Elon: FSD Beta tweets

[willow_hiller]

I always keep my foot hovering over the brake.

May be you like living dangerously, but definitely I’d not advice anyone to keep pressing the accelerator with FSD on …

I don't know anyone that keeps AP on all the time, and at the same time, it's very useful in certain situations. You learn where it works best, and use it there.

If people find FSD Beta too cautious around pedestrians, they should hit the video button, disable it, and drive as they like around them.

 

[AlanSubie4Life]

May be you like living dangerously, but definitely I’d not advice anyone to keep pressing the accelerator with FSD on …

I definitely don’t like to live dangerously; I want to be as safe as possible when driving.

I’ll need to experiment more with more situations, but in general accelerator application seems to lower the threshold for warnings. So there is a lot of beeping even for things that are not hazards (but not so common that you ignore it!).

I find that this technique tends to lend itself to a more natural human-like driving pattern, which means it is easier for drivers around to predict what you might do.

So I believe overall it is safer, if a bit beepy.

You don’t necessarily have to have the accelerator pressed all the time, but it’s good to be in the habit, for rapid reaction to (or complete prevention of) sudden slowing or jerky behavior.

Tapering accelerator application at a stop while preventing jerking is quite difficult though. Takes practice and exquisite “throttle” control. But often smoother than a natural FSD stop.

I always keep my foot hovering over the brake.

Obviously when a hazard is identified this is the default behavior. But note in the FSD state the behavior differs from manual control.

In manual control, I usually use the accelerator to brake. So I never cover the brake unless I have completely released the accelerator. So many potential hazards still mean a foot over the accelerator.

But in FSD mode when moving to cover the brake, or even partially releasing the accelerator, the slowing will often not be as much as expected. So it is important to actually press the brake in these cases. It’s a tricky interface issue. I feel that keeping behaviors as close to manual control is better for the takeover transition.

 

[willow_hiller]

Realistically speaking, how long would it take Tesla to record billions of miles of data?

https://twitter.com/x/status/1569140947396575234

40 million miles have taken them about 2 years. At that rate, they'll achieve a billion in about 50 years. They better start rolling out FSD wider!

 

[Knightshade]

View attachment 851509

I don't understand this, maybe because I don't understand NN. How can there be a "bug" of that specificity in a CNN inference engine? Maybe he means a detection response bug in the planner?

Last I knew a fair bit of FSD code was still traditional C++ and not purely NNs (though more if it is NNs now than when it NNs were ONLY used for perception)

 

[Ramphex]

Realistically speaking, how long would it take Tesla to record billions of miles of data?

https://twitter.com/x/status/1569140947396575234

40 million miles have taken them about 2 years. At that rate, they'll achieve a billion in about 50 years. They better start rolling out FSD wider!

What’s the multiplier though? 40 million miles from 1000 cars in 2 years? 25,000 cars should give you 1,000,000,000 in 2 years.

 

[thewishmaster]

It is a bug if it panic stops for someone who is clearly NOT going to walk in front of the car. Mine did this on 10.12 a few weeks ago. The J-walker was heading to walk behind my car and even shakes his head at my "ignorance" for slamming on the brakes for him.

Best is when it panic stops when maintaining the same velocity would’ve actually passed the perceived danger before the person can even become one… Sure I guess if they sprinted immediately they might get hit, but at our current velocities there’s no way they would even reach the road in time… (that was the case on my first 69.2 drive today)

 

[Bladerskb]

Realistically speaking, how long would it take Tesla to record billions of miles of data?

https://twitter.com/x/status/1569140947396575234

40 million miles have taken them about 2 years. At that rate, they'll achieve a billion in about 50 years. They better start rolling out FSD wider!

"billions of miles of real-world training"

The guy is spreading BS again

 

[Mardak]

40 million miles have taken them about 2 years

Assuming FSD Beta population has been around 100k vehicles since June, potentially Tesla is actually closer to 55M cumulative miles increasing at roughly 7.5M miles per month.

But yes, with the current beta population, it would probably take over a decade to reach 1B cumulative miles. Could Tesla 10x the FSD Beta population to reach that 1B milestone in say 1 year instead?

Additionally, would releasing single stack highway driving "count" for these cumulative miles? If so, if we say highway driving and city driving are close to equal (not that far from EPA using 55% city and 45% highway for MPG), FSD Beta 11 would potentially cut the time needed to reach 1B in half or even more if people tend to use it for more miles on highway than city.

Of course, this all depends on how capable FSD Beta is to have people feel comfortable in turning it on, e.g., "spouse approved," so that's another aspect affecting the rate of how cumulative miles driven with FSD Beta will grow.

 

[momo3605]

Assuming FSD Beta population has been around 100k vehicles since June, potentially Tesla is actually closer to 55M cumulative miles increasing at roughly 7.5M miles per month.

View attachment 851584

But yes, with the current beta population, it would probably take over a decade to reach 1B cumulative miles. Could Tesla 10x the FSD Beta population to reach that 1B milestone in say 1 year instead?

Additionally, would releasing single stack highway driving "count" for these cumulative miles? If so, if we say highway driving and city driving are close to equal (not that far from EPA using 55% city and 45% highway for MPG), FSD Beta 11 would potentially cut the time needed to reach 1B in half or even more if people tend to use it for more miles on highway than city.

Of course, this all depends on how capable FSD Beta is to have people feel comfortable in turning it on, e.g., "spouse approved," so that's another aspect affecting the rate of how cumulative miles driven with FSD Beta will grow.

Highway driving wouldn’t count imo. I assume billions of miles of software validation and training to improve city driving means you need data from city driving. You’re not going to have anywhere near the same scenarios on the freeways.

But anyway doesn’t matter, he’s just spouting nonsense. Driving a billion miles isn’t magically going to make FSD “superhuman”. No one knows what it’ll take at this time. Better hardware and more AI breakthroughs over the next many years/decades will be part of it I’m sure.

 

[AlanSubie4Life]

https://twitter.com/x/status/1569178980703502337

 

[Supcom]

https://twitter.com/x/status/1569178980703502337

Well, at least he was specific as to who will be getting this blast of updates. No ambiguous 'wide release'.

Rollout to all current beta testers imminent!

 

[MarcG]

Well, at least he was specific as to who will be getting this blast of updates. No ambiguous 'wide release'.

Rollout to all current beta testers imminent!

Yeah, got mine around 1pm today (had 10.69.1.1 since Monday). Haven’t driven it yet so let’s see tomorrow how it does.

 

[tpatana]

Why not me yet Still on old version

 

[DarkForest]

Why not me yet Still on old version

My download became available literally one minute after your comment. Did you get the download after all?

 

[nvx1977]

Totally forgot about FSDb. Went into the car for the first drive of the day and there was a software update available. Waited until I got back home before starting it (long-time rule of mine never to accept an update while not at home, just in case it borks and I'm stranded for no good reason).

Now I need an excuse to drive somewhere.

 

[2101Guy]

What’s the multiplier though? 40 million miles from 1000 cars in 2 years? 25,000 cars should give you 1,000,000,000 in 2 years.

But…but…Dojo..

 

[willow_hiller]

"billions of miles of real-world training"

The guy is spreading BS again

You missed the boat to complain about this goal about 6 years ago. It's been a clear part of Elon's Master Plan part 2 for a long while:

We expect that worldwide regulatory approval will require something on the order of 6 billion miles (10 billion km).

And I wasn't saying it wasn't feasible, just that they need to pick up the pace. As others have pointed out, 40-50 million miles driven with a base of 100k testers could be done 10 times as fast with 1 million testers.

 

[Knightshade]

But…but…Dojo..

..are you under the mistaken impression Dojo is a thing going in cars, since you're mentioning it in relation to a post about miles of car-collected data?

 

[EVNow]

Obviously when a hazard is identified this is the default behavior. But note in the FSD state the behavior differs from manual control.

The main reason to keep the foot on the brake is to be able to disengage quickly - additionally it will slow down the car when it is trying to hit another car or VR or does anything wrong. Since we need to react lot more quickly in all these cases, I keep my foot over the brake.

Need to press accelerator are rarer (in my case) and also don't have to react that quickly.

 

[EVNow]

The guy is spreading BS again

This is exactly what I say about you all the time !