Tesla's Full Self-Driving beta is out and the

There are videos on YouTube right now of Teslas driving themselves through city streets. Not highways. Not controlled environments. Actual city streets with stop signs, pedestrians, cyclists, unprotected left turns, school zones, and roundabouts.

Tesla released its Full Self-Driving beta to a small group of owners last week, and the footage is the most fascinating thing I’ve watched since the Falcon Heavy landing.

What it looks like

I’ve spent hours watching these videos. Here’s what happens.

The driver enters a destination. The car starts navigating. It reads stop signs. It waits at red lights. It turns corners. It handles intersections. It changes lanes. It makes unprotected left turns, which means turning across oncoming traffic with no green arrow, just the car deciding it has enough gap.

Most of the time, it works. The car moves through neighborhoods with a confidence that’s eerie. You watch it approach a four-way stop, yield to the car that arrived first, and then proceed, and your brain does the same glitch it does when a rocket lands. This shouldn’t be happening. But it is.

And then it tries to drive into a bush.

Or it hesitates at a green light for no apparent reason. Or it starts to make a turn and then changes its mind halfway through. Or it approaches a construction zone and the human has to take over because the neural network has never seen orange cones in that specific configuration.

The neural network is showing

What’s different about Tesla’s approach, compared to Waymo or Cruise, is the underlying technology. Tesla uses cameras and neural networks. No LIDAR. No pre-mapped routes. The car is supposed to see the world the way a human does (through visual input) and figure out what to do.

This means you can watch the visualization on the car’s screen. It shows what the neural network sees: cars, pedestrians, lane lines, signs, all rendered as colored objects. It’s like watching a brain’s interpretation of reality in real time.

And sometimes the brain gets it wrong. A shadow on the road becomes a phantom obstacle. A truck merges and the network briefly categorizes it as two separate vehicles. A traffic cone gets classified as… something that isn’t a traffic cone.

This is what learning looks like. Not in a textbook. Not in a paper. In the real world, at 30 mph, with a family in the back seat.

What bothers me

I should be clear: I find this exciting. The engineering is impressive. The progress from Autopilot (highway only, lane keeping, cruise control) to FSD beta (city streets, turns, intersections) is enormous.

But we’re watching a neural network learn to drive on public roads. With real pedestrians. Real cyclists. Real school buses.

The NHTSA hasn’t approved this. It’s a beta. The drivers are supposed to be paying attention, hands on the wheel, ready to take over. And from the videos, they are. But it’s still a beta test of autonomous driving happening on streets shared with people who didn’t sign up for it.

I trust the technology more than most people, I think. I’ve ridden in Waymos. I’ve sat in a Tesla on Autopilot. I believe this is the future. But I also think the gap between “works most of the time” and “works all of the time” is where all the danger lives. And we’re deep in that gap right now.

Where this goes

Tesla has millions of cars on the road with camera hardware. Every mile driven by FSD beta feeds training data back to the neural network. The system gets better every week. Literally every week. Each software update improves the driving.

This is what machine learning looks like at scale. Not a lab experiment. A fleet of cars, driven by real people, generating real data, training a real neural network to do something that most humans can barely do well.

I think five years from now, these videos will look like early Wright Brothers footage. Wobbly, imperfect, clearly primitive. But undeniably flight.

I’m going to keep watching the videos. They’re the best window we have into how machines learn to do hard things. The bush incidents included.

Related thinking: