No, your car does not drive itself

No, your car does not drive itself. Your Toyota may centre itself between painted lines and follow the car ahead. Useful. Occasionally pleasant. But it is controlling lateral position and following distance inside a narrow envelope. It is not driving.

Driving begins where the immediate stimulus ends. Tesla’s Full Self-Driving (Supervised) watches a semi turn, predicts where its trailer will be seconds later, selects Reverse and makes room before the conflict exists. On a narrow road it can reverse until it finds a place wide enough for two cars. It does not merely correct. It plans.

That is not autonomy; Supervised is in the name. Selected successes do not establish a safety rate. They do show planning that ordinary lane assistants do not perform.

A truck occupies the future

A lane assistant needs the present: too far left, steer right. Driving needs the future. A semi’s cab may clear the Tesla while its trailer is still bound to sweep through the Tesla’s position. The correct action can therefore look premature: reverse before anything is close.

The Cybertruck is stopped at a red light. It shifts into Reverse, backs away from a turning semi, waits for the trailer to clear, then returns to Drive. The Self-Driving indicator remains visible.

The semi needs space it has not occupied yet.

In Gary Branger’s post, space surrendered now prevents a conflict later.

Stops and reverses on a curve to clear the truck’s swept path.

In the next clip FSD selects Reverse for about three seconds but does not move. The semi clears, so FSD abandons the manoeuvre.

FSD prepares to yield, then holds when the semi clears.

Reverses uphill to clear the semi.

Stops, reverses, then holds until the trailer passes.

The relevant property is turning radius, not vehicle class.

Reverses before the semi completes its right turn.

Reverses before the supervising driver notices the semi.

Another car’s indicator prompts FSD to make room before the turn begins. A blinker is evidence about a probable future path.

In DN’s post, the left indicator prompts Reverse before the other car turns.

Sometimes forward is reverse

Stopping is the universal escape hatch of mediocre automation. It can be safe; it can also transfer the unresolved problem to everyone else. Two cars facing each other on a single-lane road do not need more stillness. One needs to find space.

On this mountain road, FSD rejects several convenient stopping points and keeps reversing until the road is wide enough to pass. The destination happens to be behind it.

Reverses past narrow sections until it reaches a pull-out.

In Lukas Mendez Duus’s post, entering was easy. Passing required retreat.

Moves laterally, then reverses for the approaching car.

There need not be a painted rule. A one-lane bridge runs on geometry and local convention: notice, infer, yield. Earlier versions reportedly failed this bridge repeatedly. Version 14 makes it look ordinary.

No lane marking contains the answer.

The car must understand itself

Other vehicles are only half the model. FSD must know its own turning envelope. At a mountain hairpin it backs up, opens the radius and takes the corner. At a tight right turn, a short reverse makes its own future path possible.

Reverse is one segment of the route.

The car makes room for its own future arc.

Plans survive contact

A semi blocks a parking-lot exit. FSD reverses, takes a side aisle, discovers a dead end, turns through a parking space and recovers its route. The first plan fails. The system makes another.

Five stages, including multiple gear changes.

A delivery truck blocks an FSD 13.2 parking exit. The Tesla reverses around it, then waits for oncoming traffic—including a Waymo—before proceeding. Tesla says the behaviour emerged implicitly from a model trained on real-world data. The underlying original post contains the same video.

In Tesla’s post, FSD clears the truck, then waits for oncoming traffic.

The expensive part is intelligence

Lidar measures depth. Radar measures range and velocity. Infrared sees what eyes cannot. Useful, certainly. None of it is thought. Give a monkey lidar, infrared and laser vision and it still cannot drive. An old deaf person with two ageing eyes and slower reflexes can. The difference is intelligence.

A modest array of inexpensive cameras supplies the surrounding scene. The car naturally measures its own wheel speed, steering and motion. But the rich external world—from a blinker to the future sweep of a trailer—arrives as video.

Cameras do not supply free space, intent or turning radius. The model must infer them, then choose an action whose value may not appear for several seconds. The cameras are cheap. Turning photons into judgment is not.

The network need not contain a neat, human-readable world model. Its behaviour shows enough internal structure to compare futures and act across several steps.

What this proves

These clips are demonstrations, not a denominator. The safety evidence comes from fleet telemetry. Tesla’s current FSD (Supervised) Vehicle Safety Report, covering nearly 13.8 billion cumulative FSD miles and a rolling twelve months of collision rates, reports 7× fewer major collisions, 7× fewer minor collisions and 5× fewer off-highway collisions than its estimated U.S. average.

Tesla counts a collision against FSD if the system was active at any point in the five seconds before impact, including a disengagement or system abort. The U.S. baseline combines federal mileage and crash data and necessarily involves assumptions; it is Tesla’s analysis, not an independent controlled study. The reported difference is nevertheless severalfold, not marginal.

The statistics measure how often FSD crashes. These clips show what it can do: predict other agents, reason about geometry, change gear, surrender short-term progress and recover from a failed route.

Safer is not infallible; the supervisor must supervise. But do not point to the lane-centering button in your Toyota and declare equivalence. That is not skepticism. It is a failure to distinguish reflex from thought.

Your car keeps a lane. This one is beginning to drive.