Man, when I was doing my PhD at CERN the most far out thing that we were generally allowed to use was boosted decision trees :) getting anything other than that approved for any Higgs-related analysis was… not a thing.
Using neural nets for signal background separation was pretty standard in ~2012 when the Higgs discovery was published. Was also pretty standard for CDF + DZero papers that came out before the LHC turned on.
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.
What they are doing here is selecting a number of events (particle collisions) and comparing with a simulation of the standard model (sans Higgs or whatever they are looking for). If there is a significant excess, then you found something.
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.
Finally! As a SWE, I was getting worried scientists would be left out of the increasingly-worried-about-getting-replaced club that SWEs and mathematicians are currently enjoying being members of.
I like this, and CERN might be closer to a potential European solution for us than you think, CERN was built on the premise that nuclear technology was too dangerous for anyone single nation to have. So they decided to collaborate.
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
HEP theory should be abandoned because supersymmetry is a failed theory that predicts particles that were not found in the 100-1000GeV range at LHC, and HEP experiment should be abandoned because supersymmetry is an excellent and natural theory that predicts no new, reachable particles under 1PeV. This interpretation is known as, "supersymmetric planning," and produces a well-behaved vacuum after causing cancelation.
I’d be surprised if that isn’t already the case with CERN itself already being a multi-national collaboration project and with things like GÉANT https://geant.org/ existing
It makes sense in a political way but I'm not sure if it's the best choice for scientific development seeing as Mistral has been pretty behind the curve it's entire existence.
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.
I'm not at all anti-AI, but I'm still holding my breath to see real scientific contributions coming from modern (LLM/CoT/agent) AI. Much of the premise behind "AI will make superabundance a reality" hinges on getting AI-driven science that affects everyday lives Anthropic's quasi-CRISPR-like discovery did not seem super compelling, and that is still the best I have seen. If anyone has more heartening stories that could convince me to be an optimist, please do share ;)
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.
The distinction for many people is that math and cs feel like very natural domains for LLMs (and importantly, domains that enable quick iteration and verification). Much less applicable with eg biology and other fields that require lots of physical experimentation.
(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.
I think there's a pretty clear distinction between math and science. The former builds theoretical models of what certain axioms and assumptions imply. The latter builds actual models of what we see and measure in the real world.
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).
I have a feeling that many more fields are utilizing AI and having some success with it and we just don't know about it in our STEM bubble.
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.
If you are limiting yourself to modern whatever, you are not going to find that a lot of science that affects everyday lives. Between the first results and practical applications, it can take many years. There is a reason Nobel prize winners tend to be old.
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.
I think maths is really the only solid case so far. I do wonder whether the big early contribution to experimental science will be in speeding up the scientists conducting it, at least until some non-negligible boundary in simulation fidelity is crossed.
AI has an incredibly long history of contributing to science and engineering. Some highlights:
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.
There’s been a tendency since well before LLMs for AI demos to focus on “here’s a new problem we made up that it solves that you could imagine will soon transfer to an existing real problem” as opposed to “AI solves X”. This has been an ongoing challenge, it does cool stuff and looks so promising but has been very hard to harness for useful work. It’s still going to take time before the hype dies down and we see what’s left.
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.
Saw a post recently about how AI agents were used to check if papers had flaws in their conclusions and it found 15% did, that kind of auditing is already a massive massive advancement for science
For me, the issue is the whole piracy and unauthorized use of intellectual property with no rewards, and the environmental damage intensifies it by 1000x. I know that this has been the case for many different situations in human history, but that doesn't make it any better.
I think there is a large group of commenters who actually just hate silicon valley tech companies, and they're fine with AI as long as it comes from someone else.
AI has been beneficial in multiple research fields across maths and science. That is to say, it's actually benefited humanity in those particular use cases.
As opposed to enshitifying the lives of the working class in most other areas.
Maturity is understanding there's an enormous range between "like" and "dislike", between "love" and "hate", between "good" and "bad", and its also possible to be neither at once or both at once.
Using AI as a tool to investigate the mysteries of the universe is a huge win for humanity. Generating slop images to confuse people online is not. AI is tool, we like when tools are used for good.
Man, when I was doing my PhD at CERN the most far out thing that we were generally allowed to use was boosted decision trees :) getting anything other than that approved for any Higgs-related analysis was… not a thing.
Using neural nets for signal background separation was pretty standard in ~2012 when the Higgs discovery was published. Was also pretty standard for CDF + DZero papers that came out before the LHC turned on.
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.
Yep, you’re right about this. I worked on some of these things, but for the actual analysis it was generally all cut-based.
> actual analysis
Um, I believe I have linked you to "actual analysis" where techniques used are more exotic than boosted decision trees.
I stand to be corrected :) I knew about the tt stuff, but I had no idea about the 7 TeV photon analysis! Every day is a school day :)
We surely came a long way since then. I recently finished my master thesis on self-supervised foundation models for jets :)
You needed approval to try different methods for data analysis?
What they are doing here is selecting a number of events (particle collisions) and comparing with a simulation of the standard model (sans Higgs or whatever they are looking for). If there is a significant excess, then you found something.
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.
They once spent 2 months to be allowed to use imaginary numbers
Say more please!
The thing a CERN is: You have some of the smartest and most educated people on the planet, and they are all super motivated to excel.
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.
Finally! As a SWE, I was getting worried scientists would be left out of the increasingly-worried-about-getting-replaced club that SWEs and mathematicians are currently enjoying being members of.
One of us! One of us!
Join us, in the pit of despair!
I like this, and CERN might be closer to a potential European solution for us than you think, CERN was built on the premise that nuclear technology was too dangerous for anyone single nation to have. So they decided to collaborate.
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
> ... high energy particle physics is a dead end...
Why that? I'm not following all that too closely
https://en.wikipedia.org/wiki/Desert_(particle_physics)
HEP theory should be abandoned because supersymmetry is a failed theory that predicts particles that were not found in the 100-1000GeV range at LHC, and HEP experiment should be abandoned because supersymmetry is an excellent and natural theory that predicts no new, reachable particles under 1PeV. This interpretation is known as, "supersymmetric planning," and produces a well-behaved vacuum after causing cancelation.
I hope European university networks and their underfunded labs will join forces as soon as possible with other European national research centers.
I’d be surprised if that isn’t already the case with CERN itself already being a multi-national collaboration project and with things like GÉANT https://geant.org/ existing
with a different title, they would have achieved much higher engagement. e.g. CERN and Mistral evaluate if large latent models can advance physics
Internet 2.0 here we come!
This is great progress in fundamental research and I’m thrilled to see Mistral leading it.
Makes perfect sense and I’m glad that CERN is focusing on EU providers
It makes sense in a political way but I'm not sure if it's the best choice for scientific development seeing as Mistral has been pretty behind the curve it's entire existence.
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.
Great news, let them cook.
Wonderful news to not see the duopoly accelerating real science.
I'm not at all anti-AI, but I'm still holding my breath to see real scientific contributions coming from modern (LLM/CoT/agent) AI. Much of the premise behind "AI will make superabundance a reality" hinges on getting AI-driven science that affects everyday lives Anthropic's quasi-CRISPR-like discovery did not seem super compelling, and that is still the best I have seen. If anyone has more heartening stories that could convince me to be an optimist, please do share ;)
Is math or theoretical Computer Science not "real science"? Whatever one may think about the process, the results are real and are coming from LLMs.
It is.
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.
I'd expect that process to be accelerated by AI more than the process of novel discovery.
The distinction for many people is that math and cs feel like very natural domains for LLMs (and importantly, domains that enable quick iteration and verification). Much less applicable with eg biology and other fields that require lots of physical experimentation.
(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.
I think there's a pretty clear distinction between math and science. The former builds theoretical models of what certain axioms and assumptions imply. The latter builds actual models of what we see and measure in the real world.
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).
Additionally, engineering is distinct still.
well they don't affect anyone's day-to-day lives, which is what the person you're replying to is talking about..
I have a feeling that many more fields are utilizing AI and having some success with it and we just don't know about it in our STEM bubble.
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.
If you are limiting yourself to modern whatever, you are not going to find that a lot of science that affects everyday lives. Between the first results and practical applications, it can take many years. There is a reason Nobel prize winners tend to be old.
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.
A lot of little things can add up to a lot but fly under the radar.
This post so perfectly captures the tragedy of how the term AI has lost its meaning to LLMs.
Because LLMs are the only thing that can be said to pass in some sense the Turing test, which is like the benchmark for AI.
I think maths is really the only solid case so far. I do wonder whether the big early contribution to experimental science will be in speeding up the scientists conducting it, at least until some non-negligible boundary in simulation fidelity is crossed.
AI has an incredibly long history of contributing to science and engineering. Some highlights:
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.
There’s been a tendency since well before LLMs for AI demos to focus on “here’s a new problem we made up that it solves that you could imagine will soon transfer to an existing real problem” as opposed to “AI solves X”. This has been an ongoing challenge, it does cool stuff and looks so promising but has been very hard to harness for useful work. It’s still going to take time before the hype dies down and we see what’s left.
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.
LLMs are amazing for engineering work. They can automate design and testing of the entire spectrum.
Saw a post recently about how AI agents were used to check if papers had flaws in their conclusions and it found 15% did, that kind of auditing is already a massive massive advancement for science
this announcement is not regarding a large language model
Wait, so we like AI now?
For me, the issue is the whole piracy and unauthorized use of intellectual property with no rewards, and the environmental damage intensifies it by 1000x. I know that this has been the case for many different situations in human history, but that doesn't make it any better.
AI != GPT
I think there is a large group of commenters who actually just hate silicon valley tech companies, and they're fine with AI as long as it comes from someone else.
Poetic considering the locale. But ya, modern tech is pretty gross, and I yearn for the open models.
1. Yes. 2. Maybe
Some people do, some people don't. Some it depends on what for and how it's used as well, as well as what kind.
Yes, for their sense of nuance.
AI has been beneficial in multiple research fields across maths and science. That is to say, it's actually benefited humanity in those particular use cases.
As opposed to enshitifying the lives of the working class in most other areas.
Maturity is understanding there's an enormous range between "like" and "dislike", between "love" and "hate", between "good" and "bad", and its also possible to be neither at once or both at once.
Using AI as a tool to investigate the mysteries of the universe is a huge win for humanity. Generating slop images to confuse people online is not. AI is tool, we like when tools are used for good.