Ryan Cory-Wright

New paper 🚨 optimization-online.org/2025/01/improv… “Improved Approximation Algorithms for Low-Rank Problems Using Semidefinite Optimization” (w/ Jean Pauphilet) Inspired by Goemans-Williamson’s success in binary quadratic optimization, we generalize to semi-orthogonal and low-rank matrices.

@mmaaz_98 I feel like when they say "it won't work" they mean that they don't think it is implementable within the time limit they have. And "it won't work" is what they have been trained to say in that situation. You can usually get around it by using a more specific prompt.

One of the funniest things the coding agents do is refusing to implement something because they think it won’t work. I’m trying to see if it’ll work, it’s called research!

there is a much more fundamental problem than 50 accepted neurips papers with made up citations. and this problem is over 5200 accepted neurips papers.

Me defending my O(n^3) solution to the coding interviewer.

Here’s a writing tip that will make me sound as old as I am, but it has saved me from costly mistakes. Never put your name on something you didn’t write. This rule will stop you from signing petitions or statements you may come to regret. But it will also save you from shortcuts that will lead to AI editing or authoring your works. It may seem convenient to have the machine write your most useless content, but that habit will accelerate over time. You will spend years developing your voice. A machine can silence it in days with a few clicks. Make it a habit not to lend your name to anyone else’s words or to think you can borrow words because they’re free. They’re never free.

The paper answers the question: "when we make a new discovery that is not explained by existing theory, what background knowledge were we missing?" in a systematic way using algebraic geometry (primary decompositions). Relevant with the advent of many new discoveries from AI

New preprint alert 🚨🚨"Bridging the Gap Between Scientific Laws Derived by AI Systems and Canonical Knowledge via Abductive Inference with AI-Noether" w Karan Srivastava, Sanjeeb Dash, Barry Tragerm @Cristina__C and @InverseProblems Link: arxiv.org/abs/2509.23004

We are recruiting PhD applicants in Analytics and Operations at @ImperialBiz To learn more: imperial.ac.uk/business-schoo… To apply: imperial.ac.uk/business-schoo… The "system" deadline is 2 December; applications will also be accepted from 3-31 December via email (see flyer)

I will be presenting this in Building B Level 2 B201 on 26 Oct at INFORMS in Atlanta. Join us!

Updated paper 🚨 Our paper "Improved Approximation Algorithms for Low-Rank Problems Using Semidefinite Optimization" now contains an approximation algorithm for orthogonality constrained quadratic optimization logarithmic in no. columns. Check it out ⬇️ optimization-online.org/2025/01/improv…

@mmaaz_98 Timely article! Verification is arguably the bottleneck when you use AI for any form of discovery arxiv.org/abs/2509.01398

The Analytics & Operations group @ImperialBiz Imperial Business School is hiring for an open-ranked position (especially interested in rookies). Deadline 8 October, please apply via the link here: imperial.ac.uk/jobs/search-jo…

@NateSilver538 Strong disagree. Attention is All You Need was published at NeurIPS (the top ML conference, with a good review process here: papers.nips.cc/paper/2017/fil…) and has nearly 200k citations. Journalists spent most of the last 20 years giving Andrew Wakefield the time of day.

Academic journals might be a lost cause but they'd probably be better if you had some non-academic practitioners serving as reviewers. Journalists have their problems too but they have much better bullshit detectors, for instance.

@mmaaz_98 Exactly! Add in “oh I make a good salary now so I deserve to go to a fancy gym etc.” and it can go fast. I feel like everyone who makes 100k+ should be required to make a spreadsheet of their expenses before they can complain. Usually it’s inflated expectations haha

@RyanCoryWright Yup for sure, I lived in a little studio for like $1600/mo but that was a COVID-era deal. Then I moved into a 2 bedroom w my friend and the place was much nicer, and I still only paid $1600/mo. But I’d likely want my own place after I graduate and not a little studio haha

Ngl I felt like this w my 50k/yr PhD stipend in Toronto Had a nice apartment, ate out every day, saw my friends all the time, traveled all the time (supervisor covered it). Idk where everyone’s money goes that they think you need 100k to live here…

@profkuang Also look at Bertsimas and Popescu (2005)—for a fixed A, b, Sigma, mu, you can compute lower and upper bounds on the probability by solving a polynomial optimization problem. Cost would quickly increase with the dimension though.

@profkuang No worries! We were recently looking at something related (expectations where you try to relate Sigma to xx’/||x||_^2 to within a constant factor via the semidefinite order) in lemma EC.3 of this paper. Could be related. arxiv.org/abs/2501.02942

Can I get your help X probability trolls? I have matrix, A, vector b, and a normal multi-dimensional normal vector X~ N(mu, \Sigma). I want some kind of large deviations principle bound on P( A x \leq b) where \leq is understood to be entry-wise. If dim(b)=1, this is just large deviation with 1-d Gaussian, but when dim(b) >2, it's measuring some probability of the normal vector satisfying some linear constraint. Looking for (tight) lower bound. It feels like the kind of problem where there's some classical results yielding a rate function that depends on some geometrical properties of A, b, mu, Sigma? @miniapeur @aryehazan @michaelchchoi any hints?

@ErnestRyu Interesting! I found with O3, chat-GPT was quite bad, but with O3-pro there was a significant improvement in proving small things, especially if you ask it to verify numerically (it self corrects if it finds a mistake via the numerics). Similar to advising PhD students!

@RyanCoryWright Gemini sometimes gives me interesting proof ideas that pan out or locates a relevant theorem that I did not know. ChatGPT also does this, but less frequently.

Two cents on AI getting International Math Olympiad (IMO) Gold, from a mathematician. Background: Last year, Google DeepMind (GDM) got Silver in IMO 2024. This year, OpenAI solved problems P1-P5 for IMO 2025 (but not P6), and this performance corresponds to Gold. (1/10)

@ErnestRyu What precisely do you mean by outperforms? Like it’s wrong less often, or it’s more creative? Asking because I have been experimenting with using Chat-GPT to prove things but it seems to make a lot of mistakes, even with GPT-pro

5. In my experience using LLMs for math research, Gemini outperforms ChatGPT. We will see if the next-gen models (which seem to be what OpenAI and GDM are using for IMO) perform at research-level math. (5/10)

@mmaaz_98 IBM Research

What are the most “academic” industry research places?