Andrej sounds totally normal if you set playback speed to 0.75x... His brain is running a third faster than most people! NeuraLink is going to be interesting when communication is no longer the current slow speed speach...
The video has been sped up, apparently by around 1.25x. Unfortunately needs 0.8x on the playback to make it correct but that's not available. 0.75x makes the playback a little bit too slow. So you either get too slow (0.75x playback) or too fast (1.0x playback). Not sure why the person who posted this video sped it up like that.
That image of the fleet data on a map at 12:24 is absolutely stunning. I'm from Cyprus, and even though we don't have Tesla showrooms, superchargers or service centres I can see that Tesla is still collecting data from the cars which were imported here.
There are some pretty big high-level takeaways here: using pure vision with a slight hint of radar, for now, to get lidar like accuracy in terms of depth perception. This is major.
What a great video on AI and neural networks. Amazing future. Tesla is well positioned to remain in the lead with autonomous vehicles because they have so many cars in the field gathering data and teaching its neural network. No one else is in the game.
Awesome presentation. Question: when he says '"fleet" that means all the autopilot enables Tesla cars currently owned by public, or a dedicated fleet (purely for testing) of Teslas driven all around the world by a group of Tesla employees or contractors?
The future of communication speed of mankind is evolving into something that eventutally will sound like R2D2... Andrej is only the first phase of this
14:56 to 15:55, how long does the iterative run of the data engine take to complete the cycle? Seconds? Minutes? Days? And how often is an update sent to the fleet? Every day? Every week? 17:32, all examples coming from the fleet will be annotated, so are all the half million examples annotated by hand by people? And, how long does it take to download an new update to the car?
Fleet examples are automatically annotated mainly by Dojo (their massive cluster) and anything with a low confidence score get human input, making the ai smarter. Asking the fleet and getting correct answers is a measure of the algorithms accuracy. Updates come to the fleet about once per month in the main updates, hard to say which of the smaller ones are streamed and when.
@@rickkay9548 You will surely get a message if an update has happened. And surely also a message about when that was. Don't you have a history to retrieve? And also what name they all have. And what size. This must be somewhere in the screen menu to be called up?
@@silberlinie The updates only come in mass stages as official releases, not individually as tiny packages. You can auto-apply them when notified of readiness, or just delay or ignore them. Sometimes they have names like "V10" or "Holiday update" but usually are labeled with their release number or git hash, and can be tracked here teslascope.com/teslapedia/software
@@rickkay9548 The updates come via WLAN? And also via cell phone network? Or can you define that yourself? Do you have to be in park mode for that? Or is it also possible while driving? Can you watch yourself when it srartet and finished, so you can see how long it takes? On the Teslascope page there is no information about the MBytes of each packet, what do you know about that?
@@malikhamza9286 Sounds like a chicken and egg problem. If ML is perfect, it will return all the correct results; but if ML is perfect, they would not need to query and label these images.
They could be using a larger, more accurate but slower network to do this automatically. Consider that FSD has to run in real time on relatively modest hardware compared to what tesla has at their disposal for actually training the network. They could also be simply looking for moments where the network is uncertain about what it's looking at, and that should generally be good enough to get desirable results
I learned a lot about AI thanks to Andrej videos, he is a talented teacher, no doubt. But never heard about that Mephisto, so I checked, and yes, sure that's his voice, especially here: th-cam.com/video/WhPjlnWbtS8/w-d-xo.html I could'nt believe it.
Tesla has it right, an AI taught to replace humans should see like humans. Not only interact with the road but also with the environment. and the evil is in the details... The long tail.
I don't understand one thing with Tesla cars is what happens when they are not connected to the network for long periods of time? Also, I'm sure these cars are going to be shipped with the Starlink receivers in the near future in some sort of miniature forms.
The car doesn’t use network for self driving, the relevant code is stored on the car’s computer (twice). The only difference is updates and route planning.
So basically you guys are utilising the fleet for getting varied data and also ensuring if the model works fine and if it fails then again quickly train the model on those groups of datasets to make the model more robust!!! Interesting However how far are we from the moment where we kind of act well as humans do with just very few datasets ??? Cuz what we are doing is statistical inference on the basis of large datasets ! So it's basically good datasets and good compute as mentioned earlier
Wouldn't it still be a good idea to include lidar devices in some of the cars with FSD so you have can have a ridiculously large dataset for unsupervised training, image -> point cloud?
That's what I'm inferring they're doing, because they explicitly said it's possible to do exactly that. However they did double back and said they could train on distance without it anyway.
Is this true, that Tesla is primarily concerned with image detection? Driving requires so many skills: the understanding of basic physics, understanding when objects are static and don't move and when they're dynamic... How does Tesla incorporate all these other concerns into their models?
I think he explained they are using drivers as training data as how the drive cars to maneuver the car based on the steering ,wheel etc. Might need more progress on that tho .19:30
Now if they could just make it so my car doesn't hit the brakes like mad when another car is turning across my lane of traffic. It seems like the AI should be able to tell at the speed the car is moving it would be long out of my way and doesn't slam on the brakes. I wish they had an option in the car to report when the AI does something wrong.
Another reason I believe that human body, brain, neuro system was actually PLANNED, THOUGHT THROUGH, DESIGNED by the SUPREME being, God! If it takes so much planning and efforts from people to design a limited version of AI software, its logic etc. I will never believe there is no God. There is God and He came to live among us in the person of Jesus from Nazareth.
The fact it works on their own chip is one of the most underrated thing Tesla has done tbh. The training is done back at Tesla and tHe nets are downloaded to the car so it can drive. Other cars have massive computers in the boot.
@@lmcclymont Mercedes-benz already took control of the EVs industry. And the faul lies in this weirdo who never could finish 100% the Self-driving autopilot. Thats why the company Tesla can not really take off.
@@edwardolvera5280 I gar unter you that FSD beta can work hands free at every single place the Merc can and then at millions of other places as well. You have zero understanding of what is happening here.
@@edwardolvera5280 umm L3 on a handful of roads is not full self driving 🤦♀️. Maybe read something other than the New York post or other FUD and you will understand the difference.
![Elon Musk says losers use LiDAR. [Explanation video]](/img/n.gif)
I had to double-check that I wasn’t on 2X playback speed
Yep. I did too!
the video has been sped up, he isn't just talking fast but moving fast too. 0.75 is better
He has the chip in his brain
I swear to God me too
I didn’t see your comment before, I did exactly the same thing
The ability to source any type of examples from the fleet is amazing!
Andrej sounds totally normal if you set playback speed to 0.75x...
His brain is running a third faster than most people!
NeuraLink is going to be interesting when communication is no longer the current slow speed speach...
A third is understatement
OMG ... So true. thanks for the tip 😃
The video has been sped up, apparently by around 1.25x. Unfortunately needs 0.8x on the playback to make it correct but that's not available. 0.75x makes the playback a little bit too slow. So you either get too slow (0.75x playback) or too fast (1.0x playback). Not sure why the person who posted this video sped it up like that.
I had to check my settings, he made me think I changed it to 1.5 !
@@zantetsu8674 seems like you haven't heard Andrej speak.
That moment when Elon reminds Andrej about CS231 😅😅
That image of the fleet data on a map at 12:24 is absolutely stunning. I'm from Cyprus, and even though we don't have Tesla showrooms, superchargers or service centres I can see that Tesla is still collecting data from the cars which were imported here.
How come?
There are some pretty big high-level takeaways here: using pure vision with a slight hint of radar, for now, to get lidar like accuracy in terms of depth perception. This is major.
This is awesome!... I wish the Teslas have the aerial camera view by combining all 8 cams to show all around the car. Like BMW does on their cars
There is no front bumper camera for doing that.
@@fredepstein that makes sense!... but i think you can project images from the front cameras with a transform to fit the road in front.
@@PierreH1968 not possible because the camera can not see just in front the bumper.
very powerful hot dog/not hot dog technologies, thanks for sharing
Jian Yang must be super proud
What a great video on AI and neural networks. Amazing future. Tesla is well positioned to remain in the lead with autonomous vehicles because they have so many cars in the field gathering data and teaching its neural network. No one else is in the game.
I like how elon chipped in when the guy got nervous. Great team!
Awesome presentation.
Question: when he says '"fleet" that means all the autopilot enables Tesla cars currently owned by public, or a dedicated fleet (purely for testing) of Teslas driven all around the world by a group of Tesla employees or contractors?
Thanks for the delivery
The future of communication speed of mankind is evolving into something that eventutally will sound like R2D2... Andrej is only the first phase of this
Is this on fast forward for anyone else?
at 0.75 playback speed, it sounds about right.
No they just live life on 1.25x
14:56 to 15:55,
how long does the iterative run of the data
engine take to complete the cycle? Seconds?
Minutes? Days?
And how often is an update sent to the fleet?
Every day? Every week?
17:32,
all examples coming from the fleet will be
annotated, so are all the half million examples
annotated by hand by people?
And,
how long does it take to download an new
update to the car?
Keen to know
Fleet examples are automatically annotated mainly by Dojo (their massive cluster) and anything with a low confidence score get human input, making the ai smarter. Asking the fleet and getting correct answers is a measure of the algorithms accuracy.
Updates come to the fleet about once per month in the main updates, hard to say which of the smaller ones are streamed and when.
@@rickkay9548 You will surely get a message if an
update has happened.
And surely also a message about when that was.
Don't you have a history to retrieve?
And also what name they all have.
And what size.
This must be somewhere in the screen menu
to be called up?
@@silberlinie The updates only come in mass stages as official releases, not individually as tiny packages. You can auto-apply them when notified of readiness, or just delay or ignore them.
Sometimes they have names like "V10" or "Holiday update" but usually are labeled with their release number or git hash, and can be tracked here teslascope.com/teslapedia/software
@@rickkay9548 The updates come via WLAN?
And also via cell phone network?
Or can you define that yourself?
Do you have to be in park mode for that?
Or is it also possible while driving?
Can you watch yourself when it srartet
and finished, so you can see how long
it takes?
On the Teslascope page there is no
information about the MBytes of each packet,
what do you know about that?
Thank you!
What did you use to query the fleet? How can you guarantee the search engine return all the correct result?
They already said machine Learning Mechanism to do this.
@@malikhamza9286 Sounds like a chicken and egg problem. If ML is perfect, it will return all the correct results; but if ML is perfect, they would not need to query and label these images.
@@X_platform doesnt have to be perfect when such influx of data is coming in
They could be using a larger, more accurate but slower network to do this automatically. Consider that FSD has to run in real time on relatively modest hardware compared to what tesla has at their disposal for actually training the network. They could also be simply looking for moments where the network is uncertain about what it's looking at, and that should generally be good enough to get desirable results
What an amazing presentation! Thank you
Did I go slow or is this video sled up?
Just wanted to turn down from 1.5x speed, but saw the video was playing at normal 1x speed...
Why use neural net to predict depth if you always already have the ground truth from Tesla’s sensor?
Oh the sensor only give sparse data points, neural net will predict every pixel and their distance
And 1 sensor is cheaper than 2 😬
The sensors can not see far enough.
Lidar is bad with object densities, works badly in the heavy rain and cannot make the difference between an opaque plastic bag and a rock...
That sounds like a startup worth launching
The fact that his last name is Karpathy
@@timurishuov5141 I mean since he’s working on AI for cars following roads (paths)
It's weird to think I've watched this guy before he got into neutral networks. Glad to see badmephisto is doing well
Same here....I learned advance cross and look ahead from his videos.
I learned a lot about AI thanks to Andrej videos, he is a talented teacher, no doubt. But never heard about that Mephisto, so I checked, and yes, sure that's his voice,
especially here: th-cam.com/video/WhPjlnWbtS8/w-d-xo.html
I could'nt believe it.
Human drivers have accelerometers and microphones built-in that should be used too for self driving cars
yes youre watching this at 1x
Still... Such a powerful talk!
Tesla has it right, an AI taught to replace humans should see like humans. Not only interact with the road but also with the environment. and the evil is in the details... The long tail.
Why does it seem like his speaking is sped up? Is this video speed increased
I put speed at .75 to hear it properly
I don't understand one thing with Tesla cars is what happens when they are not connected to the network for long periods of time? Also, I'm sure these cars are going to be shipped with the Starlink receivers in the near future in some sort of miniature forms.
The car doesn’t use network for self driving, the relevant code is stored on the car’s computer (twice). The only difference is updates and route planning.
This is how I imagine some ai dude speaking. 2x the normal speed
Lol so true
I had to slow my playback speed to 0.75 to understand him lol.
@@mattfairman8585 LMAO I always wondered what the point of that feature was .. Now I know.
So basically you guys are utilising the fleet for getting varied data and also ensuring if the model works fine and if it fails then again quickly train the model on those groups of datasets to make the model more robust!!! Interesting
However how far are we from the moment where we kind of act well as humans do with just very few datasets ???
Cuz what we are doing is statistical inference on the basis of large datasets ! So it's basically good datasets and good compute as mentioned earlier
This is genial!
Wouldn't it still be a good idea to include lidar devices in some of the cars with FSD so you have can have a ridiculously large dataset for unsupervised training, image -> point cloud?
That's what I'm inferring they're doing, because they explicitly said it's possible to do exactly that. However they did double back and said they could train on distance without it anyway.
I'm absolutely blown away. This is true AI.
I had to double check playback speed and actually put it at 0.75x
Is this true, that Tesla is primarily concerned with image detection? Driving requires so many skills: the understanding of basic physics, understanding when objects are static and don't move and when they're dynamic... How does Tesla incorporate all these other concerns into their models?
They cant
Thats why the Teslas keep crashing.
I think he explained they are using drivers as training data as how the drive cars to maneuver the car based on the steering ,wheel etc. Might need more progress on that tho .19:30
Shut up Elon musk and let the smart guy talk
Set playback speed to 2x to fully understand what is happening in Andrej's brain
Can I enroll in these courses that teaches neural networks
Cousrea, youtube, udacity, udemy
How are half a million images annotated? Is it done manually?
Normally yes, but there are also methods to make it semi-automated, you can check active learning.
This is now mostly done automatically
Imagine being an annotater at Tesla. What a mind numbing job that must be…
Another beautiful brilliant mind.
Now if they could just make it so my car doesn't hit the brakes like mad when another car is turning across my lane of traffic. It seems like the AI should be able to tell at the speed the car is moving it would be long out of my way and doesn't slam on the brakes. I wish they had an option in the car to report when the AI does something wrong.
I heard Musk say it. 11:55 "You can't. There's no way". Where is my prizemoney??
Walt and Gale vibes
Missiles making
Lidar hit job?
Understand why Tesla is going so fast
Kar-pathy .... Like psychopathy for cars... how ironic that's his last name
Uh lol
Fast talker is also fast Mathematics problems solver.
wow👏
Me with 4 GB RAM machine:
Interesting 😯😯😮
it doesn't rely much on RAM, it's more about GPU
Hey OEMs, pay us and we will license our suite of hardware and software. We will be sure your cars are just one update behind ours always.
If I had a tesla, I wouldn't care about them taking images from my car to improve fsd. Go Tesla!
There may not be sufficient number of edge case images to perfect the neural net. This limitation is crucial and may not be solvable.
Like human vision?
Another reason I believe that human body, brain, neuro system was actually PLANNED, THOUGHT THROUGH, DESIGNED by the SUPREME being, God! If it takes so much planning and efforts from people to design a limited version of AI software, its logic etc. I will never believe there is no God. There is God and He came to live among us in the person of Jesus from Nazareth.
Wouldn't your car need a freakin' super computer onboard to run this neural net??
The fact it works on their own chip is one of the most underrated thing Tesla has done tbh. The training is done back at Tesla and tHe nets are downloaded to the car so it can drive. Other cars have massive computers in the boot.
No, training of neuronet needs a gpu,s and tpu's running models already trained can on cellphone or raspberry pies
and this whole thing ended up being a dead end?
the good old days when Musk wasn't a total nutjob
Andrej always speeks 1.5x
Play this video in 0.75 your welcome
Elon musk is like /|:
does he really talk like that? he sounds like a stephen hawking's computer
This is old
Mercedes-benz is already level 3
this weirdo has no noticed it yet
It is L3 on certain roads under 30mph 😂. Tesla is almost L3 anywhere in Northern America.
@@lmcclymont mercedes-benz did it in a quarter of time Tesla did its.
This weirdo does not know much about cars.
@@lmcclymont Mercedes-benz already took control of the EVs industry.
And the faul lies in this weirdo who never could finish 100% the Self-driving autopilot.
Thats why the company Tesla can not really take off.
@@edwardolvera5280 I gar unter you that FSD beta can work hands free at every single place the Merc can and then at millions of other places as well. You have zero understanding of what is happening here.
@@edwardolvera5280 umm L3 on a handful of roads is not full self driving 🤦♀️. Maybe read something other than the New York post or other FUD and you will understand the difference.
🥵🥵🥵 How fast you can talk ?!
(lol) weak techs (taunt)
Kachinski here will fail
Elon, quit worrying! we know you're an impostor but we will still love you just like your dad does
What do u mean