Essentials of Generative AI

Optimizing GenAI Models

Quick Recap

• GenAI designs various types of creative content,

including images, text, and music, emphasizing its
role in creative endeavors.
• LLMs like ChatGPT are advanced AI models that go
beyond understanding human language. It also
processes and analyzes data in various formats,
such as images, graphs, and charts.
Engage and Think

You are working on a document and must quickly

summarize a lengthy report. Instead of reading the
entire report yourself, copy and paste it into a GenAI
writing assistant. With a simple prompt like
Summarize the following text in 3-4 concise
paragraphs, the AI generates a clear and concise
summary for you in seconds.

How does the AI system understand and extract the

key points from a lengthy text to generate a concise
and structured summary from the simple prompt you
provided?
 Learning Objectives

By the end of this lesson, you will be able to:

Generate effective prompts for optimizing interactions with GenAI

models
Apply advanced prompting techniques, such as zero-shot, few-shot,
chain-of-thought, self-consistency, and tree-of-thought, to generate
the desired output from the GenAI model
Understand the concepts of pre-training and fine-tuning the GenAI
models to enhance domain-specific performance
Introduction to Prompt Engineering
 Prompt Engineering

Prompt engineering involves instructing conversational AI models, such as ChatGPT, to generate

desired outputs.

A well-built prompt must have:

Clarity

Specificity

Format

Context
 Importance of Prompt Engineering

It is increasingly recognized as a crucial skill set in the realm of GenAI. Here are several key reasons why
prompt engineering is needed:

Maximize AI potential

Encourage AI literacy Enhance user-AI

and creativity interaction

Support ethical AI Increase efficiency

use
 Prompts: Examples

Whenever you give a generic prompt, you'll get a random response each time, but engineered prompts
will bring back a relevant response.

Generic prompt Engineered prompt

Advanced Prompting Techniques
 Prompting Techniques: From Basics to Advanced

Listed below are some advanced techniques for prompting GenAI models to achieve desired outcomes.

Zero-shot prompting Few-shot prompting

Chain-of-thought prompting Self-consistency prompting

Tree-of-thought prompting
 Zero-Shot Prompting

It is a technique that provides a quick response with minimal guidance, helps in setting the context, and
gives the desired output without any specific examples.

Its existing knowledge is used for task execution.

Do It on Your Own: Zero-Shot Prompting

Use ChatGPT to try out the prompt below:

Generating social media posts for a brand

Do It on Your Own: Zero-Shot Prompting

Use ChatGPT to try out the prompt below:

Performing a SWOT analysis

Do It on Your Own: Zero-Shot Prompting

Use ChatGPT to try out the prompt below:

Building a to-do list

 Few-Shot Prompting

It is a style of guiding GenAI models with specific examples to achieve desired responses.

Provides the model with a limited number of specific

examples and a definite context

Focuses on providing training sessions to the GenAI bot,

enhancing the relevance and precision of the output

Offers a clearer idea of the specific task or style while

leveraging the model’s existing knowledge
Do It on Your Own: Few-Shot Prompting

Use ChatGPT to try out the prompt below:

 Chain-of-Thought Prompting (CoT)

It is a method of providing sequential prompts to the GenAI model to generate sequential outputs,
rather than getting the entire output at once.

It encourages more thoughtful and detailed responses by

avoiding the overload of multiple questions at once.
Do It on Your Own: Chain-of-Thought Prompting

Use ChatGPT to try out the prompt below:

 Self-Consistency Prompting

It ensures that a GenAI bot delivers reliable and similar responses when asked the same
questions repeatedly.

Example

Repeatedly prompting "What's the capital of

France?" should consistently yield the answer Paris.

This process makes the AI more reliable, reduces confusion, enhances the user experience,
and paves the way for efficient learning.
Do It on Your Own: Self-Consistency Prompting

Use ChatGPT to try out the prompt below:

 Tree-of-Thought Prompting (ToT)

It is a method in GenAI models that helps solve complex problems.

It forms a tree with various thought paths or

steps, resembling branches on a tree.

Each branch represents a potential step or idea

toward solving a problem.

The AI evaluates these branches to identify the

best solution.

It is useful for tasks that require strategic thinking or involve multiple steps.
 ToT Prompting: Example

Use ChatGPT to try out the prompt below:

Note: The AI uses ToT to break down the problem into smaller steps, evaluate each
step, and arrive at the correct solution.
Best Practices for Creating Prompts
 Best Practices: Prompt Creation​

Dos

• Begin the requests with clear and specific framing.

• Example: List 5 Paris tourist attractions with 2-3 sentence descriptions.

• Offer context by providing background details​.

• Example: I'm writing an article on sustainable farming practices, so I need information

on the benefits of crop rotation and how it improves soil health.
 Best Practices: Prompt Creation​

Dos

• Ask follow-up questions if the response lacks completeness​.

• Example: Your explanation of photosynthesis was helpful, but could you also briefly
explain how plants use the energy produced during that process?

• Provide guidelines early on regarding tone, length, and format​.

• Example: For this blog post, I need around 500 words written in a conversational,
lighthearted tone with examples or analogies to explain complex concepts. Please use
headings to break it into 3-4 sections.
 Best Practices: Prompt Creation​

Don’ts

• Don’t provide unnecessary details​.

• Example: Just give me the ingredients list for a basic spaghetti recipe, nothing more. I
don't need details on preparation steps, cooking times, or serving suggestions.

• Don’t give it too many tasks to perform at the same time​.

• Example: First, draft an outline for the introduction section of the report. Then,
research and write the background information. After that, analyze the data and
summarize the key findings. Let's work on one component at a time.
 Best Practices: Prompt Creation​

Don’ts

• Don't provide any contradictory statements while prompting.

• Example: ​Describe a utopian society where everyone lives in harmony, but also
include details about conflicts and divisions between different groups.

Note: Always validate the information given by the tool, as the GenAI model might
generate an inaccurate response.
Demo: Refining Prompts

Duration: 15 minutes

In this demo, we will continue iterating on the prompt, demonstrating techniques like:

• Asking clarifying questions about elements that are still vague or unclear
• Providing examples to illustrate the desired type of response
• Adding or removing specificity as needed
• Setting guidelines for length, tone, structure, and more

Note: Please download the demo documents from the LMS Reference Material Section for a
detailed explanation.
Introduction to Fine-Tuning
 Pre-Training

Pre-training is the process of training a GenAI model on a large corpus dataset.

Prompt

LLM

Pre-trained model Large dataset

Generated output
 Fine-Tuning

Fine-tuning involves taking pre-trained models and training them more on smaller, specific
datasets to enhance performance in a particular task or domain.

Why is network
traffic encrypted?

Why is network Network traffic is encrypted using

traffic encrypted?
complex algorithms to prevent
unauthorized interception. This
safeguards sensitive data from
Because it protects data potential attackers, ensuring
during transmission confidentiality and integrity.
Common encryption protocols
include SSL/TLS, IPsec, and WPA2 for
web, VPN, and wireless traffic,
respectively.

Pre-trained model Fine-tuned model

 Benefits of Fine-Tuning

Fine-tuning ChatGPT offers several benefits, including:

Enhanced performance

Adaptation to domain-specific data

Reduced training time and cost

 Improved Performance and Domain Adaptation

Fine-tuning GPTs on task-specific data helps them understand the subtle details and
complexities of the target domain.

Encrypted files with the .zzz Encrypted files with the .zzz
extension, ransom note extension, ransom note
demanding payment? demanding payment?

LLM LLM

Pre-trained model Fine-tuned model Cybersecurity data

Potential malware Likely a ransomware attack using the .zzz

infection ransomware variant. Immediate actions: isolate
infected systems, disable network shares, identify
patient zero, and initiate an incident response plan.
 Considerations for Fine-Tuning Language Models

Fine-tuning isn't always the best choice. Consider this scenario: You've fine-tuned a base model using
your Slack messages dataset.

Write a report analyzing the impacts of climate change.

Sure, I can work on that report tomorrow and have it

ready for you.

No, I need you to write it now. I'm on a tight deadline.

Understood. Let me begin drafting that report analyzing

climate change impacts right away.
 Challenges of Fine-Tuning

Undesirable fine-tuning behavior where an LLM excels on training

Overfitting
data but struggles with new tasks or domain-specific data.

Fine-tuning may cause the LLM to forget all pre-training knowledge,

Catastrophic forgetting
leading to catastrophic forgetting.

Fine-tuning LLMs for specific tasks demands substantial data, which

Expensive is difficult to collect and expensive. Complex computing
infrastructure adds to the cost.

Pre-trained LLMs may carry biases, which can worsen during fine-
Bias tuning. Bias mitigation measures should be in place before fine-
tuning.

Note: The process of fine-tuning is time-consuming and expensive to implement.

 Guided Practice

Overview Duration: 30 minutes

In this activity, you will craft an effective prompt by applying advanced prompting techniques. This
activity will help you gain hands-on experience with writing prompts.

Note: Please download the demo documents from the LMS Reference Material Section for a
detailed explanation.
 Key Takeaways

Prompt engineering involves creating clear and effective instructions

to guide large language models in generating desired outputs.
Advanced techniques in prompt engineering, such as zero-shot and
few-shot methods, assist in creating effective prompts.
Fine-tuning customizes large language models for specific tasks or
fields using specialized data.
 Additional Resources

• How to Become a Prompt Engineer: Your Path to AI Mastery

Q&A