"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis

E27: “Google’s Med-PaLM and Med-PaLM2 with Vivek Natarajan”

May 21, 2023
Vivek Natarajan, a research scientist at Google Health, leads groundbreaking advancements with Med-PaLM and Med-PaLM2. He shares insights on the evolution of AI in healthcare, highlighting how these models outperform even top clinicians in certain evaluations. The conversation dives into the importance of specialized evaluation for AI in medical contexts and explores the use of soft prompting to enhance information delivery. With a focus on equitability in access to healthcare, Vivek emphasizes the need for responsible innovation and the transformative potential of AI.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
ANECDOTE

USMLE as Benchmark

  • The USMLE tests medical knowledge and reasoning through complex vignettes and multiple-choice questions.
  • While helpful for measuring AI progress, it's not a perfect indicator of real-world medical utility.
INSIGHT

Real-World Utility

  • Real-world utility in medicine is the primary goal, not just passing benchmarks like the USMLE.
  • The focus should be on practical use cases, such as answering consumer medical questions.
ANECDOTE

Flan-PaLM Shortcomings

  • Despite high USMLE scores, Flan-PaLM showed gaps in factuality, reasoning, and helpfulness on real-world questions.
  • Med-PaLM addresses these gaps by focusing on specialization and using expert physician demonstrations.
Get the Snipd Podcast app to discover more snips from this episode
Get the app