I stand by that assessment. Stoned teenagers generally drive pretty slowly and boringly, until they see a cop.
FSD is assertive, has good reaction time, and doesn’t creep along at half the speed limit. I said boring, not slow. Not remotely the same thing.
One of the Robotaxi riders caught this awful human driver:
Sort of the same thing as the aforementioned Robotaxi mistake, but way worse since there was oncoming traffic that he nearly collided with. The Robotaxi gently slows to allow the black SUV to merge back in, just as an alert human would. Boring.
And I didn’t say at half the speed limit. Seems like we’re describing the same thing.
You and I have different notions of how a stoned teenager drives. Regardless, if said stoned teenager drives like a normal person and largely avoids doing anything too out of the ordinary, while also making allowances for poor human driving, then I guess I’m in favor of FSD driving like a stoned teenager. We should all drive like stoned teenagers.
Ehhhh, maybe. They’re pretty safe drivers, to be honest. But they’re fucking irritating and do things that are sort of inexplicable.
I have what I think is an interesting question set about ordinary Teslas owned by ordinary people. I’m trying to avoid using acronyms or buzz terms here because I’m not certain we all share the same definition of terms like e.g. “FSD”. That will make what follows a bit wordy, but ideally with reduced ambiguity. To wit:
What percentage of city miles, and separately what percentage of highway miles, driven by private Teslas …
- are driven in the maximum self driving mode, where the car knows the route, makes its own turn & lane decisions, operates the turn signals, reacts to traffic lights, speed limit changes, etc.?
- are driven in a mode equivalent to that many non-Teslas have: active cruise control and dumb lane-keeping, but no awareness of route, traffic lights, etc.
- are driven totally manually. (Setting aside driver assistance features like emergency rear-end collision avoidance or lane departure warning that many cars of all brands have active at substantially all times.)
My goal with these 3 buckets is to cover all the possibilities, such that the sum of the answer to all 3 city questions would be 100% and likewise the three highway totals would sum to 100%. If I’ve left out a material intermediate case, I’ll ask a Tesla maven to fill it in.
My point, once this data is available:
Where I live, Teslas are at least as common as Nissans, probably more so. Every stop at every traffic light will have a Tesla or two in the pack. They are an ordinary part of traffic, not exotic sightings.
Punchline: How many of all those Teslas around me every day are being driven by their computer, not by their human? If the answer is 5% that’s one thing, and my in-traffic experience with Tesla’s doesn’t say much about their AI skills. If the answer is 75%, that’s something else, leading me to the conclusions below.
Because you know what? I have never seen a Tesla behave out of the ordinary in traffic. What’s noteworthy is how ordinary they drive, not how oddly they do. I’m suggesting here that the fact they drive so innocuously is proof they’ve solved most of the problems.
To be sure, what I’m really observing is whatever AI-driving is going on, as overrridden if necessary by the human driver. But at the same time what I don’t see is them zooming or careening then being saved by heroic emergency human interventions. What I see is they drive like any other car.
Among others …
Yeah. When first introduced, antilock brakes were supposed to prevent a lot of accidents. By and large what happened was the general public began driving crazier and the accident rate in scenarios ABS was relevant went down just a smidgen. The Google terms are “risk homeostasis” and Risk compensation - Wikipedia.
My rather new non-Tesla has a decent active cruise control / lane-keeping system. When I first got the car ~9 months ago I found it maddeningly incapable. As I’ve developed more experience with it, I now use it most of the time I’m not actively maneuvering. IOW, use it on surface streets, except when turning. Disable it to enter a highway but once on the highway settled into traffic, reactivate it until it’s time to enter the off-ramp. etc.
I don’t drive wasted. I do sometimes drive sorta buzzed. The thought has definitely occurred to me that “I’m glad the car is doing the small stuff so I can focus my reduced capability on the bigger picture.” I can’t be alone in that thought.
Eventually we’ll all be re-enacting the old trope of the dead-drunk cowboy folded sideways face-down over the saddle of his horse while the horse dutifully takes him home.
It’s a good question but I don’t have much of an answer. About 20% of Teslas get FSD, so that’s an upper bound for #1. You would then have to multiply by the usage rate. For me, that’s about 90-95%. But I don’t know the average. It might not be that high, especially given that it was mostly trained on CA roads and weather conditions.
All Teslas include basic Autopilot. Not useful on surface streets, but fine for highway use. Pre-FSD, I used Autopilot for again 90+% of my highway miles, basically constantly for any stretch of the same road and only disabling it for on/off-ramps. I don’t know how common that is.
I guess if I had to ballpark, I’d suggest that 10% of the surface street miles you see are fully autonomous, and maybe 70% of highway miles are at least partially autonomous.
FSD does make mistakes on surface streets. But when it does, it’s mostly stuff that others wouldn’t notice, either because I correct it immediately (like if it pauses at a green light), or it’s just something that others wouldn’t pay attention to (like the exact stopping point at a stoplight). Occasionally it does something a bit “wobbly” but I suspect you’d never notice outside the car.
Thank you. That 20% number is surprisingly low to me. Then again, I understand it’s a rather expensive option, especially as compared to its capabilities in the early days.
For sure that low upper bound says my larger tentative conclusion doesn’t really hold. There are too few Teslas driving around me in full AI mode for me to conclude much of anything from my observations of their behavior.
As to this:
Can I/we get a 2- or 3 sentence description of what Autopilot’s capabilities are and aren’t, using non Tesla specific terminology?
Autopilot is basically smart cruise control (maintains speed but won’t crash into the car ahead) and lane keeping. That’s on top of safety features like emergency braking. FSD is the whole kit and kaboodle, just punch in a destination and it will drive there. You have to look at the road but don’t have to touch the wheel.
Not sure if you have to touch the wheel in Autopilot mode. Probably you do. Technically it’s not touch it detects but a small amount of applied torque.
Autopilot lane keeping is more solid than Kia, at least. It locks on and will handle even sharp bends. Kia felt sloppy and would fail for even modestly sharp bends. Technically you can use Autopilot on non highway roads, but it’s pointless since it won’t deal with stoplights.
Thank you.
Autopilot is essentially what I have, but the lane keeping is probably more at the Kia end of the spectrum. Given that I live in Greater Miami’s rectilinear metroblob, curves are not quite as mythical as hills around here, but close. 
As to surface streets, I find it does great at traffic lights. As long as you’re not the first car in your lane approaching a red light. It’ll do a solid job of slowing to stop & wait behind the car ahead, then will go when the car ahead goes. Just don’t forget that the case of first in your lane at a red light is your job to guard, not its. Easy to do in a car like mine where it’s never traffic light aware so I’m always traffic light aware. That would be an easier mistake to make in a car with true AI driving.
Here is my guess, from a different angle, and a limited sample: Almost none of them are driving completely with the computer.
This is my reasoning. Frequently when I am next to a Tesla, I’ll try to get a look at its screen and see if it has the tell tale signs of using some form of self driving. I’ve never seen another Tesla with the screen in an undeniable self driving configuration[1].
Here are the big caveats: When I’m in my Tesla, which is most of my road time, I’m at the same height or lower than other Teslas, so my ability to see inside them is pretty limited. It’s easer from my motorcycle or truck. I also only look when it’s safe to do so, such as when stopped at a traffic light.
So with my caveats as a great big asterisk, in the time since fall 2018 I’ve been paying attention to Tesla self driving, I can say I’ve never seen another Tesla that I know was using self driving.
When in self driving the portion of the screen that shows the car (not the map) will have a blue line in front of the car. This has recently gotten easier to see, because now when going into self driving the car view portion expands to fullscreen, and the map is reduced to an overlay in the corner. It is possible to have the car view part be full screen when manually driving, so it isn’t a guarantee of self driving. The blue path line is a guarantee. ↩︎
Both of those anecstats suggests that Tesla HQ is not racking up the huge miles of in-service experience they might have wished for. Yes, I expect manually driven cars are still sending experential data, but that’s a different use case.
I can add that I know a lot of Tesla owners and only two besides me uses FSD regularly. Most are afraid to even try it when they are gifted it.
Well, that does make me think of another way to calculate it. Tesla claims 3.8 billion miles driven on FSD. They’ve sold 7.6M vehicles total, though not all of those are capable of FSD, plus many are in countries where it’s not available. Let’s say the actual working population is 3M cars. That means 1300 miles per car over their average life, which is relatively short given Tesla’s historical growth rate (not much in the past year, but historically has been pretty high). So average life is probably around 2.5 years currently. That’s 520 miles per year, which is about 4% of total miles per year.
There’s a lot of uncertainty in those numbers, and doesn’t distinguish between city/highway miles, but there you go. 4% isn’t that high, but on the other hand 3.8B is still a lot of miles. So it could be simultaneously true that you see very few people using FSD, and yet they’re nevertheless collecting quite a bit of data.
Quite a bit in an absolute sense but a drop in the potential bucket. Perhaps it’s time to gift FSD to every Tesla forever?
Interesting thought. On one hand, why would they just give away such a valuable feature? But maybe once Robotaxis have ironed out some of the wrinkles with unsupervised FSD, they can give away “FSD (Supervised)” and charge $$$ for “FSD (Unsupervised)”. Both can collect data and have similar safety benefits, but you’re still getting something for the money. They could also have a “FSD (Robotaxi compatible)” version for even more money.
If most Tesla drivers who have it don’t (or only rarely) use it, and most Tesla drivers who don’t have it a) don’t want it, and b) think it’s wildly overpriced if they did want it, then I seriously question whether it’s a “valuable” product.
Does it embody a vast amount of tech magic and hard dollar investment? You bet. Does the public want it enough to pay anything for it? Revealed preference is evidently “Meh. Not really.”
Speaking just for me I’m on your team. I’d pay something hefty for it if my favored car brands had it. But the rank and file populace does not see the value proposition. Yet.
Well, even if people aren’t using it, they still seem to be selling it. It’s an $8k feature and if 20% of people get it, that’s an average of $1600 per car extra margin. Not complete peanuts. Tesla has not counted all of those dollars to their earnings–they have some schedule based on feature completeness that I don’t have any visibility into–but as they’ve improved things they’ve counted some fraction of those dollars. And it’s cash in hand, regardless.
Tesla does occasionally give out free months of FSD usage. I presume they get a few takers but probably a bunch are fully unaware.
I would pay $8k to add FSD to my non-Tesla after about 3 seconds’ thought. That’s obviously practically impossible, but the concept is compelling.
Evidently the rest of the public disagrees.