Skip to main content
Foundations of Generative AI
Generative AI
- Large language models or LLMS encapsulates the linguistic and semantic relationships between the words and phrases in a vocabulary.
- The model can use these relationships to reason over natural language input and generate meaningful and relevant responses.
- It is often trained on internet-scale data, to generate content in response to user prompts.
Working of Generative AI
- GenAI models learn by analyzing huge volumes of data known as machine learning.
- The architecture behind Generative AI is called as transformer architecture.
- It process entire documents or conversations, maintaining context for more accurate summaries and insights.
- Transformers handle data in parallel, making them fast and suitable for enterprise workloads.
- They work across domains like text, images, code for flexible foundation for business innovation.
Use case of Generative AI
- generate personalized email campaigns
- automate responses to common inquiries
- summarize financial reports or generate insights
- draft clinical documentation or patient summaries.
- simulate product designs or optimize supply chains
Microsoft Generative AI solutions
- Microsoft Copilot
- Microsoft Foundry (Azure AI)
- GitHub Copilot
- Dynamics 365 Copilot
- Power Platform Copilot
Types of Generative AI Models
- Large Language Models - Chatbots, content creation (GPT-4.1, Cluade)
- Code Models: Code completion, debugging (codex-mini)
- Diffusion Models: generate images from text prompts (DALL_E, Stable Diffusion)
- Multimodal models: Interactive assistants (GPT-4o)
- Domain-specific models: Legal, healthcare, finance (Phi-4, DeepSeek)
Challenges in Generative AI
- It can produce information that sounds correct but can be factually wrong or entirely fabricated. This challenge can be reversed by investing in fact-checking systems.
- It may produce different answers to similar prompts, making predictable performance difficult. Inconsistent outputs can disrupt workflows and decision-making.
- Bias occurs when AI reflect prejudices or imbalances presents in its training data. It introduces ethical and legal risks that can damage brand reputation and erode customer trust.
- It raises data privacy, energy consumption issues.
Understanding Prompt Engineering
- A prompt is a simple question or an instruction to get a specific response.
- Generative AI rely on prompts or instructions. How prompts are crafted influences the quality, accuracy and relevance of the AI output.
- Poorly designed prompts can lead to irrelevant results, inaccurate information and wasted time in refining output.
Ground AI
- Generative AI can produce convincing but incorrect information. To address this issue, there are 2 key strategies: grounding and retrieval-augmented generation or RAG.
- Grounded AI uses real-time or enterprise-specific information to produce answers.
- Grounding is critical for accuracy, compliance and trust.
- RAG is a technique to ground a language model.
- The RAG pattern is: retrieve grounded data based on initial user-entered prompt, augment the prompt with grounded data and use a language model to generate a grounded response.
Data readiness
- For success with AI, we need data readiness. It includes:
- Data inventory and classification
- Data accessibility
- Metadata and labeling
- Data governance
- Feedback loops
Responsible AI Principles
- Fairness
- Reliability
- Safety
- Accountability
Machine Learning
- Its goal is to use data to create predictive models that can be used into applications or services.
Stages of Machine Learning lifecycle
- Define the task
- Collect and prepare the data
- Train and validate the model
- Deploy the model
- MLOps: Monitor and manage the lifecycle