Skip to content

Generative AI Terminology

cosmic-snow edited this page Jul 22, 2024 · 3 revisions

Generative AI has rapidly transformed how we interact with technology, from crafting compelling stories to designing stunning visuals. To navigate this exciting landscape, a solid grasp of the terminology is essential. This glossary provides an overview of key terms, demystifying the complex world of generative AI. Let's dive in!

A glossary of generative AI terminology:

  • Generative AI: AI systems capable of creating new content, such as text, images, or audio.
    • LLM: Often referred to as "AI models", a "Large Language Model" is trained on vast amounts of text data, its purpose is to understand and generate human-like text.
  • Hallucination: When an LLM generates plausible but factually incorrect or nonsensical information.
  • Inference: The process of using a trained model to make next token predictions or "generate" output.
  • Tokenization: The process of breaking text into smaller units (tokens) that the model can process.
    • Token: A token can be a word, punctuation mark, or even a sub-word unit.
    • Special Token: These are reserved tokens intended for a function of the LLM. For example <|start|> or <|end|> would determine when the model would start or end something like a reply or a users message. Each model has its own special tokens that must be used appropriately or the model will not behave as expected.
  • Prompt template: Specially formatted instructions used by the LLM to define the information passed to and from it. These require "Special Tokens" to function properly.
    • System prompts: Instructions or context given to an LLM to define its behavior, personality, or role in a conversation. These require "Special Tokens" to function properly, and the model should also be designed to expect a system prompt.
  • Prompt engineering: The art of crafting effective inputs to elicit desired output from an LLM.
    • Zero-shot: The ability of a model to perform tasks it wasn't explicitly trained on, based on its general knowledge.
    • Few-shot: Using a small number of examples to guide the model in performing a new task.
  • Perplexity: A measure of how well a language model predicts a sample of text, often used to evaluate model performance. A lower score is better.
  • Training: This involves feeding the model massive amounts of text data and allowing it to learn patterns and relationships between words.
    • Base model: The original, foundational version of a large language model. This model has not been "fine-tuned".
    • Pre-trained model: An interchangeable term for a model which may be either a base model or one modified from the base dataset.
    • Fine-tuning: Adapting a pre-trained model to a specific task or domain by training it on a smaller, specialized dataset.
    • LoRA: Low-Rank Adaptation, a variant of Fine-tuning which requires less computational resources compared to full fine-tuning.
    • Dataset: The training data used to teach an LLM patterns and information during its training phase.
  • R.A.G. (Retrieval-Augmented Generation): A technique that combines information retrieval from external sources with the generative capabilities of an LLM to produce more accurate and up-to-date responses.
    • Embeddings: You can imagine Embeddings are like digital maps of words or phrases. They turn text into lists of numbers (vectors) that computers can easily work with. These number lists are designed to capture the meaning of the words, so that words with similar meanings have similar number patterns.
      For example, in this number-based representation:
      "Cat"    [0.2, 0.5, 0.1]  
      "Kitten" [0.3, 0.6, 0.2]  
      "Dog"    [-0.1, 0.4, 0.8]
      

      The similarity in numbers between "cat" and "kitten" shows they're related, while the difference with "dog" shows it's a distinct concept. This allows computers to understand and compare words based on their meaning, not just their spelling.