
How should one design survey experiments for maximal impact in economics? Here's my slide deck from a recent workshop in Uppsala. Thanks to everyone attending for making it a great experience @EconomicsUU drive.google.com/file/d/1yN4…

AI & economics. Professor at the Norwegian University of Science and Technology

How should one design survey experiments for maximal impact in economics? Here's my slide deck from a recent workshop in Uppsala. Thanks to everyone attending for making it a great experience @EconomicsUU drive.google.com/file/d/1yN4…

Really enjoyed this conversation with @tylercowen and @eriktorenberg. Thanks for having me on!
Google DeepMind's Alex Imas says AI agents still struggle to coordinate, and that's a key bottleneck to economic transformation: "What you actually need for real economic transformation is not just competence... It's the ability of these agents to coordinate with one another." "There's this idea of agentic swarms, right? But if you unpack it, these agents are still very bad at coordinating. They're not coordinating, they're mostly just doing parallel processing." "If you have benchmarks specifically on coordination, you could actually show that 1 agent does better than 2 agents splitting the work, even if they can communicate. Whereas if you give the same thing to people, they actually have division of labor, they communicate, and they do way better than 1 person with double the resources." "But if they get better at coordinating, you can basically break the Coasean ceiling on firm size, and you can get very, very different firms to emerge." "Once coordination becomes better, the economic space of possibilities of what firms look like, what happens to wages, what happens to employment, that's when the real kind of visible changes are going to happen." @alexolegimas

What’s the evidence on rent control? I charted the findings from 112 studies: Great for rent-controlled tenants ✅ But: Pushes up rents for other people ❌ Pushes down new housing construction ❌ Pushes down building quality ❌

Our @ExpectedParrot tools also help with optimal beard planning. Please take survey here: expectedparrot.com/respond/h…. 0 = Clean shaven 1 = Full Ingar
Guys, the real decision is not whether to wear a leather jacket but how much you should grow your beard. @johnjhorton you should try to simulate that too!

And people still claim CEOs are paid too much 👀
Sobering thought: Nvidia has lost $1.8 trillion in five weeks. That is more than the EU’s five most valuable companies are worth together: ASML, SAP, L’Oréal, Siemens, LVMH. What is left of Nvidia's market cap is still worth more than every listed company in Germany.

I hope he doesn't think that cancer and mental illness will be "relics of the past" because Claude decides to wipe out humanity in the process
Anthropic CEO Dario Amodei thinks cancer, Alzheimer’s, heart disease, and mental illness will be “relics of the past” in less than a decade, thanks to AI.

That said, when the paper is written up, you should absolutely use AI to get a review on the paper. My reviewer is a good starting point and is free if you have a Codex or Claude Code subscription github.com/Ingar30/reviewer

Frontier models are actually very bad at writing academic papers. They will write them up in a very technical way that is different from how we write papers, at least in economics. If you cannot do a better than a frontier model, get a different job.

For writing up the paper, you should do it yourself, full stop. You can ask the AI for help finding relevant papers etc, but it's your job to write up the motivation, literature review, and the results section. You will learn much along the way.

Don't ask Codex for a first draft of your slides. Instead, write bullet points for each slide and then feed the bullet points to Codex and ask it to convert them into slick beamer slides without changing the underlying content (except correcting things like grammar mistakes).

When the data is analyzed, your code is audited and vetted by the AI, and you have something to present, create slides. Slides are actually a great way to structure the paper, so create a nice set of slides yourself. Creating good slides is a work of art.

It's fine to use Codex to create a coefficient plot, but you should specify your own regressions and know what is going on. Codex should also not decide how you specify your interaction terms. I cannot stress this enough: Your undertanding will be tested when you present the paper.

When you have identified a good question and collected the data, by all means use AI to explore and help you analyze the data. But you still need to know what is going on. You cannot respond "Codex suggested to cluster the standard errors this way". You will get killed in a seminar.

It is also possible to examine properties of LLM models (e.g. whether they are biased when used to screen CVs), but be aware that models are changing fast. My advice for PhD students is to focus on identifying big questions in economics and use LLMs as a vehicle to study those.

Another alternative is to use AI to create some cool software and run a software-based experiment. This could be a news app, a mental health app, or something else that you might be interested in studying. Experimental economics will be much more app-focused going forward.

You can also use LLMs to run behavioral interventions at scale (e.g., as we do in this paper: econtribute.de/RePEc/ajk/ajk…). This allows you to vary program components in a clean way to see what actually makes an intervention effective.