Generative AI


Generative AI is a class of artificial intelligence models designed to generate novel data (such as text, images, audio, or code) by learning patterns from existing data sets.

  • large language model (LLM) is one type of generative AI
  • not all Gen AI tools use LLMs
    • Diffusion models generate data by learning how to turn random noise into something realistic
    • e.g., Canva uses Stable Diffusion to generate an image from a text prompt
  • orgs use tools like Hugging Face and LangChain to implement fine-tuned solutions that support their use cases
    • Hugging Face is an open-source platform that has a collection of machine learning models, data sets, and tools, making AI models accessible to developers and organizations
      • allows you to fine tune your models so they are more specific for your use case
    • LangChain allows you to build applications with LLMs
      • offers retrieval-augmented generation (RAG) pipelines, allowing you to connect your documents, media, and data to the LLM

A Large Language Model (LLM) is an advanced artificial intelligence model trained on vast amounts of text data to understand, generate, and analyze human language for tasks such as summarization, content generation, and natural language querying.

  • is a very large transformer neural network trained on trillions of tokens
    • predicts the next token in a sequence

Foundations

  • two key technical milestones:
    1. 2014, introduction of generative adversarial networks (GANs)
      • A type of artificial intelligence model in which two neural networks compete against each other to generate realistic synthetic data, such as images or text.
      • paved the way for tools that use diffusion models
        • e.g., DALL-E and OpenAI’s Sora
    2. 2017, introduction of the transformer architecture
      • empowered us to develop large language models at scale
      • e.g., Open AI’s GPT-5
      • trained with billions of tokens, using natural language and probability to support information requests, design presentations, calculate complex equations, and even perform analysis

Pitfalls

  • Bias
    • GenAI tools use information in the model they’re trained on to predict outcomes and generate responses
      • but doesn’t guarantee the information is accurate
    • consider whether bias may exist
  • Hallucinations
    • GenAI tools are known to hallucinate
      • model produces nonsensical or factually inaccurate information but presents it as truth
    • can show up in many different ways
    • should verify any output from a GenAI tool
  • Use the right tool for the job, and add your own touch
    • Generative models create data (e.g., text or images), while discriminative models classify or predict outcomes
    • incorporate outputs into your work and data