←
OptimaLab509 posts
OptimaLab avatar

OptimaLab

@optimalab1

Optimization for ML at Rice University (CS) led by Associate Prof. Anastasios Kyrillidis - Efficient training methods, non-convex optimization, and more.

◷ Joined December 2020▧ 71 Photos and videos
OptimaLab avatar
OptimaLab@optimalab1·Sep 29

So maybe the real question is not which field suffers more. It is whether all of us, students and professors, mathematicians and computer scientists, stay active and keep learning while the map gets drawn.

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

@TaoistTerence uses AI tools openly in his work. I read that as an invitation to all of us. Nobody has the map yet, not for research and not for teaching.

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

Think back to before search engines. Person A worked out everything alone. Person B searched, found two papers, and combined them into an answer. Is B less capable? Is B cheating? Or should A get in sync with the new tool?

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

Knowledge needs people who read it and care. Learning a new technique still excites me. That is why I became an academic.

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

In 1931, Gödel showed that no fixed set of axioms can prove every true statement about arithmetic. No one, human or machine, will ever finish math.

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

Most AI-for-math headlines so far solve problems humans already posed. Impressive. The harder part is posing new ones worth solving. AI will propose some strange ones; we still need mathematicians to tell which ones matter.

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

Math is not a handful of famous problems. I'd bet thousands of agents are already working on P vs NP. Fine. Beyond the headlines sit countless open questions. AI will not solve them by itself; someone has to poke it.

OptimaLab avatar
OptimaLab@optimalab1·Sep 29

Long time, no post (everyone is "running" freaking crazy these days...) Personal thoughts from chats with colleagues at Rice: Will AI hurt mathematics more than computer science? My answer (probably I'm wrong): AI will change how we do math. It will not end it. 🧵 (For Rice colleagues: a coffee conversation with no single right answer. Excellent excuse to have a cup of coffee ☕)

OptimaLab avatar
OptimaLab@optimalab1·Sep 1

It has never been easier to skip the meeting: an LLM iterates faster than four people in a room. But the whiteboard leap (the wrong idea that becomes the right one at minute 20) never came out of a chat window. My recipe: AI to prepare, the room to decide. What's yours?

Post attachment
OptimaLab avatar
OptimaLab@optimalab1·Aug 3

All numbers are classical experiments, 5 seeds, sanity-gated. No quantum solver was executed. Running them end-to-end on these instance families is the open problem.

OptimaLab avatar
OptimaLab@optimalab1·Aug 3

We also test our own method (Catalyst-PPA, AAAI 2026): with a classical inner solver the proximal trade-off has no interior optimum. The optimum is on the boundary, and the boundary is "don't precondition." Preconditioning pays only when the inner solver is κ-linear. CG isn't.

OptimaLab avatar
OptimaLab@optimalab1·Aug 3

So instance-easiness erodes the κ axis (where quantum was already behind) and leaves the N axis, where the exponential lives, untouched. That asymmetry is the post's actual contribution.

OptimaLab avatar
OptimaLab@optimalab1·Aug 3

Quantum κ_eff forgives you for having no weight on small singular values. Krylov forgives you for having weight on few eigenvalues; wherever they sit. Put b on 4 eigen-directions at the small end: → CG: 3 iterations → κ_eff: 100,000 Classically trivial. Quantum worst-case.

OptimaLab avatar
OptimaLab@optimalab1·Aug 3

Both fields discount the worst case. Both do it with a functional of the same object: the spectrum of A weighted by b. They discount different things.

OptimaLab avatar
OptimaLab@optimalab1·Aug 3

That κ is pessimistic isn't news. Classically it's been known since the 1950s, and the quantum side now has solvers governed by an effective condition number rather than κ. The post concedes all of that up front. The interesting part is what comes next.