Ref: https://www.udemy.com/course/aws-ai-practitioner-certified/learn/lecture/44879563 and https://www.udemy.com/course/aws-ai-practitioner-certified/learn/lecture/45375103

• Contents:

GenAI - Basic Concepts

• 🔧 Def: AI focused on generating new data similar to training data
  • Subset of Deep Learning (DL), which itself is a subset of Machine Learning (ML), which itself is a subset of AI
  • Data can be text, image, audio, code, video…
• Tons of unlabeled data pretrain a Foundation Model (FM), which can then generate data
  • 💡 Foundation Models are extremely expensive, only big companies have their own
    • E.g. OpenAI owns the foundation model GPT-4o (foundation model behind ChatGPT)
  • User gives a prompt for Model to generate data
• ‼️ Generated data is non-deterministic!!
  • Same prompt can generate similar but different data
  • Generated data is determined thanks to statistical/probabilistic methods (NOT with deterministic methods)

Large Language Models (LLMs)

• 🔧 AI designed to generate coherent human-like text
  • Subset of Foundation Models (FMs)
  • e.g. OpenAI's GPT-4
• Can perform language-related tasks: translation, summarization, question answering, content creation
• Algorithm selects words from a list based on probability (randomly)

Graphical GenAI

• Generate images from text prompts
  • “Generate a blue sky with white clouds and the word “Hello” written in the sky”
• Generate images from images
  • “Transform this image in Japanese anime style”
• Generate text from images
  • “Describe how many apples you see in the picture”
• Diffusion Models are very popular for generating images
  • e.g. Stable Diffusion
  • 💡Add noise to image until it's no longer recognizable, the model then learns “what makes a cat a cat”, can then generate a cat image from noise

Advanced GenAI Concepts