Paragon AI MAG — Series 4, Episode 4: "How the AI Tools You Use Actually Work"
Verdict: WATCHING
In Austin, Texas, passengers can now summon and ride in a commercial taxi with no driver at all — and the newest prototypes have no steering wheel, no pedals, and no mirrors. Yet in San Francisco, an autonomous vehicle struck a pedestrian who had been thrown into its path by another car, misunderstood what had happened, and dragged her 20 feet toward the curb.
Both of these stories are true. That contradiction is the whole story of self-driving cars in 2026: a technology that is already working, already safer than humans where it runs — and still far from finished.
The four-step loop
A self-driving car replaces the biological chain of human driving with one automated cycle, running tens of times every second: sense, guess, decide, move.
The sensors watch the road. The software guesses what everyone around the car will do next. Then the car steers and hits the brakes. The whole loop has to finish in tens of milliseconds — because at 60 mph, a car eats up about 88 feet (27 metres) of road every single second.
The computer's one massive advantage: it never blinks, never gets tired, never drinks, and never texts while driving.
The math of an emergency
Here is the number that matters most. A surprised human takes about a second and a half just to hit the brakes — half that time noticing the danger, the rest moving a foot to the pedal. At highway speeds, a human covers 132 feet (about 40 metres) of road before braking even starts.
An automated system reacts in 0.1 to 0.5 seconds, covering only about nine feet in the best case. The physical braking distance is exactly the same for both cars — once the brake pads clamp, physics takes over identically. The entire advantage of the robot is that it notices the danger instantly.
Four sensors, and why one alone always fails
Waymo's cars stack four different kinds of sensors, because each one is blind in a different way:
- Lidar fires up to 2.8 million laser pulses a second and maps exact 3D distances down to the centimetre, day or night. But it cannot read a traffic light, and heavy rain scatters its lasers, cutting its range by 50–80%.
- Cameras are the only sensors that read meaning — red lights, speed-limit signs, lane lines. But they fail in fog and harsh sun, and independent tests at Michigan State University found camera-based systems can miss up to 20% of objects in light rain and up to 40% in heavy rain.
- Radar uses radio waves to measure speeds straight through fog and dust, down to about 0.1 metres per second. But its picture is coarse, and it cannot read a traffic light.
- Ultrasonic sensors handle short-range proximity — parking and immediate obstacles.
Stack them together and the weaknesses cancel out: if one sensor gets blinded, the others take over. Waymo's fifth-generation car carried 29 cameras, 5 lidar units and 6 radar units; the sixth generation cut the sensor count by 42% — 13 cameras, 4 lidar, 6 radar — and added microphones to hear emergency sirens.
The cameras-versus-lidar war
This is the biggest fight in the industry. Waymo spends tens of thousands of dollars on hardware redundancy, mounting lidar, radar and cameras all over a $70,000 Jaguar, so no single blind spot ever wins.
Tesla went the opposite way: it removed radar in 2021, ditched ultrasonic sensors in 2022, and bet everything on cheap cameras and neural networks. The argument is simple — humans drive with two eyes, so cameras should be plenty. Bad weather is the reality check for that gamble.
Tesla's system, Full Self-Driving, costs $99 a month as a subscription and is still officially SAE Level 2 — the driver must supervise constantly. Tesla's robotaxi pilot launched in Austin in June 2025 with about ten Model Ys, a $4.20 flat fare, and a safety monitor in the passenger seat.
HD maps: the blueprint in the robot's memory
A self-driving car does not navigate on phone GPS. Your phone is only accurate to about two to five metres — half the time it does not even know which side of the street you are on.
Before a robotaxi ever drives in a new city, Waymo pre-maps every curb, stop line and the exact 3D height of every traffic light, to 10 centimetres — roughly the width of a coffee mug. The car matches its live sensor feeds against this high-definition memory of the street. If a prankster stands by the road holding a stop sign, the car checks its map, sees there is no intersection there, and knows it does not need to stop.
Prediction: guessing three seconds ahead
The smartest step of the loop is the one most people miss: prediction. The software does not just see a pedestrian walking. It calculates where that person will step three seconds into the future. If a cyclist wobbles, or someone angles their shoulders toward the curb, the car plans an evasive move before they even leave the sidewalk. It is a relentless, high-speed guessing game on a rolling three-second horizon.
What the safety data actually shows
Road traffic killed 39,254 people in the United States in 2024 — about 108 deaths every day. A 2015 NHTSA study attributed the critical reason for about 94% of crashes to human error.
Waymo's own data, across 271.3 million rider-only miles through June 2026, shows a 95% reduction in serious-injury-or-worse crashes, an 82% reduction in any-injury crashes, and a 93% reduction in pedestrian-injury crashes, compared with human drivers.
An independent study by the Insurance Institute for Highway Safety (July 2026, about 50 million driverless miles) found a 68% lower rate of police-reportable crashes than human drivers in the same cities (1.28 versus 4.06 per million miles), an 81% reduction in injury crashes, and an 85% reduction in single-vehicle collisions.
One honest caveat: independent validation currently applies only to Waymo, because it is the only commercial operator publishing enough mileage data to compare.
When systems fail: the Cruise lesson
On October 2, 2023, in San Francisco, a pedestrian hit by a human-driven car was thrown into the path of a Cruise robotaxi. After stopping, the car tried to pull over to the curb and dragged her about 20 feet at roughly 7 mph. California suspended Cruise's operating permits on October 24, 2023, the company paid a $500,000 penalty to the US Department of Justice, and GM shut Cruise down in December 2024 after investing more than $10 billion.
To handle the cases the software cannot, fleets run remote assistance centres. Waymo has about 70 remote agents supervising roughly 3,000 cars — about one agent per 40 cars. The agents do not drive by joystick; they help the car decide when it meets something ambiguous, like a road closure it has never seen.
The practical payoff: levels, buying, and sharing the road
SAE levels 0–5, in plain words:
- Level 0–2: you are legally the driver and must supervise at all times. Tesla's Full Self-Driving, GM Super Cruise and Ford BlueCruise are all Level 2.
- Level 3: the car drives itself under strict conditions — Mercedes Drive Pilot, under 40 mph in daylight on approved highways — but you must be ready to take over.
- Level 4: no human is needed at all; the car is fully responsible. You cannot buy a Level 4 car today. You can only ride in one, like a Waymo.
- Level 5: a car that drives anywhere, in any weather, with no human needed. No Level 5 system exists anywhere in the world.
Before you buy a self-driving upgrade, ask one question: who is legally driving? It is not about how smart the car feels. It is 100% about who holds the bag when things go wrong.
Tesla's new Cybercab — no steering wheel, no pedals, no mirrors, not even a rear window — has begun public rides in Austin. It costs about $29,990, but it will not move an inch unless you pay a mandatory $299 monthly subscription, and it only works inside heavily mapped zones in Texas and Florida.
Around a robot taxi, behave as if the driver cannot see you — because it cannot, in the human sense. Stick to crosswalks. Never jump in front of one to test its brakes: if its sensors get confused, its safest programmed reflex is to stop dead in the middle of traffic. If you are riding inside a Waymo and it suddenly stops, stay in the car, keep your seatbelt on, do not touch the wheel or the pedals, and use the screen to call support — a real person will talk to you through the car speakers.
Verdict: WATCHING
Trust autonomous vehicles where the data proves them safe — like Waymo's 95% drop in serious crashes — but demand accountability where the hardware falls short, especially in nasty weather. The driverless taxi is real. The car that drives anywhere, in any weather, while you nap, does not exist yet.
Paragon AI MAG — one video and one podcast episode every day. Every fact in this article was checked against sources several times before release. Watch the video version on the Paragon AI MAG YouTube channel and listen to the full podcast episode on RedCircle.
Watch the video: https://youtu.be/Goi9kQknAI4
Listen to the podcast: https://redcircle.com/shows/cbce814a-d451-4913-a015-fe05b4d0e059/episodes/4046051b-b6c3-48f2-9d53-cd726f04b22f
For the engineers: this breakdown dissects the autonomous perception stack — lidar firing up to 2.8 million pulses a second, cameras reading signs, radar measuring speed through fog, HD-map localisation to 10 cm, and the prediction engine running on a rolling three-second horizon. Plus the architecture feud: Tesla's end-to-end vision-only FSD v12 versus Waymo's multi-sensor fusion stack. Follow Paragon AI MAG on Dev.to for daily source-checked explainers. Full video: https://youtu.be/Goi9kQknAI4
Top comments (0)