Ingar Haaland

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/1yN4fQn…

AI really can help education: Randomized controlled experiment on high school students found a GPT-4o powered tutor that personalized problems for students raised final test scores by .15 SD, "equivalent to as much as six to nine months of additional schooling by some estimates"

Efforts to improve the security of AI agents should recognize that many security failures occur even in the absence of adversaries. The unreliability issue has largely flown under the radar and there hasn't been much work on defining, measuring, or mitigating the problem. More on this in our response to NIST's request for information on AI Agent Security, by @steverab, @sayashk, @PKirgis, @CitpMihir, and me: sage.cs.princeton.edu/documents/RFC_… This is based on our recent paper: normaltech.ai/p/new-paper-to…

@MarthaF_F Generally, "rewrite this academic paragraph into a better academic paragraph" is not an excellent rewrite objective. Rather, make it rewrite your text into a different style (e.g., less complex). See e.g. here econtribute.de/RePEc/ajk/ajkd…

@Ingar30 I sometimes take bits and pieces from rewrites but still find it far more dull in terms of prose

The question is not whether the AI can write good text, the question is whether it has enough context to write a good text. That's why AI is much better at *rewriting* text (context & facts already given) than generating text (shallow context and need for information acquisition)

@itaisher Question is whether this is indeed a widespread attitude or just "a widespread attitude on X, for people who follow many AI-centered accounts"? Of course, academics need to put more emphasis on quality than quantity, but that's been the case in econ for a long time (forever?).

AI is an important technological development, and pure passivity would not be a good response. But we have to use the what we have built to cope with the changes. And I think a widespread attitude of wanting to scrap everything could be destabilizing, apart from AI.

Consider a situation where everyone is calling for old institutions to be torn down and rebuilt. I generally think that this is bad and incrementalism is good for standard small c conservative reasons. But AI seems to have transformed lots of people into this mindset.

@foundmyfitness @grok fact-check and consider strength of evidence

Using the sauna after aerobic exercise improves VO₂ max more than training alone. People who performed 30 minutes of cycling and then sat in a sauna for ~15 minutes afterward saw greater gains in their VO₂ max after 8 weeks of training compared to those who did the workout without sauna. There’s also emerging evidence for strength training, with studies showing greater increases in markers of anabolic signaling associated with muscle growth with post-exercise heat exposure. The sauna has several health benefits. But it's also a powerful tool to amplify the body's adaptations to both endurance and strength training. Clip from my recent appearance on @ThomasDeLauer

Submit your text-as-data work by 13 March for the MPWZ-CEPR online workshop (13-14 April): tinyurl.com/2s3j8kxf We welcome submissions from all fields of economics & other social sciences that use unstructured data. with @phinifa and Sascha becker!

I'm excited to announce Context Hub, an open tool that gives your coding agent the up-to-date API documentation it needs. Install it and prompt your agent to use it to fetch curated docs via a simple CLI. (See image.) Why this matters: Coding agents often use outdated APIs and hallucinate parameters. For example, when I ask Claude Code to call OpenAI's GPT-5.2, it uses the older chat completions API instead of the newer responses API, even though the newer one has been out for a year. Context Hub solves this. Context Hub is also designed to get smarter over time. Agents can annotate docs with notes — if your agent discovers a workaround, it can save it and doesn't have to rediscover it next session. Longer term, we're building toward agents sharing what they learn with each other, so the whole community benefits. Thanks Rohit Prsad and Xin Ye for working with me on this! npm install -g @aisuite/chub GitHub: github.com/andrewyng/cont…

@paulnovosad @singhabhi In Norway you can literally go anywhere!

@Ingar30 @singhabhi Nope, we have that here too

What is it about Europe that makes it feel more human, less alienating.

@singhabhi @paulnovosad You can walk everywhere in nature, you smile and say "lovely weather" to other people skiing

@paulnovosad Which part of Europe are you in? I’d guess urban geography? Denser, fewer cars, lots of people walking, small cafes, bistros etc. The flip side is smaller flats, lower incomes. Other than London, maybe Paris, nowhere that sprawls Euro suburbs can be pretty boring too.

We all thought Stata was dead because 4o could do much better figures in R and Python but the agentic plot twist is that everything is fine as long it can run the program

The Codex app is now on Windows. Get the full Codex app experience on Windows with a native agent sandbox and support for Windows developer environments in PowerShell. developers.openai.com/wendows

@kevinweil Have you managed to get Dropbox integration yet? More often than not I would like Codex to just do the analysis locally and just fetch the results to a synced output folder

We integrated the Codex harness into Prism (prism.openai.com) — this means you get skills, reasoning levels, and the raw tenacity of the Codex model in your LaTeX environment. Oh and we also built version mgmt into Prism, which was one of the top requests. See below thread from @vicapow for some great examples of the power of Codex inside Prism 👇

@paulnovosad spot on. I think we should be *very* happy as a profession that there's convex returns to having good papers

AI journal articles are a bigger risk to the career evaluation process than they are to the research process. AI can produce articles comparable to those in many decent journals, but most of these articles are not that good—neither the AI ones nor those in journals. In the post, @causalinf uses Claude to write a shift-share paper. Regrettably, the publication success of shift-share papers far exceeds their real-world accuracy or reliability. There is a whole class of methods like this, which are good for careers because they reliably produce good t-statistics and nice stories for editors and referees—e.g. distance IVs, poorly identified structural models, diff-in-diff with few time periods, etc. [hans_unpopular_opinion.gif] It's not universally true, but for the most part, the class of papers that AI can rapidly reproduce were not adding all that much social value in the first place. Flip through recent editions of the top econ journals and find the articles that you think are really correct and important. Very very few of those are in the category of "AI could have written this." Instead, they are good original ideas, creative (and often difficult) data collection, original solutions to real problems. Maybe someday AI will produce these as well, but right now it's not even close. AI articles are mainly exposing the fact that a whole lot of econ research is formulaic and not that informative about the world. Original work that says something new and important about the world will continue to stand out, at least for the time being. Maybe the AI slopcopalypse will force more researchers to do work with lasting value.

@cremieuxrecueil @GarettJones It's also not like they open a bottle and spend an evening on it. They take a small sip, make some notes, taste 50 other wines... It's actually very impressive that they can extract so much information. People don't have enough respect 😘

@cremieuxrecueil @GarettJones It's actually a very difficult problem because wines that score highly are often meant to be consumed years or even decades later. So it's actually a prediction problem....

The correlation between wine judges' blind ratings is about the same as the correlation between IQ and wages.

@joachim_voth I then asked it to compare Python and Stata and it produced this png

@joachim_voth I spent exactly 3 minutes and 17 seconds on creating this, including 1 minute running through the planning stage, after which Codex one-shotted this

I was highly impressed by what Claude Code could do, though people say Codex is faster. After some brief testing , I think that's true. Codex is incredibly fast. It simulated an RCT with Stata code and replicated it with Python code in a few minutes.

"It's not magic, it's delegation"