#AcademicRunPlaylist - 4/7/25

Author

Ben Waber
April 07, 2025

A selfie of me inside a house in front of a grey porch with a dark brown wicker bench on it, with some evergreen trees and a house beyond it behind the porch, on a bright day. I'm a bald, middle-aged, white man with a red beard flecked with white. I'm wearing glasses with a metal top rim and a black sweatshirt with Boston written on the middle in red letters.

It was pretty gross out today, but even though I spent most of the day inside I still listened to talks for my #AcademicRunPlaylist!

First was a slate of talks at the Brown University Mind Brain Research Day. I particularly liked the talk by Amanda Arulpragasam on non-invasive deep brain stimulation with ultrasound https://www.youtube.com/watch?v=B4YNBCHSVDk

Next was an intriguing talk by Kilian Weinberger on diffusion models for text generation at the Simons Institute for the Theory of Computing https://www.youtube.com/watch?v=klW65MWJ1PY

Next was a nice talk by Agata Farina on multi-dimensional information disclosure effects in economic games at the Mannheim Centre for Competition and Innovation (MaCCI) https://www.youtube.com/watch?v=kE3Db-nnvxs

Next was a great talk by Aviral Kumar on scaling test-time compute effectively at the Simons Institute https://www.youtube.com/watch?v=5hJhPxKFaus

Next was a fantastic event on social protection and gender equality in the gig economy at the Asian Development Bank Institute with Tetsushi Sonobe, Raja Rajendra Timilsina (social welfare program demand in Indian informal workers), Joseph Han (social protection of Korean platform workers), and Stephanie Dimatulac Santos (Southeast Asia gig economy gender dynamics). Highly recommend https://www.youtube.com/watch?v=2KN0TwrUfGA

Next was a thought-provoking talk by Luke Zettlemoyer on mixed-model language modeling at the Simons Institute https://www.youtube.com/watch?v=JYMXlmSM_Ew

Next was an excellent talk by Nathan Lambert on reinforcement learning with verifiable rewards at the USC Information Sciences Institute https://www.youtube.com/watch?v=MTr2KM9lK1M

Next was an amazing talk by Andrew Ilyas on predicting and optimizing the behavior of large ML models at the Simons Institute. I love the focus on data here, and the concept of "data counterfactuals" to predict the effects of training dataset changes seems promising. Highly recommend https://www.youtube.com/watch?v=iePMkTFuEW8

Next was an engaging discussion with Maohao Shen on improving LLM performance with autoregressive search on the TWIML podcast https://www.youtube.com/watch?v=LEJgwuVo7mo

Last was an illuminating talk by Nathan Srebro on weak to strong generalization in random feature models at UW. Srebro provides great intuition for why weaker models can teach things even to much stronger teacher models, then proceeds to demonstrate and actually mathematically prove this, setting up a lot of future work. Highly recommend https://www.youtube.com/watch?v=M8DL05YhU24