Simple Bayesian A/B testing
Apr 15, 2024
About
I’m Marlin Figgins, PhD.
Previously. Built automated forecasting systems for respiratory virus evolution at Nextstrain including model design and implementation, data specification, and evaluation. At Microsoft Research, I studied how people use language models in their daily lives to understand where LLMs excel and lag. At Google, I worked on a JAX model garden implementing standard and frontier models across vision, natural language, and video including several multimodal approaches.
Now. Open to technical roles building and evaluating real systems. Relevant job titles would be machine learning engineer / researcher, member of technical staff, or quantitative researcher.
Current work.
Reward and evaluation design. Designing rewards and evals that stay useful even when a model is learning to break them.
Biosecurity systems. Screening and functional characterization of natural and designed sequences. Static benchmarks lose relevance when inputs are chosen adversarially during deployment.
Representation geometry in sequence models. Investigating whether protein and language models represent underlying biological structure or use shortcuts. Developing training tasks which force the former.
Get in contact?
My personal email is marlinfiggins [at] gmail [dot] com.
Education
University of Washington, Seattle | Seattle, WA
PhD in Applied Mathematics | Sep 2020 – Dec 2024
M.S in Applied Mathematics | Sep 2020 – June 2021
University of Chicago | Chicago, IL
B.S in Mathematics | Sep 2016 – Dec 2019
Most recent blog posts
