I work on language models as a senior research scientist at Google DeepMind. I currently do both capabilities and safety/alignment work to make AI useful and deployable.
Broadly, my goal is to improve how humans and AI work together at both the individual and societal level. My work occasionally intersects with cognitive science, human-AI interaction, and computational social science. I'm also interested in applications in creativity, education, and science.
Previously, I did my PhD at MIT, advised by Deb Roy in the Media Lab and Jacob Andreas in CSAIL. I interned at Facebook AI Research with Jason Weston and Stephen Roller, and Google Brain with Peter Liu. Before that, I was briefly a data scientist at Facebook, building human-AI systems and multimodal classifiers for ads moderation.
I did my undergrad at UC Berkeley, where I tried research in biomedical imaging, protein folding, and topological data analysis.
Selected work and projects
See my Google Scholar and older website for more detail.
Machine learning, natural language processing, and generative models
- (2022) Large language models: PaLM 2, Bard, Gemini
- (2020) Big-BENCH: logical reasoning dataset ('navigate'); highlighted in LLM scaling and chain-of-thought papers as a difficult task
| Paper @ TMLR
- (2019) Improving generative models with hierarchical plans
- (2018) "MeanSum: first end-to-end model for unsupervised, multi-document, abstractive summarization"
| Paper @ ICLR
- (2017) "Learning Personas from dialogue with attentive memory networks"
| Paper @ EMNLP
- (2016) "Audio-visual sentiment analysis for learning emotional arcs in movies"
| Paper @ ICDM, ICCV movie description workshop)
- (2016) Adaptive speech synthesis for tutoring
AI safety and alignment
- (2022) Learning from human feedback @ Google: initiated and led code RLHF efforts
- (2022) Code safety @ Google: led generative code safety efforts
- (2021) "Language models trained on media diets can predict public opinion"
| Congressional Hearing
(2020) Influence functions for tracing provenance of toxic generations in language models
- (2019) "Are visual explanations useful? A case study in model-in-the-loop prediction"
- (2019) "Human-machine text classifiers may have disparate impact on minority communities"
- (2019) "Games for fairness and interpretability"
| Paper @ ICLR Trustworthy Workshop
- (2017) AI & Equality: co-created MIT class on how AI can promote / impede equality in domains
- (2015) Ads moderation @ Facebook:
led multi-office project running A/B tests on how to present model predictions and uncertainty to human reviewers;
prototyped multimodal classifiers.
Additional work and interests
I also dabble and am interested in the following:
Creativity and art:
code AI @ Google,
Dec 2021 presentation on "Creative AI: Generative Art & AI-Assisted Creativity": (slides),
"Evolving Evocative 2D Views of Generated 3D Objects" (NeurIPS creativity paper),
"Artistic Influence GAN" (NeurIPS creativity paper),
"Pablo West" (NeurIPS creativity paper),
CAD tool for designing topological sculptures
"Parents’ online school reviews reflect several racial and socioeconomic disparities in K–12 education" (paper),
literacy learning apps,
adaptive speech synthesis for tutoring children (link),
intern at Knewton adaptive learning edtech startup
Science: biology, health, math:
mathematical reasoning and numeracy in language models,
seriously considered post-PhD offers in AI + health (Stanford postdoc with Andrew Ng) and drug discovery (ML scientist at Prescient Design),
undergrad research in magnetic particle imaging, protein folding, applied algebraic topology,
started undergrad as bioengineering major