Gavin Brown

@OpenAI Not quite what the headline claims? GPT5.2 was a massive help in “simplifying expressions”? And a “scaffolded model” reasoned through the problem? So the scaffolding was contamination, seeding what researchers suspected the answer would be. It’s amazing, but not a true headline…

The authors of the preprint realized this a year ago and sought to find the correct formula for interactions involving any number n of gluons, going up to n=6 by hand but obtaining complicated expressions that they sought to simplify, without success. GPT-5.2 simplified these expressions and then conjectured a simple formula for the general case. Next, a separate scaffolded internal OpenAI model spent roughly 12 hours reasoning through the problem, independently deriving the same formula and producing a formal proof. The final result was then verified by the authors.

GPT-5.2 derived a new result in theoretical physics. We’re releasing the result in a preprint with researchers from @the_IAS, @VanderbiltU, @Cambridge_Uni, and @Harvard. It shows that a gluon interaction many physicists expected would not occur can arise under specific conditions. openai.com/index/new-resu…

My son is learning to count. He said “37, 38, 39, 40… Daddy, what comes after 40?” I said “back pain and mortgages”

@MushtaqBilalPhD @nmatiaszczyk Very interesting that you didn’t check your sources.

@nmatiaszczyk Very interesting.

If you're doing (or thinking of doing) a PhD, read this: 8-year long postdoc, paper in Science journal, 15,000+ citations. Now works as a barista. Academia doesn't care about you. You are totally on your own. www. science. org/content/article/how-chasing-high-impact-publication-nearly-broke-me

@irfnali1 @mic_caprio Very kind of you to say so Ilan :-)

@mic_caprio @profgavinbrown A brilliant paper indeed, hats off to you! 👏 Should be much more widely known, I’m guessing with time it will be.

Very pleased to hit 2^7 citations for a work I consider the most satisfying (and long awaited) of my career. We solved the “ensemble diversity” question. Look forward to seeing 2^8, 2^9…. scholar.google.com/citations?view…

@OmarRivasplata I think I can see my book on the left there….

Stopped by the Main Library The #University of #Manchester

@chrisoffner3d And to be clear also, those fields still are in “active research”, it’s just very different than before.

@chrisoffner3d I didn’t say it was done yet - the transition from academia to industry driving things took 15-20 years in each of those cases. It’s arguable the ML transition began ~2013, so we’re perhaps halfway. I know this from my colleagues in each of those fields, who lived through it.

Unless a significant paradigm shift lowers the computational demands of SOTA methods, computer vision researchers in academia will be reduced to mere consumers of models from industry labs. yashbhalgat.github.io/blog/phd-or-no…

@MushtaqBilalPhD @TheParableMan You know Cambridge is in the UK, right?

@MushtaqBilalPhD @TheParableMan Friendly guy….

A Cambridge PhD says getting a PhD is not hard. Of course, it's not hard for him. For the rest of the 99% it's impossibly hard even to get into any PhD program in Cambridge let along graduate from one.

@MushtaqBilalPhD @TheParableMan Well ok - it’s true in Denmark, but not true in the UK, or several other places that you put under the umbrella of Europe. It depends.

@profgavinbrown @TheParableMan It’s true. You can’t get into a PhD program without a Master’s in any university here in Denmark.

@MushtaqBilalPhD @TheParableMan That’s not true. It’s not uncommon to see someone in the UK starting a PhD after 3 years of just an undergraduate degree. Entirely depends on your grades and experience.

@TheParableMan In the US, you can start a PhD after a 4-year undergrad. In Europe, you do undergrad (3-4 years), then a masters (1-2 years) before you can start a PhD.

@pfau And the other one enabled was this: jmlr.org/papers/volume2… 👍

@profgavinbrown I had no idea!

New (sort of) preprint on arXiv today: a generalized bias-variance decomposition for Bregman divergences!

@pfau My favorite work that it enabled was this: openreview.net/pdf?id=4TnFbv1…

@pfau True in hindsight, but the bijection between ExpFams and Bregmans wasn’t known til Bannerjee (2006), and your paper made it clear to us that the central prediction is a left Bregman centroid, with all sorts of beautiful properties.

Really pleased to welcome our first ever external lecturer at the @EITOxford Centre for Doctoral Training in Fundamentals of Artificial Intelligence at @UniofOxford - @profgavinbrown from @OfficialUoM talking about his work on defining diversity in ensemble learning

New test of LLM intelligence. None of the big models I tried (eg chatGPT) has even come close to solving it. Pick 4 people in your family. Generate an image of scrabble tiles with their names interlocked. It’s not always possible, but not very hard for a human to do if so.

@medha279 @NotebookLM You see quizzes etc on mobile app? I’ve never seen that - just thought it was how the app worked. It only shows me audio overviews.

@NotebookLM I am using the app, not the browser, on both iPhone and iPad. @NotebookLM