And if you're faculty: I beg you, for the sake of everyone else in the world, strongly consider creating your own public webpages for your course instead of having everything locked up in Canvas.

Congrats!

First day of class for the semester. Told my 2.5-year-old that "Daddy is going to be a teacher today, just like your teachers!". Response: "Daddy you're so silly". Apparently I lack the gravitas of a daycare teacher.

Glad you liked our lightness paper! I'm pretty excited about that direction.

Incredibly well deserved!!

Awesome, congrats!!

Synthetic Data is all the rage in LLM training, but why does it work? In arxiv.org/abs/2502.08924 we show how to analyze this question through the lens of boosting. Unlike boosting, however, our assumptions on the data and the learning method are inverted.

A proof is a logical argument written to convince a skeptical audience. A corollary is that the best way to read a proof is to roleplay as a skeptical audience.

I think there’s a lot to be said for going all in on something. It was neat being at Google and doing everything on Google systems. On the other hand, faculty autonomy is one of the nice things about academia compared to industry.

We used zoom for lectures, canvas for communicating with students and organizing classes. We have a zoom license but also we’re a Microsoft campus. So classes are zoom & canvas, official stuff with admin is on teams, and most departments have an internal slack.

Yeah, for cross-department you’re stuck with whatever the university is set up on. We’re also on teams, but no professors use it, so we all just use email still.

Might be even more of a long shot in math, but I think that Zulip is actually better than slack, and it’s free for academics. I use it internally in my research group. I know the category theorists like it, so maybe that will convince other math people?

Google meet is surprisingly good too, but Teams-only is crazy. Strongly recommend getting your department to use slack if possible. We just use the free version, and even that is great. Aside from messages, having a #teaching channel and an #advising channel to ask questions is super useful.

We have a department slack that is very active for faculty. That’s now how I mainly communicate with other faculty in my department. Across departments it’s still email, though - no one I know is willing to use teams.

I may or may not have some Hagoromo chalk in my office, if you want to make the best use of those blackboards :)

Interviewer: can you explain this gap in your resume?

Networking researcher: UDP

Just saw a new result on the arxiv: polylog approximation for directed steiner tree! arxiv.org/abs/2412.10744 . Amazing breakthrough in a fundamental approximation algorithms problem. Congrats Bundit!

A statistic I've seen repeated in the discussion of health insurance is that UHC used an "AI with a 90% error rate" to deny claims. Recently I saw it repeated twice here: dailynous.com/2024/12/15/c... in which a moral philosopher struggles to figure out if murdering health insurance CEOs is moral.

Prompt: Draw a picture of an American family at Thanksgiving dinner.

Very interesting! Somehow I completely missed this self-improving literature, but it looks really neat. We're definitely *not* self-improving in this sense -- we're not "learning" the true distribution as we go and modifying our strategy, just continuing to use our fixed (but robust) search tree.

I have similar problems with activation energy :). Not sure if this helps or not, but maybe a point I should have advertised: the entire construction and proof is ~2 pages. The rest of the paper is lower bounds, experiments, discussion, etc.

I always love it when new settings (like algorithms with predictions) end up leading us back to classical problems. Hopefully there's more to be done on "robust" versions of classical optimal data structures!

I think that this is super cool. Somehow, despite optimal BSTs being a super classical question (introduced by Knuth in the 70's and taught in undergrad algorithms), no one seems to have asked the question "what if our distribution is wrong?"

And since H(p) is a known lower bound, our search trees are almost optimal when EMD(p,q) is reasonably small!