Speak EV - Electric Car Forums banner
1 - 11 of 11 Posts

·
Registered
Joined
·
118 Posts
Discussion Starter · #1 ·
I know there have been many discussions on the subject but I thought I would add mine. I have said on many occasions that FSD is many years away from working. Having now watched the latest beta testers in the USA and listen to Elon and friends on AI day I am even more convinced. The Beta currently being tested is good but is only level 3 at best. To move to the next step is huge. At AI day Elon said Hardware 4 will be required so currently no car is equipped with the right hardware. I am also a little confused about the Neural Net but 99% of AI went over my head! I understand the complexities and size of the Neural Net required for self driving but what I don't understand is how that is going to be implemented in cars. There is no way the Neural net can be installed in each car and you cannot rely on communication between the Neural Net and each car so how do you teach each car every single rule?
 

·
Registered
Joined
·
2,746 Posts
It's complicated but there are some simple principles which help:

The car has many neural nets (NNs) running all the time. This is the "hydra" architecture that Andrej discussed. Imagine each "head" of the hydra has a different speciality, like recognising pedestrians, cars, figuring out drivable space, is it raining, ⁹etc.

The in-car NNs are playback only. They don't learn.

The "training" is done in-house. Dojo will be used for this soon.

Driving policy is traditional C++ code at the moment
 

·
Registered
Joined
·
204 Posts
I suspect that FSD will be like nuclear fusion. It will always be 10 years away from fully working. I also suspect that the biggest problem that it will have is road users who don't use it and that includes people who are driving without it, horse riders, cyclists, pedestrians and free roaming animals in National Parks.

All of those groups have 'grandfather rights' so FSD has to cope with all of those without increasing the risk of injury or death.
 

·
Registered
Joined
·
158 Posts
FSD will never be finished as stated on AI Day but their aim is to keep improving compared to an "average" driver. I'm always a bit sceptical simply because I drive most days and if you actually break down what you are dealing with on a daily basis WOW is it a complicated task that changes everyday even doing the same route.

May have to be introduced on more orderly road networks to start with and I wouldn't look at it as a defeat if it was geo fenced to start with. If only to demonstrate competence to legislators.

From what I remember from AI day the FSD 1/HW3 inference hardware is sufficient for say 300% improvement compared to that elusive average driver and maybe FSD 2/HW4 would be required to continue the march of nines.

For me I'd only need the Enhanced Autopilot and not the full FSD so I'd be tempted to add that except it can't be transfer it so I'm trying to put off a purchase until my ideal Tesla comes along be it the Compact, Cyberxxxx or Model Y. Model Y wouldn't be ideal but it would likely keep it's value better than another Model 3.

I just bought my M3 outright as I didn't have the option of the Model Y and nothing else comes close IMHO.
 

·
Registered
Joined
·
3,017 Posts
In terms of level 2 autonomy, Tesla pretty much only has an advantage in the US now, where FSD is at least useable.

It seems that both the regulatory requirements and manpower required to make FSD possible in Europe have been a real blocker. We never even got properly functional Nav on AP.

Now that every 20k+ car on the market has an AP equivalent for motorways, and most premium cars offer something akin to EAP (stereo cameras, lane change), I don’t personally see a reason to choose Tesla over other brands for driver assistance tech in Europe.

The crazy thing is that for 95% of the driving we do, regular AP (or LFA&ACC) is just fine - at a basic level, it’s just a matter of staying in lane and not hitting the guy in front.

But that last 5% - tight roads with parked cars, junctions, roundabouts, school zones etc - requires a level of processing power and decision making that computers just don’t seem to have yet.
 

·
Premium Member
2014 Model S
Joined
·
8,790 Posts
Watching the Waymo car get totally bamboozled by a cone and you realise self driving cars are just not going to happen any time soon.

If a car is going to be called full self driving then it needs to be able to drive around a test route like a learner has to pass their test. Any intervention from the examiner is an immediate fail for that vehicle specification. They should be held to the same standard as human drivers are. Then after that they'd be expected to improve with practice so they'd still have to be able to drive safely without making serious errors.

If an AI is downloaded to each car then it's possible for that AI to diverge over time with other ones. You'd need to find away to merge all the learnt information so that all of them benefit from different driving environments.
 

·
Registered
2021 Tesla Model 3 LR
Joined
·
1,046 Posts
Watching cars, lorries, tiny horses and cones appear, move and disappear from the FSD visualisation in my car has made it plain to me that there won’t be genuine FSD in my lifetime. We’re a long way off, and that’s alright with me.

What I’m pleased about is that Tesla don’t set assistance on by default every time you switch on, unlike some manufacturers. Long may that continue. It’s so irritating having to turn things off before you go anywhere.
 

·
Registered
Joined
·
2,746 Posts
Because the Net presented at AI day was enormous and probably ran into millions of dollars.
Eh?

What was presented at AI day was a system for training NNs and the end result, which is a bunch of NNs which run in the car.

Because they don't "learn" independently the NNs running in the car will never diverge.
 

·
Registered
Joined
·
3,017 Posts
Yeah the models are downloaded regularly. They don’t change when on a car. The fleet sends data back, the network at Tesla is trained on these, then an updated model is shipped.

Training a model takes days and days of flat out processing. It will never be done on individual cars, and there’s no benefit to doing so.
 
1 - 11 of 11 Posts
Top