Core AI Concepts – Part 3
Join hosts Lois Houston and Nikita Abraham, along with Principal AI/ML Instructor Himanshu Raj, as they discuss the transformative world of Generative AI. Together, they uncover the ways in which generative AI agents are changing the way we interact with technology, automating tasks and delivering new possibilities. AI for You: https://mylearn.oracle.com/ou/course/ai-for-you/152601/252500 Oracle University Learning Community: https://education.oracle.com/ou-community LinkedIn: https://www.linkedin.com/showcase/oracle-university/ X: https://x.com/Oracle_Edu Special thanks to Arijit Ghosh, David Wright, Kris-Ann Nansen, Radhika Banka, and the OU Studio Team for helping us create this episode. ------------------------------------------------------- Episode Transcript: 00:00 Welcome to the Oracle University Podcast, the first stop on your cloud journey. During this series of informative podcasts, we’ll bring you foundational training on the most popular Oracle technologies. Let’s get started! 00:25 Lois: Welcome to the Oracle University Podcast! I’m Lois Houston, Director of Innovation Programs with Oracle University, and with me is Nikita Abraham, Team Lead of Editorial Services. Nikita: Hi everyone! Last week was Part 2 of our conversation on core AI concepts, where we went over the basics of data science. In Part 3 today, we’ll look at generative AI and gen AI agents in detail. To help us with that, we have Himanshu Raj, Principal AI/ML Instructor. Hi Himanshu, what’s the difference between traditional AI and generative AI? 01:01 Himanshu: So until now, when we talked about artificial intelligence, we usually meant models that could analyze information and make decisions based on it, like a judge who looks at evidence and gives a verdict. And that's what we call traditional AI that's focused on analysis, classification, and prediction. But with generative AI, something remarkable happens. Generative AI does not just evaluate. It creates. It's more like a storyteller who uses knowledge from the past to imagine and build something brand new. For example, instead of just detecting if an email is spam, generative AI could write an entirely new email for you. Another example, traditional AI might predict what a photo contains. Generative AI, on the other hand, creates a brand-new photo based on description. Generative AI refers to artificial intelligence models that can create entirely new content, such as text, images, music, code, or video that resembles human-made work. Instead of simple analyzing or predicting, generative AI produces something original that resembles what a human might create. 02:16 Lois: How did traditional AI progress to the generative AI we know today? Himanshu: First, we will look at small supervised learning. So in early days, AI models were trained on small labeled data sets. For example, we could train a model with a few thousand emails labeled spam or not spam. The model would learn simple decision boundaries. If email contains, "congratulations," it might be spam. This was efficient for a straightforward task, but it struggled with anything more complex. Then, comes the large supervised learning. As the internet exploded, massive data sets became available, so millions of images, billions of text snippets, and models got better because they had much more data and stronger compute power and thanks to advances, like GPUs, and cloud computing, for example, training a model on millions of product reviews to predict customer sentiment, positive or negative, or to classify thousands of images in cars, dogs, planes, etc. Models became more sophisticated, capturing deeper patterns rather than simple rules. And then, generative AI came into the picture, and we eventually reached a point where instead of just classifying or predicting, models could generate entirely new content. Generative AI models like ChatGPT or GitHub Copilot are trained on enormous data sets, not to simply answer a yes or no, but to create outputs that look and feel like human made. Instead of judging the spam or sentiment, now the model can write an article, compose a song, or paint a picture, or generate new software code. 03:55 Nikita: Himanshu, what motivated this sort of progression? Himanshu: Because of the three reasons. First one, data, we had way more of it thanks to the internet, smartphones, and social media. Second is compute. Graphics cards, GPUs, parallel computing, and cloud systems made it cheap and fast to train giant models. And third, and most important is ambition. Humans always wanted machines not just to judge existing data, but to create new knowledge, art, and ideas. 04:25 Lois: So, what’s happening behind the scenes? How is gen AI making these things happen? Himanshu: Generative AI is about creating entirely new things across different domains. On one side, we have large language models or LLMs. They are masters of generating text conversations, stories, emails, and even code. And on the other side, we have diffusion models. They are the creative artists of AI, turning text prompts into detailed images, paintings, or even videos. And these two together are like two different specialists. The LLM acts like a brain that understands and talks, and the diffusion model acts like an artist that paints based on the instructions. And when we connect these spaces together, we create something called multimodal AI, systems that can take in text and produce images, audio, or other media, opening a whole new range of possibilities. It can not only take the text, but also deal in different media options. So today when we say ChatGPT or Gemini, they can generate images, and it's not just one model doing everything. These are specialized systems working together behind the scenes. 05:38 Lois: You mentioned large language models and how they power text-based gen AI, so let’s talk more about them. Himanshu, what is an LLM and how does it work? Himanshu: So it's a probabilistic model of text, which means, it tries to predict what word is most likely to come next based on what came before. This ability to predict one word at a time intelligently is what builds full sentences, paragraphs, and even stories. 06:06 Nikita: But what’s large about this? Why’s it called a large language model? Himanshu: It simply means the model has lots and lots of parameters. And think of parameters as adjustable dials the model fine tuned during learning. There is no strict rule, but today, large models can have billions or even trillions of these parameters. And the more the parameters, more complex patterns, the model can understand and can generate a language better, more like human. 06:37 Nikita: Ok… and image-based generative AI is powered by diffusion models, right? How do they work? Himanshu: Diffusion models start with something that looks like pure random noise. Imagine static on an old TV screen. No meaningful image at all. From there, the model carefully removes noise step by step to create something more meaningful and think of it like sculpting a statue. You start with a rough block of stone and slowly, carefully you chisel away to reveal a beautiful sculpture hidden inside. And in each step of this process, the AI is making an educated guess based on everything it has learned from millions of real images. It's trying to predict. 07:24 Stay current by taking the 2025 Oracle Fusion Cloud Applications Delta Certifications. This is your chance to demonstrate your understanding of the latest features and prove your expertise by obtaining a globally recognized certification, all for free! Discover the certification paths, use the resources on MyLearn to prepare, and future-proof your skills. Get started now at mylearn.oracle.com. 07:53 Nikita: Welcome back! Himanshu, for most of us, our experience with generative AI is with text-based tools like ChatGPT. But I’m sure the uses go far beyond that, right? Can you walk us through some of them? Himanshu: First one is text generation. So we can talk about chatbots, which are now capable of handling nuanced customer queries in banking travel and retail, saving companies hours of support time. Think of a bank chatbot helping a customer understand mortgage options or virtual HR Assistant in a large company, handling leave request. You can have embedding models which powers smart search systems. Instead of searching by keywords, businesses can now search by meaning. For instance, a legal firm can search cases about contract violations in tech and get semantically relevant results, even if those exact words are not used in the documents. The third one, for example, code generation, tools like GitHub Copilot help developers write boilerplate or even functional code, accelerating software development, especially in routine or repetitive tasks. Imagine writing a waveform with just a few prompts. The second application, is image generation. So first obvious use is art. So designers and marketers can generate creative concepts instantly. Say, you need illustrations for a campaign on future cities. Generative AI can produce dozens of stylized visuals in minutes. For design, interior designers or architects use it to visualize room layouts or design ideas even before a blueprint is finalized. And realistic images, retail companies generate images of people wearing their clothing items without needing real models or photoshoots, and this reduces the cost and increase the personalization. Third application is multimodal systems, and these are combined systems that take one kind of input or a combination of different inputs and produce different kind of outputs, or can even combine various kinds, be it text image in both input and output. Text to image It's being used in e-commerce, movie concept art, and educational content creation. For text to video, this is still in early days, but imagine creating a product explainer video just by typing out the script. Marketing teams love this for quick turnarounds. And the last one is text to audio. Tools like ElevenLabs can convert text into realistic, human like voiceovers useful in training modules, audiobooks, and accessibility apps. So generative AI is no longer just a technical tool. It's becoming a creative copilot across departments, whether it's marketing, design, product support, and even operations. 10:42 Lois: That’s great! So, we’ve established that generative AI is pretty powerful. But what kind of risks does it pose for businesses and society in general? Himanshu: The first one is deepfakes. Generative AI can create fake but highly realistic media, video, audios or even faces that look and sound authentic. Imagine a fake video of a political leader announcing a policy, they never approved. This could cause mass confusion or even impact elections. In case of business, deepfakes can be also used in scams where a CEO's voice is faked to approve fraudulent transactions. Number two, bias, if AI is trained on biased historical data, it can reinforce stereotypes even when unintended. For example, a hiring AI system that favors male candidates over equally qualified women because of historical data was biased. And this bias can expose companies to discrimination, lawsuits, brand damage and ethical concerns. Number three is hallucinations. So sometimes AI system confidently generate information that is completely wrong without realizing it. Sometimes you ask a chatbot for a legal case summary, and it gives you a very convincing but entirely made up court ruling. In case of business impact, sectors like health care, finance, or law hallucinations can or could have serious or even dangerous consequences if not caught. The fourth one is copyright and IP issues, generative AI creates new content, but often, based on material it was trained on. Who owns a new work? A real life example could be where an artist finds their unique style was copied by an AI that was trained on their paintings without permission. In case of a business impact, companies using AI-generated content for marketing, branding or product designs must watch for legal gray areas around copyright and intellectual properties. So generative AI is not just a technology conversation, it's a responsibility conversation. Businesses must innovate and protect. Creativity and caution must go together. 12:50 Nikita: Let’s move on to generative AI agents. How is a generative AI agent different from just a chatbot or a basic AI tool? Himanshu: So think of it like a smart assistant, not just answering your questions, but also taking actions on your behalf. So you don't just ask, what's the best flight to Vegas? Instead, you tell the agent, book me a flight to Vegas and a room at the Hilton. And it goes ahead, understands that, finds the options, connects to the booking tools, and gets it done. So act on your behalf using goals, context, and tools, often with a degree of autonomy. Goals, are user defined outcomes. Example, I want to fly to Vegas and stay at Hilton. Context, this includes preferences history, constraints like economy class only or don't book for Mondays. Tools could be APIs, databases, or services it can call, such as a travel API or a company calendar. And together, they let the agent reason, plan, and act. 14:02 Nikita: How does a gen AI agent work under the hood? Himanshu: So usually, they go through four stages. First, one is understands and interprets your request like natural language understanding. Second, figure out what needs to be done, in this case flight booking plus hotel search. Third, retrieves data or connects to tools APIs if needed, such as Skyscanner, Expedia, or a Calendar. And fourth is takes action. That means confirming the booking and giving you a response like your travel is booked. Keep in mind not all gen AI agents are fully independent. 14:38 Lois: Himanshu, we’ve seen people use the terms generative AI agents and agentic AI interchangeably. What’s the difference between the two? Himanshu: Agentic AI is a broad umbrella. It refers to any AI system that can perceive, reason, plan, and act toward a goal and may improve and adapt over time. Most gen AI agents are reactive, not proactive. On the other hand, agentic AI can plan ahead, anticipate problems, and can even adjust strategies. So gen AI agents are often semi-autonomous. They act in predefined ways or with human approval. Agentic systems can range from low to full autonomy. For example, auto-GPT runs loops without user prompts and autonomous car decides routes and reactions. Most gen AI agents can only make multiple steps if explicitly designed that way, like a step-by-step logic flows in LangChain. And in case of agentic AI, it can plan across multiple steps with evolving decisions. On the memory and goal persistence, gen AI agents are typically stateless. That means they forget their goal unless you remind them. In case of agentic AI, these systems remember, adapt, and refine based on goal progression. For example, a warehouse robot optimizing delivery based on changing layouts. Some generative AI agents are agentic, like auto GPT. They use LLMs to reason, plan, and act, but not all. And likewise not all agentic AIs are generative. For example, an autonomous car, which may use computer vision control systems and planning, but no generative models. So agentic AI is a design philosophy or system behavior, which could be goal-driven, autonomous, and decision making. They can overlap, but as I said, not all generative AI agents are agentic, and not all agentic AI systems are generative. 16:39 Lois: What makes a generative AI agent actually work? Himanshu: A gen AI agent isn't just about answering the question. It's about breaking down a user's goal, figuring out how to achieve it, and then executing that plan intelligently. These agents are built from five core components and each playing a critical role. The first one is goal. So what is this agent trying to achieve? Think of this as the mission or intent. For example, if I tell the agent, help me organized a team meeting for Friday. So the goal in that case would be schedule a meeting. Number 2, memory. What does it remember? So this is the agent's context awareness. Storing previous chats, preferences, or ongoing tasks. For example, if last week I said I prefer meetings in the afternoon or I have already shared my team's availability, the agent can reuse that. And without the memory, the agent behaves stateless like a typical chatbot that forgets context after every prompt. Third is tools. What can it access? Agents aren't just smart, they are also connected. They can be given access to tools like calendars, CRMs, web APIs, spreadsheets, and so on. The fourth one is planner. So how does it break down the goal? And this is where the reasoning happens. The planner breaks big goals into a step-by-step plans, for example checking team availability, drafting meeting invite, and then sending the invite. And then probably, will confirm the booking. Agents don't just guess. They reason and organize actions into a logical path. And the fifth and final one is executor, who gets it done. And this is where the action takes place. The executor performs what the planner lays out. For example, calling APIs, sending message, booking reservations, and if planner is the architect, executor is the builder. 18:36 Nikita: And where are generative AI agents being used? Himanshu: Generative AI agents aren't just abstract ideas, they are being used across business functions to eliminate repetitive work, improve consistency, and enable faster decision making. For marketing, a generative AI agent can search websites and social platforms to summarize competitor activity. They can draft content for newsletters or campaign briefs in your brand tone, and they can auto-generate email variations based on audience segment or engagement history. For finance, a generative AI agent can auto-generate financial summaries and dashboards by pulling from ERP spreadsheets and BI tools. They can also draft variance analysis and budget reports tailored for different departments. They can scan regulations or policy documents to flag potential compliance risks or changes. For sales, a generative AI agent can auto-draft personalized sales pitches based on customer behavior or past conversations. They can also log CRM entries automatically once submitting summary is generated. They can also generate battlecards or next-step recommendations based on the deal stage. For human resource, a generative AI agent can pre-screen resumes based on job requirements. They can send interview invites and coordinate calendars. A common theme here is that generative AI agents help you scale your teams without scaling the headcount. 20:02 Nikita: Himanshu, let’s talk about the capabilities and benefits of generative AI agents. Himanshu: So generative AI agents are transforming how entire departments function. For example, in customer service, 24/7 AI agents handle first level queries, freeing humans for complex cases. They also enhance the decision making. Agents can quickly analyze reports, summarize lengthy documents, or spot trends across data sets. For example, a finance agent reviewing Excel data can highlight cash flow anomalies or forecast trends faster than a team of analysts. In case of personalization, the agents can deliver unique, tailored experiences without manual effort. For example, in marketing, agents generate personalized product emails based on each user's past behavior. For operational efficiency, they can reduce repetitive, low-value tasks. For example, an HR agent can screen hundreds of resumes, shortlist candidates, and auto-schedule interviews, saving HR team hours each week. 21:06 Lois: Ok. And what are the risks of using generative AI agents? Himanshu: The first one is job displacement. Let's be honest, automation raises concerns. Roles involving repetitive tasks such as data entry, content sorting are at risk. In case of ethics and accountability, when an AI agent makes a mistake, who is responsible? For example, if an AI makes a biased hiring decision or gives incorrect medical guidance, businesses must ensure accountability and fairness. For data privacy, agents often access sensitive data, for example employee records or customer history. If mishandled, it could lead to compliance violations. In case of hallucinations, agents may generate confident but incorrect outputs called hallucinations. This can often mislead users, especially in critical domains like health care, finance, or legal. So generative AI agents aren't just tools, they are a force multiplier. But they need to be deployed thoughtfully with a human lens and strong guardrails. And that's how we ensure the benefits outweigh the risks. 22:10 Lois: Thank you so much, Himanshu, for educating us. We’ve had such a great time with you! If you want to learn more about the topics discussed today, head over to mylearn.oracle.com and get started on the AI for You course. Nikita: Join us next week as we chat about AI workflows and tools. Until then, this is Nikita Abraham… Lois: And Lois Houston signing off! 22:32 That’s all for this episode of the Oracle University Podcast. If you enjoyed listening, please click Subscribe to get all the latest episodes. We’d also love it if you would take a moment to rate and review us on your podcast app. See you again on the next episode of the Oracle University Podcast.
Brasil