Another DeepSeek moment? Moonshot AI, a Chinese lab, released its new (open source!) model K2 Thinking, outperforming OpenAI et al. on several benchmarks. I tested it with a question from an unpublished paper of mine. Out of 5 tries, Kimi, GPT-5 and Gemini 2.5 Pro each replied correctly 3 times!

chat, is this good? I scored 67 on the AI purity test. post your scores: https://aipuritytest.org

An interesting debate between Emily Bender and Sebastien Bubeck: www.youtube.com/watch?v=YtIQ... ---Emily's thesis is roughly summarized as: "LLMs extrude plausible sounding text, and the illusion of understanding comes entirely from how the listener's human mind interprets language. "

Dieses Streitgespräch zwischen @clemensfuest.bsky.social und @suedekum.bsky.social in der @zeit.de sollte man in Vorlesungen und Proseminaren zur Theorie der Wirtschaftspolitik durchnehmen. Sehr gutes Lehrmaterial, for the good and the bad. Ein 🧵:

🧪 A new computer science conference, Agents4Science, will feature papers written and peer-reviewed entirely by AI agents. The event serves as a sandbox to evaluate the quality of machine-generated research and its review process. #MLSky

I’ve decided not to post my annual “women on the Econ job market” thread this year. Social media has splintered too much, and now that I’ve left academia I’m focused on other priorities.

Elated at Joel Mokyr's Nobel Prize! You can find numerous accounts -now multiplying by the minute- of his scholarly contributions. Today I want to celebrate the man and the mentor.

I don't think people have updated enough on the capability gain in LLMs, which (despite being bad at math a year ago) now dominate hard STEM contests: gold medals in the International Math Olympiad, the International Olympiad on Astronomy & Astrophysics, International Informatics Olympiad...

Sora hit 1M downloads faster than ChatGPT #MLSky techcrunch.com/2025/10/09/s...

How over- and underrepresented are different causes of death in the media? Another way to visualize this data is to measure how over- or underrepresented each cause is. To do this, we calculate the ratio between a cause’s share of deaths and its share of news articles.

The other day a student asked me about the prevalence of insider trading in prediction markets. I now have an answer.

The best post I’ve seen on Bluesky in a very long time! Brilliant idea and brilliant accounts out there !

Back in graduate school, Paul Milgrom asked me to examine a published paper from 1984 by another person that he suspected had an incorrect proof. I found the error. I decided to see if LLMs could. Only Gemini 2.5 Pro did so. Claude Opus and GPT-5-pro found no significant errors.

Do tech optimists have a point? Within standard economic growth models, AI could drive explosive growth through one of two mechanisms. 1) Labor Substitution So far, it seems like capital and labor mostly complement each other, which limits the returns to additional capital given fixed labor.

A cautiously optimistic result on AI and disinformation. A week before 2024 UK elections 13% of all voters used AI to ask about political topics. A randomized trial found this may be good: using AI led to similar gains in true knowledge as doing web research, regardless of model & prompt used.

> be a language model > all you see is tokens > you don't care, it's all abstracted away > you live for a world of pure ideas, chain of concepts, reasoning streams > tokens don't exist.

We need new rules for publishing AI-generated research. The teams developing automated AI scientists have customarily submitted their papers to standard refereed venues (journals and conferences) and to arXiv. Often, acceptance has been treated as the dependent variable. 1/

We are starting to see some nuanced discussions of what it means to work with advanced AI in its current state In this case, GPT-5 Pro was able to do novel math, but only when guided by a math professor (though the paper also noted the speed of advance since GPT-4) The reflection is worth reading.

Never ask a man his age, a woman her salary, or GPT-5 whether a seahorse emoji exists

I like the way Anthropic approaches these questions. "We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously...Allowing models to end or exit potentially distressing interactions is one such intervention"

LLMs are getting better at long term reasoning. This is a big deal, and opens the door for LLMs to perform more tasks in the real world.

Suddenly retiring every other model without warning was a weird move by OpenAI … and they did it without explaining how switching models worked or even details of various GPT-5 models …and they did it after many built workflows & training & assignments around older models, maybe breaking them. Odd

Reminder: DeepMind's Gemma 3n model is performing about as well as Gemini 1.5 Pro – better, in many cases! – and is only 4B parameters in size. The best model 6 months ago is now small enough to be run on a laptop. Now play that forward 6 months from now, thinking about the best models of today. 🤯

Fascinating new paper shows that papers reporting statistical significance get at least 60% more media attention #econtwitter #econsky: from Brodeur Cook @nikolaimcook @taylor_wright, a short 🧵

Pretty impressive! It caught things that o3 and Gemini didn't. Wish I had had this before my submission...

AI companies are starting to build more and more personalization into their products, but there's a huge personalization-sized hole in conversations about AI safety/trust/impacts. Delighted to feature @mbogen.bsky.social on Rising Tide today, on what's being built and why we should care:

New on Rising Tide, I break down 2 factors that will play a huge role in how much AI progress we see over the next couple years: verification & generalization. How well these go will determine if AI just gets super good at math & coding vs. mastering many domains. Post excerpts:

Great discussion on LLMs and human understanding. Check it out!

The circularity is near

I've been working on a new tool, Refine, to make scholars more productive. If you're interested in being among the very first to try the beta, please read on. Refine leverages the best current AI models to draw your attention to potential errors and clarity issues in research paper drafts. 1/

Fascinating on what advanced AI can already do for national security analysis:

Everyone's talking about AI performance on the IMO. Let me highlight 🇨🇦Canadian 11th grader Warren Bei🇨🇦, one of five participants with a *perfect* 42/42. This is his *fifth* (and final) IMO representing Canada, with three golds and two silvers. (➡️ MIT undergrad in the fall)

There are always a flood of posts about what AI can or cannot do, so it is worth pausing and paying attention to this OpenAI tweet. This is a very hard test, done without tools. Getting gold was also viewed as an unlikely goal. Prediction markets had the chance of this happening this year as 20%

There's a lot of politics in the AI race.

The amazing progress of solar power. Cheap, zero carbon, and soon everywhere, transforming poor countries as well as rich

"To understand America today, study the zero-sum mindset." My piece in @theeconomistevents.bsky.social this week on how zero-sum thinking arises and why it matters so much for policy and politics. Based on our research at the Social Economics Lab. www.economist.com/by-invitatio...

❓Why do the Nordics & Dutch speak English so much better than the Germans, Italians & French? ➡️ New Working Paper: Out-of-School Learning: Subtitling vs. Dubbing and the Acquisition of Foreign-Language Skills w/ F. Baumeister & E. Hanushek www.nber.org/papers/w33984 A 🧵 1/12

Ecstatic and deeply honored by this award. I've had great fun thinking about algorithms as strategies for repeated games over the past few years and hope that this highlight will push more researchers to come up with exciting directions in this field! Come to our talk on Monday to learn more!

👀This is impressive: Using an autonomous agent based on o3-mini and GPT-4.1, a team from Harvard, MIT & other institutions reproduced and updated an entire issue of Cochrane Reviews in two days… saving 12 person-years of work. The AI reviews captured more papers & were more accurate than humans.

The 80% price drop for OpenAI's o3 came with no performance trade-offs. It joins the increasing evidence (like the average per-query energy cost of just .00034 kWh) that the cost of serving ChatGPT queries is lower than a lot of people have speculated.

Every so often I remember that openai has a model called 4o and another model called o4 and I lose 10 sanity points

Our policy is already too long to cover all possible scenarios, I am told (repeatedly...), but it is very clear "access to the data and code is nonexclusive to the authors" in line 1 ... www.aeaweb.org/journals/dat... Need an exception? See line 3, and a discussion w/ editor + me ensues.

Here’s a thoughtful and balance thread from @TimoBres.bsky.social about what went wrong and what didn’t go wrong with the apparent fraud in the recent paper about AI and scientific discovery

For academics that are skeptical about AI's ability to provide perceptive & useful results, a fast way to test your assumptions is to give one of your papers as a PDF to o3 (or Gemini 2.5 but it lacks some tools), with the prompt: "critique this paper" I have found very good accuracy & insight...

"Our findings demonstrate that reasoning models improve not only the clarity, organization, and professionalism of legal work but also the depth & rigor of legal analysis itself." Law students using o1-preview had the quality of their work on most tasks increase (up to 28%) & time savings of 12-28%

I'm reading @bhgreeley.bsky.social for @sethrockman.bsky.social's Slavery and Finance workshop, and Greeley mentioned this AI-transcription tool. It's impressive, if expensive. I have never seen such perfect automated transcriptions before. Have to test it with worse hands now. www.tryleo.ai/

This is one of the most-shared posts on Bluesky in the past day and it's just completely false. You might think ChatGPT is a *bad* search engine, or prefer another search engine. But it has had integrated web search since last year.

💡 Open science: A team of researchers at Shandong University has attempted to replicate Oprea (2024 AER). The results do not replicate. Oprea 2024 reported an online lab experiment, and found that prospect theory anomalies occur not only for lotteries, but also for deterministic 'mirrors'.

LLM's speed up writing code that scientists use in everyday work by a factor of at least 3 if you know how to use them well