When should you trust LLMs

Large Language Models for Business

Iason Prassides

Content Developer, DataCamp

Maximizing Duolingo

  • Duolingo Max
    • Utilizes ChatGPT for human-like conversations

Image of the logo for the Duolingo Max product

1 https://blog.duolingo.com/duolingo-max/
Large Language Models for Business

Maximizing Duolingo

  • Duolingo Max
    • Utilizes ChatGPT for human-like conversations
  • Learners practice languages
    • GPT takes on different personalities
    • Gives instant feedback

Image showing the use of Duolingo Max on a mobile phone application

1 https://blog.duolingo.com/duolingo-max/
Large Language Models for Business

Understanding LLMs

  • Business leaders must truly understand LLMs
  • LLMs are not magic!
    • Often predicting the next part of a sequence
  • They do not understand language like humans

Magician with a wand performing tricks

Large Language Models for Business

Common sense in LLMs?

Man holding up a label that has "common sense" written on it

  • LLMs have no concept of what is "true"
  • They lack common sense
  • Human oversight is crucial for every use case
  • Sole reliance on LLMs can lead to unwanted implications
Large Language Models for Business

Training oversight

  • LLMs are dependent on the data used to train them
  • Bias or social injustice can be amplified by LLMs
    • Check responses to identify harmful outputs

Image of coins stacked to different heights, and different color game pieces on top of each stack

Large Language Models for Business

Is there bias?

  • LLMs use hateful comments to train on
    • Responses will also be hateful
  • Datasets containing biases
    • Create LLMs with systemic output biases
  • Can result in perpetuating prejudices, stereotypes, or misinformation
  • Developers added measures to prevent this
  • Businesses must be aware and ensure safeguards

Image showing the word misinformation

Large Language Models for Business

Flexibility and its consequences

  • Same prompt -> different responses
  • Helpful in creative pursuits, problematic for other uses
  • Marketing team creates product description using LLMs:
    • Customers see different descriptions
    • Brand image could be damaged

Man looking at two products in the supermarket that appear the same apart from the color of the packaging

Large Language Models for Business

Security alert!

  • LLMs are complex but can still be tricked
  • Attackers can embed harmful prompts on websites
    • Can distort LLMs outputs
  • Known as prompt injection
  • Could lead to reputational damage

Image of a read exclamation mark in a triangle indicating a security alert

Large Language Models for Business

The challenge with LLMs

  • Many challenges are industry-specific
  • Open Worldwide Application Security Project (OWASP)
    • Quantifying challenges
  • Businesses must be aware of concerns
  • Must understand the intricacies of LLMs
  • Ensure human oversight
  • Invest in training and security

Road sign showing a man jumping a hurdle with the words "Challenges Ahead"

1 https://owasp.org/
Large Language Models for Business

Let's practice!

Large Language Models for Business

Preparing Video For Download...