Elon Musk and Andrej Karpathy on the PAST, PRESENT, and FUTURE of AI and Tesla Full Self Driving!

'Elona' Musk discusses why Tesla's FSD is so difficult to create in a tweet, and Andrej Karpathy, director of AI at Tesla, recreates a seminal 1989 Neural Network paper by Yann LeCun et al--and discusses how things have changed in the past 33 years, and how they might change in the next 33 years. Hint: the more things change, the more they stay the same! :)
Green Hill Software, Dan O'Dowd, the New York Times, and a recent Tesla Smear Campaign are all covered as well.

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

Elon Musk and Andrej Karpathy on the PAST, PRESENT, and FUTURE of AI and Tesla Full Self Driving!

[gaydybwad]

The point is 1989 was very important! For example "Seinfeld" first aired on CRT TVs in '89 and teaches us still many things but not data set manipulation.

[richardgoldsmith]

In 1973 I was working on U.K. postcode recognition training a recognition system using what they called the caltrops scorer. U.K. postcodes were more complicated by the use of alpha characters as well as 0-9. The scorer was named after the quad spike anti-cavalry deterrent device. It referenced the more complicated than usual scanning to build up a score based recognition system. No Neural nets but lots of photos of letters with postcodes written in guide boxes which I had to pass in front of the reader camera for hopers on end!

[polarlight]

Super excellent video, John - the more things change, the more they stay the same ... fantastic review of Andrej Karpathy's blog article - thank you!

[adamrak]

You can "press" large neural networks into VLSI.
The network may need a city sized supercomputer to train, but if you convert the weights into physical wiring on a chip, it would shrink to thumb size.
This way large foundational networks can be converted into small and cheap accelerators chips, so you do not need to use the cloud for inferencing tasks.

[andrewsaint]

AI on TV from 1985.
Max Headroom was cool back in the day.
Made me laugh, Cheers Doc.

[johnyoungquist]

Much of what modern AI folks take credit for today was developed by Frank Rosenblatt in the late 50's. His paper published March 15 1961 a mere 621 pages long describes " back propagating error correction procedures" in section 13.3 starting on page 292. It is now unclassified and was published by the Armed Services Technical Information Agency. Document number AD-256 582. It is titled "Principals of Neurodynamics Perceptrons and the theory of Brain Mechanisms" He did optical character recognition with 22 square array of photo tubes using motor driven potentiometers. It worked!! Unbelievable but true !!!! I have a printed copy that originated from a PDF but don't have the file.

[kstaxman]

Excellent video as always. To peer behind the curtain with the help of your understanding of neural networks is so helpful in understanding such a complex issue. Andrej Karpathy's article would have been impossible for me to follow without your breakdown just from a jargon perspective. It's an interesting journey we are on building machines that can do what only humans could do just a few years ago.

[paulwilhelm]

Fun blast from the past. We considered neural networks for the then AutoPap Pap smear screener (developed 1989 - 1994), but found other more traditional classification methodologies did as well. We shelved efforts at using neural networks because nobody knew how to efficiently train deep neural networks. The backpropagation methodologies matched to the neuron designs of the times broke down when networks were deeper than a couple of tens of neurons.

[splashmaker]

The team is probably the most important part of the equation! Organizations often seem to forget this, and fall back to job titles.

[alexanderpoplawski]

"Hey Megabrain, make me some eggplant parmesan!". Megabrain feeds me some protein, fat goo but, thanks to Neuralink, my eyes and taste receptors tell me I am eating the best eggplant parmesan in the world. Stanislaw Lem - The Futurological Congress, anyone?

[fredhearty]

Super video! My neural net has plenty to crunch on...

[Kenlwallace]

My understanding is that hardware 3 in Tesla’s uses two custom Tesla chips that contain about a dozen A72 ARM cores … is that true? That’s not really ASACs … maybe ASICs are what the next ‘hardware 4’ is all about,

[jbarvideo]

Thank you "Elona" Musk, Dr Karpathi and Dr Know It All and providing help in discussing the creating of FSD and seeing how Yann LeCun's paper 33 years ago was so accurate then.

[montypalmer]

Add a billion miles, it's mostly a bore. Situations are the challenge for FSD or someone driving down the road. Turning left, reacting to bad drivers, make a list of situations and design responses. That's what FSD is all about, what any driver has to deal with.

[chickenp]

hey, could you share some of your ai projects, i am really interested what got you to understand ML.

[dscarty]

The difficult (hard) we do immediately. The impossible takes a little longer.

[skinnymoonbob]

Seeing previous versions of FSD beta and now 10.11, when do you think a robotaxi pilot can start?

[kstaxman]

When you have rapidly moving one to five ton vehicles moving freely closely surrounded by other similar vehicles what could possible go wrong?

[aljohnson]

Interesting discussion, explanation and interpretation . Although, not training a NN at some future 33 year point to just asking or THINKING of what one may want may appears enticing. I always remember the jest of the film "Forbidden Planet" (circa 1956) --- machines in some distant planet conjured solutions to one's thoughts, but the developers of those machines forgot or did not consider what was the tag line in the movie (ahead of its time) "BEWARE OF THE POWER OF THE ID!" That miscalculation, fulfilled every wish, thought, pleasure and deep hate in the basic part of human subconsciousness -- the ID's (normally suppressed) impulses were projected to action causing the destruction of any enemy, not matter how trivial, real or perceived. I.E:. I wish that person that just cut me off in traffic would drop dead and find your base primordial wish is granted. Ultimately, all life on the Forbidden Planet was destroyed, save an old gentlemen and his daughter.

[fredbloggs]

If a typical CompSci student were to similarly repeat the paper today as a project, what level would they likely be at? First-year? 2nd, 3rd? PostGrad?