Talks and press
Class takeaways
Turning Data Into a Superpower, Stanford GSB Class Takeaways, March 2026. Five minutes of what I try to get across in the MBA course: ground decisions in data rather than intuition, build enough technical skill to run the tools yourself, work out the right question before chasing an answer, pair AI with expert judgment instead of handing the judgment over, and treat data work as a cross-functional problem. GSB Insights, with transcript · Stanford Report
Podcast
The AI Prescription for Healthcare, Stanford GSB If/Then, June 2025. Twenty-eight minutes on what actually blocks AI in medicine: trust between clinicians and models, privacy, hallucination, data quality, and the distance between benchmark performance and clinical usefulness. Episode page with transcript · Apple Podcasts · Spotify
Recorded talks
- Estimating
Interference Effects via Distributional Dynamics
SNAPP Seminar, March 2026. The causal message passing work: estimating treatment effects under network interference without ever observing the network. - Two
Studies in AI and Experimentation: Model Behavior Alignment and Network Interference
Effects
University of Toronto Data Sciences Institute, May 2025. Also given at the Chicago Booth Applied AI Seminar, February 2025. - The Unreasonable
Effectiveness of Greedy Algorithms in Multi-Armed Bandits
Stanford HAI Weekly Seminar, February 2022. Abstract - On Worst-Case Regret
of Thompson Sampling and a General Framework to Analyze Linear Bandits
SNAPP Seminar, August 2020. - A Framework for
Personalizing and Testing Medical Decisions
Stanford Data Science AI for Health annual meeting, November 2019. - Data-Driven Decision
Making in Healthcare Systems
Microsoft Research, 2016.
Recent invited talks
- An Introduction to AI Alignment Challenges and Progress
INFORMS Applied Probability Society Distinguished Lecture, October 2025. - Aligning AI with Clinical Knowledge
Stanford Center for Digital Health Annual Symposium, October 2025. Coverage - A General Method for Experimentation Under Network Interference
UIUC Gies, Wharton OID, Columbia DRO, Koç University, USC, MIT ORC, NYU, Boston University, Georgia Tech AI4OPT, 2023 to 2025. - Challenges and Progress in Aligning AI Models for Healthcare
Applications
Cornell University, June 2024. Stanford GSB faculty flash talks, May 2024.
Coverage
Researchers Build a Virtual World to Run Experiments Over and Over
Twenty thousand language model agents, built from census attributes, replaying the 2010 Facebook voter mobilization experiment. The simulation got the direction right and the magnitude wrong, which is itself the useful finding.
An AI Copilot Can Reduce Prescription Errors That Put Patients at Risk
MEDIC, a domain-tuned model with pharmacy guardrails, was evaluated on real prescriptions at an online pharmacy and scored by pharmacists. It cut near-miss medication direction errors by about a third against the checks that pharmacy was already running, while general-purpose language models did worse than those checks. Stanford Report version
Bridging Humans and Machines: Advancing Alignment in AI
Recap of the AI alignment conference I organized at the GSB with Susan Athey, Andy Hall and Gabriel Weintraub.
Faculty Voices
A profile on the move from pure mathematics into clinical AI, and on why clinician trust turned out to matter more than model accuracy.
A Peek Inside Doctors' Notes Reveals Symptoms of Burnout
Machine learning on electronic health record traces from 233 physicians across 60 clinics, predicting which clinics are most likely to produce burnout.
Experiments Inspired by Slot Machines Promise Bigger Research Payoffs
An accessible account of multi-armed bandits as an alternative to fixed-allocation A/B testing, and honest about what the adaptive designs give up.
Is Your Business Ready to Jump Into AI? Read This First.
Where the alignment gap comes from: the distance between what a model was trained on and the task it is actually being asked to do.
When "Greedy" Is Good
Why exploitation-only bandit algorithms do far better than the theory predicted, and why that matters for clinical trials. GSB version
Many Health Care Workers Are Emotionally Exhausted, and Technology May Be to Blame
Across 15,505 health care workers in 31 hospitals, frustration with technology was the second strongest predictor of emotional exhaustion, after sleep difficulty.
Avoiding Experimentation in Online Decision-Making
Covariate diversity, and the condition under which a decision-maker can skip exploration entirely.
Why Hospitals Underreport the Number of Patients They Infect
Hospitals in weakly regulated states reclassify hospital-acquired infections as present-on-admission, affecting at least 10,000 cases a year. Earlier framing
Why Hospital ER Wait Times Are Often Wrong
The rolling-average method most hospitals use was off by as much as 90 minutes. Our method cut the error by up to a third and went into production at San Mateo Medical Center.
How Decision Software Can Improve Health Care
The congestive heart failure readmission work with Microsoft Research, later deployed in several hospitals. HBR