I have a fun new paper today w/ @joshgans: what makes an AI valuable? We noodled on this literally since 2024. Answer: AI is used by humans. They can extend, verify, get a second opinion, etc. AI is therefore part of what decision theorists call a "composite experiment". 1/10 Let using AI have some cost, and likewise those other actions also potentially be costly. To make this concrete, assume you can trust an AI prediction, try to verify it as cost, or not use the AI at all. You then take an action if you like, getting H>0 if right, L<0 if wrong 2/10

I read Stephen's argument (caveat, he lived downstairs from me in college!) as "US can treat market access like a monopsonist, extract revenue from abroad for access". Surely true just like China can insist on tech transfer in exchange for FDI but Uganda can't. But a problem: 1/4

https://twitter.com/ben_golub/status/2071245953953059042

The world integrating inputs with US economy is *good* for US. Just like Toyota doesn't want to extract all rents from suppliers: many noncontractible fixed cost investments that make both parties better off. Grossman-Hart-Moore plus relational contract logic tells you why. 2/4

On PPP, this is good. Theory is pretty clear: take domestic output or income, deflate, look at real GDP pc or labor productivity. US growing ~1% faster than W Europe per year. Converted to current or chained PPP, you don't get this. 1/7

https://twitter.com/RobertInklaar/status/2061141729743429972

Imagine preferences and product mix identical in US and Europe: say, we only consume nontradable widgets. 100 people in each. Yr 1: 200 widgets consumed in each, 1 Euro=1 USD. Yr 2: 200 in US, 300 in EU, 1 Euro=2 USD. In both years, a widget in US is $1, in Europe is 1 Euro. 2/7

My read on "normal policymaker & corp. leader on AI": mostly now they don't need to be convinced it is very important (unlike a year ago). But they still see its capabilities as today + epsilon. So just briefly, here is what even "AI is normal tech" folks in the labs believe: 1/8 1) There is basically no evidence internal to the labs that we are seeing a slowdown in capabilities. 2) Demand will grow fast enough that compute + energy will become real caps in short run, though we have ideas on what to do here. 2/8

Another two weeks in various parts of India. I have said before that India feels like China 2005, where I briefly worked, in ambition and growth. However, three major issues need to be solved. 1/5 First, the License Raj mentality is still all over government services. Almost every govt interaction is poor from the first step (joke of an e-visa site, need to scan passport for Wifi, blink your eyes on video to get a SIM, queue at office 1 then hand token to office 2...) 2/5

Wild set of stories in Canada recently: most famous indigenous author, 2nd most famous writer, most famous singer, most famous director, a federal cabinet minister, a big university president, famous law professor...all turned out to be faking their Native ancestry. 1/3


In the case of King, he's getting favorable press about how "devastated" he is to have learned his absentee father wasn't actually Native. But online you can find discussions going back a dozen years from Cherokee genealogists noting he ignored them even back then. 2/3

@joshgans and I did an internal talk on AI for research. Mostly demos, so no slides, but broadly: 1) Research should be efficient, open & replicable. 2) AI helps will all three. 3) Always use the best model. 4) Structure your processes/tools/etc. so you can continue to do 3. 1/15 What I mean by "structure your process/tools/etc" is first that everything you do - code, writing, editing, collabs, slides - should have plain text as the substrate. This means no matter what, you can always use every AI tool now and in the future to interact with it. 2/x

AI economists and AI researchers: this is *excellent*. Details below, but as I feel like I've said in every talk on this topic since my slides said "GPT-2", 1) AI technical capabilities are better and improving quicker than you think, 2) impact on economy *much* slower. 1/13 Hinton famously said (at our event!) a decade ago to not train radiologists. Today, radiologist hiring very strong w/ avg income of US$520k. What went wrong? As a technical matter, Hinton was right - anomalies on images very easy for AI today (this article undersells that!). 2/13

The "reasoning doesn't exist" Apple paper drives me crazy. Take logic puzzle like Tower of Hanoi w/ 10s to 1000000s of moves to solve correctly. Check first step where an LLM makes mistake. Long problems aren't solved. Fewer thought tokens/early mistakes on longer problems. 1/11 But fundamental problem is that there are foundation *models*, and then there is *the way we train these models to cutoff thought*. Imagine you say "hello". Try it on o3 and 4o. o3 takes longer because it "thinks" about other things you could mean. Unnecessary in this case. 2/x

Gobsmacked that the Toner-Rodgers neural network for scientific discovery paper has been withdrawn. This was a very strong claim about how 2023 era AI was improving the speed of science. I don't have details on the issues, but wanted to pass along this retraction. 1/2 That said, I am personally aware of 2025 automated lab scientific progress using LLMs and other techniques, so don't update too much in the other direction! MIT note here: economics.mit.edu/news/assuring-…

On the NBA and economic theory: dynamic mechanism design tells us how to redo the draft. A good rule of thumb (due to Reny?) is that stochastic mechanisms like draft lotteries usually aren't best. How to prevent tanking while still having worst team get top pick? Simple! 1/x Theory is beautiful. Myerson Revelation: look only at mechanisms that get teams to reveal quality "type" (we/ dynamic caveats!). Better teams will get some information rents else they tank. Let's find mechanism with no tanking that maximizes prob worst team gets top pick. 2/x

Not great for my comparative advantage, but from some experiments we have done at Rotman, I am totally convinced the vast majority of research that doesn't involve the physical world can be done more cheaply with AI & a little human intervention than by even good researchers. 1/7 Brainstorm interesting question, find all prior lit on this, grab necessary data inputs if any and clean/merge, write theory or statistical model, check extensions, prep writeup, prep replication file. This is 90% of work for papers of that type. What can't AI help with? 2/7

Back to econ in a second, but let's have fun: how about 50 correct moderate takes: 1) Ukrainians are heroes who suffered a ton. 2) Putin obviously covets Georgia and the Baltics also. 3) EU not in talks because they basically have no hard power; France even lost the Sahel. 1/x 4) The far right European parties are bad. 5) Pretending parties that win state and EU elections, and are in govt in NL and AT and IT, aren't legitimate will not end well. 6) And in fact EU & UK speech laws not on side of liberty. 7) (but EU food and kid culture is better!) 2/x

France just held a big AI summit. They want to build up their tech sector. Global prosperity would be enhanced by this. But I can't tell you how insane EU regulation is here, and how it inadvertently enhances market power of big US firms. Let's look at GDPR, for example. 1/x You may think this is just "accept cookie" banners and rules so, like, your personal medical information isn't sold. Not at all. Let's imagine you are a tiny US software company, doing a hundred bucks a month of sales in Europe. What's required? 2/x

Curious about the wide range of reactions to o3 Deep Research. I went through some cases where people found "bad results". I think I understand what is going on - part of this is just bad UX choices by OpenAI, part a misunderstanding of what the model is for. 1/x So o3 is a reasoning model - an LLM. Deep Research grafts web search, python tools, and a fine-tuned model to use those tools to gather additional data on top of o3. It can't read gated articles, has a mediocre OCR for public pdfs, can't read dynamically generated webpages. 2/x

The new OpenAI model announced today is quite wild. It is essentially Google's Deep Research idea with multistep reasoning, web search, *and* the o3 model underneath (as far as I know). It sometimes takes a half hour to answer. Let me show you an example. 1/x So related to the tariff nuttiness, what if you wanted to read about the economics of the 1890 McKinley tariff, drawing on modern trade theory? I asked Deep Research to spin up a draft, with a couple paragraphs of guidance, in Latex, with citations. 2/x

The tariffs are insane; everyone knows that if Congress voted in secret, Senate would be 95-5 against. This whole situation is based on the ignorance of one man, and the lack of fortitude among Congress to push back. It will harm all four countries. So what should Canada do? 1/x Facts: most N. America trade is intermediate goods, Canada is (much) more reliant on US intermediate than US is on Canada, "negotiation" tough b/c the policy is bad for the US as is (hard to threaten someone who isn't even maximizing US welfare!). 2/x

I have friends on that AEA committee on social media but seriously, read this report and tell me if you think it is a fair, scholarly, and nonpartisan look at evidence. There are like a half dozen potshots at people who disagree with a certain political consensus. 1/4 Evidence that doesn't support narrative is buried in appendices (eg, in every question there is no gap in harassment by race, so despite the text, no data on that factor is shown at all), "#econtwitter" is seen as neutral environment (it was very left wing applied micro) etc 2/4

Since immigration reform might be on the table: mostly family green cards and H1B via lottery is crazy. Just auction then off with age and field adjustment and a background check + some method for high grit strivers to come in numbers. Abundance! I'd also make it de facto unrestricted for developed countries where net migration isn't huge. This is status quo for Canadians, Chileans, Singaporeans, Australians by treaty (E3, TN, H1B1). No issues except less friction. If future Jensens, Katlins & Elons want to come let them!

Super excited to publicly launch "All Day TA", a product @joshgans and I have been working on with our team over the last year. Short version: if you teach in spring, you will want to use this! It's the future of higher education. A short thread: 1/x @joshgans Today, we lecture, give exams, see student gets 7/10, move on. Educ lit clear that individual tutoring, tracked to level of each student, much more effective - but costly. In Algebra 1 type courses, orgs like @khanacademy have great million $ tools. What about *your* course? 2/x

For new PhD students in social science (and some old professors - I sure needed this a few years ago!): prepared some notes on what the "2024 PhD Tech Stack" should look like. Goal: your research is replicable, accessible, and efficient. 1/9 Replicable: whether or not the journal requires a full replication file, there is *no reason* every table and result in your paper can't be created with one line of code, and *no reason* you can't find every change you or a coauthor made. It's called version control! 2/9