Professional Documents
Culture Documents
Play
● “Tell me a joke about [...]”
● “Write an essay about [...]”
● “How to write a Python script to
[...]”
● “Draft a lesson plan on Columbus
Day for 9th graders”
Brown, T. B., et al. (2020). Language Models are Few-Shot Learners (arXiv:2005.14165). arXiv.
http://arxiv.org/abs/2005.14165
Training: Generate training examples
Training: Machine learning
From GPT to ChatGPT
Reinforcement
Learning from Human
Feedback (RLHF)
From GPT to ChatGPT
Chat Interface
Reinforcement
Learning from Human
Feedback (RLHF)
AI History and Competing Paradigms
Symbolic Subsymbolic, connectionist
● Rule-based ● Stochastic
(Engelbart, 1962)
lo gy
e
u ag act odo ing
ang rtif eth rain
L A M T
H → H-LAM/T
H → H-LAM/T
Instead of asking how powerful ChatGPT is, ask:
Created with DALL-E 2 (Prompt: student and (Zhao & Frank, 2003)
artificial intelligence working together)
Proposition 2: GenAI can better augment learning if
we precisely pinpoint the intelligence that is being
augmented, and how the system is transformed.
● What learning processes are appropriate for augmentation by GenAI?
● How is agency distributed between human and GenAI?
● How should other parts of the system (such as definitions of learning)
adjust?
● …
Autodesk Dreamcatcher
Source
Work with AI black boxes
Bearman, M., & Ajjawi, R. (2023). Learning to work with the black
box: Pedagogy for a world with artificial intelligence. British
Journal of Educational Technology.
https://doi.org/10.1111/bjet.13337
Principles of “effective ethical engagement” with GenAI
Context:
● A high-school class about World Religions course
● One teacher and 10 high school students
● Students were expected to examine various religions in the world and
develop religious literacy
Design Two phases of using ChatGPT in the class:
- Exploring the problem space (Pattern 1)
● Co-design partnership between - Co-creating knowledge to be presented in the
the teacher and researchers final essay (Pattern 2)
○ Problem definition
○ New ideas
○ Promisingness evaluation
○ Meta-dialogue
○ Comparison
○ Critical discourse
○ Higher-level ideas
(Bereiter & Scardamalia, 2017)
Utility of ChatGPT
● Information search
● Accomplishing mundane learning tasks, such as grammar check
Mechanisms of ChatGPT
Strengths:
● Interpreting user prompts
● Retrieving information efficiently
● Offering quick and clear responses
● Students perceived the information provided by ChatGPT to be rich and diverse,
representing different "facts'' that inspired them to generate new ideas
● Students appreciated how ChatGPT's responses were akin to those from a human
Weakness:
● Output quality: inaccurate or dated information, “black box”
● Limited cognitive capabilities
Results: Students’ AI Literacy
Risk and Societal Implications: Relationship with ChatGPT:
● In school: Potential abuse of AI in school ● A valuable tool that greatly supported their
settings, especially for students who might learning processes.
not have the opportunities to learn and
understand how to use AI properly. ● Primary source for information.
cbd@upenn.edu
bodong.ch