Skip to main content

Gen AI

LLM​

  • A model at its core predicts the next token, it means it predicts tyh e** enxt word**
  • A Large Language Model is trained on almost all the publically available web information, books, research papers, opensource code
  • A model can generate different types of outputs (Modal)
    • Text
    • Image
    • Video
    • Code

Temperature and Data Accuracy​

When the temperature is lower, the system provides more accurate values. As the temperature increases, the deviation or gap in the values also increases, resulting in less predictable and less accurate data.

tool calling​

  • LLM’s introduced tool calling,This helped in llm deciding which tool to call.

Problems​

  • LLMs work with a principle Garbage in => Garbage out i.e. model output will be good according to your prompt

  • Models have context limits => number of tokens that can be passed and generally if you look at llm pricing we started

    • Context Engineering.
    • Problem to solve: Context Poisoning.

  • Gaurdrails: Guardrails are rules and restrictions that control what an LLM can and cannot do. They help prevent harmful, unsafe, incorrect, or unauthorized responses.

  • Use Grounding to solve Hallucinations: Hallucination means the LLM makes assumptions and generates information that sounds correct but is actually wrong or not based on facts.