Arjun Ramani
@arjun_ramani3
Econ PhD Student @MIT | Prev @TheEconomist in 🇬🇧 & 🇮🇳 | Hoosier4life
NEW ESSAY: "Dispatches from India" A year ago I was living in New Delhi, India 🇮🇳. I traveled across 10 states and 30 cities+villages writing for The Economist, including a cover story on the economy Since then I’ve been chewing on how to capture what I learned It still…
Awesome
NEW ESSAY: "Dispatches from India" A year ago I was living in New Delhi, India 🇮🇳. I traveled across 10 states and 30 cities+villages writing for The Economist, including a cover story on the economy Since then I’ve been chewing on how to capture what I learned It still…
Should be very interesting.
NEW ESSAY: "Dispatches from India" A year ago I was living in New Delhi, India 🇮🇳. I traveled across 10 states and 30 cities+villages writing for The Economist, including a cover story on the economy Since then I’ve been chewing on how to capture what I learned It still…
fantastic (and funny) read to help calibrate your AI timelines the story of the fox, the hedgehog, and the rhino (guess who the rhino represents) was especially delightful!
I wrote some fiction in the style of AI 2027. It combines the parts of AI 2027 and AI as Normal Technology that resonate with me most. Come for the predictions, stay for the animal parables!
Which tasks are going to take longest for AI to do? 4 categories: 1. The data poor The internet has trillions of words, so AI got p good at writing. But there's no equivalent database of 3D movement, and it's expensive to gather, so robots still struggle.
Benchmarks/evals suffer from selection bias -- the tasks for which you can construct a usable benchmark are precisely the tasks RL is good for An implication is we should EXPECT there to be a growing gap between benchmarks and real-world impact
I find the story of AI and radiology fascinating. Of course, Hinton's prediction was wrong* and tech advances don't automatically and straightforwardly cause job replacement — that's not the interesting part. Radiology has embraced AI enthusiastically, and the labor force is…
Excited to launch @insilicopod — a new podcast exploring how AI is transforming science! I’ll be talking to scientists and researchers, as well as investors and policymakers, about the frontiers of Science: autonomous labs, AI-driven modelling, and more. Episode 1 is live👇
First episode is out!🎊 I spoke with @Sergei_Imaging (@UTKnoxville,@PNNLab) on how AI is transforming microscopy, from imaging to autonomous experimentation. Stay tuned for the second episode next week! Listen here: open.substack.com/pub/ml4sci/p/s…
Report from the teaching trenches: I teach an advanced elective (Social and economic networks) which is difficult for top undergrads but where AI can do the homework perfectly. The main changes this year: (i) I encourage AI use for learning; (ii) closed book exams 1/
If ChatGPT can do your homework, then there's no point learning how to do it anyway. Instead, learn how to do something that AI can't do.
THREAD/ specific answers vary by field, but i've spent a large portion of my research career interested in this question. punchline: funding (the right) data is a relatively cheap but super effective way to boost progress in a field.
What are some data sets that don't exist (or don't exist publicly) that would accelerate technological progress?