It’s one of the coolest environments one can be in. You can righteously wall yourself into a room and solve a problem.
I do hope that this is the start of something cool.
Today they’re one of the biggest compute in the world, they have the strongest data interconnects, multi gigawatts of power and their deeply embedded in science thousands of the greatest scientist in the world work there including thousands on machine learning.
In my opinion, high energy particle physics is a dead end yet still they raise money like nobody in Europe for the next future circular collider they’re asking for close to €100 billion
Clearly, there is a path there where a fraction of that money, infra, and a fraction of the incredible geniouses that work there could be rerouted into a Opus/Mythos/Astra shape open / European system
Why that? I'm not following all that too closely
Makes perfect sense and I’m glad that CERN is focusing on EU providers
It makes more sense if they're the only company who was willing to work on a more custom solution for the proposed LHC model but I don't that a bit as it would be great PR for the frontier labs.
This seems a lot more like an effort to bolster European AI by moving CERN money into Mistral.
What I think the parent post is missing is that breakthroughs in science don't immediately translate into breakthroughs in engineering and even less so into economically viable products.
(in)Conveniently those are also the fields where most ordinary people would experience the most effects from accelerated progress. Dario famously promises potential cures for cancer and dementia.
Advances in math will probably result in advances in science, which in turn will result in advances in engineering (which is probably where the average person sees an actual benefit in their day to day life). But it's not automatic and a lot of the hand wringing by mathematicians about the AI math results include "is this (AI) method of finding proofs actually going to result in as many of those downstream advances as the old fashioned way?"
(I include theoretical CS as a subfield of math - and believe the S in it is a misnomer).
On the other hand, you could downplay the math findings a bit, since it didn't further our understanding of mathematics just yet, only provided proves to things we didn't have proves for.
Less modern AI (aka. machine learning) is used all the time in science. I don't know of any commercial breakthrough yet, but there are some promising developments, like in antibiotics. If it materializes, that would be a very big deal.
Gerard t' Hoofd (under Martinus Veldman) proved in 1971 that Yang-Mills gauge theories are renormalizable, using a rule-based AI (LISP-styled but manually written in assembly) called "Schoonschip" (I'm guessing called that because there were several failed attempts that went before and failed). This program was an AI in the sense of a search algorithm searching symbolic proofs. It got the 1999 Nobel prize in Physics.
This was hugely important because first, it showed that it was possible to make the Standard Model of physics and it also helped beyond that by showing for a lot of (at that point considered) impossible theories exactly how to make them work (meaning not just a proof that it could be done, but a long list of "if you have problem X, do Y" recipes). One of these became first the "electroweak unification", then the standard model.
Hao Wang proved that the logic of "Principia Mathematica" was actually correct using another symbolic proof generator. His efforts were also the start of many other such programs that contributed lots of stuff to mathematics that nobody sane was ever going to do, such as finding all the higher-order versions of the quadratic formula (there's a lot of them).
At Stanford university, tens of thousands of molecules were found in 1965 by having an AI, this one was what's called an "expert system", evaluate mass spectroscopy data and combine it with chemistry knowledge in the form of rules about what was possible or not possible and if only one option remained, spit it out. It almost doubled the sum total of our chemistry knowledge if counted by known molecular structures. It was later adapted to make mass spectroscopy useful in finding the composition of samples.
Then there's Alphafold, first a ConvNet (Resnet+Convnet to be exact), then switched to the transformer architecture, which gave use the folding of just about every protein we have ever found, and even more proteins that we've never even found. Nobel prize in Chemistry 2024.
I don’t think it’s going away, LLMs and other ML stuff will no doubt remain part of people’s workflow, but I also see no evidence for superabundance or anything close to it.
As opposed to enshitifying the lives of the working class in most other areas.
https://arxiv.org/abs/1207.7214 - The original ATLAS Higgs discovery paper uses neural nets to identify photons. It also uses a neural net b-tagger to improve the H->WW signal to noise ratio.
https://arxiv.org/abs/1303.0763 - used neural nets extra tt+jets noise https://arxiv.org/abs/1503.05066 - used neural nets for signal/background separation.
Um, I believe I have linked you to "actual analysis" where techniques used are more exotic than boosted decision trees.
One scenario, and I think what people were paranoid about, is that you could make your cuts really arbitrary, and select a region where you by accident have too few simulation events and too many data events - say "Select all events where the first muon has momentum > 15.6 GeV, but not between 20.1 GeV and 20.3 GeV, unless there are 3 jets, in which case the selection is...". You just try out so many regions until you find something. For this reason you need to keep every step of your analysis justifyable, and a BDT is somewhat more explainable than a neural net (it's essentially just a trained list of simple cuts). Another thing that is frequently done is a blind analysis, where you only look at the real data when everything is finalized and locked in.
There are other issues beyond cherry picking that determine which methods you can use. E.g. the simulation doesn't reflect all correlations between variables, and it doesn't simulate the hardware triggers perfectly - so your simulations might behave different wrt. certain cuts than your data.
Nevertheless, it is not like you are banned-banned from trying different methods. When I was in particle physics (ATLAS and D0), we tried all kinds of creative things including NNs. But you have to convince hundreds of collaborators to put their names under your paper and you have to pass internal peer-review, which can feel a bit conservative and frustrating.