Tesla Model 3 revealed

[Durzel]

Guvernator said:

Have to agree with the above, there are some fundamental unsolved problems which mean a proper FSD as most people would define it is still a decade maybe more away, therefore it's a bit disingenuous to make claims that FSD is here when it really isn't.

For reference I'd define a complete FSD as a system that can complete 95-99% of normal, everyday journeys with the only necessary input from the driver being to programme in a destination.

It doesn't help really that these things are sold as "Autopilot" and "Full Self Driving" etc. On the one hand I can understand why - they are terms that are easily understood and massively appealing, but they do not really mirror reality and won't for many years, and many iterations of hardware (and legislation that catches up).

[skwdenyer]

Tuna said:

skwdenyer said:

Simulation isn't worthless, and Tesla are simulating. But simulation can only simulate the things that are imagined by the programmer or (if we're getting to that point) by the code.

For sure. But driving in real world conditions, you can't do the same junction but 3 inches to the left, then 2 miles an hour faster, then with three more cars crossing lanes etc. etc. without an infinitely larger number of cars on the road.

Just as with the dashboard in the model 3, Tesla make a virtue out of a necessity - the reports are clear that they do not have the capacity to simulate at the scale of Google. Meanwhile Autopilot at full chat is processing gigabytes of information every minute - no-one is pretending that your car is uploading more than a few megabytes of that per day for training.

AIUI they don’t need to. That’s not the point. The data uploads seem to be broadly in 2 buckets:

1) events where the on-car S/W predicted something that didn’t happen (or vice versa); or
2) examples of imagery requested by the centre to aid training

I would presume that the threshold for 1) right now is quite high - gross differences. Over time I would expect the threshold to fall.

The clear presumption is that the rate of change is/will be exponential. Whether that turns out to be the case or not remains to be seen.

[stuckmojo]

Heres Johnny said:

They'll be the ones I list quite a lot of - I've about double the inventory in the UK compared to what Tesla publicly declare - all they do is nobble the top level listing but the individual page for the car still exists.

And the other other thing they do is sell service cars with quite a lot of miles on as new inventory with massive discounts, if you want an example...

https://tesla-info.com/detail.php?ref=62e2be8084cd...

That is a great website. Essential

[HumphAML]

Waiting for a delivery date for my M3 LR.
Anybody found insurance that’s even remotely sensible? I’ve been quoted £1200 from Admiral & £1098 from DL (oh & £5900 from Adrian Flux!). For a 56 year old with no convictions & no accidents that sounds excessive!
Interested to hear other potential owners’ experiences?

[FurtiveFreddy]

HumphAML said:

Waiting for a delivery date for my M3 LR.
Anybody found insurance that’s even remotely sensible? I’ve been quoted £1200 from Admiral & £1098 from DL (oh & £5900 from Adrian Flux!). For a 56 year old with no convictions & no accidents that sounds excessive!
Interested to hear other potential owners’ experiences?

The Model 3 still hasn't been properly assessed by the underwriters, so any quotes right now are based on worst case and guesswork. It should all be finalised very soon and then I think you'll find the quotes start getting more sensible.

[Heres Johnny]

HumphAML said:

Waiting for a delivery date for my M3 LR.
Anybody found insurance that’s even remotely sensible? I’ve been quoted £1200 from Admiral & £1098 from DL (oh & £5900 from Adrian Flux!). For a 56 year old with no convictions & no accidents that sounds excessive!
Interested to hear other potential owners’ experiences?

LV have been known to insure Teslas or theres a specialist Novo, but no idea with either on he model 3

[HumphAML]

FurtiveFreddy said:

The Model 3 still hasn't been properly assessed by the underwriters, so any quotes right now are based on worst case and guesswork. It should all be finalised very soon and then I think you'll find the quotes start getting more sensible.

Agreed, some insurers said that Thatcham had only loaded M3 Performance & not Standard or LR models so couldn’t quote.

[FurtiveFreddy]

Most are quoting but they are making it up or basing it on a Model S (in Direct Line's case).

Nobody actually has the correct information yet.

[NRS]

Tuna said:

NRS said:

I would guess a lot of the value isn't the driving itself - that is probably pretty easy and done now? It's surely the real world driving which is the critical issue - either discovering issues with your sensors, or recording the unexpected and what do you see that could help predict it in future.

Yes, it's all about correctly interpreting the environment around you and deriving an intent from that.

It looks to an external developer a lot like the AI research in the 90's - initially there were great strides, and everyone got excited that it was a 'solved problem'. I remember a few companies that were spending millions on knowledge sets with the expectation that free inference would just drop out of the system if it was sufficiently primed. It never happened.

The reports that I'm seeing (and example here: https://forums.tesla.com/forum/forums/shadow-braki... ) seem to suggest two problems - first that the 'interpretation' phase is quite sensitive, leading to either phantom events or over-sensitivity to stimulus. The second is the problem of interpreting intent - if the brake lights go on, on the car in front, is it a prelude to braking hard, or just a tap? And these are exactly in the area that is hardest to solve - how do you tune a system to interpret its environment correctly when you only have partial information on what the correct interpretation is? Watching what the driver is doing doesn't help - we often react slowly, and let potentially dangerous situations build up because we can see that they will resolve themselves. We have no way of telling the car 'that wasn't meant to happen' apart from slamming on the brakes which isn't exactly a sophisticated feedback mechanism.

Rob doesn't like me being down on Tesla, but Musk appears to be betting the company on FSD, and I think that's a huge gamble - not helped by his habit of completely over-hyping capabilities. Anyone following the nonsense around Hyperloop and the Boring company will understand that what he says he is achieving, and what is actually happening tend to be two entirely different things. And you can't really apply 'common sense' to this sort of software research - it's not a problem that you just throw stuff at until it works, there are still some properly unsolved issues that appear to require a big breakthrough to fix. In the mean time, Tesla appear to be fine tuning things to make Autopilot as consistent as possible - but that's not the same as making the step up to FSD.

As with a lot of stuff there is a slow start, then stuff starts getting money thrown at it and a lot of rapid improvements. Followed by a slow down to sort the hard issues. We've had a lot of improvements, but there is still a long way to go.

You were saying Google is better, as they have much more simulation. The whole point with Tesla is they are finding real world issues. I am not sure I'd have assumed a bridge shadow creates a braking incident in a simulation. Therefore the real world testing seems to be showing the issue, and they are doing stuff to reduce the problem - for example shadow breaking. Of course you want it to be too sensitive at first, as the opposite is likely to kill people. Could also happen with shadow breaking, but seems to be less of an issue from the lack of reports of people dying due to it.

Of course the car is not just interpreting the intent of the car ahead - it's other sensors will look for other stuff. Some will not be as good as us, others may be better. Of course the computing is way behind our brains at the moment. For example the car might start braking ahead - at night a human might not see the drunk by the side of the road. The car could use sensors to detect the person and know it needs a harder brake than just tapping the brake. Not to mention humans will not always interpret correctly either.

Of course Musk is somewhat "dodgy" - but as of now it generally has worked. The people providing cash have done so due to "believing" in him enough. It's very risky, but if it works he does really well. It's a case of fake it until you make it. It might work, it might not. And maybe it doesn't make the dream he outlines - but if it's good enough then people will accept it too. Is an iPhone worth the extra cost? Many will say no - but huge numbers go for it anyway. Or it could all blow up. Who knows?

[Heres Johnny]

I disagree that a false positive is much better than a missed positive, it depends on the situation and both can be catastrophic - following a joint in the tarmac when the lane divider disappeared resulted in a fatal accident is one I can think of.

You also have to consider the behavioural science aspects of using the public to be your testers, the better it gets the less likely they are to be ready to step in when there’s a false positive or missed positive. Then layer on the nature of machine learning where things don’t always get better, a given scenario you’ve successfully experienced successfully many times before could fall outside after a change in model so you can’t even reliably know when and where to pay attention. You hear this now with the phantom braking, but it could be a corner, a junction, a particular road marking etc

It’s not easy, and the Tesla fans will cry foul but I suspect Tesla will come under increasing pressure not less on their FSD approach.

[SpikeBmth]

HumphAML said:

Waiting for a delivery date for my M3 LR.
Anybody found insurance that’s even remotely sensible? I’ve been quoted £1200 from Admiral & £1098 from DL (oh & £5900 from Adrian Flux!). For a 56 year old with no convictions & no accidents that sounds excessive!
Interested to hear other potential owners’ experiences?

certainly look at LV, they now have a specialist EV department, and cover includes recovery if you run out range.

There is never any point getting any insurance more than 30 days before its due, with 21 days being the optimum time, as prices fluctuate just like plane tickets & hotels.

Also without the DVLA recognising a registration, then Insurance companies will be quoting higher, I see the model 3 isnt listed yet on most sites, so too early to get any reasonable quotes really..

[MrMajic]

SpikeBmth said:

HumphAML said:

Waiting for a delivery date for my M3 LR.
Anybody found insurance that’s even remotely sensible? I’ve been quoted £1200 from Admiral & £1098 from DL (oh & £5900 from Adrian Flux!). For a 56 year old with no convictions & no accidents that sounds excessive!
Interested to hear other potential owners’ experiences?

certainly look at LV, they now have a specialist EV department, and cover includes recovery if you run out range.

There is never any point getting any insurance more than 30 days before its due, with 21 days being the optimum time, as prices fluctuate just like plane tickets & hotels.

Also without the DVLA recognising a registration, then Insurance companies will be quoting higher, I see the model 3 isnt listed yet on most sites, so too early to get any reasonable quotes really..

Churchill are slightly cheaper than LV too in the quotes I’ve seen.

[NRS]

Heres Johnny said:

I disagree that a false positive is much better than a missed positive, it depends on the situation and both can be catastrophic - following a joint in the tarmac when the lane divider disappeared resulted in a fatal accident is one I can think of.

You also have to consider the behavioural science aspects of using the public to be your testers, the better it gets the less likely they are to be ready to step in when there’s a false positive or missed positive. Then layer on the nature of machine learning where things don’t always get better, a given scenario you’ve successfully experienced successfully many times before could fall outside after a change in model so you can’t even reliably know when and where to pay attention. You hear this now with the phantom braking, but it could be a corner, a junction, a particular road marking etc

It’s not easy, and the Tesla fans will cry foul but I suspect Tesla will come under increasing pressure not less on their FSD approach.

Yes, agreed. Theoretically it just needs to be safer than human drivers to make it worth it. In reality because the death is a result of the company it needs to be to a much higher standard, particularly in countries like the US (and in a way flawless - which is always likely to be hard/impossible).

[anonymous-user]

Doesn't a big step forward in autonomous vehicles come when, rather than each individual being autonomous, the whole fleet collaborates

e.g. on a motorway when any one car identifies the need to brake hard to avoid a hazard, it knows which other cars are near it and tells them that it is braking hard

So even if the driver can't see it (say it is 12 cars ahead in dense traffic), her vehicle knows and can start to brake much sooner than any driver ever could

Average speeds kept up, road use efficiency up

Similarly, if 4 cars are approaching a crossroads, 1 from each direction, they can tell each other in advance which way they intend to go at the crossroads and, with no need to stop, they can sort it out between them so that all 4 can flow across the crossroads with the minimum of fuss. Maybe that just needs a couple of them to ease the power a touch and that none has to slow sufficiently to work out where the other 3 are going - they all know already.

[RobDickinson]

Thats one aspect, though vehicle to vehicle coms for that kind of stuff has been around for a long time there are a few standards already, and I guess its in use on those volvo road trains for trucks

[ntiz]

Direct line are meant to be one of the best for Tesla insurance. I’m with NFU but they told me they aren’t taking them on anymore at all.

Insurance will be high though for a while I imagine. The model S and X are high because they cost an insane amount to fix plus take a very long time. Time will tell if it is the same for the M3.

[RobDickinson]

Interesting video on the 3 safety, thought it would be fluf but its pretty decent

https://www.youtube.com/watch?v=bkayYiwrjyQ

[Tuna]

JPJPJP said:

Doesn't a big step forward in autonomous vehicles come when, rather than each individual being autonomous, the whole fleet collaborates

e.g. on a motorway when any one car identifies the need to brake hard to avoid a hazard, it knows which other cars are near it and tells them that it is braking hard

It's the 'last mile' problem writ large. It's easy to do the big bits (motorway driving for instance), but just getting out of your driveway without incidence can be difficult. I can't remember the exact statistics, but the majority of car accidents happen within a quarter of a mile of your home(*). It's the small feed roads, odd estate layouts, strange driveways that cause all the problems. Driving on a big straight road at 70mph along with a bunch of people all driving in pretty much exactly the same direction is comparatively easy.

And the idea of safety is difficult. To guarantee you're going to miss the drunk staggering by the side of the road, you need to brake hard - but doing so will guarantee the car behind you is going to rear end you. Humans tend to delay reactions until they're sure something bad will happen. That's fine, because ninety nine times out of a hundred, the drunk will just continue staggering along the pavement. We juggle the risks. But to meet that one in a hundred incident record we expect of FSD, you have ninety-nine cars slamming on their brakes and being rear-ended. And the problem here is that communicating with the car behind you isn't instant... so it brakes a little later, and still rear-ends you.

It seems a much more likely bet that we'll see FSD on heavy goods vehicles, allowing motorway convoy driving long before we have 'consumer' FSD driving you to work.

(*) To solve the problem, I've moved to a different house, half a mile away.

[Witchfinder]

Heres Johnny said:

60k is pushing it but there are 66k cars around the now discontinued pano roof and EAP (which is to all intents and purposes 95% of the current FSD package and much better than the AP they now bundle)

Tesla are now listing 100D/Long Range in-stock cars on the website for ~£67k with unlimited Supercharging. £750/month with a £10k deposit.

Looks like they're really trying to shift that stock.

[Heres Johnny]

Witchfinder said:

Heres Johnny said:

60k is pushing it but there are 66k cars around the now discontinued pano roof and EAP (which is to all intents and purposes 95% of the current FSD package and much better than the AP they now bundle)

Tesla are now listing 100D/Long Range in-stock cars on the website for ~£67k with unlimited Supercharging. £750/month with a £10k deposit.

Looks like they're really trying to shift that stock.

Yep, the prices are tumbling, that 66k is now 63k