Elon Musk takes end-to-end FSD V12 for a rush-hour spin

Elon Musk just took Tesla's unreleased FSD Beta V12 out for a 45-minute spin during rush hour in Palo Alto, livestreaming the whole thing to show off what happens when you delete hundreds of thousands of lines of code in favor of pure AI.

Cruising to Vivaldi's Four Seasons, the Model S handled heavy traffic, sharp turns, and unprotected lefts with barely an intervention or two. Once it reached its first stop, the car automatically pulled over and parked by the curb on its own. As a joke, Musk then plugged in Mark Zuckerberg's house address, and the car navigated there without fuss.

Ditching rules for pure video

Under the hood, V12 marks Tesla's transition to an end-to-end neural network. Instead of engineers hardcoding rules for roundabouts or scooters, the system learned how to drive purely by watching millions of video clips.

Musk noted that this shift allowed Tesla to scrap hundreds of thousands of lines of C++ code, making the system nimbler. Crucially, the software can operate entirely offline without an active data connection, even in unfamiliar terrain.

Hardware 3 owners also got good news. The demo ran on HW3 hardware with its eight cameras feeding 36 frames per second.

Musk explained that pure AI calculations can process inputs much faster than traditional code, with a theoretical cap near 50 FPS, though 24 FPS is plenty for real-world driving. Tesla already has V12 beta testers testing roads globally, including Norway, Japan, Thailand, and New Zealand.

The 0.5% stop sign dilemma

Joining Musk in the car was an engineer from Tesla's autopilot team, and their on-air conversation highlighted a major regulatory headache. While discussing stop signs, Musk guessed that barely 5% of human drivers come to a dead stop.

The engineer corrected him with an even wilder stat: real-world data shows only 0.5% of human drivers actually stop completely at all stop signs. Most people just creep through at low speeds.

That creates an awkward clash. If your neural network learns directly from humans, it copies human bad habits, but federal safety laws mandate a full stop. Meanwhile, the NHTSA is wrapping up its probe into Tesla's driver-assist tech, with rulings expected shortly after past recalls specifically targeted rolling stops.

Why it matters

For Tesla owners, the demo proves that older HW3 cars aren't obsolete just yet. Running a true end-to-end vision model locally without network reliance or lag shows how much performance Tesla can still squeeze out of existing silicon.

For the autonomous vehicle race, Tesla is betting the farm on throwing away modular pipelines. If end-to-end learning works at scale, every competitor relying on heavily hand-crafted rule engines and HD maps will look hopelessly outdated.

However, we think this approach introduces a glaring blind spot. When you replace transparent logic with an opaque neural network, edge cases become harder to police. A car that mimics human driving will naturally mimic human rule-breaking, putting Tesla on a collision course with safety regulators like the NHTSA, who want predictable guardrails rather than a black box that drives like an impatient local.