Researchers from Harvard, Keyon Vafa (Keyon Vafa) and MIT, Peter Chang (ᴘᴇᴛᴇʀ ɢ. ᴄʜᴀɴɢ), Ashesh Rambachan (Ashesh Rambachan), and Sendhil Mullainathan (𝐒𝐞𝐧𝐝𝐡𝐢𝐥 𝐌𝐮𝐥𝐥𝐚𝐢𝐧𝐚𝐭𝐡𝐚𝐧) have published what I consider the most interesting study on the abilities of AI models in 2025. They wanted to address

Hannah's absolutely goated!!! I'm so proud of her for working extremely hard the past several years, making it to the IMO, and bringing home the silver 🎉🥈🎉. And she's a great friend too!

Rivka Lipkovitz yeah! like on top of being so incredibly goated shes also just one of the kindest human beings we know. so blessed to know her

🚨❗ Results of the FIRST EVER trial comparing ultra-processed food (UPF) and minimally-processed food (MPF) diets following healthy dietary guidelines on health (and the longest trial of UPF to date) published today in Nature Medicine What did we find? A Thread… 🧵 1/

But I asked AI and it said it was true!! 🤡

Really great eval LLMs are already great at next token prediction in the training corpus, so the test for truth-seeking is now whether they can answer correctly when the correct next token is at odds with the next token in the most similar example in the training corpus.

I fine-tuned GPT-4o to outperform Latin BERT at restoring ancient Latin texts—for under 2% of Latin BERT's training budget. This suggests that researchers can fine-tune modern LLMs to emend other texts or even languages at a fraction of the cost. rivka.me/edging-out-lat…

I think about this post once a month

Spaced repetition is truly amazing! Using Anki, I learned all of the MIT biology curriculum in under 6 weeks and passed the advanced standing exam, earning credit towards one of my graduation requirements. Highly recommend Anki if you're studying a memorization-heavy subject.

If you're a senior in high school, I highly recommend applying for this

I wonder how things would be if we learned real analysis before calculus, or if our first introduction to algebra was linear algebra. Maybe we would go slower in the beginning but it would pay off later?

Why don't we band together on social media and get words like "overmorrow" (day after tomorrow) and "ereyesterday" (day before yesterday) added to the Cambridge dictionary?

Thrilled to share that **fect** has won the 2025 Best Software Award from the Society of Political Methodology. We're honored! polmeth.org/statistical-so… To celebrate, we've just released fect v2.0.5 on CRAN & Github 🎉

Apparently taking the stairs is faster than taking the elevator at my college dorm (and if you run, it's >10s faster!) I took these measurements when the elevator was already at the right floor and didn't make any stops, so in practice, taking the stairs is even more efficient.

Guys, I found an arbitrage opportunity

Is it kosher to write up chemistry and economics problem sets in LaTeX... I swear it actually saves a lot of time!

18-year-old Rivka Lipkovitz used statistical modeling to show that states adopting strict voter ID laws saw a 2.4% drop in voter turnout. Her research helped her win fifth place at this year’s #RegeneronSTS and a $90,000 prize. Forbes forbes.com/sites/kevinand…

The U.S. #Physics Team won all five #gold medals at the 2025 International Physics Olympiad (IPhO) in #Paris, #France, with #team members Agastya Goel, Allen Li, Joshua Wang, Feodor Yevtushenko, and Brian Zhang earning.