when we talk about reasoning, when we talk about creativity, all of these aspects of a human that makes a human a human will be incorporated in this level of artificial superintelligence. Now, these are the three levels. So artificial narrow intelligence, artificial general intelligence, and artificial superintelligence. Now, if you come to the technical side of it, right? In terms of the technical aspects, you can think about AI on three different fronts. One is machine learning, second is deep learning or neural networks, and the third is generative AI. What is machine learning? Machine learning basically means learning from the data patterns. So, they act as the backbone for any kind of AI system. Machine learning is also is also divided into three different parts. So, supervised learning, unsupervised learning, and reinforcement learning. Now, supervised learning basically means you have labeled data sets. So, you have a certain input, and then you have a certain output. And each input is mapped to a particular output. This kind of classification is used maybe to understand whenever you're scanning with your phone, you're looking at an image, and you classify a dog is a dog and a cat is a cat. Same way, when you classify something, let's say, a fiction book, non-fiction book, something that is a spam email and not a spam email. All of these classification methods that are learned under supervision is known as a supervised learning model. Now, supervised learning model also can be regressions. So, this can be predictions of stock prices, prediction of house prices in the future, whether you're trading, those kind of predictions also come under supervised learning. The second aspect or the second kind of machine learning is unsupervised learning. So, here there are no labeled data patterns. Instead, there are clusters. So, it can be categorizations. So, for example, there could be a book that could be a romantic fiction. So, it is a cluster or a grouping where there is not a one-to-one mapping, but there is a mapping of a cluster, right? So, news categorization, customer grouping. And the second kind could be association. So, for example, a market basket analysis. So, bread and milk, right? These are the kind of unsupervised learning models that exist. The third kind is the reinforcement learning. So, here we look at agent-driven interactions, right? So, it's a feedback loop. Every time you do something good, you get a reward. Every time you do something bad, you get a penalty. So, it acts as a feedback loop where the agent is interacting with the environment. So, something like a self-driving car knowing not to bang into a pole because the self-driving car knows for the fact that there's a pole ahead, and whenever you bang into a pole, you cause some kind of collision impact on the car. Just uh another example of this could be chess engines, right? So, all of these are again aspects of different machine learning pathways. Then you've got deep learning and neural networks. Here, what it basically means uh or how it's defined is brain-inspired multi-layer models. So, what happens is in the deep learning and neural network segment, all the models are multi-layered, so there's no single layer, so there are multiple layers to it, and all of them are inspired with how the architecture is in our brain. So, in terms of architecture, there's an input layer which basically receives all the data. Then there are some hidden layers that that extract the hierarchical features, and then there is an output layer that delivers the final prediction. The components in this are known as nodes and synapses, the same way it works in our head. So, different neurons in our head, they're connected by synapses. Same way here, each and every input layer and hidden layer has different nodes or neurons that are connected to different synapses or the links between them. In terms of the training mechanism, uh deep learning networks and neural networks, they work on weights. So, weights are basically how important a connection is between two nodes. On forward propagation, which is basically making a guess, so kind of like a prediction. Back propagation, which refers to error correction or learning from the mistakes that you're making, and activation functions, which are basically mathematical functions that let you understand what's the threshold for a signal to be activated. So, how do I get from the hidden layer all the way up to the output layer? Or how do I start from the input layer and have some kind of threshold-based signals that lets the activation occur from the baseline all the way up to, let's say, the output. All right? And lastly, for the different neural networks, there are certain subtypes, right? There are CNNs, so these are convoluted neural networks, which basically work for image processing and vision. Then there is RNN, which works on sequential data and time series. And lastly, there are transformers, which is the GPT, so the in GPT, the T stands for transformers, where modern LLMs and um language models, so large language models, use these transformers to basically interact with people through different languages, right? And translation is also involved at this stage. Next, lastly, the segment related to AI in this AI Foundations video that we want to cover today is generative AI. Now, what is generative AI, right? Generative AI simply is models that generate new kind of content. So, this could be examples such as text. So, this um could be your chat GPT, Gemini, Claude doing that. It could be images, so things like Midjourney or Nano Banana from Google, they're generating images or child GPT image and they're generating images. Then you've got video. So like you've got the Google video 3 or you've got the Google Flow which does video creation, Higgs Field. All of these tools that are generating videos from text. So basically what it means is that these are AI models that generate new content from existing content. And the mechanisms they work using are large language models or diffusion. This is basically all about how the technical aspects of AI are divided. Now, let's just quickly look at the applications of AI. So AI can be applied in three different ways, right? So you've got everyday usage, you've got professional domains and then you've got advanced fields. Everyday usage basically means something that you're using everyday. So your Google Maps. Your Google Maps is using data to predict how much traffic is expected on a particular road, right? It sometimes happens that when you take a particular road, you see that it's completely red on Google Maps, but the road is actually completely free. So it's supposed to be marked blue, but because of past patterns, it is marked as yellow or as red. In terms of streaming, if you look at YouTube or Netflix recommendations, that's also an application that's here. On social media, whenever you interact with a certain kind of content, you get recommended that content more. So algorithmic feeds is also an application of AI. In terms of professional domains, you can think about health care. So diagnostic imaging or drug discovery. Um AlphaFold is a very famous project that's currently being run by Google, which uh is related to genetic understanding of the human body. So health care has a lot of applications when it comes to AI. Then there's finance, so whether it's fraud detection, trading, um or it's predicting which stock is going to go up, which stock is going to go down. Those kind of financial predictions can be done to an extent using AI models. Then you've got commerce, so you look at supply chain management, you look at inventory optimization. And lastly, you've got industry, so you look at predictive maintenance or automation. So for example, if you are sending a let's say request to on our website, how is it going directly into my CRM or my customer relationship manager. In certain advanced fields, there are also applications of let's say artificial intelligence such as robotics. So we look at autonomous movement and sensing. Then we look at conservation in terms of wildlife and forest monitoring. So whenever you want to keep track of multiple animals being dispersed in different areas, you use this kind of detection where you look at and monitor using AI systems based on the pug marks of a tiger or the let's say feet mark of an elephant of what kind of path they took, what is the path they can take. And that kind of predictive analysis in conservation letting environmentalists decide what area needs to be closed off, what area needs to be let out open is dependent on AI systems now. Then lastly, cybersecurity, something that we also teach her at UpGrad Nexus where we look at anomaly and threat detection in different logs that are available. For us, cybersecurity is also a double-edged sword because here you can use AI to detect, at the same time you can use AI to attack. So with cybersecurity, you have to be extremely careful. But overall, if you look at this, this is the way that AI is currently functioning. And if you want a quick road map of how you can skill yourself in AI. Lastly, before I close for today, don't forget to subscribe to our channel if you want to stay updated with the AI and automation updates. And if you're interested to explore more courses, we at UpSkill Nation currently offer courses in AI and automation. One as a self- paced course, that's a complete stack, so a full-stack AI and automation course. Second, we have live hybrid courses. So, these are courses where the classes are taken on the weekend for working professionals. And third, we offer specialized industry-based certifications powered by AI search, where based on your industry, you can do self-paced courses, learn about the AI landscape, and keep upskilling yourself. Thank you so much for your time, and I will see you in the next one.
Assistir ao vídeo original no YouTube
Assuntos: AGI, ANI, ASI. Publicado no YouTube em 2026-09-07.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.