Novelty of LLMs

Concetti sui Large Language Models (LLM)

Vidhi Chugh

AI strategist and ethicist

Using text data

An image showing sentiment analysis, spam detection, and digital assistants

  • Unstructured data - messy and inconsistent
1 Freepik
Concetti sui Large Language Models (LLM)

Machines do not understand language!

An image showing a computer's inability to interpret language as is

1 Freepik
Concetti sui Large Language Models (LLM)

Need for NLP

 

An image showing NLP as the bridge between text and numbers which are then understood by the computer

1 Freepik
Concetti sui Large Language Models (LLM)

Unique capabilities of LLMs

 

  • Linguistic subtleties
    • Irony
    • Humor
    • Pun
    • Sarcasm
    • Intonation
    • Intent

A person standing with a clipboard point at some generic text and a big question mark

1 Freepik
Concetti sui Large Language Models (LLM)

What's your favorite book?

 

  • Natural response: "Oh, that's a tough one!"

 

  • Personal opinion: "My all-time favorite book is To Kill a Mockingbird by Harper Lee."

 

  • Supporting statement: "It's a powerful story about prejudice, justice, and the human experience."

 

  • Follow-up question: "Have you read it?"
Concetti sui Large Language Models (LLM)

Linguistic subtleties

Sarcasm: "Oh great, another meeting."

Traditional language model:

  • Response: "What's the meeting about?"
    • Neutral
    • Does not pick up sarcasm

 

Large language model:

  • Response: "Sounds like you're looking forward to it!"
    • Playful
    • Engaging
    • Matches the sarcasm
Concetti sui Large Language Models (LLM)

How do LLMs understand

 

  • Trained on vast amounts of data

  • Largeness of LLMs: parameters

  • Parameters represent the patterns and rules

  • More parameters -> complex patterns

  • Generates sophisticated and accurate responses

Concetti sui Large Language Models (LLM)

Parameters

 

  • Small number of bricks -> limited structures

An image showing small number of blocks

 

  • Larger number of bricks -> complex and detailed structures

An image showing large number of bricks

Concetti sui Large Language Models (LLM)

Emergence of new capabilities

 

  • Emergent abilities
    • only present in large-scale models

 

  • Scale:

    • The volume of training data
    • The number of model parameters

 

Graph showing sudden improvement in model performance when the scale of the model increases beyond a transition threshold

Concetti sui Large Language Models (LLM)

Emergence of new capabilities

 

An image showing new capabilities like music, poetry, code generation, medical diagnosis and treatment groups

Concetti sui Large Language Models (LLM)

Building blocks of LLMs

Building blocks of LLMs

Concetti sui Large Language Models (LLM)

To recap

LLMs:

  • Overcome data's unstructured nature

 

  • Outperform traditional models

 

  • Understand linguistic subtleties

How?

  • LLMs' "largeness"
  • Extensive training data
  • Many parameters
  • Emergent abilities
Concetti sui Large Language Models (LLM)

Let's practice!

Concetti sui Large Language Models (LLM)

Preparing Video For Download...