UA Libraries
Orientation to Generative AI
A workshop for faculty and instructors

Nicole Hennig & Michelle Halla
October 2024

ChatGPT news is everywhere
- Entrepreneur
- The Byte
It can feel overwhelming at times.
Experts don't seem to agree.

AI will cure cancer and solve climate change.
No, it will destroy humanity!
No, it won't destroy humanity, but it will lead to a flood of disinformation!
Some say educators are going through the five stages of grief
We were grieving the demise of the “old way” of doing things, caught in a cycle of denial, anger, bargaining, and depression.
Later, acceptance.
Developing AI literacy can help you cut through the hype.
Knowing what's possible can help you decide what's practical and ethical for your use cases.
We'll cover
  1. Basic definitions & terminology
  1. Use cases & examples
  1. How to learn more
Time for questions at the end
  • We'll record this session.
  • Link to these slides:
    unknown link
  • Add your questions, we'll answer at the end.
About me
Worked at the MIT Libraries for 14 years, first as webmaster, then head of UX department
2013 - 2018
My own business: technology training for librarians.
2019 to the present
E-learning developer at UA
I collaborate with my colleagues to create online tutorials.
Michelle Halla, Assistant Librarian
First-Year Experience Librarian
  • Instruction & outreach with an emphasis on underserved groups
  • Work with the Writing Program, Honors College, and UNIV 101/301
Previously
  • Started at University Libraries in 2015 in Access & Information Services
  • Provided reference in person and via chat and email
  • Oversaw 3D printing service at Health Sciences Library
  • Manager of Reference/Unit Lead for Reference Services 2017-2023
Poll
How frequently do you use ChatGPT or similar tools?
Some history
1956
1958 book for kids
"The computer can reason," he went on. "It can do sums and give information and draw logical conclusions, but it can't create anything.
"Ha! I feel better, Joe said. "People still have something the machine hasn't got."
"That's right. It's a wonderful, complex tool, but it has no mind. It doesn't know it exists."
– Danny and his friend, Joe
Terminology
These days AI is part of many services
  • Google and Alexa voice assistants
  • Netflix or YouTube recommendations based on your viewing habits
  • Gmail's spam filtering
Those are examples of narrow AI
  • AI designed to perform a specific task
  • AlphaGo beat the human Go champion in 2017.
General AI
Artificial General Intelligence (AGI)
Experts don't agree on a precise definition.
You've see all the AI in science fiction. ChatGPT is nothing like this! 😀
ChatGPT is not AGI
  • It's a general-purpose technology, but not AGI
  • Designed only to generate text in a conversational style
Where does ChatGPT fit?
What is machine learning?
A subfield of AI that involves developing models that can automatically learn from and improve their performance based on input data, allowing computers to make predictions, recognize patterns, and solve problems without being explicitly programmed for each task.
Without being explicitly programmed!
They learn patterns, instead of following rules.
They are probabilistic, not deterministic.
This is very different from explicit software programming and rules.
It's about patterns and probabilities.
There's no "if this, then that…"
What is deep learning?
It's a particular type of neural network known as a deep neural network, which consist of multiple layers of interconnected nodes that enable the learning of more complex patterns.
a model inspired by neural networks in the human brain.
It consists of processing units, known as nodes, organized into layers.
These nodes receive input, perform math computations, and produce output.
Knowledge is not explicitly represented
With billions of parameters in their networks, they don't have any easily interpretable "rules" or "knowledge base" that engineers can directly inspect or modify. Their knowledge is embedded in the connections between nodes in neural networks.
"a black box"
175 billion parameters in ChatGPT
It's not words and sentences.
It's a big network of math operations.
(don't call it a "database")
“In AI, the role of the programmer isn’t to tell the algorithm what to do. It’s to tell the algorithm how to train itself what to do, using data and the rules of probability.”
— Nick Polson and James Scott
AIQ: How People and Machines Are Smarter Together
Terminology so far
  • Narrow AI
  • AGI (artificial general intelligence)
  • Machine learning
  • Neural network
  • Probabilistic models
  • Black box
Why it's useful to
know these terms
  • Helps you to understand news stories (and journalists sometimes get things wrong)
  • Knowing that it's probabilistic helps you get what it's good or not good for
  • You'll know why no one understands exactly how answers are arrived at
A bit more history
Transformer architecture
Looks at the whole context at once, can pay attention to certain words.
LLMs are also called "foundation models"
They are becoming multimodal. Not only about language or text.
ChatGPT now has image creation, computer vision for uploading and talking about images, and works as a voice assistant (talk to it and it replies verbally) in the mobile app.
CEO: Sam Altman | Began as a non-profit | 2019 added a for-profit division
Microsoft invested $$$ | Open AI's vision | They are now planning to be for-profit.
Versions for researchers
  • GPT 1 (2018)
  • GPT 2 (2019)
  • GPT 3 (2020)
Added an easy chat interface and made it free for the public. (GPT 3.5)
ChatGPT
"Generative Pre-trained Transformer"
What was ChatGPT (3.5) trained on?
Training happens only every few months or maybe a year.
So the training data is not completely current.
Open AI hasn't made public what GPT-4 is trained on.
What is Common Crawl?
Founded in 2007
Common Crawl is a nonprofit organization that crawls the web and freely provides its archives and datasets to the public.​ ​​
Common Crawl's web archive consists of petabytes of data collected since 2008.​ I​t completes crawls every month.​​
When training is done: the data is set aside.
Once the training is done, it's not needed for the model to work.
Saved as patterns
It learned from all that data. It uses that to generate new text.
Instead of human-labeled examples, it's trained on a large amount of unlabeled data,
and it has sorted it according to patterns and categories it detects all on its own.
It then came to its own conclusions about how language works.
Geometric relationships
Every language has a shape
that AI discovers
Words that mean similar things are placed near each other, words that share a relationship, share a distance and direction.
175 billion dimensions
Distance and direction
man is to king as woman is to queen
The shape of English
The shape of all those relationships among the English language’s 10,000 most common words.
Words have different meanings depending on the context.
Semantic searching
  • Searching by meaning (from the geometric space)
  • Similar concepts don't need to include the same words, like in keyword searching.
After pre-training —>Fine-tuning
Changes from a word predictor to a conversation assistant
It learns from examples of conversations.
  • It's fine-tuned with high-quality, specific datasets like Q&A documents.
  • This sets the specific behavior instructions for the AI assistant.
How to prevent harmful use? Guardrails
Systems to guide behavior of the model. Aimed at preventing harmful or misleading content.
  • Content filters are added
You can give it a thumbs down
and submit a comment to OpenAI.
What is "generative AI?"
AI that can generate new content:
  • text
  • images
  • video
  • music
  • speech

This image was generated with Adobe Firefly.
It's important to know the difference between generative AI & other types of AI.
Discriminative AI
Classify or recognize patterns in existing data.
Examples
  • Spam filtering
  • YouTube or Netflix video recommendations
Generative AI
Creates new content based on learned patterns from existing data.
Examples
  • ChatGPT, Microsoft Copilot, Google Gemini
  • MidJourney, DALL-E
  • Runway, Eleven Labs
Today we'll focus on text generation.

Image generated with MidJourney
Why know the difference?
  • These types are often lumped together in news stories, but they are very different.
  • The strengths and weaknesses of discriminative AI are different from those of generative AI.
Facial recognition is discriminative AI.
Time to Ban Facial Recognition from Public Spaces and Borders.
Terminology so far
  • Transformer architecture
  • GPT: Generative pre-trained transformer
  • Large language model (LLM)
  • Common Crawl
  • Training data
  • Semantic search
  • Fine-tuning
  • Guardrails
  • Generative AI vs discriminative AI
Why it's useful to
know these terms
  • Gives you a bit of historical perspective
  • Helps you understand arguments about training
  • Helps understand safety issues (guardrails)
By the way, there is a history of scary words!
  • spiders
  • crawlers
  • bots
  • scraping
All of these terms were around long before ChatGPT… they are common technologies of the web for more than 20 years.
(The world wide web: WWW)
Beyond ChatGPT
ChatGPT is not the only LLM
All have both free and paid versions.
Grounding gives the model a tool
Grounding is an important concept in AI research. It is about connecting the model to external sources of knowledge - such as:
  • web search results
    (like Microsoft Copilot)
Which models are
grounded with search results?
Un-grounded models
Un-grounded models rely only their training data
These models will never be useful for very current information.
  • ChatGPT: up to January 2022
Tasks for un-grounded models
  • Generate ideas (for anything)
  • Come up with examples
  • Generate keywords for searching in research databases
  • Summarize long documents
  • Revise your writing in different styles
  • Revise your writing for different audiences or levels
  • Copy or upload parts of a research paper you don't understand and ask it to explain it in simpler terms
Think of these as wordsmiths and idea generators, not search engines.
What is "hallucination?"
The official term from the field of machine learning for outputting inaccurate information.
"making things up"
Always fact check.
There is ongoing research on
ways to mitigate hallucinaion
Grounded models
They have a search engine as a tool to work with.
Models with grounding
Models that answer based on their training PLUS …. a source of facts
Examples: web search results, or a database lookup.
  • Microsoft Copilot
  • Perplexity AI
  • Gemini
  • Elicit (uses Semantic Scholar database)
You can use grounded models for web search
  • Finding current information (past the training dates of ChatGPT or Claude).
  • Summarizing and linking to web search results
  • Searching a database of research papers, like Semantic Scholar (Elicit)
My favorite tool for this is Perplexity
perplexity.ai
Hallucination can still happen, even with grounded models.
But it's easy to recognize when it links to a source that's not relevant.
And or course, web search results can contain misinformation.
Remember that LLMs are probabilistic
It's not a database. It's a network of probabilistic math.
The exact same process will give slightly different results each time, since it’s based on probabilities.
It's worth learning a bit about prompting
Prompting: how you talk with it
It makes a huge difference in how it responds.
Give it context.
Role
“You are an expert in child development.”
“Act as an architect."
Context
"Using the theory of [insert a theory], create a …"
Give it clear instructions, including format.
Example:
"Use Mike Caulfield's ideas for how to be an effective fact-checker. Create a lesson plan for college freshmen to help them learn to fact-check information they find on the Internet. The outcomes should be that students become familiar with Caulfield's "Four Moves" for effective fact-checking, and that they gain practical skills to fact-check information they find on the Internet. Format the output in the way a typical lesson plan is formatted."
Formats
  • tables
  • emojis
  • ascii art
  • code
  • quiz questions
  • and more…
Don't expect the perfect answer right away.
Ask for changes, clarifications, or improvements. Tell it exactly what you want and what format you'd like it in.
Keep conversing — "ask and adjust."
Use the results as a starting point.
Use your own expertise to modify these. It's great for getting ideas. But you are the expert.
"Prompt engineering?" or "Prompt crafting?"
  • It's not like coding, you get slightly different responses each time.
  • It's more like talking to your human assistant and giving them detailed instructions.
  • People who are good at this often come from a background of writing or teaching.
Learn more at our workshop on Monday
Michelle - use cases
Learning more
Remember
  • It's a big network of math, not a database of words.
  • Ungrounded models are wordsmiths, idea generators, writing style changes.
  • Grounded models can summarize search results from the web or from research papers (specific tools like Elicit).
One page for all of our gen AI related materials
Our libguides
Tutorials
Library tutorials in D2L
FAQs about generative AI
Workshops
More workshops coming this fall.
  • A workshop for undergrad students on using AI to help with research tasks, Nov. 5 (Yvonne Mery)
Future workshop ideas:
  • Beyond ChatGPT (intermediate skills)
  • Data analysis with ChatGPT Plus
  • more…
Questions & discussion
Thank you!
nhennig@arizona.edu
https://nicolehennig.com
Michelle Halla
michellehalla@arizona.edu