We dont have to compare human drivers to anything. I fail to follow your logic about investment, since you seem to be implying that there was some "expected" sum to develop autonomous driving? Expected by whom? This sounds suspiciously like the old argument "This can't be a good idea, because if it was someone else would already have thought of it."
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
It's happening....FSD v9 to be released (to current beta testers)
- Thread starter Canefan456
- Start date
Daniel in SD
(supervised)
No, it's the exact opposite, making a self-driving car is a very good idea, that's why an obscene amount of money has been spent trying to realize it.
When we're talking about autonomous vehicles we have to compare them to humans because that's the benchmark. Saying humans are "bad" at driving is a very popular sentiment but I contend that humans are actually very good at driving. I'm saying that if 100's of billions of dollars have been spent trying to make machines drive as well as humans with minimal success then maybe humans aren't so bad at driving after all.
As you said, all that really matters are the statistics. How close is unsupervised beta FSD to human safety? Can you give an estimate? What is the bar?
Daniel in SD
(supervised)
Obviously I don't expect an exact number, just an order of magnitude. I'm just curious what people think from watching the videos.
rxlawdude
Active Member
First, I'm in general agreement overall. This is an interesting discussion, with a lot of competing concerns and challenges. Not really pushing back on the general idea, but I do have some different observations on some details:
1:2M miles for an accident and 1:100M for a fatality is not *that low*. It's pretty remarkable how good humans are at such a complex task actually. In fact, they are so good at the task, that it may be decades before computers can do it as well. If 1:2M is a "low bar", what do you expect an acceptable bar for a computer is, and why are computers still very far away from it? If humans are so crappy at driving cars, why are we all willing to hop in an Uber or co-workers car without a second thought?
Maybe. But today, we're nowhere near being able to do this. We struggle to get a system to not drive into cones, or not think a picture of a car is a real car. So it's not like because we suddenly have data for an "edge case" that we could instantly program it in for next time. Sure, this is a hopeful end point, but it's really, really far away. Many of these edge cases will require complex, complete understanding of the world with context, and saying it will be trivial to do this is completely ignoring how hard it's been to even get good image recognition working today, and just assuming that we'll be so much better in the future. If the car has this complete understanding of the world like a human does, why did it crash in the first place?
Plus, that's the thing with edge cases- you may never see it again, so fixing it *after* it happens is too late. Your next case will be different. This is why safety critical systems are not designed by sending them out into the field, waiting until they hurt someone, fixing that failure, and then waiting for the next one.
If we get to prove-able 1:20M accident rate, and 1:1B miles fatality, then an accident will be news- because it will be so rare, like an airline accident is today. But below that, there will be a reasonable discussion as to if the system is actually safer than a human. It will also matter what mistake the vehicle made, and if it was one a reasonable human driver would have made or not.
This is exactly the issue with airplanes today- they are so safe that any accident is a news story, which as you say, makes people bad at statistics, as they dramatically over-estimate the crash rate of aircraft, ironically pushing some to more dangerous forms of transport. But it also overlaps because airplanes are things you *ride* in, and have no control over, and that is a powerful mental shift vs a car you drive yourself. Self driving is going to have that same weird issue of handing your fate over to a machine that we do as passengers in an airliner.
1:2M miles for an accident and 1:100M for a fatality is not *that low*. It's pretty remarkable how good humans are at such a complex task actually. In fact, they are so good at the task, that it may be decades before computers can do it as well. If 1:2M is a "low bar", what do you expect an acceptable bar for a computer is, and why are computers still very far away from it? If humans are so crappy at driving cars, why are we all willing to hop in an Uber or co-workers car without a second thought?
What percentage of accidents could be solved with these skills? These are skills needed *after* something else has gone wrong, and aren't generally needed by attentive drivers. The majority of accidents in the USA happen at low speed, in urban environments. Giving everyone great emergency braking or swerving skills won't make all that much difference if attention deficit is the issue. And, like I explained before, these are things tech already is solving- ABS, stability control, and panic braking assist do all help here even without full FSD. We can (and are) making humans superhuman in these areas with much simpler tech than FSD.
Maybe. But today, we're nowhere near being able to do this. We struggle to get a system to not drive into cones, or not think a picture of a car is a real car. So it's not like because we suddenly have data for an "edge case" that we could instantly program it in for next time. Sure, this is a hopeful end point, but it's really, really far away. Many of these edge cases will require complex, complete understanding of the world with context, and saying it will be trivial to do this is completely ignoring how hard it's been to even get good image recognition working today, and just assuming that we'll be so much better in the future. If the car has this complete understanding of the world like a human does, why did it crash in the first place?
Plus, that's the thing with edge cases- you may never see it again, so fixing it *after* it happens is too late. Your next case will be different. This is why safety critical systems are not designed by sending them out into the field, waiting until they hurt someone, fixing that failure, and then waiting for the next one.
To be aware, Tesla themselves says that an accident occurs on AP every 1:4M miles today. That's multiple accidents a day on AP already. These do not all make the news for the very reason that they are so common.
If we get to prove-able 1:20M accident rate, and 1:1B miles fatality, then an accident will be news- because it will be so rare, like an airline accident is today. But below that, there will be a reasonable discussion as to if the system is actually safer than a human. It will also matter what mistake the vehicle made, and if it was one a reasonable human driver would have made or not.
This is exactly the issue with airplanes today- they are so safe that any accident is a news story, which as you say, makes people bad at statistics, as they dramatically over-estimate the crash rate of aircraft, ironically pushing some to more dangerous forms of transport. But it also overlaps because airplanes are things you *ride* in, and have no control over, and that is a powerful mental shift vs a car you drive yourself. Self driving is going to have that same weird issue of handing your fate over to a machine that we do as passengers in an airliner.
I think Tesla abuses the label while at the same time other companies don't honestly admit how many bugs their systems have. Other companies are especially problematic with their media systems, and charging systems. Just look how long it took Ford to sort out issues with the Mach-E. Heck in the early days some people couldn't even start their vehicles.
Another example is the Subaru Eyesight system. The first few generations were problematic, but it never had the beta tag. In fact I'd say that the first few generations of most ADAS systems left a lot to be desired.
The industry is as a whole is moving to OTA SW updates which is a two sides sword. On the one hand it allows for easy fixing of issues, and adding addition functionality (if the HW supports it). But, it creates major SW headaches because then you have a mix of hardware variations in the field where its hard to test all the possible scenarios.
The rain sensing wipers is a great example of something that really should be given the freedom to evolve if its truly their as a convenience to a human driver. I can't say I particularly liked the AP1 Rain sensor despite Tesla using a proven one back then, I also don't particularly like the Sprinter Van rain sensor. Sometimes it just gets confused and will aggressively start wiping. Do I like the one in my Model 3? Sometimes. It's better than it was.
As to L4/L5 I'm not sure why I even mentioned L5. Sure I know L5 is the goal for Musk, but I firmly believe he picked the impossible goal because it's impossible. If he had picked some limited L4 functionality a lot more people would be livid at the lack of progress to it.
The problem with the second coming of Christ is that's as a real as hostile aliens. If you were a hostile alien looking to take over the planet your best bet would be to disguise it as a the second coming of Christ. You could wipe out entire areas without anyone putting up a fight. But, I'm pretty sure even aliens couldn't figure out how to get L5 to work safely alongside human drivers. So maybe L5 after the second coming of Christ.
But, is it phantom braking if a human driver slows down when being mooned?
Folks I bought car with FSD ( not a subscription) bit still today no any features of FSD ( and my software even not FSD). I am new to Tesla and what is this? If it's not ready ( don't put on FULL subscription or sell it) please return money back and if you charged people for features which is not provided this is a PURE FRAUD.
Terminator857
Active Member
Where is Kelowna?
Kelowna in British Columbia, Canada. OK if it's only for BETA testers ( upgrade or sell it' ONLY for testers) and if it's not ready for public don't sell it TO PUBLIC, otherwise this is not provided feature what was paid for considered a fraud. I understand if this is very NEW features and I can wait for couple of months, but not 6 month -1 year. Moreover I called Customer Service and they say Book Mobile service appointment and they can engaged it for me, because in my case it's PART of Vehicle.Where is Kelowna?
Check the furthest right column on the above table that says FSD / EA for FSD features.
Last edited:
It has been in developed and for sale since 2016. 2019 it was coming “later this (i.e. 2019) year”.
I found my answer directly from TESLA SERVICE TECH DEPARTMENT. I need service appointment with Mobile service tech and he can config and engage my FSD. I am already pass 1000km and my cameras must be calibrated. Nothing with this is BETA TESTERS. I have integrated full FSD, Premium connectivity and it's MUST works ( probably a bug issue). BETA TESTERS very limited Tesla owners who using absolutely different software available only for them.
rxlawdude
Active Member
Wow, pretty vitriolic. I'm sure that you'll be courteous to Tesla's employees, because if you come in with an attitude like "FRAUD!" you will be treated in kind.
Off topic, you live in a beautiful area. We took our Model 3 on a road trip to Banf, and stopped in Kelowna for a B&B and to check out the wineries.
I am very new on this is forum and this is word "Fraud" not mine . Many topics were people( experience people with Tesla, not like me) complaining about this is system from 2018, 2019... using this word even more harder!!! With my BMW I take a phone and after 5 min I found all my answers, but with Tesla this is not so easy. I love my Model 3 very good vehicle, and this is up to me I will use FSD or not, but if I bought this is feature I want this is works. Very simply.
rxlawdude
Active Member
And you will indeed get it to work after service does their thing. Your area has a few divided highways (trans-Canada Rt 1 and Rt 5) and the current functionality of FSD will work very nicely on those. City or town FSD is the one in very limited beta testing.
(In 2019, TACC/FSD speed limits were based on Google maps and were frequently 5km/hr lower than posted; now, since our cars read signs, I expect that minor frustration will be gone.)
May be I am very naïve, but I believe Tesla and Elon. I will patient and wait when it will be added cities and streets. What Tesla says configuration what I have right now in my Model 3 must works.
While you are right to be skeptical, just arguing from “common sense” is rather meaningless. Virtually every major scientific discovery in history has gone against the “common sense” of the time. Your argument is that an autonomous vehicle requires sentience and that we don’t know how to create this in a computer. The latter, for the foreseeable future , in certainly true, but not the former. On what do you base your argument that an autonomous car requires sentience?
Certainly some exceptional circumstances require reasoning, but most day to day driving is mechanical and based on following simple rules. Which is all that is required for the autonomomy that Tesla is currently aiming for (I won’t comment on their earlier more ambitious goals). Sure there are a lot of rules, some conflicting, but that again is just a scale issue rather than one requiring sentience (and no, I’m not going to debate endlessly on what sentience means).
I think Tesla (or at least Elon) have been way too optimistic with timelines, but I don’t think the problem is technically impossible, nor is a solution improbable.
You are muddling things up. First, the low fatality rate is mostly related to the vastly safer cars driven today. Crumple zones, seat belts, air bags etc all make accidents more and more survivable. It's important to distinguish accident avoidance from accident mitigation.
Let's look at serious accidents. What are the causes? Some are mechanical, to be sure, but most (look up the stats) are human error. Someone doesnt look in their mirror before a lane change etc etc. These are lapses in concentration (or just bad driving). How would an autonomous vehicle reduce these serious accidents? First, by avoiding causing the accident in the first place since the car isnt going to stop paying attention while it checks a text message (or get drunk etc). Second, by reacting faster and more safely when it does find itself faced with an emergency. Avoidance and mitigation. Sure, numerically there are far fewer of these serious accidents, but so what? They are the significant ones because they are serious. Arguing that they are not relevant because they are numerically infrequent compared to fender-benders is an invalid argument. I'm also unclear as to why you argue we dont need FSD because we have airbags. Really?
What about those minor accidents? As you note, the majority of accidents fall into this category, mostly rear-ending someone at a stop light and similar. Again, the cause here is a lapse in concentration. And again, an autonomous car wont suffer from that.
The fact is, both humans and autonomous cars are going to get things wrong. Sometimes badly. Sometimes fatally. But the nature of those errors will be fundamentally different. Any decent responsible human driver knows what they should do when driving, but lapses, distraction, laziness and fatigue mean they often do not do those things. We've more or less given up trying to make humans better drivers (dont drink and drive? dont text while driving?) .. as I noted above most improvements over the past decades have been mechanical mitigations not driver education.
The problem facing autonomous car development isnt solving lapses in concentration, that comes essentially free of charge by the very nature of the system. Instead, its advancing the cars ability to understand what it should do. We are essentially seeing scientists and engineers teaching a car to drive. If/when that teaching reaches a comparable level to an average human driver, then Its inevitable that the autonomous car will be safer because it never has lapses of concentration.
Of course, the big "if" here is can that autonomous driving level be achieved And no-one knows this .. sure we can all speculate and argue about it (it's part of the fun). But that's all it is .. speculation. People here have used peculiar arguments about Lidar, Radar, sentience and who knows what else. But until we get hard numbers no-one knows. But if it is solved, then humans are definitely going to be in the back seat .. both figuratively and literally.
Daniel in SD
(supervised)
You: Humans are bad drivers, the bar is low.
Us: Humans get into collisions (>12mph) every 2 million miles. It doesn't look like FSD beta is very close to that.
You: Humans are bad in these ways, machines are good in these ways. Autonomous vehicles will replace human drivers when they achieve greater than human safety.
Me: Yes. Are we arguing about something?
Similar threads
