Prequasar

1.Introduction: A Social Experiment with No Right to Withdraw By 2026, the tide of autonomous driving has become unstoppable. Entrepreneurs proclaim the arrival of "large-scale trials," declaring machines to be 200 times safer than humans, and move to eliminate steering wheels and brakes. Capital markets are boiling over, and the public, amid dazzling propaganda, gradually accepts a presupposition: that technological progress inevitably leads to a safer future. But the safety issues of AI have never been truly resolved. We are forcing every ordinary person to participate in a gamble they never consented to. The stakes are their own lives and physical integrity, and the house is a statistical model with fundamental flaws. Chapter 1: The Innate Curse of Statistical AI At the core of all current mainstream autonomous driving systems are statistical models based on empirical risk minimization. These models have a fatal mathematical prerequisite: Training data and the real world must be independent and identically distributed. However, the very nature of traffic scenarios is open, infinite, and even adversarial. The Inevitability of Out-of-Distribution Samples: No matter how many billions of miles of data are collected, the real world will always contain scenarios never covered by the training set – a drunk pedestrian climbing over a guardrail, a mattress falling off a truck, road markings blurred by a sudden blizzard. When a model encounters these "out-of-distribution" samples, its output is unpredictable, and the model itself is unaware of its own "ignorance." The Curse of the Long Tail: Traffic scenarios follow a long-tail distribution. High-frequency scenarios at the head (going straight, stopping at red lights) account for 99% of the data volume, but it's the "Corner Cases" in the tail – those with extremely low probability but infinite variety – that are the true killers. Statistical models can only fit the patterns they have seen; for