Stenobird

Episode

Yejin Choi: teaching AI common sense and morality

Podcast
The Robot Brains Podcast
Published
May 17, 2023
Duration seconds
3398
Processing state
processed
Canonical source
https://www.therobotbrains.ai/who-is-yejin-choi
Audio
https://sphinx.acast.com/p/open/s/6053a29a0d11b0148adcfc96/e/646415dbbe93c900116fc524/media.mp3
JSON
/v1/public/podcasts/robot-brains-podcast/episodes/yejin-choi-teaching-ai-common-sense-and-morality
Markdown
/podcast/robot-brains-podcast/yejin-choi-teaching-ai-common-sense-and-morality.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/robot-brains-podcast/episodes/yejin-choi-teaching-ai-common-sense-and-morality/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/robot-brains-podcast/yejin-choi-teaching-ai-common-sense-and-morality.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

AI researcher Yejin Choi explores why large language models still struggle with fundamental common sense and social norms. She discusses the potential of using AI-generated data for training and the shift from symbolic logic to neural reasoning.

Topics

  • Natural Language Processing
  • Large Language Models
  • Common Sense Reasoning
  • Artificial Intelligence
  • Machine Learning
  • Neural Networks
  • AI Ethics
  • Data Quality

Highlights

  • Main idea: Large language models act as powerful unified engines for language-based reasoning and pattern recognition
  • Failure mode: Models frequently fail at simple, intuitive tasks that even children can solve, revealing a lack of true common sense
  • Practical takeaway: High-quality, curated data and iterative AI-to-AI criticism can be more effective than simply increasing training volume
  • Research direction: Moving toward 'truth-seeking' pre-training to prioritize correctness over mere linguistic probability
  • Future outlook: The evolution of LLMs opens new possibilities for personalized educational tools and intelligent tutoring systems

Chapters

  1. 5:05 The Power of Unified Models: An exploration of how LLMs use massive internet datasets to achieve unprecedented capabilities in language and reasoning.
  2. 9:20 The Quest for Truth-Seeking AI: Discussing the need to reformulate pre-training to focus on correctness and reliability rather than just pattern matching.
  3. 21:55 Hidden Behaviors in Neural Networks: Insights into unexpected empirical results and degenerate behaviors discovered within large-scale language models.
  4. 26:30 AI-Generated Data and Self-Correction: How using one AI to critique another can act as a form of reinforcement learning to improve data quality.
  5. 35:15 The Common Sense Gap: Defining common sense as shared human knowledge and identifying why models still fail at basic social norms.
  6. 43:50 New Frontiers in Education: The potential for neural language models to power transformative applications like intelligent tutoring and student assistants.
  7. 48:25 Personal Journey in AI: Yejin Choi reflects on her path from studying math and science to becoming a leader in the NLP field.