lecture 6 Flashcards
creativity is…
- original
- useful
GPT-3
o Very deep neural network
o Trained on text
o Very good at predicting the next word or fill in the blank
creativity score
(e.g., different ways to use a brick)
based on semantic distance
> how likely is a word to be the next word
how often the word occurs in a huge database of responses
score from 1-5
ways to measure creativity
Self-report
Creative insight problems
* Come up with something new (alternative uses test)
Remote associates test
* What word connects a number of concepts
Divergent thinking test
* Draw something with a circle in the middle, what are creative uses for a brick
creative thinking in humans
Originality-utility trade off
o The more original, the less usefulness
o truly creative = useful + original
Serial order effect
o Ideas get more creative over time in people
Optimal foraging
o People tend to search in categories
Humans vs GPT-3 – alternative uses test
- Semantic distance
- Humans have more semantically distant responses - Originality & utility
- Humans more originality, GPT-3 more usefulness - Surprise
- Humans score higher on surprise - Categories
- follows more of a category pattern
Conclusion
o Gpt-3 is close, but humans still win
4 Ps of creativity
- Product
output - Process
how does one come up with ideas - Person
how creative is the person - Press
effects of context/environment
GPT-3 easily ticks all these boxes, so we need new definitions and methods
Raven’s progressive matrices
Goal is to create models that can solve the raven to show AI is as intelligent as humans
Doesn’t reach adult level at this point but does outperform children
ARC
Abstraction reasoning corpus
Humans solve about 80% of the problems, the best AI model gets to about 20%
o Uses more of a brute force kind of algorithm
o Intelligent AI that solves this is yet to come up
what lies at the heart of human intelligence and creative thinking
Abstraction and analogy
