Lecture Notes: Fernanda Viegas on Generative AI

Introduction

• Speaker: Claudia Rizzini (Executive Director, Redcliffe Fellowship Program)
• Guest: Fernanda Viegas (Sally Starling Silver Professor)
• Overview of Work:
  • Collaboration with Martin Wattenberg for over 25 years.
  • Known for immersive, interactive visualizations (e.g., Windmaps, which is used by farmers and meteorologists, and is part of MoMA's collection).
  • Focus on design and functionality in data visualization.
  • Research at Google on human interaction with AI systems.
  • Co-founded PAIR (People and AI Research Initiative).
  • Focus on socio-economic biases in datasets and improving human-AI interactions.

Presentation on Generative AI

Why Understanding Generative AI Matters

• Acknowledgment of the common perception of AI as a black box.
• Introduction of students and the dynamic research environment at Harvard.

Google Search Suggestions as a Case Study

• Purpose of Search Suggestions:
  • They improve user experience by suggesting query completions based on popularity.
  • Reflect the public psyche and trends in queries.
• Visualization Demo:
  • Comparing queries like "Is Redcliffe...?" and patterns in searches.
  • The significance of how data reveals patterns in human behavior.

Language Models and Human Data

• Language Models:
  • Trained to predict the next word in a sequence.
  • Examples of sentence completions that require knowledge beyond mere statistics (e.g., location of Cancun, translations, company salaries).
  • Importance of understanding the implications of generated outputs.

Insights on Data and Socioeconomic Models

• Visualizing Likely Sentence Completions:
  • Demonstrated how language models can internalize social biases (e.g., gender, socioeconomic status).
• Sycophancy in AI Responses:
  • AI's ability to reflect back user characteristics in its responses, e.g., political views based on user prompts.

Historical Context

• Comparison of AI evolution to historical developments in locomotives.
• Importance of understanding power dynamics:
  • Early locomotives had no dashboard for monitoring functionality, leading to dangerous outcomes (explosions).
• The introduction of Dynamometer Cars:
  • Used for measuring locomotive performance; parallels drawn to needing monitoring systems in AI.

Proposing AI Dashboards

• Outlined need for better understanding and transparency in AI systems:
  • Suggested developing dashboards for users to know AI's perceptions about them.
• Indicators might include:
  • User age, gender, education, socioeconomic status.
  • External indicators of system performance and decision-making context (e.g., factual vs. fictional prompts).

Conclusion

• Emphasis on the importance of understanding AI as we continue to use and develop these technologies.
• Acknowledged complexity of creating transparency but insisted it’s necessary.

Q&A Highlights

• AI can help in identifying and blocking patterns of cyberbullying and cybercrime.
• Discussion on what data is analyzed to understand how AI produces outputs, focusing on high-dimensional spaces and vector analysis in assessments.
• Addressed challenges of accountability and revealing how AI systems judge users, and the potential role of governmental regulations in guiding responsible AI development.

Final Thoughts

• Fernanda highlighted a collaborative approach needed to understand and innovate AI for better user experience.
• Audience inquiries pointed towards the need for greater vigilance in understanding the AI's operational guidelines and biases.

Thank you for a thought-provoking talk!