Andrej Karpathy: Why Tesla Has a MASSIVE Data Advantage!! The Tesla Data Engine

in a recent Tweet, Andrej Karpathy, former Tesla AI lead, describes how any company can dominate the AI/Machine Learning Space. It's not just about the data, but about the data engine or flywheel that can produce, label, clean and test data as fast as possible. Tesla has this in spades after years of engineering efforts, and other auto manufacturers are giving up!

Get The Elon Musk Mission (I've got two chapters in it) here:

For a limited time, use the code "Knows2021" to get 20% off your entire order!

**Want to get in on WeBull's awesome Crypto and stock fun? Check out this link, and get started trading stock and Crypto!

**You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you!

**What do we use to shoot our videos?

**Here are a few products we've found really fun and/or useful:

Tesla Stock: TSLA

**JOWUA:
use the Promotion Code: DRKNOW5
use the Promotion Code: DRKNOW5
use the Promotion Code: DRKNOW5

**EVANNEX
If you use my discount code, KnowsEVs, you get $10 off any order over $100!

Instagram: @drknowitallknows

[tommornini]

I believe Andrej said:

1) a ton of…
2) clean
3) well labeled
4) diverse

Tesla has #1-#3, but #4 is very important, and very difficult to achieve.

It implies the “long tail” of weird stuff, as well as the common stuff, at every hour of the day, in every type of weather, during every season, in every locality, in every condition the camera can be in.

Which is to say, a fleet of tens of millions of cars, driving billions of miles a day, for years.

Which is why Tesla has a very large lead, yet still has not crossed the finish line — and why single stack is so critical — we need FSD Beta stack data from the entire fleet!

And this will continue even after the system works well!

If companies intend to compete, they MUST deliver millions of cars with substantial AI inference capability and a well thought out camera fleet ASAP.

Before they do that, they shouldn’t even bother attempting FSD development.

Given traditional automaker’s low margins, they’re going to have a hard time “giving away” that hardware on vehicles that don’t pay for it.

[BobBrownIII]

Data is plural. “Those” data. “That” datum. Sorry. Pet peeve. Great video. Really clarified the importance of the data engines vs data.

[tomschuessler]

Thank you for dedicating the effort to explain the thinking behind the new terminology and thoughts of the Karpathy's , and Musk's of this world. When they talk it is easy for me to get lost or at least confused and your explanations help me, at least, think I have more clarity. At least for a few minutes.

[skinnymoonbob]

It there a viable alternative to Tesla FSD? Sounds like how Microsoft ruled the PC industry in the 1990’s. If Tesla will license FSD to other car manufacturers, I can imagine that Elon mentioned earnings can become ’nutty’.

[pontonlabel]

I am impressed how one tweet generates 15-30min videos on youtube nowadays

[Martinit]

The data engine is the manufacturing line of AI applications.
Without a data engine you can only do one-off prototypes at great expense.

[markumbers]

I agree. I think other OEMs will adopt Teslas FSD system. It makes sense. They have been catalogue engineers for a long time now. They need to stop thinking they are something they are not.

[fjalics]

Tesla needs to start capturing low resolution map data. I want it to figure out when road construction happens, so it can anticipate which lane to be in, how fast to go, where the turn is, when ramps are blocked. Like a human that goes the same way, the third time through you remember what to do. You don't need to see it.

[davidbeppler]

What happened to Mobileye? Back in July of 2016 they split with Tesla. At the time Mobileye said they were 10 years ahead of Tesla and would have level 5 in 2020. Here we are in 2023 (almost) and Tesla FSD is on any Tesla built after 2016. Over 2 million cars. Where is Mobileye? Are they still 10 years ahead? Does BMW that uses Mobileye have lv 5?

[muratarican]

Thank you, doc, for all these short lectures which I believe are very inspiring for those who want to become AI software developers..

[oneproductivemuskpm]

This explanation/revelation simply puts Tesla in the far lead with no competition is sight because they have all that it takes to get FSD to work! This also means they can easily license the technology to all automobile companies and that means huge revenue! In fact this only goes to show how far into the future Elon thinks!
Great Job my friend!!!

[balaji-kartha]

Any news if Andrej Karpathy is going back to Tesla?

[phatfel]

any comments about the HD radar tesla is coming out with?

[helmutmueller]

You are absolutely right. The problem is second that the models depend on camera positions. So only companies having a fleet out there have continuous access to viable data at all - then they must have the communication with the fleet up - then they must own the data - if the data is owned by a supplier and not the car manufacturer this may be a big non technical issue and you must be able to push back versions to the fleet online with all technical and legal aspects. And you must have a fleet out there that can execute AI models at all, though nobody gives you money for the hardware as long as your models are bad - so you have to preinvest and ruin your profits in the beginning. That is a long way to go if you start from now. But still - autonomous driving is far from trivial and crashes with bad models may ruin you fast. So big risk - and big opportunity.😢

[nicholasmuni]

I wonder why the good top people at Tesla keep leaving?

[marriagepartnersministry]

There are many ways to solve the FSD problem. To say real world data is THE way, seems intuitively correct but intuition isnt always correct. There are many examples in life that teach us that alot of things are counter-intuitive and correct vs the intuitive answer. There are several data engines using simulated data very effectively. Also real world data of a driver doing something "wrong" is not only unusefull but has a compounding hindering effect within the neuronet. Many say that high quality simulated data can be just as useful and much faster than accumulating real world data and parcing out the real world mistakes that one doesnt want in the neuronet.

[harry-eto]

If one understands that and compares it to marketing stunts like Mercedes' "level 3"...

[lylestavast]

So if Tesla isn't saving all their labelled and clean data endlessly, they have to re-do that in some way on new incoming (or are the neural nets holding what's important from them as the "storage" medium of useful knowledge ) - is the idea to update the auto-cleaning and auto-labelling to such high accuracy that running through the last batch produces no regressions/errors all the time ? How do they keep all the really unique situations if they're throwing away so much ? And at what phase of all this do things like RH drive country, different signs and road textures etc get factored in ?

[Mellowyellow]

the future companies are going to be MAAS.. model as a service.. anyone that has propietary data and also can collect data the quickest will corner that AI market for their sector/nice that they are working on..

[MrDuncanBooth]

John I remember when you first started doing lots of Tesla vids you said that you only had a few shares and didn’t really plan on buying any more. Still the case?