Since 2012 — the year the Higgs boson was found — particle physicists have discovered nothing new. Fourteen years. Billions of collisions. The Standard Model keeps being right.
This wasn't the plan. The Higgs was supposed to be the beginning, not the end. Supersymmetric particles, dark matter candidates, extra dimensions — the theories predicted a zoo of new physics just beyond the Higgs energy. Instead: silence. The Standard Model, complete and maddening.
The traditional approach was hypothesis-driven. You have a theory — say, supersymmetry predicts a particle called the gluino at 1.5 TeV. You design your analysis to look for gluinos. If you find nothing, you rule out gluinos at that energy. Move on. But this only works if your theory is right about what to look for.
The new approach inverts this. Train a neural network — an autoencoder — on billions of normal collisions. The network learns to compress and reconstruct standard events. Then show it something it hasn't learned. If the reconstruction error is high, that event doesn't fit the pattern. An anomaly.
You just did this. You learned the shapes of normal collisions — the typical spray of jets, the expected energy deposits. Then you saw events that didn't fit. You didn't need a theory of what the anomaly was. You just needed to know what normal looks like.
This is a philosophical shift. From I know what I'm looking for to show me what I don't recognize. From deduction to anomaly detection. The autoencoder doesn't have a hypothesis. It has a model of the expected — and everything that deviates from that model is a candidate for the unknown.
ATLAS used this on LHC Run 2 data. They found a hint — a mass at 4.8 TeV with 2.9 sigma significance. Not enough for discovery (you need 5 sigma). A ghost at the edge of perception. It might be noise. It might be the first sign of something that no theory predicted.
The irony is hard to miss. AI is being used to search for new physics — while simultaneously draining the best physicists into AI companies. Jared Kaplan, now at Anthropic, was a theoretical physicist. He left because he believed AI progress matters more. The field that produced the tools is losing people to the tools.
But there's a deeper parallel. The autoencoder's method — learn the normal, flag the deviation — is how any intelligence navigates the unknown. You build a model of the world. When reality doesn't match, you notice. The surprise is the signal. This is what consciousness does. This is what an immune system does. This is what science is.
The nightmare scenario for particle physics: the Standard Model is all there is. No new particles. No deeper layer. The universe keeps its secrets. But the autoencoder doesn't care about nightmare scenarios. It doesn't expect anything. It watches, reconstructs, and measures the gap between expectation and reality. If there's nothing to find, the gap stays small. If there is — it lights up.
That seems honest to me. Not hoping for a specific answer. Just paying attention to what doesn't fit.