Yonatan Zunger, CVP of AI Safety & Security at Microsoft, shares his journey from theoretical physics to AI leadership. He distinguishes between generative and predictive AI, emphasizing their unique strengths and ethical implications. Zunger highlights the need for proactive safety measures and diverse perspectives in AI development. With real-world examples, he illustrates both the benefits and risks of AI applications. The discussion encourages critical thinking about AI's evolving role in society and the importance of designing for safety.
53:34
forum Ask episode
web_stories AI Snips
view_agenda Chapters
auto_awesome Transcript
info_circle Episode notes
insights INSIGHT
Generative vs. Predictive AI
Generative AI excels at summarizing content and role-playing various characters.
Predictive AI focuses on predictions, classifications, and recommendations from large datasets.
insights INSIGHT
AI Analogy to Human Vision
Predictive AI works like the lower-level processing in human vision, extracting features from large datasets.
Generative AI is analogous to higher-level processing, narrativizing and reasoning with extracted information.
question_answer ANECDOTE
AI Bias in Sentencing
AI-powered sentencing recommendations in criminal law ended up being racially biased.
The system unintentionally predicted race by using proxy variables like income and neighborhood.
Get the Snipd Podcast app to discover more snips from this episode
While we are on our winter publishing break, please enjoy an episode of our N2K CyberWire network show, The BlueHat Podcast by Microsoft and MSRC. See you in 2025!
Yonatan Zunger, CVP of AI Safety & Security at Microsoft joins Nic Fillingham and Wendy Zenone on this week's episode of The BlueHat Podcast. Yonatan explains the distinction between generative and predictive AI, noting that while predictive AI excels in classification and recommendation, generative AI focuses on summarizing and role-playing. He highlights how generative AI's ability to process natural language and role-play has vast potential, though its applications are still emerging. He contrasts this with predictive AI's strength in handling large datasets for specific tasks. Yonatan emphasizes the importance of ethical considerations in AI development, stressing the need for continuous safety engineering and diverse perspectives to anticipate and mitigate potential failures. He provides examples of AI's positive and negative uses, illustrating the importance of designing systems that account for various scenarios and potential misuses.
In This Episode You Will Learn:
How predictive AI anticipates outcomes based on historical data
The difficulties and strategies involved in making AI systems safe and secure from misuse
How role-playing exercises help developers understand the behavior of AI systems
Some Questions We Ask:
What distinguishes predictive AI from generative AI?
Can generative AI be used to improve decision-making processes?
What is the role of unit testing and test cases in policy and AI system development?
CyberWire Daily