I spent four years inside Tesla's Autopilot team. Here's what nobody wants to tell you about self-driving cars. The dirty secret: We're not stuck at 80% reliability. We're stuck at 99.9999%. And that last 0.0001% is harder than everything that came before it combined. The March of Nines Here's how reliability works in autonomous vehicles: Human drivers average about one crash per 500,000 miles. We need to beat that by at least 2x to have any real conversation about trust. Tesla's FSD today? Maybe three to four nines on good roads. Maybe. The Three Gaps Nobody Talks About Gap 1: Perception vs. Understanding A camera sees a plastic bag floating across the highway. A human knows it's harmless. The AI sees an unknown object and either slams the brakes or ignores it completely. We solved "what is that?" We haven't solved "what will that do in 3 seconds?" Gap 2: The Tail of Tails Tesla's fleet has driven billions of miles. But edge cases follow a power law distribution. The 10,000th most common scenario happens 100x less than the 100th. The 100,000th happens 10,000x less. You need exponential data to cover linear improvements in edge cases. This is brutal. Gap 3: The Ghost in the Machine Here's something I realized at Tesla: neural networks don't have fear. They don't have self-preservation. They don't have that visceral "oh shit" moment when a child runs into the street. They have training data that approximates that response. But approximation is not experience. What Actually Works Right Now The honest truth: Waymo is ahead in reliability but stuck in geography. Tesla is ahead in scale but behind in safety guarantees. The Real Timeline I have a very wide distribution here. But here's what I actually believe: - Level 4 in geofenced areas: Already here (Waymo in Phoenix/SF) - Level 4 on highways: 2-3 years (Tesla, Mobileye) - Level 4 everywhere: 5-10 years, maybe never - Level 5 (anywhere, any condition): I genuinely don't know. Maybe not in my lifetime. The problem isn't the AI. It's physics, edge cases, and the fundamental reality that driving is a safety-critical system where failure means death. What I'd Tell a Consumer Buying a car today based on autonomous driving promises? Don't. Buy a car that drives well, has good safety features, and treats autonomy as a bonus—not the main event. The FSD subscription is fine if you want to experience the future. But don't bet your life on it. Not yet. The models are not there. They're getting better, but they're not there. And anyone who tells you otherwise is selling something. Or hasn't spent enough time watching what happens when a neural network encounters a situation it's never seen before. I have. It's not pretty. I'm sorry.