Compare AI vs ML #

What is Artificial Intelligence?

  • Definition: A branch of computer science focused on building systems that can simulate human intelligence
  • Goal: Enable machines to see, understand, learn, and make decisions

AI is All Around Us

  • Self-driving Cars: Sense environment and make driving decisions
  • Spam Filters: Detect and block unwanted emails
  • Fraud Detection: Analyze behavior to flag suspicious transactions
  • Recommendation Systems: Suggest music, movies, products based on past activity
  • Email Sorting: Categorize emails into Primary, Promotions, and Social tabs

How is Traditional Programming done?

  • Scenario: You want to predict house prices
  • Traditional Programming: You write explicit rules — “IF location is X AND size is Y THEN price is Z”

How does Machine Learning work?

  • What Happens?: ML system that learns from data instead of rules
  • Machine Learning: Give examples → ML finds patterns → Predicts outcome
  • Power: ML improves as more data is provided

How Machine Learning Works

  • Step 1: Feed Data – Provide historical examples
  • Step 2: Train Model – ML learns patterns in data
  • Step 3: Predict – Use trained model to predict on new data

Aspect Artificial Intelligence (AI) Machine Learning (ML)
Purpose Simulate human intelligence Enable machines to learn from data
Scope Broad — includes reasoning, learning, and problem-solving Narrower — focused mainly on learning from data
Relation ML is a subset of AI Part of AI
Data Dependency May or may not use data Strongly depends on large amounts of data

What are the Key Steps in Machine Learning Lifecycle? #

Step 1: Data Ingestion

  • Collect data from real-world sources
  • Examples: User clicks, sales records, logs, images, transaction systems, big data systems, ..

Step 2: Data Preparation

  • Clean and transform data
  • Examples: Remove duplicates, handle missing values, ..

Step 3: Model Training

  • Use training data to build a prediction model
  • Use frameworks like TensorFlow, PyTorch, scikit-learn

Step 4: Model Deployment

  • Make model available to real-world systems
  • Examples: Integrate model into mobile app or website

Step 5: Model Management

  • Track model performance over time
  • Actions: Retrain with new data

End-to-End Example: Spam Detection

  • Ingest: Pull email history
  • Prepare: Label spam vs not spam
  • Train: Learn patterns from labeled data
  • Deploy: Filter incoming emails
  • Manage: Keep accuracy high as spam tactics evolve

What is Generative AI? #

Why Generative AI?

  • Scenario: Traditional ML models can detect spam or predict house prices. But what if we want the machine to write an email, generate an image, or create a video?
  • Generative AI: A special kind of AI that creates new content like text, code, images, and audio — not just make predictions.
  • Definition: A subset of Machine Learning that creates original content from patterns it learned
  • Goal: Move from consuming data to generating data
  • Examples: Writing poems, answering questions, generating software code, creating art

How is Generative AI Different from Traditional ML?

Aspect Generative AI Traditional ML
Purpose Create new content (text, images, code, etc.) Make predictions or classifications
Output Type Original and creative outputs Structured outputs like labels or numbers
Examples ChatGPT, DALL·E, GitHub Copilot Fraud detection, stock price prediction
Training Data Requires massive datasets and compute power Can work with smaller, structured datasets
Use Cases Content generation, summarization, coding help Forecasting, classification, ..

How does Generative AI Work?

  • Foundation Models Power Generative AI
    • Definition: Huge pre-trained models trained on broad data (text, code, images)
    • Reusable: One model can handle many tasks (chat, translate, summarize)
    • Customizable: Many apps fine-tune foundation models for their specific needs
  • Training Foundation Models is Complex
    • Massive Datasets: Learn from books, code, web pages, image captions
    • Expensive Training: Needs GPUs, TPUs, distributed storage
    • Takes Time: Training can take weeks or months, depending on size
  • Key Limitations of Foundation Models
    • Data Dependency: Garbage in, garbage out
      • Model quality depends on training data
      • If trained on biased or outdated info, it will produce flawed results
    • Knowledge Cutoff: Models freeze in time
      • Cannot access real-time data
      • Cannot tell you what happened after their training period
    • Hallucinations: Confident but incorrect responses
      • May generate fake facts, URLs, citations, or statistics
    • Bias: Inherits bias from training data
      • Can unintentionally favor certain groups over others
    • Edge Cases: Uncommon inputs often fail
      • Models may behave unexpectedly with unusual or rare queries
  • Reducing Hallucinations with Grounding
    • Grounding: Link AI outputs to verified data
      • Example: Connect chatbot to trusted company documents
    • Benefits:
      • Reduces hallucinations
      • Provides sources
    • Popular Technique – RAG (Retrieval-Augmented Generation)
      • Step 1: Search trusted documents
      • Step 2: Add findings to the prompt
      • Step 3: Let model generate based on that data

How did the ML landscape change after Generative AI? #

ML Landscape Before Generative AI

  • 3 Typical Approaches: Prebuilt ML APIs, AutoML tools, or full custom models
  • ML APIs: Use ready-made APIs for text, speech, image, or video tasks
    • Use Case: Convert customer calls to text using Speech-to-Text API
    • Use Case: Detect objects in product images using Vision API
    • Use Case: Classify customer reviews as positive or negative
  • AutoML: Train custom models without ML expertise by uploading labeled data — platform handles training and deployment
    • Use Case: Classify support tickets into categories using labeled examples
    • Use Case: Categorize pet images (e.g., dogs vs cats) by uploading a labeled dataset
  • Custom Models: Full control for ML teams using frameworks like TensorFlow, PyTorch, scikit-learn — bring your own data and code
    • Use Case: Build a custom model to forecast daily sales using historical data
    • Use Case: Train a deep learning model for detecting anomalies in financial transactions

Generative AI Brought in New Capabilities

  • Foundation Models Introduced: Pre-trained on massive datasets to generate text, code, audio, and images
    • Use Case: Generate product descriptions automatically for e-commerce listings
    • Use Case: Write code snippets based on natural language prompts
  • New Capabilities: Summarization, chat, translation, and content generation via simple APIs
    • Use Case: Summarize long customer feedback into short highlights
    • Use Case: Translate product manuals into multiple languages
  • Models Can Be Tuned Easily: Foundation models can be fine-tuned with your own data to adjust tone, context, or behavior
    • Use Case: Fine-tune a chatbot to use your company’s tone and answer product-specific questions

ML Landscape After Generative AI

  • Use Pretrained APIs (NO CHANGE) — Use ready-made APIs: text, speech, image, or video tasks
  • Use Foundation Models via API (NEW) — New capabilities: chat, summarize, translate, generate code
    • Use Case: Summarize long customer feedback into short highlights
  • Customize Foundation Models (NEW) — fine-tune on your own data
    • Use Case: Fine-tune a chatbot to use your company’s tone and answer product-specific questions
  • Train with AutoML (NO CHANGE) — build models without writing code
    • Use Case: Categorize pet images (e.g., dogs vs cats) by uploading a labeled dataset
  • Build Custom Models (NO CHANGE) — use your own infra and tools if full control is needed
    • Use Case: Train a deep learning model for detecting anomalies in financial transactions

Key Shift After GenAI

  • Platforms now support both traditional ML and GenAI paths

Managed Services In Different Cloud Platforms for AI and ML #

Category Description AWS Azure Google Cloud
Use Pretrained APIs Use ready-made APIs: text, speech, image, or video tasks Amazon Rekognition (image & video analysis)
Amazon Comprehend (text analytics),..		 Azure AI Vision (image analysis)
Azure AI Language (text analysis), ..		 Google Cloud Vision AI (image analysis)
Google Cloud Natural Language AI (text analysis),..
Use Foundation Models via API New capabilities: chat, summarize, translate, generate code Amazon Bedrock Azure AI Foundry Vertex AI Model Garden
Customize Foundation Models Fine-tune foundation models on your own data (cloud-specific & open-source models such as Anthropic Claude, Meta Llama, & Mistral) Amazon Bedrock Azure AI Foundry Vertex AI Model Garden
Train with AutoML Build models without writing code Amazon SageMaker Autopilot Azure Machine Learning Automated ML Google Cloud AutoML
Build Custom Models Platforms for building, training, and deploying custom ML models using TensorFlow,... Amazon SageMaker Azure Machine Learning Vertex AI Custom Training