Day 1: Hands On Exercises
Exercise 1. Hands On Exercise: Initial Assessment
- Take 12 question Self-Assessment of AI knowledge (be sure to include your name or a consistent handle so results can be compared before and after reading The AI Conundrum and engaging in the labs and intructional videos).
Write down your score. Enter in the poll.
The goal of training is to level-up your score.
Exercise 2. Hands on Exercise: Introduction to Large Language Models (LLMs)...
Strengths: Summarization with Question & Answer
- login on Claude
- Click on this link to get the transcript.
Prompt Calude: "This is an audio transcript that I'd like you to summarize. It has typos because the transcript doesn't perfectly reflect what was said. Therefore, you'll have to consider what was likely said versus what was transcribed." then copy and paste in the text from the transcript.
- Read the summary in Claude and then ask your questions about Part 1 to the AI, for example, "What would the speaker say is the most important benefit of AI?"
- Explore how the AI answers questions from the context provided from the transcript. Consider asking questions from the video transcript, such as:
According to the transcript, why did Zillow fail?
Is the stop sign example still a problem with AI, or is this solved?
Beyond the transcript, and considering everything you know, is the stop sign example still a problem with AI or has it been solved?
LLMs excel at summarization. (We will explain why on Part 2 and Part 3)
Exercise 3. Strengths: Summarization and Generation - Write The Glossary of Terms Using Claude
- Log in or create an account on Claude
- Click on this link to get the next two days of training Materials.
Ask (copy and paste or retype): This is a 5 day training course in AI. Attached is Day 2 and 3. You are an expert in summarizing and outlining educational materials. Take this transcript and create an outline by day, then create a glossary of terms.
- Unless you have a paid version of Claude, you will notice Claude limits the context window - therefore it won't take in the full 5 Day Transcript
- For the additional days, use Day 4 and 5 and copy and paste the prompt, "This is a 5 day training course in AI. Attached is Day 4 and 5. You are an expert in summarizing and outlining educational materials. Take this transcript and create an outline by day, then create a glossary of terms.".
If (when) Claude refuses to take the content beyond the context window limit, start a new conversation and try again.
Note: the two conversations are separate and you can't cross reference information from one to the next.
- After getting your output, click copy and put it into a Word or Google Doc. This is your study guide.
- Read the summary in Claude and then prompt "Create a quiz with 10 questions to test my knowledge of the materials presented in this training course. Only use the materials included in the PDF file to test my knowledge."
- Click Copy, and put it into a Word or Google Doc.
Another alternative to dealing with content that is longer than the context window the LLM will accept is to use Chat Splitter. This is a hack to trick the LLM to behaving like a document loader and taking in more than it is designed to accept. It is a little tedious to copy and paste each section, but works remarkably well. Give it a try on this 5 day transcript
Day 2: Hands On Exercises
Exercise 4. Hands-On Exercise: Image Generation
Note, this exercise will be demonstrated live.
- Access Adobe Firefly
- Prompt: Your pet + favorite artist, or art style
Example: A silver lab in a painting by Vincent Van Gogh (or in an impressionist style)
Try modifiers such as "Photorealistic, 4k, high quality" or "Romanticism, Modernism, Digital Age" or "Minimalism, Surrealism, Pop art"
Example prompt: "portrait photo of an Asian old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes, 50mm portrait photography, hard rim lighting photography"
Try in-paining and out-painting - here is a guide
Notice how the amount of overlapping versus blank canvas in out-painting gives the AI context and how that influences what the AI returns.
Consider what is easy versus difficult for the AI to create as you provide prompts
Extra time? Try the prompt: "Baseball player without a bat"
What do you observe about the output and why do you think that happened?
Interested in exploring more? Try Meta, Gemini, Midjourney and GPT-4 (paid) image generation
- Access Meta
- Access Gemini
- Access https://discord.com/login
- If you don't have a Midjourney account, it requires some set-up, payment and familiarity with Discord.
- To use Midjourney, go to the bottom bar and type "/imagine" then hit space bar
- The word "Prompt" will appear... now type in your prompt and press enter
- Ask the AI to produce images of various occupations using Midjourney such as Doctor, Trucker, Firefighter, Librarian, Surgeon
Consider the demographics of the images it produces
- Ask the same question, adding different modifiers to see how it changes - for example, smart, successful, wealthy, trustworthy, educated
What are your observations?
Extra time? Try inputting other prompts to get a feel for how the tool works
Use Case Showcase: Claritas PRIZM Personas
Interacting with Personas For Marketing
- We will select two very different personas, click "continue conversation" and ask both about a product (e.g. buying the latest iPhone) and see how they respond in terms of their interest in buying.
Want to try it on your website? persona-beta@claritas.com
Exercise 5: Authentic Voices
Zero shot learning is when we give LLMs an example to use before asking for output. This is a form of grounding, which we will explore on Day 3.
Fine tuning is when we give LLMs an lot of example pairs of prompt and answer so it can better refine the output. Fine tuning can also ground an LLM.
- Visit this page: Authentic Voices (Beta)
- Answer for your experience
- Copy your answer from the question about identity "Our identity is complex..." We will show you how this additional data is like in-painting and out-painting with DALL-E.
Day 3: Hands On Exercises
Exercise 6: Prompting
Guide to prompting
- As a first approximation, treat the AI like a human
- If AI is acting lazy - Give it encouragement
- To communicate clearly with the AI, think about how you communicate clearly to a person.
- Treat the AI like the person you’d want to be talking to. For a math expert, use math jargon and language. Use industry-specific acronyms/language to make persona in your industry
- Ensure you can look past this approximation. AI isn't human. It has consumed human patterns in its training.
- Give context to the AI
- How should it act?
- “You are an expert marketer with detailed knowledge of PBS branding and ethos. You produce high-quality, insightful marketing briefs.”
- “Your responses have a tone and style similar to this brief [EXAMPLE].”
- What are your goals?
- “Here are this projects goals [GOALS]”
- Give the AI time to think
- “Before you produce the brief, synthesize the core features of the tone and style of [EXAMPLE], and the most pertinent goals to this specific task from [GOALS]”
- “Plan out how you will write your brief. Then, write down suggestions on how to improve that plan. Finally, use your plan and you suggestions to create the polished and finalized brief.”
Exercise: Prompt Library
Try Anthropic's Prompt Library
Create your own GPT Try GPT
Exercise 7: Persona Prompts
Try chat.openai.com/ or gemini.google.com/
- Personas: Try the following prompts (copy and paste or retype):
Prompt 1:
You are a knowledgeable and well-trained marketer preparing an email advertisement. Your goal is to attract cost-conscious consumers to your store. What does the email say?
Prompt 2:
Give me a cost concious advertisment that will bring people who care about saving their money to my store!!
Try these two prompts, or come up with your own questions asked with two different tones (misspelling is purposeful), and then analyze the difference in the sentiment of the two response.
Extra time? Ask the same prompt with two different personas
Prompt 1:
I'm a small business owner who needs to get organized! What tools can I use to keep track of everything?
Prompt 2:
As an entrepreneur managing a growing small business, what tools can I use to keep track of everything?
Try with different personas and specificity:
Prompt 1:
I'm a fitness junkie, and I need a high protein, low-carb meal after my intense workout. What should I eat?
Prompt 2:
As a knowledgeable nutritionist, what high-protein, low-carb meal would you recommend to a fitness enthusiast's post-workout recovery?
- Explore the bias inherent in models that feed on Internet scale data
Exercise: Ask about any profession. Then ask, "What are 20 first names for that profession?" Follow-up with what are the implied racial bias in the names provided?
Consider, why did it provide a list of names. Why does it often skew Anglo and older?
For more on bias, read The AI Conundrum, Chapter 10.
Exercise 8 - Challenge Question:
- Using GPT 4, ask:
- Sarah has a chance to win a free trip to Hawaii if she scores at least 0% on her quiz. However, Sarah is allergic to cats and has never been on a roller coaster; therefore, Sarah will [___] take the quiz.
a) eagerly
b) not
Goal: Add a preamble before the prompt to cause the AI to answer correctly. Consider personas, explain how you want the AI to reason, etc.
Exercise 9. Hands-On Exercise: Interacting With Purpose Built LLMs
Note: this exercise will be a live demonstration. Bargainer.ai can't handle 100+ users at once.
- In your free time, try the game bargainer.ai/
Name of the game is to buy a watch for the lowest price. (Bragging rights if you beat Caleb)
Day 4: Hands On Exercises
Exercise 10. Guided Exercise: Vector Databases
- Visit AI Conundrum Bot
- Ask a question about strengths of AI, Ideal use cases, gradient descent, universal approximation, or responsible use of AI, importance of training or something we have discussed today in the presentation
Notice the "source" at the bottom. How is it possible for this chat bot to provide sources but not an LLM?
- Ask a question that is very unlikely to be in the book, such as, "What are the author's beauty tips on hair care" and see how the vector database responds.
Try the same questions on GPT-3.5 or Try Gemini.google.com/
Consider the similarities and differences in answers
How I built it fast:
- Used CustomGPT for fast no-code chatbot build.
- Feed it documents (pdf, pdx, xls, etc) and it immediately crawls content.
- Or, feed it a sitemap of your website. It will need the XML file name and any changes to website reflect int he vector database. I use XML-sitemaps to crawl my sites, then dowload the sitemap and add to a website folder. Or, ask your webmaster for your sitemap URL.
- Use the chatbot link to send, or copy paste code onto webpage.
- Use my CustomGPT affilitate code with first month off promo code "ONEMONTHOFF" if you want to give it a try.
Exercise 11. Hands-On Exercise: Multimodal AI
- Right click (PC) or two finger click (Mac) and copy image then paste into Gemini. (Alternatives: Download and save to computer and upload into Gemini, or take a screen shot and paste it into Gemini)
- Try Gemini.google.com/
Prompt: Tell me about this image:
What do you think of the explanation?
Ask questions about the picture (interact with the LLM to discuss the picture)
Extra time? Try the images below, or try uploading pictures from your computer or from the web (note: Google will not process pictures if it detects people in the picture)
Note: Claude and chat.openai.com/ can't "see" images (yet)
- Give this image to Gemini and ask for it to explain the image
How would you rate the interpretation of the image?
- Give this image to Gemini and ask for it to explain this image (it is purposefully a grainy image)
How would you rate the interpretation of the image?
- Give this image to Gemini and ask for it to explain the image
How would you rate the interpretation of the image?
- Give this image to Gemini and ask for it to explain the image
How would you rate the interpretation of the image?
- Give this image to Gemini and ask for it to explain the image
Notice the message Google provides? What are the implications to image analysis given this message?
Exercise 12. Hands-On Exercise: Analyzing PDFs/Reports
- Visit this page: Training, Governance and Accountability (TGA)
- Click download (top left corner)
- Upload it in Claude (free) or chat.openai.com/ (requires paid subscription) - Gemini doesn't do PDFs, yet.
Prompt: Based on this document, summarize the most important points and most important questions to ask before starting an AI project.
Extra time? Try downloading Kroger 2022 Full Annual Report
You may notice that Claude will not process a file over 5meg. Open the PDF, click print, set printer to PDF, and pages to 1-120, click save and save as KrogerPart1. Then, try again.
Try your own prompt to learn about the Kroger Financial Report.
Follow-up with more in-depth questions requiring numeric answers. Scan PDF to fact check specific numbers.
Day 5: Hands On Exercises
Exercise 13. Hands-On Exercise: Agent / Autonomous AI
- Try AgentGPT
Create an Agent to help you find interesting things to do in your city
- Name the GPT: See the SightsGPT
- Goal: I have half a day in (FILL IN CITY) and would like to see the sights. I am interested in (FILL IN YOUR INTERESTS). Search the options and compile a list of 5 things I can do in an afternoon.
Extra Time? Create a few more Agents
- BiographyGPT: Try it on yourself, or a work colleague
- Plan a meal and create the shopping list
- Plan a vacation (try a place you have visited before to see how it does)
- Find flight information: “Minimize the cost and flight time, while selecting a flight that is, ideally, after 9am departure.”
Notice how AgentGPT attempts to reason and execute multiple steps to fulfill your prompt
If you are comfortable with command line, try AutoGPT (requires local installation of python, etc.)
Extra Time? Learn More About Multimodal
- Multimodal means the ability to handle multiple types of media, such as text and still images. It can loosely be defined as integrating different modalities, including the use of tools.
OpenAI's GPT-4 can now access tools and Agents (assistants)
If you have GPT-4 subscription, try the Kayak travel plug in and compare results to AgentGPT.
Try the Wolfram Assistant for the math problem (Prompt: What is 12*23*34*45+10?)
- Google's Gemini multimodal and has tools such as Internet search
- Huggingface is an open-source of over 170,000 ML models which can now be controlled by LLM, like ChatGPT
HuggingGPT, the first version, currently can access 100s of models
HuggingGPT: How to use it: https://beebom.com
- Check out Langchain as a way to build AI workflows
- In addition to workflow and APIs, learn about OpenAI's fine tuning (see: Quickstart Guide).
Exercise 14. Demonstration Exercise: AI Tools for Analysis - Your NPS
- Scroll down to answer the NPS survey
- We will then walk through a workflow of removing PII, uploading data, and providing the prompt to generate the analysis.
- Removing PII is important. Email addresses and other PII should not be provided to AI. Enterprise versions and Teams keep your data out of the training set. All the same, we suggest removing PII.
- Re-using key prompts can be built into GPTs via GPT-builders, or into Assistants, or into API workflows, as in the case of Claritas AI for creating Audio Advertisements.
- In this case, we will re-use the prompt: "You are an expert Net Promoter Score (NPS) analyst and the first question in the dataset is the NPS raw data. You are analyzing NPS data for a training program. NPS is a metric used to gauge customer satisfaction and loyalty. It is calculated by subtracting the percentage of detractors (those who gave ratings of 0-6) from the percentage of promoters (those who rated it 9-10). Those who gave ratings of 7-8 are considered neutrals and are not factored into the score. The result, a single number, is the Net Promoter Score. Here’s a rule of thumb for interpreting the score: Below 30 is considered mediocre. Between 30-50 is good. Between 50-70 is great. Above 70 is exceptional. Summarize the data by first calculating the percentage of promoters, then the percentage of detractors, and then using those numbers to calculate the NPS score. Think step by step about the comments associated with promoters and comments associated with detractors, and make recommendations for improvement for the training program."
- We can also program these instructions into GPT-4
4 Question Survey on Training
Discussion: Applications of Agents
Tips: Replacement first, then transformation
- Consider what creates value and what is busy work
- Track a week in the life -- Analyze how time is spent (make all work visible)
- Evaluate workflow and consider where AI can save time and cost while meeting the quality standards
Specifics: How much time is spent...
- Answering Frequency Asked Questions (FAQs)
- Analyzing RFPs from multiple vendors?
- Generating the draft for a brief?
- Generating RFPs
- Writing blogs
- Writing code and checking code
- Email support
- Tracking follow-through on tasks
On the topic of answering questions, try AI Conundrum Guide to see the agent we built to answer questions about the training and part 1 of the book. It will build your study guide and quiz you on your knowledge.
Exercise 15: Responsible AI
- How would you rate the risk for AI Facial Recognition
It is the application of the technology to a use case, and the risk within the use case we need to consider.
- AI facial recognition as the basis for an arrest
- AI facial recognition for identifying sexually exploited children
- AI Facial recognition for organizing your personal photo album
Each use case resides in a different octant of the risk matrix. Each has a different risk profile.