Just finished a SLAC Public Lecture [1]. Dr. Arkani-Hamed mentioned that the LHC produces about 1 billion collisions every second. Of those, about 10 per second are top quarks. Certain theoretical particles are expected to be detected at a rate of one per minute or hour or even day.
Blindly data mining amidst those kinds of frequencies doesn't make sense. There is a reason the scientific method starts with hypotheses.
Ex-physicst here. JumpCrisscross is absolutely right. It's a hard sell and it should be. Saying "I want a better machine but I don't know what I'm looking for" is only worse than "I want to put objects together but I don't know what machine I'm building."
The space of what's concievably possible is too large to explore if you're not guided by an hypothesis, even if you didn't have money constraints. It's not a human problem, reality is just complex.
Regarding what I'm doing now, let me know if you want to talk in private. Nothing too exciting though.
The main reasons were: I am not that good of a physicist, feudalism in the academia (socially speaking), I fell in love with IT, extraordinary opportunity in industry. All the stars were aligned for a week and I took the opportunity.
I love physics (and science in general) but what I do now in industry is fantastic (in broadly speaking IT). The opportunities for advancement other than age based are better too (not that this is a rule in academia, but it was where I was).
And money is such that my family is secure no matter what happens to me.
Suppose I seek funding from the NSF for the following experiment: I'm going to smash peas into a wall (or collide peas traveling in opposite directions), and note down the splatter patterns. Using Advanced Machine Learning (TM) I shall extract correlations in that data, and make new discoveries!
It is only theory that can decide whether this experiment is worth pursuing or not, and unavoidably so. That's why I would be laughed off for wanting to perform the peas experiment, but we fund billions of dollars of experiments which do the same thing but with protons instead of peas!
One needs to have a theoretical framework which guides the experiments we perform and the way in which we interpret data. There is always this subtle symbiotic interplay between theory and experiment, where each one guides the next step of the other.
Wanting to get rid of theory and just do experiments is pointless and stupid. If anyone believes otherwise, I have a bridge to sell, cuz you know... empiricism is all about experiments... you can do experiments on the bridge to generate data... and data is awesome and will lead to discoveries! :-)
The LHC only got funding because it was guaranteed to answer the question 'is there a Higgs Boson'.
It's good to have a clear hypothesis for an experiment (especially an expensive one), and the predictions for the mass of the Higgs boson was a good starting hypothesis for the LHC.
But there was also the expectation that, colliding at higher energies than ever before, there would be the opportunity for unexpected discoveries.
The proof that the brief for the LHC was more than just "find the Higgs" is that the LHC found the Higgs, yet is still operating--and is just about to get a time-consuming and expensive upgrade in luminosity.
Also, theory first is barely working nowadays too. The entire field has a problem, and that problem looks much more like "all the easy problems are solved already" than "we must think things through before testing them".
Anyway, I do agree that experiment-first is much harder to sell. People dig for a good looking theory, even if it's worthless.
It's easier to get money to build a new collider if there's something specific to look for. "We're expecting to find the Higgs, but we need a better tool to look for it."
Trying to do the same with a data-driven model is a hard sell. "We haven't found anything, please give us a better machine so we might find something, but we don't know what"