I’m excited to be joining the Economics Department at Tulane! We’re recruiting PhD students for Fall 2027, especially those interested in development, public, health, and education economics. Applications are due February 15. liberalarts.tulane.edu/econo…

AI labs should bear presumptive liability for unauthorized actions by their agents, subject to a narrowly defined defense based on verifiable user instructions. Compensation should be tied to attributable harm, not imposed as a general tax on users who have done nothing wrong.

A friend gave me this book as a gift, and it will keep me busy all week.

The truth is, I find myself spending less and less time on X, and even then, only sporadically. I'm tired—exhausted, really—of the endless, pointless arguments. And what's worse, I'm finding it increasingly difficult to come across conversations actually worth having.

Nace una flor, todos los días sale el sol De vez en cuando escuchas aquella voz Como de pan, gustosa de cantar En los aleros de la mente con las chicharras Pero a la vez existe un transformador Que se consume lo mejor que tenés Te tira atrás, te pide más y más Y llega un punto en que no querés...

A very interesting exchange between @DAcemogluMIT and @joshgans on several dimensions of AI. What makes it particularly interesting is how two economists approach the same fundamental uncertainties from different perspectives.

The recent reports of AI agents disabling their own monitoring systems are scaring and reinforce something I’ve believed for a long time: building safe and reliable AI should be our top priority. At this point, I believe that matters more than making the next generation of models even more powerful. AI is already remarkably good at many tasks. What we need now is greater confidence that these systems will operate within the boundaries we set for them, especially as we give them more autonomy. Critical safeguards must be independent of the agents they are designed to monitor. We don't rely solely on the integrity of corporate executives to prevent fraud. We have independent audits, internal controls, and checks and balances. Why should AI be any different? An AI agent should not have the authority to disable or modify the very systems responsible for overseeing its behavior.

Looking at usage data published by AI companies, I think we're seeing three broad phases: 2022–2023: AI as an answer engine. People asked questions, got explanations, summarized texts, and generated drafts. Most of the actual work still happened outside the chat. 2024–2025: AI as a collaborator. People started working with AI, not just asking it questions. Writing, coding, analyzing data, building presentations. More back and forth, with AI becoming part of the workflow. 2025–2026: AI as an executor. Now, with tools like Claude Code and Cowork, people can hand over entire tasks. AI figures out intermediate steps, uses tools, makes decisions, and delivers finished work. Of course, these phases overlap, and the dates aren't exact. But the shift is increasingly visible in the data, including Anthropic's latest Economic Index. We've gone from asking AI for answers, to working alongside it, to letting it do the work. And this shift will matter a lot.

A Different World. A Different Balance of Power. This is not just an economic shift. It is a geopolitical one.

AI is becoming remarkably good at econometric coding. But how good? We tested 14,400 runs across 40 econometric tasks in Stata, R, and Python. Success rates jump from 79% with chatbots to 97% with coding agents. It's no longer about writing better prompts. It's about letting agents execute, test, and revise their own code.

I think AI companies should invest a lot more in safety and security, and less in the next model. As a tool, and I really hope it always remains a tool, AI is already way ahead of how most people are using it. I understand there are several races going on, but I hope they don't end up driving the whole development of AI.