I'm reading Holmstrom 1979, and I'm a lazy student who doesn't want to actually do the work, so I'm making ChatGPT do it for me. It is very simple, I just give o4-mini screenshots of parts and ask, wtf is going on. It tells me exactly what Holmstrom is saying, much more clearly

Congratulations to Denis Auroux, Herchel Smith Professor of Mathematics and one of five faculty members awarded a Harvard College Professorship for excellence in undergraduate teaching! #harvardmath news.harvard.edu/gazette/story/…

Looking forward to seeing those who are attending the JSQPS 2025 Summer Meeting! 登録しました。7月のJSQPSでご参加の皆様にお目にかかるのを楽しみにしております！ sites.google.com/view/jsqps/202…

I spoke at a high school yesterday and someone asked what I'd tell my 18 year old self if I could. It would be not to be tricked by prestige. It is at best a trailing indicator, and often simply wrong. Curiosity is a better guide.

A few years ago I wrote something about all the different things I've worked on over the years, and I was amazed how often I'd been tricked by prestige: paulgraham.com/worked.html

Tim Rocktäschel yes exactly, "training" an LLM on a target domain should output a manual not a weight diff.

There are two approaches to learning things: 1. Bottom-up: read textbook, learn foundations, and build up knowledge ground-up 2. Top-down: "How do rockets work?" Start from a question; backfill physics knowledge until the answer to the question makes sense

A few simple tricks: 1. Make it explain and document clearly what its code is doing 2. Make it build a lot of tests/debugging Because it's so fast, you can make an LLM do 1. and 2. well beyond what a human would do very quickly

Debugging/tests which would be overkill, and not worth the dev time, for human devs, become worth it when the error rate is higher, but the rate at which code is produced is also much higher. Similarly, it becomes extremely cheap to ask for very clean code

The thing I'm realizing is lots of people don't get how revolutionary the o3/o4 class of models are, for (econ) research. 4o and the like models are borderline useless for econ research-level math. Deepseek is borderline useful. o3/o4 are at the level of solid PhD students

Without its international students, Harvard is not Harvard. hrvd.me/IntStudents25t

Brings me no pleasure but,

東大がハーバード留学生を一時受け入れへ　「学びの継続に貢献を」 asahi.com/articles/AST5V… 一部の授業の受講を認め、将来、復学や進学をした際に、東大で受けた授業の単位が認定されるように履修証明書の発行などもするといいます。

「ハーバード大学の政治学部と統計学部で教授を務める今井耕介さんは…『今はとにかく集中していい研究をすることだけに力を注ぐようには言っているが、本当に毎日状況が変わっているので難しい』」 www3.nhk.or.jp/news/html/2025…

「日本の社会も10年、20年後にこういう問題が出てくるかもしれない。これを機会に、日本社会をどういう方向に動かしていけばいいかを考えてほしい。」

IMO, a major driver of PhD attrition is: do you have the willpower to finish a paper A full published econ/finance paper is like well over 100 pages of work after revisions and such, it's longer than the longest thing many students have ever done at that point in their lives

Our school at Waseda University is recruiting for open-rank positions in Political Science (open subfield). We welcome applications from scholars at any career stage. Come join our dynamic and internationally oriented academic community in Tokyo. waseda.jp/fpse/pse/news-…

int’l students are the heart of our department - it is absolutely ridiculous that they have to make the choice between visiting family and continuing their degrees

Harvard adds Trump's travel ban to its lawsuit in The Harvard Crimson via Samuel A. Church, Dhruv Patel, Cam Srivastava, Grace Yoon: thecrimson.com/article/2025/6…