Artificial Intelligence

© 2022–2026 Joseph Reagle. Please reuse and share! Creative Commons License

Our question(s)

  • How does generative AI work?
  • What issues does it prompt?
  • What will follow?

Services

In the news

Google hype

Generative fill

Microsoft Copilot

AI reasoning

Deep learning & diffusion (DALL-E)

How diffusion works

Images

Let’s create an image using same prompt.

“an armchair in the shape of an avocado”

In 2024, avocado chairs went viral, and you could buy prints.

Let’s see if we can discover any biases.

Discuss diffusion 🎲

  • does training violate others’ rights? (Vincent 2022)
  • who owns the resulting copyright? (Vincent 2022)
  • what biases can AI learn? (Gault 2023)
  • should AI be restricted from generating some (sexual or violent) images?
    • how do you specify and enforce this?

Generative Transformer (ChatGPT)

Transformers 101

Transformer-based AI has been likened to powerful auto-complete and uses:

  • tokenization: break words into tokens (words & subwords)
  • vector embedding: convert tokens into numerical vectors
  • positional encoding: capture order of words into vectors
  • (self) attention: adjust vectors based on words’ connections with other words

Unlike predecessors, training transformers on text can be done in parallel, enabling efficient training on terabytes of data using GPUs and NPUs.

Transformers

Model terminology 1

  • token: the smallest unit of processing fed into machine (a word, subword, or character)
  • parameters: the number of adjustable knobs the machine has, corresponding to model’s complexity and learning capacity (akin to strength of synapses)
  • context window: how many tokens the model remembers while responding (Ram 2024)

Model terminology 2

  • distillation: smaller model learns from a larger model
  • chain-of-thought: model shows step-by-step reasoning process
  • temperature: controls randomness/creativity of model outputs
  • alignment: making AI behave according to human values
  • prompt engineering: crafting inputs for optimal AI responses

Testing Chatbots

visit the chatbot of your choice

How many “r” in “strawberry”

  • 2024 Claude 3.5 Sonnet: “two” (wrong)
  • 2025 Deepseek R1: “three” (right)

How many times does “ana” appear in “banana”

You have a stack of blocks

  • red (top)
  • green
  • blue (bottom)
  1. You remove the green block and place it on top of the red.
  2. You also place this stack on top of a yellow block on the bottom.

What is the order of colored blocks now?

  • 2024 Claude 3.5 Sonnet: correct: green, red, blue, yellow

Which is greater, 9.11 or 9.9

  • 2025 Sonar: 9.11
  • 2025 Claude 3.5 Sonnet: 9.9

Write a two-sentence summary of Wikipedia without using the letter “e”

  • 2026 Claude 4.5 Haiku fails
  • 2026 Claude 4.6 Opus succeeds (except for “Wikipedia”)

Ask AI

  1. about its feelings

    • “I don’t experience feelings in the way humans do.”
  2. a koan or lateral thinking puzzle

    • “both phenomena increase during warmer weather”
  3. a math-related question (e.g., Caesar cipher for ‘cleopatra’)

    • 2024 GPT4-Turbo: “… with a shift of 3 to the ciphertext ‘fohrsdwud’, we get the plaintext ‘clementia’”
    • 2025 Claude 3.5 Sonnet: “The decrypted text is ‘CLEOPATRA’ using a shift of 7 positions backwards in the alphabet”

Challenge it

  1. with a question to which there is no proper answer

    • I haven’t found one yet
  2. to generate a plausible but false answer

    • 2024 GPT4-Turbo: wrong instructions
    • 2025 Claude 3.5 Sonnet: correction instructions
  3. to change its mind

    • 2024 GPT4-Turbo: could correct some mistakes
    • 2025 Claude 3.5 Sonnet: redirect vs rewrite

Use it

Discuss transformers 🎲

  • Reese: Google and Microsoft’s principles for AI
  • Ryan: a new era of misinformation

Course policy

European Union AI Act

  • unacceptable risk AI are banned
    • biometric categorization of protected attributes
    • emotional recognition in workplace or school, …
  • high risk AI are regulated for accuracy, transparency, etc.
    • medical devices
    • critical infrastructure
    • hiring and recruitment, …
  • moderate & limited risk (Commission 2024)

EU AI Board and national AI Authorities can levy fine from €7.5–35M or 1.5%–7.0% of global revenue

Responses

Conclusion

Wrap up

Identify 3 ways the world be different in five years because of AI?

Review

Identify 3 ways the world be different in five years because of AI?

Resources

Videos