Tesla FSD - A Five Year Perspective

My Model 3 was delivered in mid-August 2018. While I originally just had Enhanced Autopilot (EAP), I quickly paid the extra $2000 to get FSD as rumors swirled about Tesla charging more in the future. What's my opinion after 5 years of testing and playing with Tesla EAP and FSD? TLDR: It's pretty much still just a gimmick and hardly worth the extra $2K I paid for FSD, much less the $15,000 they charge today.

So a little about my experience so you understand my evaluation. Since receiving the car in August of 2018, I have done everything possible to have the latest firmware, latest features, and test and provide feedback on the latest software and capabilities. I am an engineer by training and education, and have participated in a lot of software Beta Test programs for lots of products, and I take the whole thing fairly seriously because to me it's lots of fun. So I don't feel like I have (or had) unrealistic expectations of what FSD would be able to do, what the testing process would be like, or when the features would become polished and truly usable.

All that said, let's start with what works: Navigate on Autopilot (NoA) and Auto Lane Change. NoA is an L2 ADAS feature that combines TACC and Autosteer to drive the car on limited access highways, and includes features like automatic speed-based lane changes and highway interchange transitions. IIRC, the initial version was released in October of 2018 and, since then I have put thousand of miles on it. Now a part of FSD-beta on highways, while it has improved somewhat, it basically the same functionality as NoA on release five years ago. Phantom braking is noticeably better (happens less frequently but still happens) and the system now has some subtle safety improvements, like cheating to the outside of the lane when passing a truck. But passing behavior is still wonky and can be harrowing, and lane selection for passing and exit/transition ramps is still really bad - maybe even a bit worse than in original NoA now that it doesn't use maps as much. In addition, when it came out in 2018 I felt like NoA was "best-of-breed" for these L2 ADAS highway-driving systems, but I feel like it has now been surpassed by BlueCruise, SuperCruise, Drive Pilot, and the like, many of which offer some level of hands-off and/or eyes-off driving for a truly L3 autonomous driving experience. But NoA is still a very useful feature, IMO, offering reduced workload, fatigue, and stress on long highway trips, especially if you do your own passing with Auto Lane Change instead of the automated speed-based lane changes.

Another thing that works is Summon. Remote control your car into and out of a garage or tight parking space. Only useful in very limited circumstances but it does what it does well when you need it (as long as you still have USS, evidently).

Everything else: AutoPark, Smart Summon, Autosteer on City Streets, and Traffic and Stop Sign Control? Gimmicks. Barely usable and hardly ever useful. Setting AutoPark and Smart Summon aside (because they just plain don't work), Autosteer on City Streets is a massive undertaking and has improved immensely since "beta" testing began almost three years ago now. But to engage it, you have to be ready to be extremely alert and ever vigilant - hardly a relaxing way to drive. And while it can get you from point A to point B - even sometimes without intervention (which has never happened to me) - it is most often still a harrowing experience to use it. As an ADAS feature, it has zero utility, IMO. Fun to show off to your friends, and fun to test the latest and greatest to see what it can do. But from a practical standpoint, it can't get me anywhere I need to go faster, more efficiently, more relaxing, or, regardless of how many people quote statistics to try and prove otherwise, safer than I can get there by driving myself.

And now looking back 5 years and thinking where we've been and how far we've come, it's actually quite sad to me that this is all we got. I don't know how much more appetite for "testing" FSD beta I have, and often think I would be much better off or newer firmware just driving myself or on plain, old AutoPilot (I, personally, would still have my NoA in my EAP). But if there's anybody out there thinking whether they want to invest $15,000 in Tesla's autonomous driving systems, just take a tip from me and go into it with (very) low expectations of what you will be getting.

 

[diplomat33]

In order to accommodate a robotaxi scenario FSD will have to be solved to 100%. It has to account for everything it could encounter. As anyone in software can tell you solving a software problem to 100% is an exponential difficulty curve. If they are currently at 80% (being extremely generous there) towards level 5 autonomy that might not even be half way to a robotaxi solution. Folks we are decades away from robotaxis with Tesla's chosen solution. The race isn't over but the horses are coming down the final stretch and Tesla is dozens of lengths behind Waymo at this point.

While I agree that Tesla is still far from unsupervised L5 and Tesla is behind Waymo in terms of deploying reliable driverless, I don't think it is correct to say that robotaxis need to be solved to 100%. For one, there is no tech that works 100%. There will always be some failures, even if they are rare. So the best you can hope for is to get as close to 100% as possible, hence the famous "march of 9's". You can get to say 99.9999% and keep adding more 9s on the end, but never actually 100%. Eventually, you will have enough 9s to be "good enough" because you will be close enough to 100% to be safe enough. The question becomes how many 9s do you need to deploy safe robotaxis. And of course, human drivers are far from perfect. So AVs don't need to be 100% to be better than humans. AVs only need to be better than the average human driver, in order to offer a net positive in terms of safety.

It should also be noted that robotaxis can pull over to the side if they encounter a case that they cannot safely handle. So they do not need to solve every single case, just be able to reach a safe state and call for help when they cannot handle something. Of course, to be useful, you will want the cases where it pulls over and asks for help to be as rare as possible since you don't want to inconvenience the passengers. You also want to minimize "stalls" where the robotaxi stops in the middle of the road to ask for help which can block traffic and cause problems to others. You want those "stalls" to be as rare as possible. But no, robotaxis don't need to solve everything to 100%. They just need to solve enough 9s to be useful and safe.

 

[kvandivo]

My hot take is that in silicon valley everything is perfect. In the parts of the world known as everywhere-else, people have generally crappy experiences with FSD. If upper management (aka Elon) actually sat in a car with FSD almost anywhere else, they would see how poorly it is at actually being trustworthy for a complete drive (and not just 20 second snippets where it performs as it should).

We bought the car in 2017 (with hw2.5) for one reason: FSD, and I was in the front row of the 'believer' section. Times were early, I was naive, and I foolishly believed a software engineer's predictions on when something would be 'done' (what was I thinking?!). It's still so bad now (again, not in silicon valley) that my wife will never even consider enabling it anywhere other than 4-lane driving-across-country. It would be an extremely bizarre event for her to even think about enabling it in-town, because it isn't remotely close to replicating human driving behaviors. I'm a bit more tolerant of erratic actions from the car, in the name of "helping out the algorithm (now known as 'training')" but it is a long way from being enjoyable.

If we believe the hype, v8 was going to change all of it. Then v9 was going to change all of it. Then v10 was going to change all of it. Then v11 was going to change all of it. Will v12 change all of this? I hope so.... I hope so....

 

[BigNick]

And it never will. Even Waymo, the leader in L4 AVs, and bristling with an advanced sensor package, can't do it. The only thing it can possible do is satisfy your "relaxing" criteria. AVs and L2 ADAS systems will never be faster or more efficient than you driving manually. Safety is debatable. There are people who claim FSD Beta saved them from something they didn't see, and there are people who claim FSD Beta would have caused an accident if not for their intervention. Even the more advanced AVs like Cruise are crashing into buildings, and they have LIDAR, RADAR, USS, and cameras all over them.

IMO, what we need to drive research on this... Formula E-SR. Formula E (electric) Self-Racing cars.

Once they can get through a full season at speed without a shunt, it should be good enough to use in NYC traffic!

 

[EatsShoots]

Full stop, right there. V12 is not a marvel. It's an alpha test version of some new software that shows some interesting capabilities. That's it. Everything else is speculation. V12 is so preliminary that we shouldn't even be thinking about it. The Tesla engineers have just started fooling with it, so they don't have any serious experience with it and don't know what kind of legs this approach has. They thought they could make V10 work. They thought they could make V11 work. Now they think they can make V12 work.

I haven't seen the V12 demo.

But there is a hint there; it's not V2 or V3. It's V12 because Tesla keeps dropping new versions because Tesla is clearly spinning its wheels and has no idea how to actually achieve what it's still selling as "available soon". In other words "you can't get there from here".

I would bet good money on the fact that HW3 will never have FSD driving as has been promised, and I'd bet slightly less good--but still good--money on the fact that HW4 won't, either. That's also part of the reason I didn't care to get HW4 in my July-purchased model 3.

A Tesla with 2023 FSD right now cannot reliably drive several miles on surface streets in typical traffic without having a more than marginal chance of doing something glaringly illegal or unsafe. Given that, how can one possibly expect we're anywhere but a solar system away from a tesla competently driving through an area with unmarked or incorrectly marked lanes with pylons (or maybe not pylons!) and/or doing the same thing in the snow, etc. In many ways this whole thing has helped illustrate some of the severe limitations AI still has when attempting to replicate human output.

As I said in another thread a completely untrained teenager can still in a matter of hours be taught to be a better driver than FSD with all of its supposed millions of training miles.

For pete's sake I can't even have a tesla drop me at the grocery store and safely park itself in a highly-predictable parking lot, driving at 8 mph. Tesla is still years from FSD if indeed tesla is ever able to achieve it.

What absolutely shocks the hell out of me is that Tesla is even *talking* robotaxi. (And apparently Elon is still pushing for a controller less car for some insane reason...) In order to accommodate a robotaxi scenario FSD will have to be solved to 100%. It has to account for everything it could encounter. As anyone in software can tell you solving a software problem to 100% is an exponential difficulty curve. If they are currently at 80% (being extremely generous there) towards level 5 autonomy that might not even be half way to a robotaxi solution. Folks we are decades away from robotaxis with Tesla's chosen solution. The race isn't over but the horses are coming down the final stretch and Tesla is dozens of lengths behind Waymo at this point.

The software engineer's answer to how close Tesla is on FSD would be "80% done, 80% left".

It should also be noted that robotaxis can pull over to the side if they encounter a case that they cannot safely handle. So they do not need to solve every single case, just be able to reach a safe state and call for help when they cannot handle something. Of course, to be useful, you will want the cases where it pulls over and asks for help to be as rare as possible since you don't want to inconvenience the passengers. You also want to minimize "stalls" where the robotaxi stops in the middle of the road to ask for help which can block traffic and cause problems to others. You want those "stalls" to be as rare as possible. But no, robotaxis don't need to solve everything to 100%. They just need to solve enough 9s to be useful and safe.

This is true if the robotax has enough self awareness to recognize when it doesn't know something. I can find FSD videos of Teslas not being aware, and instead driving with great aplomb into another lane, with the driver saving it from a crash.

 

[kpanda17]

Only one red light run in a 45-minute drive is impressive to me as someone who only has experience with FDSb 11.3.6 and 11.4.4.

Who has gone in front of a judge with an FSDb ticket? Red light, speeding, etc

 

[cucubits]

And now looking back 5 years and thinking where we've been and how far we've come, it's actually quite sad to me that this is all we got

Oh this, so much this! It makes me both satisfied and sad.

Satisfied because about 4-5 years ago I was regularly getting into heated arguments with folks who were 100% sure that in 5 years FSD would be out, released, working and driving elon's robotaxi fleet. The deep "I told you so feeling" is quite good.

Sad, because well, I had hopes that it'll get at least somewhere in that time... now to come back and see that it's basically a $15,000 mediocre lane keep assist. I am honestly baffled that there are actually people out there who keep gifting elon $15k now for this...

 

[Goose66]

BTW, I have no idea where that picture attached to my original post came from. That’s not my car and the map in the screen is not my area. I certainly don’t remember attaching a picture to the post, nor does that picture appear to have any relevance to the post content.

 

[mgs333]

BTW, I have no idea where that picture attached to my original post came from. That’s not my car and the map in the screen is not my area. I certainly don’t remember attaching a picture to the post, nor does that picture appear to have any relevance to the post content.

Sounds like your thread got promoted as a 'hot' thread. I sometimes see unrelated generic Tesla images get added to posts that become promoted.

 

[TheLocNar]

Tesla will die before FSD becomes a real thing. That whole v12 demo was BS. Elon knows it will never happen in our lifetimes but still puts people’s live at risk for the $. All he has to do is make empty promises every year to keep the cycle going.

 

[cucubits]

Tesla will die before FSD becomes a real thing. That whole v12 demo was BS. Elon knows it will never happen in our lifetimes but still puts people’s live at risk for the $. All he has to do is make empty promises every year to keep the cycle going.

That's true and sadly it works for him because of all hist cultists blindly believing all the nonsense he keeps promising.

It's quite a weird feeling, as much as I dislike him and see through all of his BS, I still miss my old Model S with AP1 from time to time...

 

[PhillyCJ]

My Model 3 was delivered in mid-August 2018. While I originally just had Enhanced Autopilot (EAP), I quickly paid the extra $2000 to get FSD as rumors swirled about Tesla charging more in the future. What's my opinion after 5 years of testing and playing with Tesla EAP and FSD? TLDR: It's pretty much still just a gimmick and hardly worth the extra $2K I paid for FSD, much less the $15,000 they charge today.

So a little about my experience so you understand my evaluation. Since receiving the car in August of 2018, I have done everything possible to have the latest firmware, latest features, and test and provide feedback on the latest software and capabilities. I am an engineer by training and education, and have participated in a lot of software Beta Test programs for lots of products, and I take the whole thing fairly seriously because to me it's lots of fun. So I don't feel like I have (or had) unrealistic expectations of what FSD would be able to do, what the testing process would be like, or when the features would become polished and truly usable.

All that said, let's start with what works: Navigate on Autopilot (NoA) and Auto Lane Change. NoA is an L2 ADAS feature that combines TACC and Autosteer to drive the car on limited access highways, and includes features like automatic speed-based lane changes and highway interchange transitions. IIRC, the initial version was released in October of 2018 and, since then I have put thousand of miles on it. Now a part of FSD-beta on highways, while it has improved somewhat, it basically the same functionality as NoA on release five years ago. Phantom braking is noticeably better (happens less frequently but still happens) and the system now has some subtle safety improvements, like cheating to the outside of the lane when passing a truck. But passing behavior is still wonky and can be harrowing, and lane selection for passing and exit/transition ramps is still really bad - maybe even a bit worse than in original NoA now that it doesn't use maps as much. In addition, when it came out in 2018 I felt like NoA was "best-of-breed" for these L2 ADAS highway-driving systems, but I feel like it has now been surpassed by BlueCruise, SuperCruise, Drive Pilot, and the like, many of which offer some level of hands-off and/or eyes-off driving for a truly L3 autonomous driving experience. But NoA is still a very useful feature, IMO, offering reduced workload, fatigue, and stress on long highway trips, especially if you do your own passing with Auto Lane Change instead of the automated speed-based lane changes.

Another thing that works is Summon. Remote control your car into and out of a garage or tight parking space. Only useful in very limited circumstances but it does what it does well when you need it (as long as you still have USS, evidently).

Everything else: AutoPark, Smart Summon, Autosteer on City Streets, and Traffic and Stop Sign Control? Gimmicks. Barely usable and hardly ever useful. Setting AutoPark and Smart Summon aside (because they just plain don't work), Autosteer on City Streets is a massive undertaking and has improved immensely since "beta" testing began almost three years ago now. But to engage it, you have to be ready to be extremely alert and ever vigilant - hardly a relaxing way to drive. And while it can get you from point A to point B - even sometimes without intervention (which has never happened to me) - it is most often still a harrowing experience to use it. As an ADAS feature, it has zero utility, IMO. Fun to show off to your friends, and fun to test the latest and greatest to see what it can do. But from a practical standpoint, it can't get me anywhere I need to go faster, more efficiently, more relaxing, or, regardless of how many people quote statistics to try and prove otherwise, safer than I can get there by driving myself.

And now looking back 5 years and thinking where we've been and how far we've come, it's actually quite sad to me that this is all we got. I don't know how much more appetite for "testing" FSD beta I have, and often think I would be much better off or newer firmware just driving myself or on plain, old AutoPilot (I, personally, would still have my NoA in my EAP). But if there's anybody out there thinking whether they want to invest $15,000 in Tesla's autonomous driving systems, just take a tip from me and go into it with (very) low expectations of what you will be getting. View attachment 972974

I really appreciate this post; I just bought my first Tesla, a Late Birthday 2020 model S plus and am very happy (although its currently in customs shop getting tints, ppf wrap, and a couple other things but after only 4 days i’ve been going back and forth on if i should spend the 200 per or 12k to try out these features, i picked mine up off a lease return to accommodate my budget rather than buying brand new and im not in a position to purchase 5 figure experiments until i can pay off the 40k loan i jus took out along w/ a 35% down payment so i super appreciate ya putting this up when you did; Im not saying i wont try it out, but it definitely wont be a proority (i was disappointed tesla doesnt at least offer a trial w/ a purchase regardless if
From their lease return program
Or new..but as im quickly learning
All these features come w/ a premium pricetag; Thanks for the thread

 

[PhillyCJ]

I apologize for the punctuation..(or lack there of) in advance; Txt to talk malfunction

 

[hooty]

Imagine posting this weeks after the FSD 12 demo. Not to mention it's now $12k, not $15k.

In 2023 people are still fooled by the Chief Carnival Barker

 

[Tony Y]

What absolutely shocks the hell out of me is that Tesla is even *talking* robotaxi. (And apparently Elon is still pushing for a controller less car for some insane reason...) In order to accommodate a robotaxi scenario FSD will have to be solved to 100%. It has to account for everything it could encounter. As anyone in software can tell you solving a software problem to 100% is an exponential difficulty curve. If they are currently at 80% (being extremely generous there) towards level 5 autonomy that might not even be half way to a robotaxi solution. Folks we are decades away from robotaxis with Tesla's chosen solution. The race isn't over but the horses are coming down the final stretch and Tesla is dozens of lengths behind Waymo at this point.

I disagree with your "has to be 100%". We need to educate the public though because we do have a tendance thinking of "has to be 100%". It doesn't have to be. Think of flying, train, people driving, even cycling, nothing ever be a "100%", but we are doing it. I think we need to bring on data, for example, incidents on each 100 million miles of driving. I don't know the propriate number to "be safe". Insurance companies might know it.

 

[Quamera]

So after all this time with positive, neutral and negative opinions on FSD can anyone comment on its ability to avoid people and/or animals on the road. I won't claim to have read every thread on FSD and EAP but I have seen a fair few of them and the subject rarely comes up.

 

[EatsShoots]

So after all this time with positive, neutral and negative opinions on FSD can anyone comment on its ability to avoid people and/or animals on the road. I won't claim to have read every thread on FSD and EAP but I have seen a fair few of them and the subject rarely comes up.

I saw a video five years ago with Volvo advertising, their ability for the cars to avoid deer. I have seen multiple videos of Teslas from years ago doing the same thing, yet, as far as I know, it’s not something that’s really available now. Strangely to me they seems like it should in fact be easy to put in and pretty safe. Dear can definitely look like dogs or cats from the wrong range but surely the car cut if nothing else slow down and alert you? We only got a Tesla this year because my daughter totaled her leaf a couple months ago on a deer.

 

[Quamera]

I saw a video five years ago with Volvo advertising, their ability for the cars to avoid deer. I have seen multiple videos of Teslas from years ago doing the same thing, yet, as far as I know, it’s not something that’s really available now. Strangely to me they seems like it should in fact be easy to put in and pretty safe. Dear can definitely look like dogs or cats from the wrong range but surely the car cut if nothing else slow down and alert you? We only got a Tesla this year because my daughter totaled her leaf a couple months ago on a deer.

Yes, the major cause of accidents in Australia's rural areas is animal impact, mainly kangaroos, walabies and emus but feral pigs, deer and camels also feature and even domestic animals are a problem in some areas. If FSD or EAP could reasonably predict the path of a moving animal and take evasive action it would make them more attractive financially. A stopped animal is a more difficult proposition because it is impossible to know if/when it will move but if it is close to the road it is certainly prudent to slow down so evasive action is easier if required.

 

[mjwojtukiewicz]

Full stop, right there. V12 is not a marvel. It's an alpha test version of some new software that shows some interesting capabilities. That's it. Everything else is speculation. V12 is so preliminary that we shouldn't even be thinking about it. The Tesla engineers have just started fooling with it, so they don't have any serious experience with it and don't know what kind of legs this approach has. They thought they could make V10 work. They thought they could make V11 work. Now they think they can make V12 work.

I was being facetious. Musk was saying it was a marvel in so many words. They're going to be saying the same thing about V13

 

[JB47394]

I was being facetious. Musk was saying it was a marvel in so many words. They're going to be saying the same thing about V13

Kabin picked up on that and pointed it out earlier. Unfortunately, sarcasm usually doesn't convey well in the written word, particularly to people who don't know your mannerisms. As for your prediction about V13, well, I ascribe the same prognostic abilities to you as I do to Elon. Let's just wait and see what happens.

 

[mjwojtukiewicz]

Kabin picked up on that and pointed it out earlier. Unfortunately, sarcasm usually doesn't convey well in the written word, particularly to people who don't know your mannerisms. As for your prediction about V13, well, I ascribe the same prognostic abilities to you as I do to Elon. Let's just wait and see what happens.

Yeah, as I told my wife the other day I have a Jerry Seinfeld sense of humor