Autopilot is getting worse, not better.

[Battpower]

The ‘FSD’ product is Autopilot (which IS level 2 and always has been) plus auto park, smart summon and autopilot on city streets.

I find it hard to get behind any notion that FSD was going to be anything other than level 2 when the core of the product is level 2 and makes no claims that it would be otherwise.

Not quite sure what the point is.

That FSD is delivering what was promised?

Or that 'of course FSD was never going to be more than L2'.

Since FSD is supposedly not a term that we should take to actually imply full self driving (like the car can reliably do something on its own) - just Tesla's current view of whatever FSD means which at some points has been identical to EAP - then how can it be claimed that FSD has been delivered? Or rather that at any point Tesla could claim it has been delivered!

 

[qwickshot]

Not quite sure what the point is.

That FSD is delivering what was promised?

Or that 'of course FSD was never going to be more than L2'.

Since FSD is supposedly not a term that we should take to actually imply full self driving (like the car can reliably do something on its own) - just Tesla's current view of whatever FSD means which at some points has been identical to EAP - then how can it be claimed that FSD has been delivered? Or rather that at any point Tesla could claim it has been delivered!

I’m not sure which part of my post you are responding to either.

As set out, the U.K. product hasn’t been delivered so it’s almost elementary. Get a refund if you want one.

FSD in the USA has essentially been delivered as the product was described in terms of functionality.

It was only ever going to be a level 2 ‘self driving system’, e.g. the driver was always legally in control.

FSD is just a marketing name/term for a product, it’s not its full description of its functionality. It may not specifically say that its level 2 on the page but Tesla has always been clear that the driver remains in control at all times, that’s basically the definition of level 2

 

[Battpower]

clear that the driver remains in control at all times.

That kind of clarifies things. I would agree with that statement that Tesla has never printed a spec that suggested otherwise, although they have sailed very close to the mark.

'Driver remaining in control' isn't quite as clear cut as you might wish!

 

[MrBadger]

Transcript from Oct 19, 2016 Elon Musk press conference about the release of Autopilot 2.0.

IMHO it’s pretty clear what he was saying about HW2 capabilities, let alone HW3.

Putting the date into context, it’s around the time that reservations were being taken on Model 3 so would be fresh in the mind of some early Model 3 (and S and X) adopters.

Transcript: Elon Musk's Autopilot 2.0 Conference Call

Full Transcript: Elon Musk answers all questions about Level 5 Self-Driving and Hardware 2.0 to details about the future of latest Tesla developments.

www.xautoworld.com

 

[Billbrown1982]

Doesn't make sense, you expect that all hardware development should be stopped while the software is developed for your version of hardware because it should be the complete answer to everything forever? Thats not how technology works.

The fact that there will be future hardware iterations means nothing at all about the software. Also your car from 2021 isn't going to be the best ever Tesla make.

I truly don't know if you don't understand, or refuse to. Either way, your putting words in my mouth and I can't explain it any clearer so lets just agree to disagree.

 

[EatsShoots]

Tricky one. In theory once they have this all to a level they are happy with, they might only have to run it a few times through their training computers to handle the UK and RHD. As you say though, the further out this is the less likely it is to happen for HW3 cars. They might just do it to avoid more lawsuits, don’t know. I’m not expecting any form of Tesla FSD in this country for 4 - 6 years.

Guess my question was more to someone that has taken Tesla to court or plans to, would they still do the same if what was available in the US was in the UK. Some seem to think and wouldn’t be happy until it’s Level 5 but nothing is going to get to level 5 that quickly. You’ve got to have a system that’s been through so much human testing and not needed manual intervention before you can safely ship that. Those driverless taxi’s in the US still have a human that can remotely take over when it gets stuck for some reason.

Maybe this is the answer. Sell you “FSD” for a £6k a year subscription. A person in a Tesla driving centre in India will actually take over and drive your car for you. When they are all busy it’ll claim the cameras are blocked and “FSD” isn’t available You’ll know this is the case when the car tells you it is going to do the needful.

I've never seen FSD in the UK. I've seen it in the USA On my 2023 Model 3 when they had the free pilot in April.

It's extremely impressive in that it can more or less drive around by itself just with cameras. It's a cool tech demo. I was able to make several multi-mile (like 5-8 mile) trips with zero interventions. That's not easy and I don't want to dump on it too bad....

With that said, each of those trips the system exhibited substandard behavior that would be dinged on a driving test. And some other trips I had to intervene often. It drives like a very drunk person, for the most part. Very drunk people generally make it home, but they often don't. Additionally, there are a number of occasions during that month that I had to intervene to stop it from creating a major accident. Perhaps its greatest issue is that even when it has no flipping idea what is going on (like going through a two-way, one-lane tunnel in which people know to stop and take turns), it exhibits the utmost confidence in its inability and continues on anyway.

Furthermore I genuinely have no idea how it's ever going to be able to take an unprotected left as safely as a person, considering it is using a camera that is in the center of the car, instead of offset like a human's eyes (and that human can move their head to the left a pillar if needed).

So I can say the USA doesn't have anything close to what that 2019 quote from musk promised.

 

[FastLaneJB]

My response was based on questioning what the issue is. Hardware, software (or both). What they have in the US strikes me as being the result of a lengthy coaxing of software to try and make something work somewhere. My ownership experience suggests that in the UK there are so many (may be small) differences (road layout conventions, position and type of speed / warning signs, condition of road markings, frequency of changes to lane markings at roundabouts, narrower roads with close / tall structures, bus stops etc etc), that to have 'what they have in the US' would likely not be any use in the UK.

If your question is 'if the functions apparently available in the US were available and worked in the UK, would I be satisfied that Tesla has met its claims' I would say 'I doubt it' because the demands of UK roads most likely need a much higher level of performance.

If Tesla had even demonstrated a solid commitment to getting even certain FSD features working reliably in the UK, I might be prepared to accept that. But with NOTHING to show in the UK and even EAP features that don't work......




Elon doesn't admit stuff. He just manipulates!

I don’t think that answers the question I asked though. If you had what they had in the US right now, would you be unhappy and want to take Tesla to court?

It’s not a it won’t work here and their system is not good enough. It’s theoretical and in this theory it’ll drive as well in the UK as it does in the US currently.

However to add to your point. When it was 300k lines of code and I’m sure a lot more to make it work in the US, that would take a long time to bring to the UK and RHD. However it’s mostly just playing it training clips of humans driving around now is how it learns to drive. If they have enough footage of people driving Tesla’s in the UK they already have all the data they need, they merely just need to get their supercomputer cluster to crunch it into a trained model.

Yes it’ll need tweaks and improvements but actually I don’t think it’ll take long to bring here when the law allows it and when Elon feels we are worthy of the time and effort.

 

[FastLaneJB]

Of course not.

But it is the way honesty works. Thousands of customers were promised that their cars would function with FSD with HW3 and HW4. Turns out that was a lie aimed at boosting stock prices and making the company seem like something it is not. A company that could provide Level 5 FSD at the end of 2019 would be very very impressive, and would be leading the industry.

But they did not do that. They lied.

2020: more lies
2021: more lies
2022: more lies
2023: more lies
2024: more lies.
2025 ?

Tesla is a company that is good at making a profit on an adequate if slightly-behind-the-times car with poor ergonomics. But what they are great at is lying.

Lies work just fine, if you are really good at it you can get elected to be the president of the USA

And of course any other country inhabited with humans who are easy to manipulate, which is all of them.

 

[FastLaneJB]

I totally get that this is the 'theory'.

And it has always been the theory.

'Tesla just doesn't get the differences' has been said many times, but as someone who has driven many miles in the States as well as the UK, the UK is pretty much one big anomaly!

I’m sure once they’ve finished training it, it’ll drive only in the middle lane, almost rub its bumper against the car in-front it’s following, not indicate on roundabouts and drive 40mph in a 30 / 40 / 50 and 60. It’ll be one of us in no time.

 

[Battpower]

Perhaps its greatest issue is that even when it has no flipping idea what is going on.......

That's interesting given your perspective. I have always maintained that there should be a notional 'current confidence meter' giving the driver a clear indicator of how 'confident' the self driving system is at a given point in time. Tesla's implementation doesn't even appear to acknowledge such a concept and approaches everything as though it is 100% correct until some point where it throws up it's hands and gives up completely.

As a human driver, I try to always be aware of my confidence - given road condition, traffic congestion, visibility, temperature etc. and I adjust my driving style appropriately.

 

[FastLaneJB]

That's interesting given your perspective. I have always maintained that there should be a notional 'current confidence meter' giving the driver a clear indicator of how 'confident' the self driving system is at a given point in time. Tesla's implementation doesn't even appear to acknowledge such a concept and approaches everything as though it is 100% correct until some point where it throws up it's hands and gives up completely.

As a human driver, I try to always be aware of my confidence - given road condition, traffic congestion, visibility, temperature etc. and I adjust my driving style appropriately.

I go Days of Thunder style, stay on the right and power through

Joking aside. Tesla would likely say that your in the drivers seat to correct it if it’s going wrong. So it’ll do its best but it’s your responsibility beyond that. That’s what Level 2 is.

 

[parsec]

Because as I already showed you his next sentence in the same quote was "However, people sometimes will extrapolate that to mean now it works with 100 percent certainty, requires no observation, perfectly. This is not the case.” - this is precisely him saying it won't be Level 5.

You are misinterpreting what he said. His word"now" meant "at the beginning of 2019". Of course it did not work then as level 5, nor does it now.

But in the paragraph before, he states emphatically that it will work as level 5 is defined: no need for intervention. Every year he repeats the slight of hand and the lie : "Doesn't work now, but it will by the end of the year," In his devious way, every year he has been saying, via wobble words: "It's level 2 now but it will be Level 5 by the end of the year." And the stock price goes up.

 

[Transformer]

The reason Tesla has gone down the end-to-end neural network route is for precisely this reason. They don't need to teach it what all these bits and pieces are, they're implicitly included in the training. So in general it'll do the same as a human does, which is to improvise carefully. None of the training data from the US will be used here. ... The great thing about this is that with sufficient training data and computing capacity all the weird British bits of driving will be accounted for out of the box (in theory...)

Let's evaluate Tesla's theory. Could it work to train monkeys to drive by just showing them an immense amount of driving video + control inputs? No DMV/DVLA driver's handbook. They can't read a handbook or road signs. No driving instruction.

The E2E neural net is even more limited than monkeys: no stereoscopic vision, no experience with the physics of motion and collisions, and no ability to feel the forces of acceleration. No high level problem solving skills.

 

[Fursty Ferret]

Let's evaluate Tesla's theory. Could it work to train monkeys to drive by just showing them an immense amount of driving video + control inputs? No DMV/DVLA driver's handbook. They can't read a handbook or road signs. No driving instruction.

Is what we do ourselves that much different? Basically getting in a car and practising on public roads? They can have FSD running on the 2nd Autopilot computer and every time its decision varies from the driver's, you get a training point for your data. Or you can have it go around in a simulator to see more interesting stuff.

Not sure I'd use the lack of stereoscopic vision as an argument in this case. It has close to full-time 360 degree vision, we have two eyeballs. Stereopsis is useful for parking and nothing else (you can drive 500 tonne passenger jets without stereopsis, so one assumes that it's not necessary for trundling around town).

 

[Battpower]

Is what we do ourselves that much different?

It is possible to construct an argument that say we (humans) don't do much different.

But such an argument would ignore that generally, humans have got sensor fusion sorted!

 

[Battpower]

train monkeys to drive by just showing them an immense amount of driving video + control inputs? No DMV/DVLA driver's handbook.

You can still have some starting points that could be handled separately from NN. Also map data. But basing everything solely on (mono scopic) vision is much different even from monkeys.

And cameras that so far have various sight impediments in some (potentially quite common) circumstances.

At some point Tesla will have to add in some other senses / sensors imo.

 

[FastLaneJB]

It is possible to construct an argument that say we (humans) don't do much different.

But such an argument would ignore that generally, humans have got sensor fusion sorted!

As adults, yes but as babies? Do you not need to teach someone how to catch for instance where they then learn to time the movement of their hands based on the input from their eyes. We don't even understand or know how to control most of our body as babies. We cannot speak, we cannot walk, etc. We learn by watching and listening to the world around us... kind of like how machines learn by watching videos. It's incredibly similar in many ways.

They don't need to be perfect, just better in that they don't get tired, don't get drunk, might know from the eyes in the back of the car that if they move just half a car forward the car behind that's struggling to stop won't hit you, etc.

Is Vision Only ideal? No I don't think so. If you want them to be even better, they need to be able to see in the dark, not get blinded by the sun, etc. However I think if you haven't even cracked the problem with one set of sensors that in fairness look like they can do a good job most of the time, why complicate the problem more. You've got to have camera's because you need to be able to read road signs, see road markings and so on. Camera's are a must, others can just help enhance beyond that.

I bet once Tesla feels happy and things like Lidar's are so damn cheap, they'll pop some in their cars. I think if they want Robotaxi's, they will do and maybe even HW5 / AI5 will have them added. Sensor fusion won't be as hard when they learn the machine learn it on its own, it was hard when they were trying to get humans to code to achieve it. They won't have the data though, AI5 will have to ship and then they can gather video and lidar data from cars at the same time. Then the training can know that when I see this in the cameras, this is what I'm seeing on my Lidar. It'll start to put 2 and 2 together between them.

 

[GeorgeSymonds]

Not sure I'd use the lack of stereoscopic vision as an argument in this case. It has close to full-time 360 degree vision, we have two eyeballs. Stereopsis is useful for parking and nothing else (you can drive 500 tonne passenger jets without stereopsis, so one assumes that it's not necessary for trundling around town).

Stereoscopic vision is useful for depth perception and changes in distance but is most effective at shorter ranges, I’m not sure passenger jets ever get close enough to an object for it to matter. Somebody with one eye find it much harder to catch a ball for instance than somebody with 2. We allow drivers with 1 eye and other disabilities to drive as society has decided it’s ok to accept the higher risk, and I suspect most people with disabilities drive in a manner that’s more cautious to balance the risk, I know a few elderly people who won’t drove at night because their vision isn’t so good. You could argue they have opted for L4 driving.

Irrespective of what we’re getting, and I wouldn’t be surprised if it’s not actually changing much as the effort is in the US FSD space, but even that seems to have stalled. FSD supervised v 12.4.1 was briefly seen a few weeks ago, Musk as usual singing its praises and since then there’s been talk of HW5.. makes you wonder if they’ve hit a compute barrier in the car irrespective of any sensor changes they may seek to make

 

[Transformer]

We learn by watching and listening to the world around us... kind of like how machines learn by watching videos. It's incredibly similar in many ways.

Mostly we learn by hypothesis testing: If I move this way, will things happen as expected? If I say this, will people react as expected?

We also learn with higher level reasoning and instruction, e.g. rules of the road, following distances, slippery surfaces, weather impact on stopping & turning, how to handle a skid, shadows & mirages, other drivers' motivations and likely actions.

Is what we do ourselves that much different? Basically getting in a car and practising on public roads? They can have FSD running on the 2nd Autopilot computer and every time its decision varies from the driver's, you get a training point for your data. Or you can have it go around in a simulator to see more interesting stuff.

Not sure I'd use the lack of stereoscopic vision as an argument in this case. It has close to full-time 360 degree vision, we have two eyeballs. Stereopsis is useful for parking and nothing else (you can drive 500 tonne passenger jets without stereopsis, so one assumes that it's not necessary for trundling around town).

Practising on public roads is a lot of hypothesis testing, not passive observation.

Stereo vision gives depth perception for a range that depends on the distance between the eyes. We can extend that by moving the head left and right. Big passenger jets need parking helpers on the ground.

 

[parsec]

The only thing you can say is that it’s ambiguous. I don’t think Tesla have ever made the specific claim that the system will be level 5 in any of their materials.

I'm not sure if you are being intentionally deceptive or pedantic here by using the words "any of their materials." I consider Musk's recorded words from investor conferences to be "Tesla" making a specific claim. (You may say that a video of him speaking is not their "materials".) Otherwise, if saner people in the corporation realized that Elon was merely confused, you'd expect a responsible person in the corporation to walk back Elon's ludicrous claims the day after the conference call. That has not happened. He has used the words level 4 and 5, and it is hard to believe that he is so stupid that he does not know what those levels mean: every person with serious interest in autonomous vehicles knows exactly what those level designations mean.

Level 3 means that driver interventions are required at times. Levels 4 and 5 mean that driver interventions are not required. If a CEO says that his corporation will make cars that can operate in full autonomy, anywhere, with no interventions required, then he is saying that those cars operate at level 5. That is unambiguous. Elon has made such claims repeatedly since 2016. The quote from 2019 above in this thread is incredibly unambiguous.

If you genuinely do not understand the SAE levels, this article includes a chart that describes them well:
Elon Musk believes Tesla will have 'level 4 or 5' self-driving this year - what does that mean?