Learning Bayesian Statistics

#132 Bayesian Cognition and the Future of Human-AI Interaction, with Tom Griffiths

6 snips
May 13, 2025
Guest
Tom Griffiths
In this discussion, Tom Griffiths, a Henry Luce professor at Princeton, bridges psychology and computer science. He reveals how Bayesian statistics can enhance our understanding of human cognition and learning. The conversation touches on the importance of individual responses over averages, and how generative AI mirrors human cognitive processes. Griffiths explains the fundamental differences between human and machine intelligence, emphasizing the potential for AI to improve human decision-making while navigating challenges in language learning and alignment.
Ask episode
AI Snips
Chapters
Books
Transcript
Episode notes
ANECDOTE

Intuitive Prior Elicitation Demo

  • Tom Griffiths illustrated prior elicitation through a fun example about predicting movie grosses and lifespans.
  • He showed people implicitly use different prior beliefs shaped by real-world distributions in their judgments.
INSIGHT

Humans as Probabilistic Samplers

  • Humans can act as stochastic samplers, revealing their implicit probability distributions through their varied responses.
  • This human Markov chain Monte Carlo approach uncovers complex priors beyond simple numeric quantities.
INSIGHT

Bayesian Analysis of AI Priors

  • Large language models reflect strong prior influences, sometimes overly so, even on deterministic tasks.
  • Bayesian frameworks help compare human and AI priors to understand their distinct constraints and cognitive architectures.
Get the Snipd Podcast app to discover more snips from this episode
Get the app