Seen a Good AI Explainer?

OEGlobal Plaza
1

Stirring up the AI pot again here… surely there’s not too much talk going on? I for one am curious for what people have seen / read that help an educator,student, administrator understand how it works.

What’s your current go to “explainer”? Tell us!

DALL·E generated image: A wise owl standing on a large rock with hand in air explaining a complex topic, digital art
DALL·E generated image: A wise owl standing on a large rock with hand in air explaining a complex topic, digital art1792×1024 124 KB

Generated image, using the extra generation frame, spawn by Alan Levine using DALL-E with prompt “A wise owl standing on a large rock with hand in air explaining a complex topic, digital art” Who knows how to license this stuff?

The most recent is a well crafted New York Times piece "Let Us Show You How GPT Works — Using Jane Austen ($wall link but I find you can get a readable version if you click from a google search result).

It’s impressive in showing the interactions involved using a smaller scale version of AI (“BabyGPT”) running on a laptop, trained on content from sources you can pick such as Jane Austen, Shakespeare, Star Trek, Harry Potter, and a few more. It’s illustrative to see how it goes from pure random characters, and iterates many times, to first produce basic random words, and eventually something similar to the source material.

Helpful, but still-- what is it doing?

A few others I have tracked include:

I meant to list only 3 but here are many more… what are you using to explain this rapidly moving spectacle to your colleagues, family, kids? And then what the heck does this mean for our open education perspective?

2

Oh I will admit in January I did the classic “Ask ChatGPT” (v3.5) to explain itself.

Alan: I do not understand how ChatGPT actually works. Please explain in non-technical terms exactly what happens after I enter a question here and press return.

When you enter a question and press return, the question is sent to the ChatGPT model which is running on a computer. The model then analyses the question and generates a response to it. The response is based on the patterns and relationships it has learned from the large amounts of text data it was trained on. The response is then sent back to you, where you can see it displayed on your screen.

Alan: That is too simplistic and general. Describe in much more detail step by step what you mean by “analyses the question”- how are you analyzing? And what exactly are you doing step by step when “generating a response”

When the ChatGPT model “analyzes the question”, it is essentially breaking down the text of the question and trying to understand its meaning. Here are the steps it goes through:

  1. The model first tokenizes the input text, which means breaking it down into individual words or phrases.
  2. Next, the model converts each token into a numerical form, known as embeddings. These embeddings are a way of representing the words in a way that a computer can understand and process.
  3. The model then uses these embeddings as input to feedforward neural network, which is used to analyze the meaning of the input text.
  4. The model uses a transformer architecture which is a method of processing sequential data like text, which allows the model to understand the relationships between words in the input and their context. The transformer architecture is a type of neural network architecture that has proven to be very effective in natural language processing tasks.
  5. Finally, the model uses the output of this analysis to generate a response.

To generate a response, the model:

  1. Uses the output of the analysis to understand the intent and meaning of the input text.
  2. Selects a subset of the training data that is most relevant to the input text.
  3. Uses this subset of the data to generate a response that is similar to what it has seen before.
  4. The model uses a decoder network that generates the output text word by word.
  5. The model uses a technique called beam search to generate response by considering multiple options at each step, and choosing the one that is most likely based on the input and the model’s previous predictions.
  6. Finally, the response is generated and sent back to you.

Please note that the above steps are a high-level description of the process and the actual implementation can be quite complex and might vary among different models.

Alan: List examples of the embeddings created for my previous question.

I’m sorry, but I am not able to list specific examples of the embeddings created for your previous question because the embeddings are numerical representations of the words, and it’s not human-interpretable. However, for each word in your input, the model creates an embedding vector that represents the word in a numerical form. These embeddings are learned during the training phase and are used to represent the words in a way that the model can understand and use for its analysis and response generation. The embeddings are learned by the model based on the patterns and relationships it finds in the training data, so each word’s embedding will be unique and specific to the context in which it was learned.

So these numerical representations of words as data vectors is beyond us as humans, eh? Harrumph.

3

Hi Alan
I’m preparing an AI webinar, for faculty and students I work with, using complexity theory and the cautionary principle. I’ll add " ChatGPT Is an Ideology Machine" and “There Is No A.I.”.
I agree with Kate Crawford (Microsoft Research and USC): “AI is neither artificial nor intelligent.” It’s just another tool.

4

I came across this article by Mark Riedl on Mastodon: A Very Gentle Introduction to Large Language Models without the Hype. It says it’s a work in progress, and I still have yet to read through it in detail, but it looks useful on a first and second glance. It talks about AI, Machine Learning, Neural Networks, Large Language Models, and more.

It does get fairly technical, but is meant to be understandable by people w/o a computer science background. it’s also pretty long, just FYI!

1 Like
5

Thanks Mark. Lanier’s position in the New Yorker piece really resonated with me. It’s practical thinking but also far from the loud hyping.

Will be interested in hearing response from your webinar, I’d guess most people are trying to sort the unknown near future, which was usually best when it was tinged with excitement, but now?

Thanks again for sharing here

6

I agree, Christina, Mark’s writing and approach to explain in plain language is commendable-- I like the opening metaphor of the self driving car with the concept of gates. It’s still a lot of my own internal neural network to grok!

7

Hi Alan,
Similar to others here I am preparing an AI webinar series. (Starting this Friday and 30 mins every Friday in May in case anyone is interested)

Here’s some good stuff I’ve found:

  1. AIandyou has a great mission to teach marginalized communities about AI and offers some really great free resources on their site. Their founder is brilliant and is part of the National Artificial Intelligence Advisory Committee (NAIAC).
  2. Engage AI Institute has this glossary of AI terms for teachers
  3. I’ve heard from teachers that are actually going through Google’s Applied Digital Skills Discover AI in Everyday Life project-based lesson to learn more.

There’s so much out there! These are just three that came to mind. At World Education, we’re hosting this four-part webinar series and then looking ahead to develop some educational resources (open, of course!) primarily for adult educators in the U.S. who teach adult basic education or English for Speakers of Other Languages (ESOL).

8

Thanks Rachel for sharing the webinar series and your set of 3 resources… looking forward to seeing what World Education develops for related OERs (and come back here to ask for suggestions).

I know @annarmills has been casting a net for potential OERs to help advance AI critical literacies-- her Writing Across the Curriculum collection on AI Text Generators and Teaching Writing: Starting Points for Inquiry is supremely useful.

9

Here comes a pile of links I yanked from the chat at today’s Internet Archive webinar on - Generative AI Meets Open Culture: Opportunities, Challenges, and Ethical Considerations

AI @ IA: https://blog.archive.org/tag/ai-research/

Extracting Words Sung on 100 year-old 78rpm records: https://blog.archive.org/2023/04/12/aiia-extracting-words-sung-on-100-year-old-78rpm-records/

https://github.com/tesseract-ocr/tesseract

TV Third Eye: https://mastodon.archive.org/@tv

https://blog.archive.org/2023/04/28/internet-archive-weighs-in-on-artificial-intelligence-at-the-copyright-office/

https://twitter.com/MarkGraham/status/1640510980714283009

A task we would love help on: https://blog.archive.org/2023/04/24/ai-audio-challenge-audio-restoration-based-on-expert-examples/

Interesting article from Canada: https://www.theglobeandmail.com/arts/art-and-architecture/article-as-ai-debate-swirls-artists-are-torn-between-embracing-it-and-trying/

This is very experimental: https://newsum.sawood-dev.us.archive.org (e.g. it does not yet filter out ads)

https://www.techwontsave.us/episode/151_dont_fall_for_the_ai_hype_w_timnit_gebru

https://www.popsci.com/technology/chatgpt-conspiracy-theory-misinfo/

Copyright Law https://crsreports.congress.gov/product/pdf/LSB/LSB10922

Wikilegal/Copyright Analysis of ChatGPT: https://meta.wikimedia.org/wiki/Wikilegal/Copyright_Analysis_of_ChatGPT

https://mlc.ai/mlc-llm/

https://huggingface.co/bigscience/bloom

https://arxiv.org/pdf/2302.13971.pdf

https://popula.com/2023/04/30/yakkin-about-chatgpt-with-david-roth/

Creative Commons blog post: Better Sharing for Generative AI: https://creativecommons.org/2023/02/06/better-sharing-for-generative-ai/

All CC AI blog posts: https://creativecommons.org/tag/ai/

Introducing Democracy’s Library: https://blog.archive.org/2022/10/19/announcing-democracys-library/

https://openml.fyi

https://archive.org/details/democracys-library

Curious to know whether any of the panellists have been following the discussion amidst Indigenous communities (aka Native American) and their representation on Wikipedia. https://capacoa.ca/en/2023/01/better-representation-of-indigenous-artists/

Useful article from Maria Bustillos: Just Because ChatBots Can’t Think Doesn’t Mean They Can’t Lie https://www.thenation.com/article/culture/internet-archive-publishers-lawsuit-chatbot/

Holden’s work on SIFT for factchecking: https://hapgood.us/2019/06/19/sift-the-four-moves/

Movement for a Better Internet: https://www.movementforabetterinternet.org/

Generative AI Report: https://www.movementforabetterinternet.org/resources/mozfest2023

https://lightning.ai/pages/community/community-discussions/the-ultimate-battle-of-language-models-lit-llama-vs-gpt3.5-vs-bloom-vs/

https://yalebooks.yale.edu/book/9780300209570/atlas-ai/
https://whoseknowledge.org/

a good list of (more or less) Opensourced LLM’s : https://medium.com/geekculture/list-of-open-sourced-fine-tuned-large-language-models-llm-8d95a2e0dc76

CC has been working on how to help ensure indigenous knowledges/cultures can participate in the commons on terms that meet what we call “better sharing” principles: sharing that is inclusive, just, and which inspires reciprocity. I bet you know https://localcontexts.org/labels/traditional-knowledge-labels/

https://blog.papareo.nz/whisper-is-another-case-study-in-colonisation/

https://www.amazon.com/Manna-Two-Visions-Humanitys-Future-ebook/dp/B007HOH67U

How Copyright Law Can Fix Artificial Intelligence’s Implicit Bias Problem: https://digitalcommons.law.uw.edu/wlr/vol93/iss2/2/

About the Densho / Wikipedia project: What Happens When Everyone who Experienced an Event is Gone?: https://blog.archive.org/2019/09/26/what-happens-when-everyone-who-experienced-an-event-is-gone/

Try this: http://milweesci.weebly.com/uploads/1/3/2/4/13247648/mannapdf.pdf

Or this: https://marshallbrain.com/manna1

Here’s the article that Jacob was mentioning earlier: https://blog.papareo.nz/whisper-is-another-case-study-in-colonisation/

A year ago on MIT Review: https://www.technologyreview.com/2022/04/22/1050394/artificial-intelligence-for-the-people/

Luis’s newsletter: https://openml.fyi/

FWIW, I was really impressed with the materials/bibliography of this undergrad course called AI Images by Erik Salvaggio: https://www.cyberneticforests.com/ai-images

AI for Education is about to get very interesting: TED: The amazing AI super tutor for students and teachers - https://www.ted.com/talks/sal_khan_the_amazing_ai_super_tutor_for_students_and_teachers/

2-sigma problem - https://www.jstor.org/stable/1175554

AI replacing stock images at shutter stock → https://www.bensbites.co/p/ai-art-officially-replaces-stock-images

https://www.businessinsider.com/ceo-buys-chatgpt-plus-accounts-all-employees-sees-productivity-boost-2023-5

This isn’t fun for everyone… this is life and death for many lived realities. I’m not sure if everyone is aware the background of OpenAI, owned by technocratic billionaires such as Musk, invested by Peter Thiel of Palantir, responsible for such shameful oppression such as documenting migrants for deportation: https://www.vice.com/en/article/pkeg99/palantirs-ceo-finally-admits-to-helping-ice-deport-undocumented-immigrants

have we mentioned already the problem of deep fakes etc. ?? https://www.voanews.com/a/research-deepfake-news-anchors-in-pro-china-footage/6953588.html

So many self-aggrandizing myths are being pushed. See: What Tech Calls Thinking by Adrian Daub: https://www.adriandaub.com/books/what-tech-calls

I’m following https://contentauthenticity.org/ and other initiatives closely.

Generative AI Report: https://www.movementforabetterinternet.org/resources/mozfest2023

10

I’m leading a discussion on AI & HE in the context of
COMPLEXITY, TECHNOLOGY, RESEARCH, & LEARNING
at the Western Institute for Social Research
June 24th 10 am to noon Pacific time (-07:00 GMT)
Everyone is welcome!
Zoom link: Meeting Registration - Zoom