Tesla, Software 2.0, and Why Andrej Karpathy Left Tesla! Interview with Lex Fridman

As part of his wide-ranging interview with Lex Fridman, Andrej Karpathy, former Tesla AI lead, discusses his vision of Software 2.0, a new, Neural Network based software, and how it can and likely will allow things like Tesla FSD, but also could allow a new age of computing. This new paradigm can be a powerful tool, but people need to think differently about how to create and use it, and Andrej explains how to do this.
Additionally Karpathy says that he is open to returning to Tesla. I speculate on why he left, and why he would return.
Read my companion blog about Software 2.0 here: www.artimatic.io/post/what-is...
Get The Elon Musk Mission right here:
Paperback: amzn.to/3TQXV9g
Kindle: amzn.to/3U7f7Hr!
**To become part of our Patreon team, help support the channel, and get awesome perks, check out our Patreon site here: / drknowitallknows . Thanks for your support!
**Want some awesome Dr. Know-it-all merch, including the AI STUDENT DRIVER Bumper Sticker? Check out our awesome Merch store: drknowitall.itemorder.com/sale
For a limited time, use the code "Knows2021" to get 20% off your entire order!
**Lift pucks for your Tesla: amzn.to/3vJF3iB
**Emergency tire fill and repair kit: amzn.to/3vMkL8d
**If you're interested in a CO2 Monitor, you can check it out here: amzn.to/3PsQRh2
**Check out Artimatic: www.artimatic.io
**For 30% off the amazing OneCarStereo instrument display for Tesla Model 3/Y, use this link: bit.ly/3vd7NRv
**Get 20% off an awesome Ekster wallet and/or wallet locator(!), check out my affiliate link here: shop.ekster.com/DrKnowitall
**Want to get in on WeBull's awesome Crypto and stock fun? Check out this link, and get started trading stock and Crypto!
a.webull.com/i/DrKnow-it-allK...
**If you are looking to purchase a new Tesla solar roof, solar tiles or PowerWall, just click this link to get up to $500 off! www.tesla.com/referral/john11286. Thank you!
**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!
* USA: amzn.to/39n5mPH
* Germany: amzn.to/2XbdxJi
* United Kingdom: amzn.to/3hGlzTR
* France: amzn.to/2KRAwXh
* Spain: amzn.to/3hJYYFV
**What do we use to shoot our videos?
-Sony alpha a7 III: amzn.to/3czV2XJ
--and lens: amzn.to/3aujOqE
-Feelworld portable field monitor: amzn.to/38yf2ah
-Neewer compact desk tripod: amzn.to/3l8yrUk
-Glidegear teleprompter: amzn.to/3rJeFkP
-Neewer dimmable LED lights: amzn.to/3qAg3oF
-Rode Wireless Go II Lavalier microphones: amzn.to/3eC9jUZ
-Rode NTK Studio Microphone: amzn.to/3vjMomW
-Focusrite Scarlette 2i2 audio interface: amzn.to/3l8vqDu
-Studio soundproofing tiles: amzn.to/3rFUtQU
-Sony MDR-7506 Professional Headphones: amzn.to/2OoDdBd
-Apple M1 MacBook Pro: amzn.to/3wPYV1D
-Docking Station for MacBook: amzn.to/3yIhc1S
-Philips Brilliance 4K Docking Monitor: amzn.to/3xwSKAb
-Sabrent 8TB SSD drive: amzn.to/3rhSxQM
-DJI Mavic Mini Drone: amzn.to/2OnHCEw
-GoPro Hero 9 Black action camera: amzn.to/3vgVMrH
-Suction car mount for camera: amzn.to/3tcUfRK
-Extender Rod for car mount camera: amzn.to/3wHQXsw
**Here are a few products we've found really fun and/or useful:
* NeoCharge Dryer/EV charger splitter: amzn.to/39UcKWx
* A totally fun Tesla coil toy--awesome for the whole family: amzn.to/30rvzbI
* A camping mattress for your Tesla model S/3/X/Y: amzn.to/3chSRrQ
**Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or KZbin - / @zenlee_music
Tesla Stock: TSLA
**JOWUA:
*For an awesome wireless charging phone holder mount, check out: bit.ly/3tRMSQR
use the Promotion Code: DRKNOW5
*Tesla USB Hub: bit.ly/2PXHPit
use the Promotion Code: DRKNOW5
*Powerful wireless iPhone/Android charger for your Tesla Model 3 or Model Y, check this link: bit.ly/3b5fHSW
use the Promotion Code: DRKNOW5
**EVANNEX
Check out the Evannex web site: evannex.com/
If you use my discount code, KnowsEVs, you get $10 off any order over $100!
**For business inquiries, please email me here: [e-mail]
Twitter: / drknowitall16
Also on Twitter: @Tesla_UnPR: / tesla_un
Instagram: @drknowitallknows
**Want some outdoorsy videos? Check out Whole Nuts and Donuts: / @wholenutsanddonuts5741
Lex's Interview: • Andrej Karpathy: Tesla...

[@daveblack8831]

From my read, he will probably be back at Tesla. He has shown strong interest in Optimus and he was the one who brought up the idea of returning first. What a cool opportunity it would be for him to lead the training team for all of those different useful tasks Optimus will be able to do!

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@jeromecastonguay]

I know, how lucky are we to have that amazing interview

[@MattOGormanSmith]

From what I gathered, it wasn't the implementation that didn't hold his interest, it was managing the team he built in the long term. He liked being a scientist, then he liked being an engineer. He even liked being a manager for a while, but he'd reached the point when his job was turning into being mostly a politician. Very few people have Elon's ability to be a CEO while still maintaining a handle on the nitty gritty. Bill Gates for instance, was at his most effective as Chief Software Architect before he gave that up and became Chairman of M$. He couldn't (or perhaps wisely didn't try to) divide his attention between the product and the overall business.

[@KevinDriedger]

Love your explanations of what Karpathy said. Would love to see more of this type of video.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@mayetesla]

Yrs ago Jensen Huang said "AI is software that writes itself" ... interesting perspective in my opinion.

[@jbarvideo12]

Fascinating discussions in understandable terms. Thanks for explaining and sharing your immense software knowledge.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning....

[@richardgoldsmith7278]

The value of data sets for training: “how to turn a beginner into an expert?” the answer is add experience. The data sets are analogous to the experience.

[@ztechrepairs]

He is a brilliant dude for sure and anybody or any company that has him is definitely lucky!

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@r.a.monigold9789]

Thank you for the insightful and educational overview.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@budgetaudiophilelife-long5461]

Fun ,interesting ,thought provoking ☑️☑️☑️

[@TheEroxion]

I see a parallel to conscious Vs unconscious learning, we first learn a new skill, like driving or tennis, with conscious planning of each movement, but we only master the skill when we subsume it into our unconscious, where patterns are recognised without needing 1.0 style step by step analysis in conscious thought. We cannot explain in detail how we hit a tennis ball so accurately, it is an emergent result of the training

[@pooglechen3251]

Generative AI is the AI killer app. Whether it's Google, Meta, OpenAI or someone else..I expect generative art, music, environments, worlds, stories..etc to be here very soon. It's going to be exciting

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@bobharrington8063]

John, Your focused critique of Andrej’s explanations of software 2.0 and neural nets is brilliant, and immensely helpful.

[@michaelcoughlin8238]

I'm missing a foundational piece of this. Can someone point me to a good explanation of "neural network architecture" please? I've been in IT in one form or another for almost 40 years but never a programmer.

[@famnaff5136]

Good work. A lot to look forward to in the coming years.

[@lylestavast7652]

Thanks for the breakdowns of these complex things with technology., terminology etc that are evolving ...

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@randomsonmymind6919]

I was like this 🥴listening to the podcast because everything was way over my head but it's interesting to hear him talk about neural nets etc.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning....

[@balaji-kartha]

Well, this 2.0 is like teaching a child instead of a robot! You tell kid what to do and what not to do, and the rest is done by the kid as it learns by itself! (And you know what, I wrote the above comment just before you mentioned the same thing at 22.00 !!) By the way, Karpathy never really left / resigned from Tesla. He is just on a break / leave and he sure would be expected to join back whenever he is ready (an honour he has earned)

[@martinbeaumier7172]

It was a great interview!

[@LuisSalazar-mq4lv]

great explanation, thanks a lot

[@snarkyboojum]

I reckon he’ll go and work at OpenAI and work on the evolution of LLMs and AGI.

[@NickMackenzieMD]

John, excellent and insightful as always, thank you! Still listening to your book which I really enjoy (using Kindle's Text-to-Speech feature) !

[@thatscottishengineerguy9606]

I had forgotten all about Mad Libs. Most of the Mad Libs that my friends and I created would've gotten us in trouble if our parents had found them. A lot of trouble. LOL

[@nelsonmacy1010]

A+. Thanks for a great analysis. This is a 10x summary vs Solving the Money Problem rants and invectives. Funny, entertaining but little analysis and explanation. John Gibbs will rule as FSD gets ready for GO TO MARKET

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@JayDeitch]

To bad I can only give you one thumbs up.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@robkeshav800]

Excellent video. Even with your parsing it is way over my head but very interesting. Andrej mentioned that he was leaving Tesla because as their AI director he was more into management role that he did not relish. Maybe he will be back (Elon has open invitation for him) as a pure researcher, big idea man.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@randywelt8210]

Determinism will win. Backprop is just a good strategy to find that optimal determinism.

[@CorwynGC]

It isn't a matter of "determinism" as philosopher's define it, but whether humans can determine outcomes. All computers (maybe all of the Universe) are deterministic.

[@randywelt8210]

@@CorwynGC I watched the whole talk. Basically Andrej doesn't seem to believe in the random dice either. I love it when others come the same conclusion.

[@flwi]

Thanks a lot for adding your great explanations to this interesting interview! You have a great way of explaining things! I'm just starting to learn more about machine learning and that made me even more aware of the sophistication of Teslas AI team.

[@TheWinstn60]

How can consistency be maintained? Is it possible for example with a car using 2.0 to drive differently to an identical car with the same software. Just like humans

[@robt8869]

I expect the conditions will never be exactly the same. Light level, shadows, Temperature, traffic conditions, etc. So the car will appear to drive slightly differently every time.

[@jtjames79]

Stable diffusion already has a sort of ide. I'm not sure if it's fully featured yet but it is web accessible. Run it on your PC, from your phone while sitting on the couch.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@WolfgangMartinViernheim]

exciting times..👍

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@nelsonmacy1010]

John please enumerate your top 10 lex interviews and 1-3 minutes on each. Good 30 show for you. Then do a top 10 lex CS people PLEASE: kernighan, Knute, LLVM, guido, Brendan Eich, Patterson, Keller, etc

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@michaelweber3836]

Andrey talks only supervised learning, cause with a car it is very risky to do anything else. Just as an experiment, wonder what would happen if in a VR simulation a car could learn unsupervised, how to drive. It would take tons of virtual car crashes, but soon the neural network would be able to drive like when babyes learn how to walk???

[@michaelweber3836]

Maybe the supercomputer does not exist for this. Yet. 🤣

[@1flash3571]

You can use AGI for other things than just Full Self driving.....

[@qkktech]

The next step is get rid of those nn weights. And that is software 3.0. The job Andre is doing is amazing but still not enough for 2.0 - 1.8 maybe

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning....

[@dewaynehasty]

I've been in IT for 35 years and have programmed a little. Call me stupid but I just can't seem to wrap my head around this. I've watched several other videos on this subject but just seems like smoke and mirrors. I'm not saying it is smoke and mirrors. lol I have the same problem when I think about how there isn't an end to the universe!

[@newenglandbarbell4647]

Code typically creates data, with machine learning the data creates the code.

[@klauswagner7380]

Thanks!

[@noleftturns]

Great recap of Software 1.0 and 2.0 - liked the summary. I'm no AI expert just a retired guy who spent 35 years programming and watched Object Oriented Programming evolve to where it is today. I just don't think FSD is a Software 2.0 project - no need for it. A robot sure - where moving through our 3D world is going to have to be radically different. But FSD can be mapped - just send a Tesla down the road and the driver's path is recorded and a map is generated and so on. In no time, the most detailed maps of roads is available to all Teslas and how seasoned drivers handle the route. Every pothole is known, every bush that blocks a view is noted, tricky intersections are marked and how 1,000 Teslas handled the turn in various traffic conditions. Sure, you still have to still detect what's on the roads and where they are going but AI has no place doing that - simple equations predict the future. After using FSD for 2 weeks now I smell a software project in a lot of trouble and doomed to failure. Better to start a "skunk works" group off campus and start over or this is going to become a death knell to Tesla.

[@lengould9262]

You don't ubderstand the project, nor neural nets. What you describe is 1.0 with lidar, centimeter maps of entire earth, not how Tesla FSD works. You're describing how things work with lidar and precision maps. FSD works like the human brain with no task-memory (like where potholes are), and avoids potholes by recognizing them from camera images by appearance every time 2.0. Lidar systems don't do that, they work the way you described ONLY using preloaded detailed maps which then never change until the next major download, with no recognition from camera images at all 1.0. Ideal would be the 2.0 system FSD uses now, with a relatively small amount of medium term memory to store useful features of local routes, like potholes etc. Shared only if deemed useful to others.

[@noleftturns]

​@@lengould9262 I have used neural nets in a robotic application - like I said, robots are a different problem, and AI can help here. But a car driving on a known map doesn't have to use AI to navigate - it just needs to read how others have handled the situation a thousand times before. It should know that a pothole will harm the car 1/2 mile ahead in the left turn lane. By the time FSD detects the hole, it is in the left turn lane, and moving out of the way is asking for an accident to happen. FSD looks ahead and renders the environment every time it drives along a known road - this is wrong for this application. Robots can go anywhere, even upwards and downwards and don't follow a set path repeatedly. I get the impression that a bunch of AI geeks caught Elon's attention and convinced him to go down a path that is very interesting to them but the wrong approach to make FSD Levels 4 and 5. It's just my feeling - take it for what you paid for it.

[@lengould9262]

@@noleftturns Your preferred solution sounds great UNTIL you have to implement it.

[@noleftturns]

@@lengould9262 Can't argue with you on that - GPS has a radius of 3 meters so a minimum number of drives would be required before the car would attempt it. I knew FSD is doomed from Level 4 when I 1st double clicked the right stem stalk and the car started to shutter to the left and right and several times picked the wrong side of the road to drive on - this is an Alpha testing stage - too wild to be let loose on drivers who aren't hunched over the steering wheel with white knuckles.

[@arondaniel]

@@noleftturns You seem to be describing the self-driving approach used by Waymo, Cruise, and others. Centimeter level maps, planning, lidar, slow speeds, and skip tricky things like unprotected left turns. To scale up this approach to work safely almost anywhere requires tons of money and time but possibly also modifications to roads or elimination of human drivers. Assuming the Tesla approach works, it's probably the only one that can scale.

[@Roguescienceguy]

So basically we let the computer write the "if then else" itself.

[@mayetesla]

Maye the force

[@rockycata6078]

The same reason J. B. Straubel returned; being part[...recycling/dry-cell technology] of the acceleration of 'future' innovation, "at the speed of thought". Recycling lithium is not new, but 'vertical-integration' into the production stream is faster/cheaper. From resource acquisition, production speed/costs, and product delivery. One structural pillar versus a slow/unreliable/costly supply chain. Tesla Semi fits into the same 'principle-for-productivity'; ..."the best delivery is no delivery, but safer/faster/smarter delivery is key to cutting TOC". How much is Tesla spending on 'leasing/operating' Class-8 semis? So, if you thought it was about moving 'other' peoples products, Tesla's internal semi requirements are where demand for the first 100K units are scheduled. Everything else is profit to satisfy the "deductible" cap-ex cost of production.Throw in a few 'incentives', to prime-the-pump, and self-sustainability of the enterprise is achieved. A 'regenerating' acquisition/delivery cycle. Time-over-money[t/$] then becomes money-over-time[$/t].

[@richb2229]

The Robotics question is enough to interest a researcher. Think as an analogy, the modified car software as “software 1.0” and the next level robot software as “software 2.0”. The next level of robot software should be differentiated that much from what is running today.

[@lengould9262]

You're not understanding what they're differentiating with 1.0 and 2.0.

[@bootiemacarthur9182]

Wow…well done….educator and engineer!

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@magic76767676]

1:35: Yeah FORTRAN! That brings back happy memories of my youth.

[@magic76767676]

Articiial Neural networks... Late 1980's. My Master's degree and my 1st ex-wife. (sadness)

[@magic76767676]

16:02: Loss function. One of my papers I wrote at Hopkins was how do you have a neural network determine, "I don't know?" A logic gave seeing Multiple answer at the same time only works in a simple scenario. But Tesla is light Years beyond my time. (Go team!)

[@willcooper7651]

John, Lex and Andrej talked about "emergent" phenomena that arise from neural nets. Can you give an example and explain what's going on? I know that it's related the the complexity of the system, but please explain how that produces surprising and unexpected results. Thanks.

[@roberts932]

autonomous cars mating.

[@MattyNiceZM]

You will have to read a book. Read Ray Kurzeril.

[@rogerstarkey5390]

Self learning

[@treestandsafety3996]

What do you think of soft robotics..Teslabot 5.0..?

[@brunoheggli2888]

Its all about the adoptionrate of the integrated parts of an artifcial Brain in the quantum computing space!How long will that take?

[@treestandsafety3996]

@@brunoheggli2888 The human brain and body has been developing for billions of years. What Teslas data collection and hardware development is doing, is expediting that process. What is perhaps different, is the Union of humankind with the bots body. This is best served if the bot utilises so called 'soft robotics', to make a more biologically themed yet durable system...

[@treestandsafety3996]

human brain and body has been developing for billions of years. What Teslas data collection and hardware development is doing, is expediting that process. What is perhaps different, is the Union of humankind with the bots body. This is best served if the bot utilises so called 'soft robotics', to make a more biologically themed yet durable system...

[@brunoheggli2888]

@@treestandsafety3996 I just used some buzzwords and constructed a meaningless sentens!

[@treestandsafety3996]

@@brunoheggli2888 Perhaps this is characteristic of youtube comments...but mine was not. What do you think?

[@BrooklineElectric]

Karpathy said that he left Tesla because he didn’t want to be a supervisor he wanted to work on the actual project.

[@BboySnake71]

Why do you think they have so much trouble with Flashing Yellow lights? Hasn't worked for me since day one of FSD. Hard brake, go, hard brake, go all they way through the traffic 🚦.

[@martincossette7420]

Software 1.0: Tesla Autopilot, Waymo, Blue Cruise, MobilEye, Argo, etc. Basically, all and every ADAS function. Software 2.0: Tesla FSD. Step 3: Profit.

[@rogerstarkey5390]

Step 3 A) Reduction of global vehicle fleet. B) Reduction of materials (and pollution) required to build the fleet. C) Reduction (eventually elimination) of pollution running that fleet. D) Refurbishment rather than replacement (best use of resources..... It will be all fleet owned!) Etc. It's not for profit, it's for "the mission"

[@martincossette7420]

@@rogerstarkey5390 True. But not as effective as the intended joke's punchline. ;-)

[@EddyKruissink]

My big question is how the connection works between all this intelligence in supercomputers and my car. I can not understand why my car does not need to carry the supercomputer. It can not possibly hold all the decision-making code/algorythms nor is it’s hardware fast enough to do all the compute.

[@MattyNiceZM]

The machine learning is what takes the huge data sets and computation power. What goes into your car is the software version that the machine learning has 'created' from it's solving the data. The software version that gets downloaded to the car is the best solution to the driving problem at any given time.

[@rogerstarkey5390]

Do you own a dog? Dogs are "smart"...... Once trained. You are the "Supercomputer". . Remember that the main enemy of this system is energy consumption. You don't want that in the car.

[@OlMossBack]

Just a thought AI needs a fifth dimension--location. So it knows which side to walk on, the meaning of certain lights, etc. these all change with geographic location.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@treestandsafety3996]

human brain and body has been developing for billions of years. What Teslas data collection and hardware development is doing, is expediting that process. What is perhaps different, is the Union of humankind with the bots body. This is best served if the bot utilises so called 'soft robotics', to make a more biologically themed yet durable system...

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning....

[@happyfarang]

only people over 70 still code in Fortran... 😅

[@nowsc]

I watched the Lex Fridman interview of Karpathy. Are you re-posting it in order to get revenue from KZbin for the work Lex Fridman and Karpathy did? Diamond

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@rogerstarkey5390]

It's informed opinion. A step between the intelligent and knuckle draggers. A chance for one or two to stop dragging?

[@colinkelley6493]

When I was programming main frames we used to have a thing we called the 90% rule: "The first 90% of the project takes the first 90% of the time. The last 10% of the project, the exceptions code, takes the other 90% of the time." There was always a lot more important "what ifs" than people expected. The truth is we never finished any of our projects, they just declared them finished and handed them over to the IT maintenance people. Then, in three to ten years the world we had built into in our code changed enough to make our underlying foundational structure obsolete, and they had to start all over with a whole new project from the ground up. People got blamed and fired, but nobody could have done a better job. So here is a hypothetical Question: What if some state, let's say North Dakota, decided to adopt all of, say, California's moving vehicle code -- drop their code and adopt California's code. How hard would it be for Tesla's FSD AI to adjust to the new reality -- or any new meta reality?

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning.

[@Digital-Dan]

Maybe all this explains why the current FSD is apparently pretty safe most of the time, but also very often behaves in a way that is annoying, distressing, or even uncomfortable for those riding in the vehicle, including the "driver." If the powers that be decide that the software has evolved to an acceptable point from a safety and road rules standpoint, they may not push forward to making the behaviors intuitive and pleasant for those occupying the vehicles. Since there's no point at all of cars moving around without people in them (let's say), this is an issue. I have some generic statements, based on a continuous journal of my use of the Model Y since I took delivery in March 2022, about what I think would make the overall driving better, but haven't found a way to communicate them to those who might benefit from it. The high order bit is that my view is the system is underdamped, reacting too quickly and more quickly than is necessary for safety to events detected in its environment. This could be wrong, but that would explain the behavior.

[@ptrsrrll]

You obviously don't drive a Tesla with FSD (beta).

[@CorwynGC]

Improvements are already being made in making FSD more comfortable. I don't see that stopping as long as people have a choice about paying money for it.

[@Digital-Dan]

@@CorwynGC It sure would be nice if there were a better way to report issues that could be useful beyond what they can observe from the raw data.

[@rogerstarkey5390]

Your opening paragraph described "drivers" in general, not just FSD. It's only a question of degree.

[@Digital-Dan]

@@ptrsrrll I wish you could have experienced the FSD left turn from First Street in Los Altos, CA onto Lyell St. (neck will be sore for a week), and then the left turn onto San Antonio, where the car couldn't figure out which left turn lane to be in, so it split the difference. It was a classic example of the wheel twitching right before it jerked left, then corrected a couple more times. Experienced drivers don't do these things. The turn acceleration would definitely have triggered the Tesla safety score detector.

[@nicosmind3]

When it comes to colours your brain learns what they are, and how to perceive them. Which means in different lighting conditions your brain makes corrections so that you still perceive those colours, as those colours, even when the frequency has changed to hit your retina as a different colour!!! That way you're not always perceiving something as it appears at that moment. Think of it this way, a snake that has red as a warning colour, as the sun is setting will appear orange or even grey. If you seen those colours as they truly are you may not be worried as to any danger and die! However your brain making automatic adjustments so that you still perceive it as red, despite your retina not seeing the red, gives you survivability. Your brain isn't even making the adjustments cause it recognises the snake, rather it recognises now shadows and light will change during the course of the day, so it doesn't matter the animal, time, circumstances, you perceive those colours much closer to good conditions than actual current conditions.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@blengi]

the irony is software 1.0 is creating a lot of the data nowadays to train software 2.0 because it's more efficient. ie simulated driving to create statistically unlikely scenarios. Heck even adversarial play is predicated on the application of consistent rules giving rise to statistically reliable sets of possibility. The notion that Neural nets are something "special" is hardly the reality, as simple formal abstractions and linear/algorithmic logic via scientific method and mathematical axioms changed the world/outer space a lot more than trial and error multi hundred billion neurons engines that had pre existed civilization for millions of years. Heck the physical formal substrate of software 1.0 via say transistors did all this using a fraction of complexity. AGI will do the same that is, the actual intelligence will come from emergent non statistical AI algorithms - AI meta cognition - to abstract over statistical hierarchies beyond the limits of human executive function by leveraging the digital advantages of speed, density and scale. The key point is integrated/symbiotic higher order logics not pattern discrimination will be what makes any AI system truly powerful even if is only .00001% of the process, just like one formally oriented human like newton organized all that pattern recognition in his skull to "methodically" deduce that gravity makes the moon dance just like falling apples. Statistical thinking doesn't make revolutions like paradigmatic algorithmic elaboration often does....

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@airbreath]

The subscribe ding sound can be lowered 90%

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@fredkelly6953]

Andrej's a more user friendly version of George Hotz.

[@vwasson6725]

So basically, the neural network has to learn quite like humans have to learn. Exposure to the world.

[@59seank]

First, thanks for explaining what Karpathy was saying. It was very helpful. Second, please put the links to interviews at the top of your description section rather than at the bottom of your insanely long description section. You description section reminds me of the junk drawer in my kitchen.

[@lexscarlet]

That place has to be a playground by now. Like Tony said, top six floors, all R&D. You can either play some run-down high school gym or you can play MSG No place is gonna let him explore like Tesla is. Tesla's a place where he can actually challenge the bounds his genius, instead of coasting on being well above average. Intellects like him don't stay away from a fully-stocked, always-running lab brimming with generational polymaths who never seem to leave the office for long. Bolt wanna run with the Bolts. Aretha's not gonna slum it with the karaoke peasants for long. I really hope he does come back. I think he and that kid who's project lead on FSD now would make a KILLER team. That boy knows his shit. And the team seem very comfortable now, both with each other AND with their exquisitely-refined tools. Man can you imagine how many idea candidates they can iterate through now? How much new stuff on ArXiv they could try? Here's hoping!

[@RavikumarTulugu]

software 2.0 is data intensive and software 1.0 is code intensive for solving the same problem. wonder what software 3.0 will look like ??

[@MattyNiceZM]

That is when AI takes over the creation process

[@rogerstarkey5390]

@@MattyNiceZM Beat be to it. Was going to say "Software 3" will write (is writing?) itself.

[@MacEwanRobert]

Hopefully he returns to Tesla.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@daveblack8831]

Has Elon ever talked about Software 2.0 in the context of Twitter? My guess would be not since an open source algorithm would have to be fully understood by humans.

[@wolfgangpreier9160]

"since an open source algorithm would have to be fully understood by humans" says who?

[@daveblack8831]

@@wolfgangpreier9160 I say that because Elon mentioned making the Twitter algorithm open source for transparency and so that it could be modified by others as needed. I had thought that this would mean it was using standard programming rather than neural nets (but someone please correct me as I am not a programmer).

[@fteoOpty64]

Why is the ? overlay placed right over the head of the Romulan Bird-of-Prey ?. Should be behind, should it not ?. My take from the Lex interview. Andrej might be back for Optimus some time later due to his belief that "embodiment" is important for the humanoid robot to behave very human like in future. I do not subscribe to belief, as I think simulation via video would do the job much more efficiently from a training perspective. Real-time corrections via actuators will be effected on every move, and learning on each of these joints will be dynamic without any need for calibration. I would call it "auto calibration via vision feedback".

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@Cybertruck_69]

Andrej please come back to Tesla. ❤

[@40jart]

💯

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@rogerboone1609]

Sounds like we/he is evolving into the machine that builds the machine....so bots will be building bots!!!

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@richardwolf6269]

How the hay are the going to make a bot 🤖 remotely as capable as a human being?? Just the mechanical aspects of it make it impossible to replicate the human touch and all that a human can do. Most robots are one to two task wonders. Can it prep a house for painting, climb ladders, chop vegetables and wipe your ass without tearing you a new one?

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@CorwynGC]

None of your examples seem very hard. Picking up a cat (without annoying it) is a really hard task.

[@richardwolf6269]

@@CorwynGC you’re kidding me! So the Optimus will do all the prep work and paint 🎨 my house for $20,000? And when might this happen??

[@johnbirk843]

Elon Musk is a nerd! And nerds are the ones who built civilization. Forget the Emperors, Warriors, Queens, Kingst, etc. in history books. They built on top of what nerds made and yet even today most people do not realize this. Scientia Non Domus (Knowledge has No Home) antiguajohn

[@capslock9031]

Scientia non habet domum. Still true, though.

[@brunoheggli2888]

He will never tell the truh anyway!

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning

[@edwardhackett-jones8126]

Who?

[@brunoheggli2888]

@@edwardhackett-jones8126 Karpathy!

[@ogzombieblunt4626]

​@@brunoheggli2888 then tell it

[@brunoheggli2888]

@@ogzombieblunt4626 If he would tell the truh he would be killed from the Elon fanboys!:)

[@michaelweber3836]

At some point there will be Bots raised in VR simulation like children, so their neural network is trained to behave like humans?

[@rogerstarkey5390]

If they behave like humans we're in trouble.

[@knightrider54]

I can't stand Fridman. He is such a hindrance to the interview

[@rogerstarkey5390]

Bye then!👏

[@scottbreseke716]

Better FSD has come out since Karpathy left Tesla.

[@martincossette7420]

Most likely the result of all the work he did there in the years prior.

[@rogerstarkey5390]

Much of the conversation related to him explaining that "they" (humans) concluded that they were not the most efficient "system" to train "the machine". But They could teach it how to train itself. Hence we have "Better FSD since he left" It's a measure of his success! (Duh?)

[@ian_bruh1]

I’m sorry dude but I can’t take you seriously when you say “software 2.0” unironically.

[@niederrheiner8468]

@Dr. Know-it-all Knows it all You commit a copyright infringement! Only sitting there, placing your face in other peoples video and saying nothing over a longer stretch of time is not fair use or a citation!

[@capslock9031]

Bist du hier die Zitatepolizei, oder was?

[@tomcollins1427]

With him breaking it up and commenting on the video Dr.know it all is protected under fair use act….ALL big KZbinrs do it some more than others.

[@niederrheiner8468]

@@capslock9031 Englisch wohl nicht gelernt, was?

[@niederrheiner8468]

@@tomcollins1427 You are only allowed to use other videos as a citation that you comment on. If you play longer parts of the video without commenting that is not fair use.

[@KCautodoctor]

@Nieder rheiner this commentary/reaction video complies with US Copyright law per US Federal Court Case -- Hosseinzadeh v. Klein No. 16-CV-3081 . This specific court case is actually used as an example by the US Copyright Office to explain what is fair use under US Copyright law related to commentary/reaction videos exactly like this one. A key point that the federal judge found was that the commentary/reaction video “does not serve as a market substitute” for the original video.

[@gxd4b1]

Isnt this the guy part of the team that hasnt been delivering FSD since 2014??? By the end of 2017 you can summon your car from New York when you are in LA ....... 5 years later, you can find videos of cars smashing into planes or getting out of a parking space then stopping in the middle of the road forcing the driver to run to the car and manually override it while frustrated drivers behind have to wait. Good job Andrej. (and to anyone that says 'well could you do better?' ... my response is .. I never promised to do better .. thats the difference .. I dont make promises that Im not even close to keeping.)

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...

[@richardwolf6269]

It’s pretty obvious that Musk is full of shit now. FSD looks more and more like a long way off if ever. The only way I see it working is if ever vehicle is equipped with it. Musks timelines are becoming a joke. The boring company is going nowhere. Solar is a joke, the power wall is nothing special. Their cars are well just cars. Their true advantage is in their supercharger network and that probably won’t last too much longer. We have an electric car and home solar with battery backup for less than the cost of the least expensive Tesla. With head twit going off the rails I don’t see Tesla keeping any advantages they have.

[@filip_ji]

👆Thanks for watching, Get-in-touch on whats 'App Lets Discuss on F inancial Planning...