Gradient optimization methods: the benefits of instability — Peter Bartlett, UC Berkeley www.youtube.com/watch?v=wEgT... #MathSky #SMRISeminar

I really need to switch to a newer Obsidian Zotero connector, but I've got such heavy scripting debt in this ancient abandonware plugin that I cant fathom how long it'll take me to port everything over

Accepted at #ITCS2026: congratulations to Abigail and Vikrant! itcs-conf.org @abigailgentle.com

Just caught myself writing D(P\|Q) on paper

I'm running a mock-conference for students using OpenReview, and have developed an appreciation for how nuanced that software is. But also an appreciation for conference chairs because of how much of a headache that software is

My petition to the 🇦🇺 Australian government: make part-time PhD students' stipends tax exempt! 📋 Read and sign here: www.aph.gov.au/e-petitions/... ⏰ Deadline: October 1

with great power (arXiv endorsement ability) comes great responsibility (quack emails)

New (solo) #privacy paper by @abigailgentle.com, PhD student at @sydneycompsci.bsky.social: "Necessity of Block Designs for Optimal Locally Private Distribution Estimation," to be presented at the Information Theory Workshop (ITW'25)! arxiv.org/abs/2508.05110

Thank you differential privacy bot :)

Workshop paper up on arXiv! Bringing together a collection of necessary conditions on optimal LDP frequency estimation algorithms, and showing they restrict us to a specific class of combinatorial objects. arxiv.org/abs/2508.05110

submitted a paper to arXiv, which has already passed peer review, but I'm still staring at a wall wondering if there's anything obviously dumb I've missed

P=BPP is so interesting a problem it almost makes me want to study complexity theory

lunch with a colleague

Greatest unsolved problem of our time: creating nice tables in a document

Does anyone know of a good collaborative pdf mark-up website or program? Kami works ok, but is a bit broken, regularly asks for permissions, and needs aggressive access to either Google drive or OneDrive

God bless the LLMs: "I added a new macro for X late in writing, please identify where I've used X and replace it with the macro" is saving my life regularly (I plan my writing poorly)

Annoyed that vscode has turned out to be the best LaTeX editor I've used so far. I'm sure nvim is competitive, but I haven't used it since undergrad

Quite unbelievable density of cool talks going on, even while its past midnight in the US where most of these researchers are based

Fiddling with llms: Gemini seems to hallucinate a lot more obviously than any other models I've tried when asked for literature reviews

Optus (phone provider) doesn't have our current address in it! I was so shocked I had to download the national address database and make sure our fairly new apartment is in it, and of course it is. Optus, put me in charge of your db I'll remember to download the quarterly updates for a modest salary

I've been to one baseball game in my life and it was to see the Sox fail badly against the second worst team, which of course made me an instant fan

Best Chicago controversy of the new Pope is the Cubs trying to lay claim to him and his brother having to speak up to clear up that he's a Sox fan. blockclubchicago.org/2025/05/08/p...

The 70s and 80s had way more of a randomised response boom than I ever imagined. These authors imagined the topic being taught in introductory statistics

I've been trying out Typst a little and I'm glad to see them taking a definitive stance on the superior epsilon

Being a researcher is so hard! I'm still working through the 3000 ICML papers from 2024 and now there's going to be another 3000?

In highschool our English teacher gave us an exercise to condense a paragraph into as few words as possible while retaining all of its meaning. It was one of the hardest things I've ever done. Relatedly I'm writing an abstract right now

Feel very proud of my first sole author paper, hope I get a chance to share it soon!

The cost of not paying attention in linear algebra class is higher than I ever could have imagined

Fun fact! Hadamard matrices have been used in Randomised Response procedures since 1979! (Raghavarao and Federer), and they sometimes even shuffled all the responses in a box! Not exactly Shuffle DP but I feel like they deserve a little credit

It is remarkable how few quality of life features overleaf has given its total market dominance. If good multiplayer comes to @obsidian.md, and I figure out how to use pandoc properly I vow never to return

I've graduated from "foundational 60s monograph" to "80s graduate textbook" and soon I may even read "2000s journal article" levels of combinatorial competency

you know things are going well when they break out the turnip photos

Doing discrete statistics: "No, it's fine, i just... i didn't think it would be combinatorics."

This form of writing out derivations in full sentences is so charming