data/transcripts/slides 8-31.txt

start_time	end_time	speaker	transcript
7.98	81.02	SPEAKER	Now that you have, now that you understand what foundation models are, let's explore how they're being used in real -world applications. Enterprises across industries are leveraging generative AI to build intelligent systems that enhance productivity, automate tasks, and also drive innovation. These applications rely on what we call generative AI resources, which combine two main components, a foundation model, and a set of supporting tools. The foundation model, such as the large language model, provides the core intelligence. This is what enables the system to understand and also to generate human -like language. However, to be practically useful, this model is typically integrated with external tools that expand on its capabilities. Now, these tools might include things like search engines, APIs, database connection, content management systems, or other domain -specific resources. Together, the foundation model and the tools form a complete generative AI resource that's capable of handling a wide range of business tasks.
86.64	273.16	SPEAKER	Let's start by taking a closer look at how generative AI is applied in classification. Now, classification tasks involve assigning labels to data based on its content or characteristics. In a business context, one practical example of this is automatically categorizing support tickets into different priority levels. So, when a customer submits a support request, the system can analyze the content off the ticket, and then it can determine whether it should be treated as a high, a medium, or a low priority. Now, this would really eliminate the need for the manual triage. It can help speed up the response time and also ensure that the critical issues are addressed promptly. To perform this task, a generative AI resource is used. Now, the model processes the incoming ticket text, it understands the context, urgency, and sentiment, and then assigns the correct priority label. Now, over time, the model can be fine -tuned or augmented with feedback from support teams, and this is what it allows it to continuously improve its accuracy. The advantage of using generative AI for classification lies in its flexibility. So, unlike traditional classification systems that rely on hard -coded rules or limited keyword matching, generative AI models understand the language at a deeper level. They can detect nuance, they can also understand the implied urgency, and also complex intent, even when the input is unstructured or if it's written in informal language. Another high impact application of generative AI is in the domain of question answering. In many enterprise environments, employees and analysts frequently need quick and accurate answers based on internal documentation, databases, or proprietary reports. Now, traditional search systems often fall short because they return links or documents instead of just giving the plain and simple direct answer. Generative AI changes this by enabling natural language question answering capabilities, and it can use that to extract as well as to synthesize relevant information in real time. Now, at the core of this capability is the generative AI resource, which, again, combines a foundation model, typically a large language model, and all of the essential tools. Now, these tools might include things like connectors to the proprietary document repositories, it could include things like APIs to knowledge bases, or indexing systems that organize internal reports for rapid access. Now, when a user poses a question, the system doesn't just search for the keywords. Instead, it understands the intent behind the question, navigates the appropriate sources, and then generates a precise context -aware response.
274.72	337.60	SPEAKER	So, for example, when an analyst might ask a question something like this, like, what were the main market trends that were noted in our Q2 industry analysis report? Instead of having to manually read through the report, the generative AI system can provide a succinct answer by identifying some of the key insights directly from this content. Now, this is especially valuable when the reports are long, technical, or contain information distributed across multiple sections. Now, this kind of question -answering functionality significantly improves the decision -making speed, as well as the accuracy. It also reduces the overall time that spends searching for information, and then it also ensures that the knowledge that's embedded in the proprietary documents is more accessible across the entire organization. And more importantly, the model can also be fine -tuned or prompted to respect data privacy and also source integrity, and that ensures that all of the sensitive content is handled responsibly.
339.52	464.68	SPEAKER	Another valuable use case off -generative AI is the structured text generation. Now, this application focuses on producing well -o -go... Let's say that again. Edit that part out. This application focuses on producing well -o... This application focuses on producing well -organized context -aware text that follows specific formatting or data entry rules. It's particularly useful in domains where large volumes of semi -structured or unstructured content might be converted into structured formats. So it could be things like legal, financial, or regulatory environments, but structured data generation typically involves extracting all of the relevant information from source documents and then using it to populate some predefined template or Excel sheets or SQL database or whatever it is. Now, a practical example is automatically extracting required fields from a legal document to generate a sort of regulatory filing. Instead of manually scanning through contracts or filings to find names, dates, clauses, or financial figures, a generative AI system can identify and extract these data points with high accuracy. They can then format them into a structured output that's ready for submission. Now, to accomplish this, the system relies on a generative AI resource. The foundation model is responsible for understanding the content. They can use it for understanding the structure and its semantic meaning of all of the text. Meanwhile, the tools enable the integration with document repositories. They can use it for validating rules and also output generation systems. When combined, they create a highly efficient pipeline for transforming free -form documents into structured machine -readable outputs. Now, what sets generative models apart in this context is their ability to understand variations in the language as well as in formatting.
466.46	492.68	SPEAKER	Because, for instance, many different legal documents might express the same clause in multiple ways. A well -trained LLM can recognize the intent and extract the correct field even when the phrasing or the structure differs based on the documents. Now, this drastically reduces the need for manual review and manual data entry, while also improving the consistency and reducing the risk of human error.
496.48	584.00	SPEAKER	Summarization is another highly practical application of generative AI that can significantly improve efficiency and also clarity in information -dense environments. In many business contexts, stakeholders need to quickly grasp the essential points from lengthy and complex content. It could be things like meetings, reports, interviews, or maybe research papers, or even earning calls. The generative AI models enable the automated summarization, allowing you to distill large volumes of text into concise coherent outputs that retain the original meaning of the entire text. Now, in this use case, a generative AI resource is used to analyze full documents or transcripts and then produce human -readable summaries. Now, as an example, after a company conducts something like an earnings call, a foundation model can then review the entire transcript and then generate something like a press release that's highlighting the key points. Now, some examples of these key points could be things like financial performance, strategic announcements, and also forward -looking statements. Now, this process that would traditionally require hours and hours and hours of manual work can be completed in just minutes, with high levels of consistency as well as accuracy.
587.00	699.78	SPEAKER	generative AI is also making a significant impact in the area of developer productivity. In modern software development, speed and accuracy, and also automation are critical for maintaining high performance and also minimizing time to market. generative AI can enhance these outcomes by automating all of these repetitive or even cognitively intensive tasks that can really help to free up some of the developers to focus on the problem solving and the innovation side of things. Now, one compelling example of this is doing things like automatic generation of test cases. Now, writing tests is really essential and important, but it's often quite time -consuming, especially when we consider it as a part of the overall software development lifecycle. Now, tests validate all of the code, and it also makes sure that the code behaves as expected under all of these various conditions. It can help to catch bugs early, and especially when you're working with it in the development cycle. Now, with generative AI, a developer can simply provide a snippet of code and the system can automatically generate a suite of relevant test cases. Now, these might include things like unit tests, integration checks, edge case validations, or even security and performance -related scenarios. Now, in addition to generating tests, these systems can also offer things like real -time code suggestions. It can offer things like documentation drafts and maybe things like refactoring advice as well as explanations of maybe things or parts of the code that are a little bit unfamiliar. Now, this can be especially valuable for onboarding new developers, for maintaining legacy code bases or even for scaling development teams.
701.52	785.98	SPEAKER	Now, generative AI is also revolutionizing creative content generation, a space that traditionally required deep domain knowledge, originality, and also the ability to tailor messages to specific audiences. Now, this application area known as creative generation leverages AI's capability to produce things like compelling and customized content at scale. Now, enterprises are using generative models to support marketing, advertising, branding, and even product communication in ways that are faster, more targeted, and cost -effective. Now, one practical example is doing things like crafting a script for a short -form video such as a social media reel. A generative AI system can take a product description, they can take the brand guidelines, and also the details about the target audience, and then generate a ready -to -use script that's optimized for that specific format. Now, the script may also be tailored for different tones, so for things like professional, for humor, or trendy, or it can really be adjusted for all of these different platforms like Instagram, TikTok, LinkedIn, or whatever the new coolest social media platform is. But this really allows the marketing teams to produce personalized content quickly and consistently across all of these channels. Now, the final application that we'll be...
787.54	860.82	SPEAKER	Let's see that again. Now, the final application that we'll explore is data analytics, where generative AI is unlocking new ways for both technical and non -technical users to explore and understand data. Traditionally, extracting insights from data sets such as CSV files required knowledge of data wrangling, required knowledge of things like statistical methods, and also programming tools like SQL or Python. With the help of large language models, much of that complexity can now be abstracted away, and we can use that through natural language interaction. Now, in this use case, a generative AI resource composed of a foundation model, and some sort of an analytical tool is used to analyze structured data formats like CSV or Excel files. You can simply upload a file and then ask questions like, what are the top performing regions this quarter? Or can you summarize trends in customer churn? And the model is going to interpret the request, process the data set, and then deliver a human -readable response or visualization. Now, this enables faster insights and also broadens access to data -driven decision -making across different roles.
866.54	942.38	SPEAKER	Let's take a look at a real -world success story that highlights the value of generative AI in enterprise settings. Now, we'll look at it specifically within the financial services industry. So financial institutions such as Goldman Sachs have adopted generative AI technologies to enhance how their teams interact with and also extract value from both structured and unstructured data. One of their use case is document -based question answering at scale. Now, in practice, this means that employees, particularly financial advisors, can now ask natural language questions about various sources of information, such as internal as well as external analyst reports for financial news articles and even regulatory filings. Now, these documents are stored in some sort of a centralized system, and when a query is submitted, the generative AI model essentially scans the document store. It interprets the content and then generates a clear and relevant answer. The overall system doesn't just return the keyword matches or document links. It actually provides summarized insights. It answers specific questions,
944.18	971.86	SPEAKER	editing, just cut that last part, I'm going to say it again. It answers specific to the question and also often, it provides answers specific to the question and also often includes context or rationale to support. Let's completely repeat this part again. I'm going to start with the overall system does more than, so if you're editing this, I'm going to start the sentence here again with the overall system.
975.11	1029.74	SPEAKER	The overall system does not just return keyword matches or document links. It provides summarized insights. It provides answers specific to the question and often includes context or rationale to support the response. Now, this fundamentally changes how knowledge is accessed and used in high -stakes decision making. For Goldman Sachs, the impact has been substantial because these generative AI applications have really helped to empower financial teams to perform deeper, faster, and more reliable analysis. They no longer have to spend valuable time searching through all of these dense documents or maybe manually cross -referencing sources. Instead, the AI does that work for them, which leads to higher productivity, it leads to faster turnaround times and also more informed client engagements.
1041.26	1087.64	SPEAKER	Another compelling success story comes from the insurance industry, where generative AI is being used to address one of the most document -heavy time -sensitive functions, claims processing. Now, insurance companies have started deploying generative AI solutions that dramatically improve the efficiency and as well as the accuracy of their claims workflows by enabling intelligent question -answering over large volumes of unstructured data. Now, claims teams routinely work with a variety of document types, including things like call transcripts, doctor notes, medical reports, and even legal paperwork. Now, these documents are often lengthy, complex, and also written in highly specialized language.
1089.24	1156.72	SPEAKER	Traditionally, reviewing and also extracting relevant information from these materials requires significant manual effort. With generative AI, that process is being transformed. At the center of this transformation is a document QA system that's powered by a foundation model that's trained to understand medical, legal, and insurance terminology. Now, this model is connected to some sort of a document store that's containing all of the relevant materials. Now, when an agent submits a question and the question could be something like, what diagnosis was given in this initial consultation? Or maybe something like, has the claimant previously filed for some sort of a similar injury? What happens is that the system automatically searches the document repository, it understands the context, and then provides a precise natural language answer. Now, these generative AI capabilities significantly reduce the time that's required to process each claim and also minimize the risk of overturning key details.
1158.44	1204.96	SPEAKER	Let's say that again. And also minimize the risk of overlooking key details. In turn, this accelerates the entire claims life cycle from intake to resolution, leading to faster service for customers and also reduced operational costs for many providers. In many cases, these tools also assist in compliance by ensuring that responses are grounded in proper documentation and also audit trails. Now, additionally, this same technology is laying the groundwork for more advanced automation in underwriting, because by enabling underwriters to query policy documents, risk assessments, and also regulatory rules in plain language, generative AI can help to streamline decision making and reduce turnaround times on policy approvals.
1212.10	1278.02	SPEAKER	While generative AI offers a wide range of powerful capabilities and transformative applications, it's equally important to understand its limitations. In this section, you will explore some of the key shortcomings that are associated with generative AI systems, especially those that are based on large language models. Now, recognizing these risks will help you to design, deploy, and also to manage AI applications more responsibly and effectively. Now, one of the most pressing concerns is the vulnerability to injection attacks. Now, these occur when malicious users manipulate the input data to trick the model into producing unintended or harmful outputs. For example, specially crafted prompts may cause a model to bypass the content filters or maybe reveal some sort of confidential information. Now, this makes prompt injection a growing area of focus in AI security. Another major issue is data privacy.
1279.68	1328.88	SPEAKER	Generative models are trained on vast data sets that may inadvertently include some sort of sensitive or personally identifiable information. Without careful data curation, there's also going to be a risk that models could memorize and also repeat some of the private content, and this could lead to serious privacy breaches in production environments. Intellectual property or IP concerns can also arise when models generate content that closely resembles copyrighted materials or proprietary styles. Now, since the models are trained on publicly available data, it can be difficult to ensure that outputs don't actually infringe on any existing IP rights. Now, this is a particular challenge in industries such as publishing, design, and entertainment.
1330.49	1432.54	SPEAKER	A fourth shortcoming is the lack of explainability, because unlike some of the rule -based systems, generative models really operate as a black box. It makes it a lot more difficult to understand how they actually arrive at some sort of a given output. Now, this lack of transparency can really undermine trust, especially in high -stakes domain like healthcare, finance, or law, where explainable decisions are often required by regulation. Now, one of the most well -known risks is hallucination, where the model generates content that sounds possible, but it's actually factually incorrect or entirely fabricated. Now, these outputs can lead to the spread of misinformation, and if not properly verified, they can also present serious challenges and tasks that require factual precision. And finally, generative models are also inherently probabilistic, meaning that the same input can produce different outputs every single time, depending on randomness and also sampling methods. Now, this can make the behavior unpredictable and limit the model's usefulness for a certain task, where things like determinism is expected. Now, as you work with generative AI, it's important to balance its strengths with these limitations. And in this course, you'll learn about some of these best practices for mitigating these risks, and we can use things like prompt engineering, model evaluation, content filtering, and understand things like human -in -the -loop validation to ensure that your AI solutions are not only innovative, but also safe.
1442.84	1461.98	SPEAKER	Now, this chart illustrates how generative AI spending is being distributed across different departments within organizations, offering valuable insights into adoption trends and also strategic priorities. As you can see here, the highest proportion of generative AI expenditure is currently being driven by IT department.
1463.58	1468.74	SPEAKER	Oops, sorry about that. Let's repeat this entire slide.
1472.32	1479.08	SPEAKER	This chart illustrates how generative AI spending is being distributed across different departments within organizations.
1481.10	1495.08	SPEAKER	Now, this really helps to offer some valuable insight into adoption trends and strategic priorities. Let's repeat this slide one more time. Editing, sorry about that.
1497.54	1679.30	SPEAKER	This chart illustrates how generative AI spending is being distributed across different departments within organizations. This really helps to offer valuable insights into adoption trends and also strategic priorities. Now, as you can see here, the highest proportion of generative AI expenditure is currently being driven by IT departments, which account for 22 % of the total spending. Now, this suggests that infrastructure, systems, integration, and also some key automation or internal automation remain an important focus area for early investment in generative technologies. Now, falling closely behind is product and engineering teams that contribute to 19 % off the spend. Now, these departments are likely leveraging generative AI to do things like maybe accelerate product development. They can enhance software capabilities and also integrate AI features directly into digital offerings. Customer support comes next at 9%. And this really helps to reflect some of the growing use of generative models in things like chatbots, virtual assistants, and maybe automated response systems that really help to improve the service efficiency and also user satisfaction. Sales and data science departments each account for about 8 % of the spend. In sales, AI is being used to or being applied to lead to editing. I'm going to say that again. In sales, AI is being applied to lead qualification, content personalization, and also deal support. While in data science, it supports tasks like model prototyping, data enrichment, and also inside generation. Marketing and human resources and also accounting finance each follow with about 7%. And this indicates a strong cross -functional use case ranging from content creation and also recruitment assistance all the way to financial reporting automation. Design and legal functions are also investing in generative AI, each making up about 6 % and 3 % of the budget respectively. Designers are using tools for creative ideation and also asset generation, while legal teams may be exploring document, review, and contract drafting support. And then finally, we have the other category, which includes departments that aren't really explicitly listed, but they account for about 6 % showing that generative AI is actually being explored across virtually every corner of the enterprise. Now, as you study these trends, you will be able to identify which departments within your own organization might benefit most from generative AI and also tailor your solution to meet their specific needs and investment levels. Now, this knowledge will be crucial for prioritizing use cases for justifying ROI and also gaining stakeholder buy -ins for your AI initiatives.
1693.12	1744.52	SPEAKER	When it comes to working with large language models, one of the first decisions that you'll need to make is how to access them. LLMs are available through both paid and open source options, and they can be accessed via public cloud platforms or directly through LLM vendors. Now, understanding the landscape of providers will help make the editing, I'm going to say that again. Now, understanding the landscape of providers will help you make informed decisions based on your project's needs for cost, performance, security, and control. Paid LLMs are typically offered by major AI companies and integrated with cloud service providers. Now, some of the leading paid providers include Google, and Google offers the Gemini family -off models which are integrated with the Google cloud services and are known for multimodal capabilities.
1746.28	1856.94	SPEAKER	OpenAI is another one. OpenAI provides access to GPT40 and earlier versions like GPT35. Now, these are commonly accessed through Azure or OpenAI's own platform. Anthropic, the creators of the cloud model series are designed for safety and also aligned reasoning, and they're often used in enterprise and productivity settings. Cohere is another provider which offers the command model that's known for language understanding tasks and also integration flexibility. Now, you can access these paid models through major cloud providers such as Microsoft Azure, Amazon Web Services, and Google Cloud Platform. And these platforms provide APIs, SDKs, and manage infrastructure that simplify the model deployment, scaling, and security. Now, on the other hand, Open Source LLMs offer flexibility, transparency, and often lower costs, especially when you're deploying it in private environments. Now, some prominent open source providers include MetaAI, which has released the LAMA series that's widely used in research and customizable for specialized applications. Mr. AI, that's known for its performance models like Mistral and Mixtrel, that's offering lightweight and efficient options for developers. Alibaba Cloud, which provides the Quen family of models that support multi -lingual and also enterprise specific use cases. Now, these open source models are supported by a growing ecosystem of hosted platforms like any scale, grok, and together AI, which offer infrastructure and tooling for scalability and secure deployment.
1865.14	1919.56	SPEAKER	In addition to accessing the large language models to crop, in addition to accessing the large language models to cloud providers or commercial APIs, organizations also have the option to deploy open source or open weight models on their own infrastructure. Now, this self -hosted approach provides greater control over the data privacy, cost management, and also performance tuning, making it an attractive option for enterprises with specific deployment requirements. Now, to begin with, companies can access and run open models using frameworks and libraries that are designed for efficiency, as well as customization. Popular platforms such as HuggingFace, LAMA .cpp, and Unslut provide streamlined implementations of open source models like LAMA or Mistral.
1921.18	1990.00	SPEAKER	Now, these tools are optimized for specific hardware configurations, including CPUs, GPUs, or even mobile devices, making them ideal for edge deployments or low latency applications. Now, by using these access libraries, teams can fine -tune models, they can optimize memory usage and also serve use cases with greater flexibility than just hosted APIs. Now, this is very useful because this allows you to really deploy everything at your own local level. Now, another common setup involves deploying open models through local or virtualized server environments using tools like VLLM and OLAMA. And these platforms offer efficient inference engines and API layers that are compatible with standard open AI interfaces, which really helps to simplify the process of integrating models into existing applications. Now, with this approach, you can host models internally, you can manage load balancing and also ensure compliance with security and governance policies, all while preserving the ability to scale as your demand grows.
1993.82	2111.20	SPEAKER	One of the most common ways to interact with large language models is through APIs, or it's short for application programming interfaces. APIs define a standardized set of rules that allow different software components to communicate with one another over a network. In the content editing, I'm going to say that again. In the context of LMS, APIs enable you to send requests from your application to a server where the model is going to be hosted, and then you receive a response back that contains the model output. Now, here's how the process works. The client, typically your application, sends an HTTP request to the server using the post method. Now, this request usually includes a payload that's formatted in some sort of a JSON format, which usually contains the prompt or the user input text for the language model, along with any other configuration parameters. It could be things like the temperature, the max tokens, or the stop sequence. Once the request is received, the server processes it using the hosted LLM, and then generates a response back. Now, the server, it sends a response to the client also in a JSON format. And this response, it includes the model -generated content, along with some other metadata. It could be things like some status code, like 200, like, oh, this was successful. And it really helps to determine whether the request was successful, or if some sort of an error had occurred. Now, this interaction is stateless and efficient. It allows your application to access powerful LLM capabilities without having to manage the model infrastructure yourself. And this method of access is widely used in real -world applications, including chatbots, content generation tools, for summarization platforms, and also code assistance.
2113.04	2142.86	SPEAKER	Cloud providers and LLM vendors typically offer robust APIs that can support scalable, secure, and low -latency interactions. Now, as you move forward in this course, you'll gain hands -on experience with working with all sorts of different LLM APIs. And this will really help you to learn how to structure requests, how to handle responses, and also how to integrate these services into larger software systems. And this knowledge is essential for building production -ready AI applications.
2149.38	2165.58	SPEAKER	This chart shows you how the market share of major LLM providers has shifted from 2023 to 2024, offering some insight into the competitive dynamics of the generative AI space. OpenAI, while still in the market leader, so a significant drop in share from 50%.
2167.78	2288.34	SPEAKER	OpenAI, while still the market leader editing, I'm going to say that again. OpenAI while still at the market leader, so a significant drop in share from 50 % in 2023 to 34 % in 2024, representing a 16 % year -over -year decrease. Now, this decline suggests that the market is becoming more distributed as more players gain traction and also enterprise customers diversify their usage across multiple models. Anthropic experienced the most notable increase, doubling its share from 12 % to 24%, again of 12 percentage points. Now, this sharp rise reflects the growing adoption of cloud model family and also likely signals confidence in anthropic safety -focused approach. Meta maintained a stable market share at about 16 % across both years, showing consistent use of the llama models, particularly within the open source and research communities. Google saw a moderate increase in share rising from 7 % to 12%, getting about 5 .0 over year, and this growth is likely due to the evolution of its Gemini models and a deeper integration with the Google cloud ecosystem. Now, Mril AI, it held steady with a minor drop from 6 % to 5%, showing a slight decline of 1 % point. Go here internal model and other unspecified providers all remain consistent at around 3%, and this indicated steady but limited adoption at the broader market level. Overall, these shifts suggest that while OpenAI remains dominant, the landscape is diversifying. There are many players, especially anthropic and Google and even deep -seek. All of these are gaining market presence. Now, as you continue through this course, understanding these trends will really help you to make the informed decisions when selecting the LM providers for different use cases. Now, balancing factors like performance, pricing, openness, and alignment with enterprise needs are all going to be crucial, and those are all things that we will be studying about.
2294.20	2325.58	SPEAKER	To understand how large language models are trained, it helps to begin with the concept of language modeling. At a high level, training an LM involves predicting the next word in a sequence of text. The model does this by assigning probabilities to each word in a fixed vocabulary based on the context of the words that came before it. Now, let's walk through an example. Consider the sentence, the movie is a visually stunning action -packed and emotionally resonant thrill ride.
2381.08	2383.30	SPEAKER	Thank you.
2411.08	2413.30	SPEAKER	Thank you.
2441.08	2443.30	SPEAKER	Thank you.
2471.08	2473.30	SPEAKER	Thank you.
2501.08	2503.30	SPEAKER	Thank you.
2531.08	2533.30	SPEAKER	Thank you.
2561.08	2563.30	SPEAKER	Thank you.
2575.48	2578.28	SPEAKER	Thank you.