Elon Musk’s Live FSD V12 Demo Shows AI‑Driven Full‑Self‑Driving in Action

Elon Musk recently conducted a 45‑minute, unedited live broadcast on X (formerly Twitter) showcasing Tesla’s highly anticipated Full Self‑Driving (FSD) V12 system. The stream attracted over 10 million viewers and featured a single manual intervention.

The demonstration used a Model S equipped with HW3 hardware; HW4 was not employed because it does not yet support FSD V12. Musk explained that the end‑to‑end architecture processes raw sensor inputs (camera video) and directly outputs steering, braking, and acceleration commands, which required re‑optimising the software for HW3.

During the test, Musk randomly set the destination to Stanford University, then to Mark Zuckerberg’s home, and later to another location, allowing the vehicle to navigate complex scenarios such as construction zones, lane changes, speed‑bump deceleration, and a round‑about without any programmed rules.

Only once, about 20 minutes into the drive, a colleague had to intervene when the car incorrectly responded to a left‑turn signal at an intersection.

Musk emphasized that no explicit code was written for traffic‑rule handling; the system learned behaviors such as slowing for speed bumps, yielding to cyclists, and navigating roundabouts purely from video training data.

"Look at the speed bump – the car slows down. We didn’t write any code for that; it learned from video."

"We never coded a rule to yield to a cyclist; the model handled it on its own."

"There’s no code dictating how to stop at a stop sign or wait for other vehicles; the model learned from examples."

"We never programmed the concept of a round‑about; we simply showed the model many round‑about videos."

The new FSD V12 no longer relies on high‑definition maps or traditional rule‑based logic. Instead, eight forward‑facing cameras feed video at up to 36 fps (potentially 50 fps) into a neural network that performs all driving decisions locally, consuming only about 100 W of power.

By replacing the previous explicit control stack, Tesla reduced code size by more than 300 k lines of C++.

Musk also highlighted the importance of high‑quality driving data for training the large model, noting that mediocre data does not improve performance.

All inference runs on the vehicle’s AI computer, so the system operates without an internet connection and can handle unfamiliar terrain.

FSD V12 is still in a testing phase, with beta testers worldwide (e.g., New Zealand, Norway, Thailand, Japan). Musk hinted that once the experience is smooth and safe, Tesla may offer a one‑month free trial to all North American owners, followed by global rollout where regulations permit.