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.
This sounds quite frustrating. I'm not in particle physics, I'm just a lowly SWE, but my workplace doesn't have a problem with me burning $30k of tokens on some one off idea.
Why did CERN require so much selectivity? Was it a lack of funding? Are particle collisions particularly expensive? Are you competing for "collider time"? I am curious what the limiting factor was.
I think I would lose my mind if I needed hundreds of collaborators to support an experiment.
I don't think the comment you react to is talking about physical limitations (budget, computing power, ...), but scientific methodology policy.
It was perfectly fine to work on any crazy ideas you wanted to. But if you wanted to publish as part of the CERN collaboration (and if you use CERN non-public data and CERN infrastructure, I'm not sure it was possible to publish as an "independent" anyway), there is a lot of protection to ensure that your article is at a good scientific standard.
As explained in the previous comment, scientists were nervous about "black box spurious pattern", "experimental artifacts" or "elsewhere effect". This was a good thing. The limitations were not there because of budget or access to infrastructure, but because people were conscientious of doing good science.
Remember the NN examples where the NN is given pictures of horses and pictures without horses and had to classify them, and it turned out one criteria was to detect the signature, in the bottom right, of the photographer that specializes in horse pictures? It meant that if you were giving a new picture of a bird but with that signature, the NN would say "it is a horse". Now, how do you make sure that something similar does not occur when treating billions of collision events? The categorized data, especially for the signal, is coming from simulation. What if your NN detects that in the signal, the pixel number 231 is not "hot" like the one in real data (because it turns out that, as sensor can be sometimes, this pixel is hotter in the real detector than it is in the simulation), and uses that to select the signal? Or that the NN overfits part of the data (the majority is classified on sensible cuts, but for more extreme cases, it classifies as signal events where there is an electron with pT between 24.5678 and 24.5685, but as background if it is before or after, because it happens to be what the training data looks like)?
Because of that, there were a lot of cross-checks and precautions being taken. One common approach was to compare more black-box-y methods with other methods that are more easily explainable. That was certainly a good thing. And this was not being seen as a chore: no one was approaching it as if people being good scientists where being personally against your new method. Proving that your new method was checking all the boxes of the scientific qualities was not seen as a bad thing, it was seen as the fun part of building new robust scientific advances.
But aside from the accuracy, I’m finding requests take much longer to action. 15 seconds to add a reminder yesterday. Not sure if that’s just me or normal now?
Remember when you asked 2012 Maps for directions to the supermarket, and it’d tell you to drive off a bridge? Looks like they finally integrated it into 2026 Siri!
One of my friends is an engineer named Claude. I randomly send him my prompts as a joke. For a while he’d just feed my messages straight to Claude, until he found it easier to respond with … curse words :P
In South Africa Toyotas (Hilux in particular) get stolen all the time because they are so reliable (and you get parts throughout Africa). So then you end up making new sales when people replace their stolen cars. One way to get around the problem for them I guess? /s :)
Yep. I'm feeling the same way... even though I'm hardly able to use AI at work, the push is still there for us to use it, making it even more frustrating. Anyway, I hate it, and I'm just sorta phoning it in now and trying to find something else to do. I'm cynical too and feel there is no point to upskilling and using AI for everything, 'cos it's not going to help you keep your job for very much longer than the people that don't use it.
It's not really what you're looking for, but I made this lichess BOT that makes stupid (but relevant) comments on your moves from time to time: https://lichess.org/@/Annie_Archy/all
It's been tried, they feed the evaluation into the LLM and get it to commentate.
The results are absolute garbage (as you would expect) since the LLM's really don't understand chess or how to play it, someone posted a benchmark of the current SOTA ones playing Chess and the best was estimated at 1480.
I think OP meant trash talking. "Nice knight on the rim you got there buddy." "How long do you think you'll be able to hold onto that pawn?" "The Latvian gambit, really?"
Having a weird experience with this at work. We are being told to use it, but the amount of flip flopping on whether or not we can use ChatGPT or Codex or anything else has left a lot of us without any clear direction. One day it’s allowed, the next day IT security guys have blocked it. Local models are also not allowed. I’m carrying on coding everything myself, because I can’t keep changing my workflow when IT changes their mind every second day. Despite this, the directive from the CIO is that by next year 70-80% of code must be AI generated. Last I checked only 2% of the company was actually given a license for codex. So idk man
reply