#AcademicRunPlaylist - 7/11/24

Author

Ben Waber
July 11, 2024

A small swamp, followed by forest on a hill, on a bright, sunny day. Trees crowd the near bank, fading into bushes, followed by water and leafless trees reaching their fractal branches to the sky. White clouds float overhead, while deep green evergreens cover the hill on the left

The kids (and by extension me) had another busy day, but I still made some time to listen to talks for my #AcademicRunPlaylist!

First was an excellent talk by Kyle Mahowald on using language models for linguistics at UQAM | Université du Québec à Montréal. LLMs are considered to be very good at syntax and grammar, but is this due to memorization, heuristics, or building up a decent model of language? Through a systematic review controlling training data on smaller models, Mahowald convincingly demonstrates that it's the latter, exploring implications for linguistics. Highly recommend https://www.youtube.com/watch?v=z8SwDlv0jx4

Next was a fascinating talk by Alexis Litvine on lighthouses and safety on the British seas from 1747-1911 at the Faculty of Economics, University of Cambridge https://www.youtube.com/watch?v=OjszycDHTKM

Next was an interesting talk by Alessandro Lenci on why LLMs don't meaningfully understand language at UQAM https://www.youtube.com/watch?v=-ukVtAFQ83U

Next was a fantastic pair of talks at the National Bureau of Economic Research by Moritz Lubczyk (the role of gender bias in the misallocation of talent) and Robert Johannes (construction technology and medieval urban development). Johannes' talk was a mind blowing effort, combining archaeological records, geological data, and historical construction records to quantify the effect of the diffusion of brick technology in Germany (starting around 1150) on a whole variety of outcomes. Highly recommend https://www.youtube.com/watch?v=9OQQDvsO5sw

Next was an essential talk by Dylan Hadfield-Menell on the challenges of specifying goals for algorithms at the Allen Institute for AI (AI2). Hadfield-Menell expertly illustrates why reward systems tend to be brittle due to difficulties in measuring/specifying everything using practical, theoretical, and empirical approaches. He concludes by describing the issue with aligning goals with those of a single individual, advocating strongly for inclusion and more diverse strategies moving forward. Highly recommend https://www.youtube.com/watch?v=ZmhmQzmK3OQ

Last was a thought-provoking pair of talks by John Moore on micro (talk 1) and macro (talk 2) perspectives on the intersection between interest rates, corporate funding, entrepreneurship, and productivity at the University of Cambridge Faculty of Economics
Talk 1: https://youtube.com/watch?v=FqnhaXrGsUc
Talk 2: https://www.youtube.com/watch?v=JsqdVr5GTtk