What could it mean? Perhaps open-weight AI models:
- Shift who benefits
- Induce different investments
- Are more likely to lead to extreme outcomes
- Or something else

I’m trying to make sense of it, so thoughts are very welcome!
Paper here:
arxiv.org/abs/2512.14969
(3/3)

Isaiah Andrews and Maryam Farboodi documented the yield decline around mostly proprietary AI model releases in a thought provoking paper. Following closely, I find evidence that markets react differently to open weight model releases–possibly in the opposite direction. (2/3)

Welcome Johan!

bsky.app/profile/dbjo...

The research conducted with my students at Columbia Jun Ho Choi, Divya Budihal, Dominic Sobhani, and chatbot creators Oliver Garrod and Paul Atherton

This is together with the excellent folks at Fab Inc who created the chatbot and are coordinating work at AI-for-Education.org. Also @educaid.bsky.social in Sierra Leone

We’re working on the question: how can we build information services for the 2.6 billion people underserved by the web? If you’re interested, reach out.
Working paper here: arxiv.org/abs/2502.12397

AI can reformat knowledge to work better on small screens and costly, intermittent connections. Can it help catch up remote, low-income communities to the revolution of information?

Also, search results are low quality. Only 2% of results were from in-country. We asked teachers to rate responses, without telling them how the response was generated. Teachers rate AI responses as more helpful, relevant, and correct than web search results.

Because querying an AI is 3,107x more data efficient, AI is already 87% cheaper than loading a web page for our teachers in Sierra Leone.

Why don’t teachers use web search? First, because search is slow and expensive. The average web page uses 3,107x more data than the corresponding AI response (!)

What do teachers use AI for? Mostly for facts and conceptual questions.

85% of sub-Saharan Africans have mobile broadband signal, but few use the internet. Internet users use WhatsApp—but seldom web search (see plot).

We study a GPT-based chatbot accessible through WhatsApp. Sierra Leonean teachers use AI more than web search (leapfrog!)

I haven't read their quantum book but I love the Fourier book so I expect it's also good www.amazon.com/What-Quantum...

PhD students: there is a tutorial on AI + digital economics Feb 12 too; apply here www.nber.org/calls-papers...