The net effect of strong LLMs will be to make it easier to find information and learn.

Dimensional lumber is out.

Distributional lumber is in.

If δ=0, we write ε-DP and call it “pure” DP. If δ>0, we call it “unclean.”

I wrote a survey article on computationally efficient methods for "robust" mean estimation, including robustness to contamination, heavy-tailed data, or in the sense of differential privacy.

The same ideas are useful for all 3 (seemingly-different) forms of robustness! 1/2
arxiv.org/abs/2412.02670

It’s bad when reviewers are extremely wrong, but it feels worse when they’re extremely right.

Instead of a special poster session, NeurIPS should use physical spotlights to identify exceptional work.

Excellent: “Energetic Aliens”

stephenmalina.com/post/2021-07...

Review: Serious issues with presentation meant I could not interpret the results.

Rebuttal: Great, you’re saying we made a breakthrough and just need to write it up better.

Ceci n'est pas une Annoucement Sign.

“We went to a restaurant once, and it made a huge impression on us.”

William Sealy Gosset developed the Student’s t-test as part of his work as Head Brewer of Guiness.

Pearson was already working with big data.

A few of us are going to corner the market on efficient differentially private mean estimation in Mahalanobis norm, really drive up the price.

Never ask a learning theorist about their algorithm’s run time. If it’s good, they’ll bring it up themselves.

Not a big fan of things changing. I’m still secretly hoping everyone will get back on AIM and Xanga.