{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/ccb9593dbd354fd9904b18972211d6a8\" frameborder=\"0\" width=\"1708\" height=\"1281\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1281,"width":1708,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1281,"thumbnail_width":1708,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/ccb9593dbd354fd9904b18972211d6a8-8c677349f8255517.gif","duration":97.381,"title":"StreetMath - 1 minute","description":"In this video, I explore whether language models can perform everyday math like humans do, particularly in terms of approximation. We conducted tests with 1,000 multiple-choice problems involving basket sums, discounts, taxes, unit prices, and tips. Our findings reveal that models tend to compute exact answers before rounding, which contrasts with human estimation methods. I encourage you to check out our code and data if you're interested in building models that can effectively choose when to approximate."}