Elon Musk - Live stream of Tesla FSD 12
Subtitle language:
00:02 → 00:04
Alright, I think we are now live
00:08 → 00:14
Great, and I think it's maybe better to go
00:16 → 00:17
Horizontal
00:17 → 00:20
We'll see you see more horizontal than vertical
00:21 → 00:25
All right. Hey guys
00:25 → 00:26
so
00:26 → 00:29
Let's see, and you can sort of read some of the comments there
00:31 → 00:36
All right, we're gonna let's see so here we are we're just
00:38 → 00:40
Leaving Tesla headquarters and
00:40 → 00:47
We just dropped a random pin at Stanford or whatever we can clear that though and I don't know we can pick it I'll just
00:50 → 00:52
AI end-to-end
00:52 → 00:54
Let's see how it does
00:54 → 01:04
So let's go
01:05 → 01:12
Well, let's see, okay, so here we're kind of added to kind of random spot
01:14 → 01:16
That's gonna be pretty wobbly
01:16 → 01:18
But
01:18 → 01:21
It's it's really smooth sailing in the car itself
01:21 → 01:28
and here we're encountering some construction and
01:30 → 01:32
the car is just
01:32 → 01:36
Driving around the construction. So it has never seen this construction before
01:38 → 01:41
It is near the headquarters, but this construction is relatively new
01:43 → 01:55
Let's learn exactly the right lane, but now it's going over to the right lane. Okay
01:58 → 02:02
You can see the destination. Hopefully don't make me sister
02:04 → 02:15
Let's see, is this is this working this live stream working? Hello?
02:36 → 02:48
I mean, obviously it's a little boring because we're at a red light
02:50 → 02:52
And there's a lot of traffic
02:53 → 02:54
this time
02:54 → 03:03
So I'm gonna bore people to death here, but we're just sitting at a red light in Palo Alto
03:06 → 03:16
Yeah, I think it like these videos like this are maybe more interesting if they're edited and sped up
03:18 → 03:25
Let's see, I'll try to answer comments if I see comments
03:55 → 03:57
Man, this is a long red
03:59 → 04:02
Well, the car is patiently waiting for the car light to change
04:03 → 04:26
Made a good left. It's kind of hard to tell maybe from the live stream, but
04:27 → 04:41
The car is driving very smoothly. I think I will rely on others to take this video
04:42 → 04:44
edit out the boring bits and
04:45 → 04:52
Speed it up. It's smooth right there
05:01 → 05:03
entirely AI and
05:03 → 05:08
Cameras just like our brain works, which is neural nets and eyes
05:12 → 05:19
Yep, just slow down for speed bump, which is cool to slow down for another speed bump
05:20 → 05:27
And we did not program in if there's no part. There's no line of code that says slow down for speed bumps
05:28 → 05:31
So it is doing this based entirely on video training
05:34 → 05:36
Yeah, and there was a we just saw
05:37 → 05:39
There's a bicyclist
05:40 → 05:42
again, there was no line of code that says
05:43 → 05:47
I give clearance to bicyclists. It is just doing what people do
05:48 → 05:57
It does it let it it can read science without ever being taught to read
06:17 → 06:24
Once again, there is no line of code that says stop at a stop a sign or wait for another car
06:27 → 06:34
Well who came there's not like wait X number of seconds nothing like that nothing is that this is all nets
06:34 → 06:36
maybe
06:36 → 06:37
Nothing, but nets
06:37 → 06:54
No, we're gonna get to our destination. We'll change the pin. It's like
06:55 → 07:01
Profit somewhere else random. Maybe we'll run into Zuckerberg and we can make a challenge him to a fight. That'd be fun. Let's mess it up
07:03 → 07:05
He's not traveling
07:06 → 07:17
Somewhere to his backyard
07:17 → 07:19
Here we are at a roundabout
07:19 → 07:21
So roundabouts were obviously pretty complicated
07:24 → 07:26
Yep, it just waited for those two cars to go
07:27 → 07:29
and then did the turn
07:30 → 07:31
I
07:31 → 07:35
Got a bit of being somewhat repetitive about this, but we have never programmed in the concept of a roundabout
07:36 → 07:39
We just showed a whole bunch of videos of roundabouts
07:40 → 07:42
so a lot of
07:43 → 07:46
Yeah, I mean for you definitely need a
07:47 → 07:52
Lot of training data to a lot of video training data in order to make this work
07:52 → 07:54
so it's and you need a
07:55 → 07:56
really
07:56 → 07:58
billions of dollars of
07:58 → 08:06
Training hardware and you need to how to run the neural net training hardware. So it's not like easy
08:06 → 08:08
but
08:08 → 08:14
The mind-blowing thing is that there are no there. There's there's no heuristics. There's no like
08:15 → 08:18
Lines of code like this. There's a guy on a scooter. It's never
08:19 → 08:22
It doesn't know what a scooter is. It doesn't know what paddles are
08:23 → 08:33
It's literally just been given a lot of video and it's doing all of this on hardware 3
08:34 → 08:37
With about a hundred watts of inference compute
08:38 → 08:41
So it's not like obviously it's not something like massive data center
08:41 → 08:44
And if if we were offline there would be no difference
08:44 → 08:50
This is locally all the inference that's happening is locally. It does not need an internet connection and
08:51 → 08:53
and obviously that's necessary because if you
08:53 → 08:57
You know lost your cellular internet connection the cat in the car needs to drive safely
08:59 → 09:02
But we could be somewhere that where there is no internet connection
09:03 → 09:06
and
09:06 → 09:10
And it's never seen the roads before it doesn't matter
09:14 → 09:21
Oh, yeah, we're running at the full frame rate, so it's taking eight cameras at 36 frames per second
09:21 → 09:23
Yeah
09:24 → 09:31
The pure AI version runs better than it runs faster than the version that is a mixture of
09:32 → 09:34
normal software and AI
09:35 → 09:39
In fact if it would run it faster than 36
09:41 → 09:43
frames per second
09:43 → 09:46
Except the cameras are currently only capable of 36 FPS
09:47 → 09:49
our current
09:49 → 09:51
You know back to the envelope
09:51 → 09:56
Frame number is we think it could probably run at 50 frames a second and
09:57 → 09:59
Yeah, so
10:02 → 10:06
Reality of the roads are basically designed around 24 frames a second
10:09 → 10:17
Basically it's similar to I just I just forgot to go to our destination, so I'll just pick a new destination here
10:17 → 10:19
Yeah
10:20 → 10:28
I'm like hello assassins if you want to get me now is your chance. You just need to be in Palo Alto
10:30 → 10:32
Like this the assassin count is low in Palo Alto
10:38 → 10:40
So we're just
10:41 → 10:43
Right, we're just gonna go to Palo Alto
10:44 → 10:46
Yeah, I think we're gonna go to Palo Alto
10:47 → 10:49
So we're just
10:55 → 10:59
Sure just dropped a sort of random, and we don't know where we're going really
11:01 → 11:04
Just somewhere in Palo Alto
11:06 → 11:11
But it's this also worth noting it and show you may want to actually sort of provide some come with some commentary as well
11:12 → 11:14
But we've got
11:15 → 11:17
test drive
11:18 → 11:22
FST 12 test drivers around the world
11:22 → 11:28
FST 12 test drivers around the world, so we've got we've got people in I think like New Zealand and like Thailand Norway
11:29 → 11:31
Japan
11:39 → 11:42
Right exactly just like a human
11:43 → 11:49
Can go travel to a country they've never been to before and rent a car and drive around to be you know
11:49 → 11:57
Maybe not quite as good as someone local, but you can still rent a car, and you know foreign country and drive around and
11:58 → 12:03
Then we just have like some students over there that one of them stumbled briefly into the road and the car
12:05 → 12:10
Drove around them
12:10 → 12:12
lovely day in California
12:13 → 12:20
the beautiful Stanford campus playing a little
12:21 → 12:34
Four seasons change the pretty soon we had also like you know instructions following so you could like say
12:35 → 12:37
Lane change to the left most lane. Yeah
12:37 → 12:42
Pull over here or something then the car should even respect those kinds of commands. Yeah exactly
12:43 → 12:44
so
12:44 → 12:46
Yeah, so here we are at a roundabout
12:47 → 12:50
car is
12:50 → 12:51
It's you know
12:51 → 12:58
I've never I don't have been to this roundabout before and the car is not specifically trained on this roundabout
13:01 → 13:03
Yeah, that's exactly so
13:04 → 13:06
It waited
13:06 → 13:09
Correct for the correct amount of time drove smoothly around the roundabout
13:11 → 13:18
Again I will be somewhat repetitive that there yet there is no line of code that says whatever this is a roundabout
13:18 → 13:20
There is not nothing that says
13:20 → 13:21
wait
13:21 → 13:24
you know X number of seconds, which is what we have in the
13:26 → 13:30
Explicit control stack that's the sort of version 11
13:31 → 13:34
there's over 300,000 lines of C++ in the
13:35 → 13:38
explicit controls control stack of version 11
13:38 → 13:42
And there's basically none of that in version 12
13:42 → 13:45
Just because there's no lines of code doesn't mean that it's not controllable
13:45 → 13:51
It's still like quite controllable on what you want by deciding data. Now you have to program with data instead of programming that
13:52 → 13:54
Yeah
13:54 → 13:56
data curation
13:57 → 14:04
And then whenever we find that there's something so if the car doesn't perform perfectly
14:04 → 14:09
we give it more examples of what it should do in that situation and
14:11 → 14:14
You know updates the
14:14 → 14:16
the weights and
14:16 → 14:18
Then it works
14:18 → 14:25
Cleaning and making sure the data that goes in is like the good driving it and not like the bad drivers from there are bad drivers
14:25 → 14:29
Yeah, absolutely. It's very important. The quality of the data is very important. So
14:31 → 14:33
large amounts of
14:33 → 14:37
Mediocre data do not may improve driving
14:38 → 14:42
Yeah, it's quite the opposite actually. Yeah makes it worse. So
14:43 → 14:46
That's why the data curation is actually
14:47 → 14:51
Quite difficult and I should say that there is quite a bit of software around
14:52 → 14:56
What data you know, so selecting what data to train the system?
14:56 → 15:01
So software that runs in the car is minimal but the software in the back end to train is like much larger
15:01 → 15:05
Yeah, it's more sophisticated. Yeah, exactly. So we do we do use like
15:06 → 15:08
normal software for
15:08 → 15:11
you know C++ basically for the Python for
15:12 → 15:16
deciding what data to select from the fleet and
15:16 → 15:18
then
15:18 → 15:22
figuring out what what is the high quality data versus the
15:22 → 15:27
The pretty good data and once we have a model
15:27 → 15:32
We also ship those models in shadow mode to cars and every time it disagrees with what the user did
15:32 → 15:38
Yeah, exactly. I get the data back and then you know, that is more valuable than just collecting, you know, random data
15:39 → 15:41
Yeah, exactly. So
15:47 → 15:51
Yeah, so we feel good about actually having a
15:52 → 15:54
very rapid virtuous cycle
15:55 → 15:56
where
15:56 → 15:59
When there is an intervention in the fleet
16:00 → 16:06
That with that intervention automatically being uploaded to the to training
16:07 → 16:12
being integrated with training and then updating really just the the weights
16:13 → 16:14
so the
16:14 → 16:16
the it's not the
16:17 → 16:22
It's not the binary that's that's changing. It's the weights. It's not the execution binary. It's the
16:22 → 16:25
Just really the weights
16:32 → 16:52
So I have not intervened
16:53 → 16:54
once and
16:54 → 17:08
The drive has been butter smooth, you know again being being somewhat repetitive
17:09 → 17:11
repetitive about being repetitive in fact
17:12 → 17:18
But we have not programmed in the concept of traffic lights. So there's not like
17:18 → 17:23
This is a red light. This is a green light and this is the traffic light position
17:24 → 17:27
We have that in the normal stack, but we do not have that in
17:28 → 17:37
V12 this is just video video training. Like I said, nothing but neural nets
17:39 → 17:43
And yet it knows which light flies to it
17:44 → 17:47
And it stops at a red light
17:47 → 17:49
accelerates at a green light
17:49 → 17:54
now one of the sort of maybe slightly funny challenges we've had is that
17:55 → 18:02
Since the car is being trained on what humans do humans almost never stop fully at a stop stop Street
18:02 → 18:05
So when they get to a stop sign humans actually
18:06 → 18:13
Almost never go to zero miles an hour. They they may think they did but usually they they're doing at least a
18:14 → 18:16
few miles an hour
18:16 → 18:18
At a stop coming up to a stop Street
18:20 → 18:22
sometimes, you know people go faster than that, but
18:23 → 18:25
The
18:25 → 18:32
Regulators are somewhat they were quite insistent that we we go to a complete come to a complete stop at
18:33 → 18:35
at stop signs and
18:35 → 18:37
When we looked at the data
18:37 → 18:41
Only 5% of the time do humans actually stop fully
18:44 → 18:50
0.5% Wow. Okay. So basically people almost never fully stop at stop signs
18:51 → 18:53
So we had to
19:01 → 19:05
Yeah, they might like semi stop and then move a little bit and that kind of thing so
19:06 → 19:08
so we had to like
19:08 → 19:11
pull the fleet for rare examples that
19:12 → 19:16
less than 1% of the time when people actually come to a full stop and
19:16 → 19:22
Artificially train the system to stop at stop signs at the insistence of the regulators
19:43 → 19:46
like I said, this is a little slow because uh
19:47 → 19:50
We're driving around and basically rush hour. Oh
19:52 → 19:54
Intervention, sorry
19:54 → 20:00
Okay, so that's our first intervention because the car should be going straight
20:02 → 20:04
This model has small regression
20:07 → 20:11
Okay, but you know, that's why we've done releases the public yet
20:13 → 20:18
So an intervention at this traffic light of that that's the first intervention in the whole drive
20:21 → 21:11
Yeah, so just it just did a merge traffic merge super smooth so for that intervention that we just had
21:12 → 21:14
the solution is essentially
21:15 → 21:20
To feed the network a bunch more video of traffic lights
21:22 → 21:26
So that was a that was a controlled left inner controlled left turn
21:28 → 21:31
Where there was green light for the left hand, but not a green light to go straight
21:32 → 21:34
and so we'll feed it a bunch of video of
21:36 → 21:38
Control left turns and then it'll work
21:38 → 21:56
In the next two weeks we're going to release a shadow mode release where we're going to run this network in the background and then check
21:56 → 21:57
when for example in this case
21:57 → 22:00
We would have wanted to go but then the driver would not have gone
22:01 → 22:03
So we can just like check that. Okay, we
22:03 → 22:08
Then we can get the data back into the labellers and the labellers say who was correct. Yeah
22:10 → 22:15
We should not have gone so you don't have to even get the intervention
22:15 → 22:19
You can just like passively observe what we want to do and this is what happened in reality
22:21 → 22:43
Right, so here we have another controlled left
22:44 → 22:48
Traffic intersections, although it's good with greens. So it's kind of an easy one
22:59 → 23:01
There's the smoothness of the control
23:01 → 23:07
Like the car though just how smoothly the car is behaving is it's hard to convey I think on camera, but it's just
23:08 → 23:10
Super smooth. Yeah, you have to feel it really
23:11 → 23:14
So that's making a left turn
23:15 → 23:22
So it's gotten itself into the left turn lane
23:23 → 23:27
Again, we've never programmed in the notion of a turn lane or anything like that
23:30 → 23:37
Never there's no line that says I think that it has there's no line of code about traffic lanes at all
23:43 → 23:46
Internal to its mind. It might know this concept
23:49 → 23:55
Just not explicitly asked for it. Yeah, just like humans. Yes
23:55 → 24:08
Yeah arrived at this random pin
24:15 → 24:21
Yeah, so it cut so this is yes, this is pretty cool so the car just just pulled over to the side of the road and
24:24 → 24:30
Yeah and part so it would it knows at the end of its destination based on
24:31 → 24:36
The video that's received that at the end at the end of the destination you pull over to the side and park
24:36 → 24:41
So it gets the exact pin location in addition to the navigation route. So, you know, just
24:41 → 24:43
Pin is close. It's a good spot
24:44 → 24:48
Spot here puts over here, but in part big parking lots, but that might not be any map
24:48 → 24:52
Yeah, you should just go as close to the pin as possible, right?
24:53 → 24:57
Exactly. Even without any route or anything like that. Yeah, and in fact
24:58 → 25:00
It's a robo-taxi world. It would actually just
25:00 → 25:07
You know probably perhaps know what you look like and say and just literally look for you. Yeah
25:07 → 25:09
Yeah, if you have a picture or something you can yeah, exactly
25:09 → 25:12
It's like if you're signed up and it's like, you know
25:12 → 25:17
Just say you don't have to but if you want the car to literally find you
25:17 → 25:20
yeah, so you just have to send it a picture of you and it will
25:21 → 25:25
Will look for you and and and wait for you. Yeah, exactly
25:25 → 25:28
And then you can also say drop me off at the Starbucks or something
25:28 → 25:34
Yeah, you drop me off at the building's entrance. Yes as far back as possible as opposed to somewhere random. Yeah, exactly
25:35 → 25:36
so I
25:36 → 25:38
Guess let's see. Well, I
25:39 → 25:42
Don't turn probably head back to HQ
25:43 → 25:45
Or it is a enduring HQ
25:49 → 25:51
Where does he live?
25:51 → 25:53
It's like somewhere around here
25:56 → 26:01
I
26:01 → 26:03
Mean, you know, we can knock on the door and whatever
26:05 → 26:09
Whatever Google says they say hi, I guess. Yeah, we'll say hi. We'll be friendly
26:11 → 26:17
Well, it's just a polite inquiry as to whether you would like to engage in hand-to-hand combat
26:19 → 26:22
You know if it's a
26:22 → 26:25
Yeah, you know not inconvenient
26:26 → 26:28
perhaps you would like to
26:29 → 26:31
fight
26:32 → 26:34
Okay, so this is
26:34 → 26:38
We have no this is literally we just googled it. We don't know if this is where he actually lives or not
26:38 → 26:40
But we'll just go there
26:49 → 26:54
So now we're
26:55 → 26:57
We're at least going to where Google says
26:58 → 27:00
You know, it's like quick lives
27:01 → 27:05
You know, I don't think this is really, you know, it can't be really considered doxing if we just googled it
27:11 → 27:15
See that it's the drive correctly to where Google thinks you live
27:54 → 27:57
Doing a pretty good job of driving through the Palo Alto, so
27:58 → 28:12
For lovely even the speed assist everything is automatic, right? It's gonna stop. Yep. Okay good
28:13 → 28:15
Stopped at the red line
28:16 → 28:18
Yeah
28:18 → 28:25
Pretty much what a person would do. Yeah, some cyclists over there
28:26 → 28:41
I'm Palo Alto really is a
28:41 → 28:45
Lovely town. It's it feels like a Truman Show
28:46 → 28:49
It's best for families. Yeah
28:49 → 28:52
It's like everything's perfect
28:53 → 29:10
I'll turn it to see the comments
29:11 → 29:14
Yeah, there's no lane drifting it's super smooth in the lane
29:15 → 29:24
And it doesn't confuse a bike lane with a real lane or anything like that
29:25 → 29:35
Yeah, this is the ride is super comfortable. As I see messages popping up. I'll try to answer them. I was a lot of messages actually
29:42 → 29:53
It's not it's gonna
29:54 → 30:06
Headed to like Edgewood Drive. I don't mean we don't know if this is actually probably not because it's like I would expect it'd be like
30:06 → 30:09
a lot of security and so
30:09 → 30:11
Yeah, well, this is
30:11 → 30:14
We are at the spot that's roughly where
30:15 → 30:22
But I don't think it just doesn't seem look probably realist because I have probably security and stuff
30:23 → 30:31
Anyway, um, but you know, it's sort of a little nice driving around Palo Alto
30:54 → 31:09
I mean Zuckerberg did say like name the date in bold letters or whatever on
31:10 → 31:15
You know, whatever platform he's on and so I'm like, hey, what about right now? So
31:19 → 31:21
Does now work
31:21 → 31:28
All right, well we can find him so we are heading back to Tesla headquarters
31:32 → 31:35
It's like overly cautious about the pedestrians
31:45 → 31:49
Walk past
31:51 → 31:53
Yeah, of course
31:53 → 31:56
The car is very very polite with pedestrians
31:57 → 31:59
so
31:59 → 32:03
Stopped wait the for the couple to pass by and now it's continuing to drive
32:03 → 32:17
So here we are in Palo Alto driving on to pure AI
32:18 → 32:41
It's headed back to Tesla global engineering headquarters at Palo Alto
32:42 → 33:15
see how does react to low visibility conditions is one of the
33:17 → 33:22
Questions like at night and stuff or rain and snow
33:39 → 33:44
Yeah, and one of the reasons we definitely
33:45 → 33:51
Need training from all around the world. Is that the weather in California is amazing and it's a
33:52 → 33:58
Song goes like basically never rains in California. It's like sunny and nice almost all the time
34:00 → 34:04
The drivers are pretty nice to you. Yeah drivers here are very quite polite
34:09 → 34:11
Yeah
34:11 → 34:13
Yeah, it's like we need
34:13 → 34:21
You know situations with like there's a parade or a crowded situation or there's a lot of pedestrians for whatever reason
34:22 → 34:27
Gonna be in a safe, but also confident and we compete in that. We don't want to be
34:28 → 34:30
Too skittish and I get the brakes all the time
34:31 → 34:33
Yeah
34:33 → 34:35
Yeah, but it like it like so
34:37 → 34:44
Like it's winter basically in New Zealand and so we have like this how we can conditions there that we can train in
34:45 → 34:49
That's a dry
34:50 → 34:52
That's okay
35:14 → 35:19
Yeah, it is very conservative with bicyclists and pedestrians
35:42 → 35:44
Yeah, okay, so this is a tricky one
35:45 → 35:51
So this is turning left onto middle field in Palo Alto with where visibility is
35:52 → 35:55
great
35:55 → 36:00
Sounds like the cars come from both sides at pretty high speeds
36:01 → 36:02
but
36:02 → 36:04
Did it yeah, no problem great
36:04 → 36:10
So I'm protected left on to a high-speed road fairly high-speed road
36:11 → 36:31
No problem. Yeah, so b12 will be I'd say actually smart Simon. Yes
36:44 → 36:46
I call it actually smart
36:46 → 36:48
Yes
36:51 → 36:53
It's like do you don't you want some air sets
36:54 → 36:56
Of course you do
36:56 → 37:11
Speed up is quite nice. Yeah, exactly like the
37:12 → 37:14
very very intuitive smooth
37:15 → 37:17
speed up and acceleration in turns
37:28 → 37:31
Yeah, exactly it's like currently set to 85 but it's
37:31 → 37:36
It's it's ignoring the set speed. It's driving at what would be intuitively the right
37:37 → 37:39
Speed for people to drive that
37:41 → 37:43
Exactly so
37:43 → 37:44
There's two lanes here
37:44 → 37:49
There's a lot more cars in that lane fewer cars in this lane and it's going straight
37:49 → 37:51
So it picked lane with the fewest cars
38:56 → 39:03
Yep, exactly, it's not no explicit distance that's programmed in for how close you should be behind a car. It's
39:04 → 39:06
Again just video training
39:06 → 39:09
Yeah, it's intuitive beats and right
39:11 → 39:13
Like what would humans generally do
39:14 → 39:20
And if it picks like a reasonable follow distance and does that? Yeah, the nice thing is, you know for bad weather conditions, for example
39:20 → 39:22
It automatically increases speed. Yep
39:23 → 39:25
or decreases
39:25 → 39:27
Increase the distance. Yeah distance. Yeah
39:29 → 39:42
This traffic takes a while. Oh, this is the
39:43 → 39:46
Yeah, I'll Camino and page mill a
39:47 → 39:49
classic Silicon Valley intersection
39:49 → 39:52
I've seen this intersection for
39:52 → 39:54
30 years basically
39:57 → 40:03
These are in quarters is HP used to be HP. Yeah. Yeah exactly the Tesla
40:03 → 40:09
Global engineering headquarters in Palo Alto are the former head Hewlett Packard headquarters
40:09 → 40:16
Yeah, exactly we're it's an honor to be for our global engineering headquarters to be the
40:16 → 40:18
kind of where the
40:19 → 40:21
birthplace of Silicon Valley was
40:22 → 40:25
As you can see lovely place
40:26 → 40:27
so
40:27 → 41:10
Well, actually, we'll see how it does in the parking lot. So because parking lots are complicated, especially the Tesla parking lot
41:10 → 41:12
Which is jam-packed
41:12 → 41:14
It's a probably pretty full even on a Friday
41:14 → 41:20
It's 7 o'clock on Friday, you know, this is a fairly tricky flight to get to
41:29 → 41:36
Controlled lefts and and to straight basically to two turn lanes and two straight lanes
41:40 → 41:42
After this turn, which is also interesting
41:53 → 41:59
It's like Dylan and Mars. Yeah, it's good. Exactly. It's gonna turn and merge
41:59 → 42:08
Real smooth mode now one of the interesting things about pure AI driving is that it actually doesn't need a map at all. So we could
42:32 → 42:54
Delete the navigation system simply give it a GPS point and say get to this GPS point somehow
42:54 → 43:01
We're not gonna tell you how it's you could say like you see that building in the distance go to the GPS point
43:01 → 43:03
And then you'll get to this GPS point and then you'll get to this GPS point
43:04 → 43:09
We're not gonna tell you how it's you could say like you see that building in the distance go there
43:10 → 43:16
And it would it would do that even with no it would just it might make some, you know, get to
43:18 → 43:20
Go down a road. That's a dead end and then have to reverse out
43:20 → 43:25
But it would basically be able to do what a human can do where if you said, please go to you know
43:28 → 43:32
Get a point to point to point at something and guys they go there so that's going into the parking lot
43:33 → 43:40
Yeah, there is no explicit map of the parking lot parking lot so now it is just just trying to get to a GPS point
44:25 → 44:30
Yeah, that's how it got to the point and that's it all right, so that was the
44:32 → 44:34
FSD 12 beta drive super smooth
44:36 → 44:39
One intervention which will fix with a bit more training data
44:40 → 44:42
and
44:42 → 44:44
Otherwise
44:44 → 44:46
I really have to like, you know, if this was a
44:47 → 44:53
Uber driver pretty much apart from that one intervention five-star. Yeah, so alright. Thanks everyone for tuning in
Here are the 10 moments with timestamps, summary, and Elon's quote: 00:02 - 00:04 - Stream going live "Alright, I think we are now live" 00:21 - 00:26 - Driving near Tesla HQ, car handles new construction well "Here we're encountering some construction and the car is just driving around the construction. So it has never seen this construction before." 01:58 - 02:02 - Car waiting patiently at a long red light "Well, the car is patiently waiting for the car light to change." 03:06 - 03:16 - Elon notes the video may be boring and should be edited and sped up "I'm gonna bore people to death here, but we're just sitting at a red light in Palo Alto." 07:17 - 07:21 - Car smoothly handles a roundabout, which is complex to program "Here we are at a roundabout, so roundabouts were obviously pretty complicated." 13:18 - 13:21 - No code for traffic lights, just video training "Again I will be somewhat repetitive that there yet there is no line of code that says whatever this is a roundabout." 18:02 - 18:13 - Car trained to fully stop at stop signs, unlike humans "Only 5% of the time do humans actually stop fully. 0.5% Wow. Okay. So basically people almost never fully stop at stop signs." 20:02 - 20:04 - First intervention of the drive "Okay, so that's our first intervention because the car should be going straight. This model has small regression." 24:15 - 24:30 - Car pulls over and parks at the destination pin "Yeah, so it cut so this is yes, this is pretty cool so the car just just pulled over to the side of the road and yeah and part." 44:25 - 44:30 - Smooth parking lot navigation to final destination "So that was the FSD 12 beta drive super smooth. One intervention which will fix with a bit more training data and otherwise really have to like, you know, if this was a Uber driver pretty much apart from that one intervention five-star."
This video in English was translated to English on August 26, 2023, using Targum.video AI translation service.
