First regular user receiving FSD 12.1?
The release notes for 12.1 were surprisingly simple, stating that v12 has single-stack end-to-end neural nets trained with millions of video clips for the driving controls. This replaces the previous 300k lines of C++ imperative programming. Essentially, we now have to "trust the nets". So, how do I feel about FSD v12 after driving 500 miles?
Here is a quick rundown:
Positive Surprises
The car drives more like a human. My wife couldn't tell whether it was me driving the vehicle or the car itself.
Highway situations:
FSD v11 (single-stack highway and locals) already handled highway driving quite well, but you could still sense the mechanistic nature of the C++ code in the control decisions. FSD v12 feels so natural.
Here's one scenario that really surprised me: You're driving in the fast lane (left) of a two-lane highway because slower cars stay in the right lane. Then a faster car approaches from behind. FSD v12 signals, safely switches to the slower lane, lets the faster car pass, then switches back into the fast lane and stays there.
Speed control is much smoother and appears to adjust itself smoothly with the surrounding traffic flow.
FSD v12 is more patient and assertive during lane-changing maneuvers. There's no more "middle-of-the-change hesitation" (changing mind in the middle of a lane change).
City steets driving:
One of the "hardest" problems that FSD v11 and earlier versions failed to solve in my nearly three years of testing FSD beta is a surprisingly simple setup – what I call "neighborhood laneless road snaking". It's very common in neighborhoods, where there are single-lane roads wide enough to accommodate roadside parking, or simply single lanes that gradually diverge into more lanes, or vice versa. All previous FSD versions struggled and tended to snake left and right within what the car perceived as a "wide" lane. Because of this single defect, I could never convince my wife to trust FSD driving. Well, that's finally gone in v12 with the end-to-end neural networks for driving controls – it simply learns how a good human driver would handle such a situation – just stays the course.
v12's handling of bumps is excellent! It reduces speed very smoothly to about 10 mph while going across bumps, making the ride super smooth.
Areas for Improvement
STOP signs: The car really doesn't have to wait a full 5 seconds (I know it's less than that, but it definitely feels that way) at every STOP sign. Every time, I have to push the accelerator to make it go a bit faster. Even if I had the patience, I'm sure the driver behind me wouldn't – they'd be thinking, "What the hell, you're driving a Tesla?!"
Perfect speed control is challenging because some speed signs are simply incorrect. You can't have a 40 MPH speed limit right in the middle of a highway, or try to accelerate to 70 MPH during a ramp onto the highway. It's definitely better in v12, but this still remains the main input I have to adjust from time to time.
Road conditions can sometimes be dangerous. There may be potholes, foreign objects that a good driver would constantly stay alert for and safely maneuver around with fine steering adjustments. I haven't tested FSD v12 enough in such situations, but I believe it will need continuous training to accommodate all these hazardous road situations and learn how to safely handle them.
As stated by Tesla, it is now mainly trained for good weather conditions (such as in California), and still needs a lot more training in areas with heavy precipitation, including rain and snow.
Conclusion
FSD v12 with single-stack neural networks for driving controls is definitely the (ONLY) right path forward. In fact, I think Tesla should have taken this approach much earlier rather than wasting time and effort tuning the C++ code for driving controls, which would have made it practically impossible to realize true FSD.
Now with FSD v12, I see a step change that fundamentally solves those "hard-to-solve" issues – just mimic humans! The rest is just more data and more training.
"try to accelerate to 70 MPH during a ramp onto the highway."
I thought that's why we get EVs.
The Twilight Zone on Pluto TV
The Twilight Zone on Pluto TV | A Thing About Machines | 30 min | Bartlett Finchley's paranoia about the machines around proves true.
pluto.tv
