ImageNet's 2012 breakthrough by Alex Krizhevsky's deep learning model triggered a seismic shift. His 9.8% victory over rivals sparked a gold rush: Google, Facebook, and Baidu hired pioneers like Geoffrey Hinton and Yann LeCun, while venture capital flooded neural network research. Within five years, deep learning eclipsed older methods not because they failed, but because financial and intellectual momentum abandoned decision theory's rigorous frameworks.

Disciplinary silos trapped decision theory in fragmented academic realms. Bayesian statistics languished in statistics departments resistant to its principles, while operations research hid in business schools. Reinforcement learning abandoned its decision-theoretic roots for deep learning collaborations. At NeurIPS 2024 (28,000 submissions), these methods remain invisible compared to OR conferences with mere thousands of attendees. The mathematics exists but lacks unified advocates.

Specification convenience made deep learning irresistible. Decision theory demands explicit utility functions, cost models, and probability distributions - a Herculean task requiring domain expertise. Deep learning sidesteps this: feed data, define loss functions like cross-entropy, and let models self-organize. This mirrors the 20th-century Bayesian-frequentist shift, where computational feasibility trumped theoretical purity. The market prioritizes deployable solutions over optimal ones.

Commercial reality explains the persistence gap. Systems optimizing for quarterly metrics often sacrifice long-term adaptability. Autonomous agents querying every tool without cost awareness exemplify this short-sightedness. Yet probabilistic methods resurge: Bayesian neural networks and probabilistic programming languages now blend pattern recognition with uncertainty modeling. The cyclical pendulum swings as limitations of pure pattern-matching become evident, though decision theory's revival hinges on solving its specification quandary.