Це відео не доступне.
Перепрошуємо.
Andrej Karpathy on Tesla Bot
Вставка
- Опубліковано 31 жов 2022
- Lex Fridman Podcast full episode: • Andrej Karpathy: Tesla...
Please support this podcast by checking out our sponsors:
- Eight Sleep: www.eightsleep... to get special savings
- BetterHelp: betterhelp.com... to get 10% off
- Fundrise: fundrise.com/lex
- Athletic Greens: athleticgreens... to get 1 month of fish oil
GUEST BIO:
Andrej Karpathy is a legendary AI researcher, engineer, and educator. He's the former director of AI at Tesla, a founding member of OpenAI, and an educator at Stanford.
PODCAST INFO:
Podcast website: lexfridman.com...
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com...
Full episodes playlist: • Lex Fridman Podcast
Clips playlist: • Lex Fridman Podcast Clips
SOCIAL:
- Twitter: / lexfridman
- LinkedIn: / lexfridman
- Facebook: / lexfridman
- Instagram: / lexfridman
- Medium: / lexfridman
- Reddit: / lexfridman
- Support on Patreon: / lexfridman
Full podcast episode: ua-cam.com/video/cdiD-9MMpb0/v-deo.html
Lex Fridman podcast channel: ua-cam.com/users/lexfridman
Guest bio: Andrej Karpathy is a legendary AI researcher, engineer, and educator. He's the former director of AI at Tesla, a founding member of OpenAI, and an educator at Stanford.
ELON and Tesla invented the star wars DROIDS.
TESLA will ship Optimus bot by the Tesla semi full load.
I can't help but really like Andrej. There is a childlike excitement that exudes out of every part of his demeanor. His obvious passion is contageous and that is truly endearing & critical in moving this field forward in a positive direction. Not too many people you could say that about.
Andrej sounds like his voice is always set to 1.5 speed lol
If the Tesla Bot is capable of driving then essentially Tesla would have made every vehicle in the world capable of "self driving".
before that could happen they would have to have solved FSD anyway.
So you're saying it'll be solved by the end of the year?
Exactly. Plug the bot into the cigarette lighter for power. Maybe a set of cameras bolted on the roof, wireless to the bot. Voile. The entire vehicle fleet of the world can be driven autonomously.
LOL The auto-dim & wipers on a TM3 work kind of pseudorandom & your talking about a Tesla-bot driving your car? Why should the bot drive if the car is capable of it by itself? Beside it will never ever do so on anything near current hardware.
@@ChipMIK Nah, I mean the bot drives older vehicles that have no smarts at all. And yes, it seems unimaginable when I ant even get google to reliably turn on the lights in my house.
What is LOVE other than Caring & Doing more for Another than Self
I don't believe it's all that often jealousy or straight-up hate. I think it's often dislike, the reason being that there are limited resources and that it is felt (I feel that way too) that robots and space missions are, presently, a rather significant misallocation. Or as they say in Chinese: Don't go fishing if you're house is on fire.
I do wish we had time for that stuff but the planet is on fire, so no.
"I don't believe it's all that often jealousy or straight-up hate" it's very often just hate, because any of these people, like you, don't have many logical points about why we shouldn't focus on interplanetary travel or such other somehow controvertial points, and when you're just in a logical fallacy, you thinking you're right about what you're saying doesn't turn you from a hater to a constructive critic. We have the resources to do all of these things, especially considering things like space missions make up a fraction of what most governments allocate their resources for, and yet even with these clearly limited resources the rocketship industry comes with innovation after innovation. We have 8 billion people on this planet. Just let people focus no what they're passionate about, and you'll have enough people to care about everything.
I guess he will just get fired if he says anything negative about tesla.
💩🙈 He left Tesla already so in the hell he will get fired 🙈🙈.
@@rj8u I meant canceled. They will both get canceled if they say anything negative about Tesla. That was my bad.
@@fundingsecured7072 Chat me up
Let's talk
Chat me up
Teslabot: " I identity as a car" 🏳️🌈
You transphobic piece of sh*t.
That's what woke mob will call you Suraj. Not me. It's funny to me
Maybe this is what the whole self identification system is about. Paving the way for AI once it’s upon us to let it decide for itself what it wants to be 😅
@@ToshaKhan lol 😂
Please fix all these bots spamming the comments on Lex's channels..
Lex really wants to get physical with humanoid robots. Not knockin him, it's just wild to think about such a future.
Let's talk
Chat me up
At this point it’s inevitable seeing how many lonely men there are
99% of men have zero chance of sleeping with a very attractive human woman. I mean: Would you sleep with a 10/10 humanoid robot or a 4/10 real human woman?
Of course we are still far away from being there yet. Current robots are all weird and uncanny. But in 30 years things might have changed quite dramatically. Even to the point that humaoids are more attractive than real humans (both in terms of looks and personality).
@@jimj2683 yeah I think about it if you wanted to sleep with anybody you technically could using the robot. Although that brings up some ethical situations as in you can't really you know 3d print a full person's body without their permission. Although I think it's possible if anyone wanted to sleep with their waifu they could.
This guy sounds like he’s talking on 1.5x speed
Dear Lex, Re Negativity [7:33] on Twitter etc.: From what I've heard, apps like Twitter, Facebook & Instagram are designed to be addictive. As a result, users get tired, they let their defences down -- and start to say what they really think. They don't realise they're being rude. Best wishes, John.
Yes, and it's about time that they hear what they're saying.
I recommend people watch Thunderf00t’s videos on Musk’s recent announcements. It is a cynical take, but offers a more grounded view as opposed to a rose-tinted vision.
Chat me up
Yup, Tesla bot it’s a joke.
I used to buy in to the hype too, sorry to say now. Thunderf00t opened my eyes on the matter a few years ago with his first debunking videos. Electric cars are the only thing that separates Tesla from Nikola, Theranos, etc.
Saw the falcon heavy launch this morning. It’s a joke and impossible until it’s not
I’m sure you have better things to do then watch thunderfoot, he’s cynical about everything.
This vid should have a disclaimer. I bet they both currently hold Tesla stock.
The way so called Intellectual Dark Web people froth over Elon Musk is fucking cringe.
That’s useful when there’s a conflict of interest.
You win the stupidest comment prize.
I think humanoid robots are basically a solution looking for a problem. Just because humans are bipedal, it doesn't mean robots have to be. Tracked and wheeled drivetrains would probably be better at traversing most kinds of terrain. And articulating robotic arms are only optimised in their capability, when they're purpose built and customised for a specific end-use.
If you want a robot to vacuum your house, you don't get an Optimus or Atlas robot, you get a Roomba. Andrew also says: They could drive a regular car. The same principle applies. You wouldn't get a humanoid bot for that, you would get a Tesla or Waymo or some kind of dedicated autonomous car.
The Boston Dynamics "Atlas" robot has virtually perfected bipedal walking and it makes the Tesla Optimus bot look like a toy by comparison. Yet what is the Atlas bot good for? Basically nothing. Apart from dancing and playing hopscotch that is.
Boston Dynamics however don't use any exaggerated hyperbole to claim that their robot is anything more than a research platform, so at least they're honest about how they characterise its end use.
the benefits of humanoid design are obvious: the modern world is designed for human interface. Doors, stairs, ladders, switches, shelves, appliances... a humanoid robot can fit into this environment without specialized physical infrastructure while a wheeled or tracked robot would have trouble.
The major point you are missing is that everything in our world is made for the bipedal form factor. Doorways, chairs, stairs, elevators, sidewalks, factories, tools, etc. This constrains the problem much more, as you don’t need to fit a new form factor to the world. Sure, there might be more optimal designs for certain tasks/terrains, but the bipedal form will be the easiest and most general to adapt to our world.
Also I guaranfuckingtee someone is going to want one with a fleshlight mount and for that application, humanoid is indispensible.
@@appa609 With great respect, can we agree to disagree.
Re: "The benefits of humanoid design are obvious".
No, they are not. Doors, stairs, ladders and shelves can be traversed in a much more resource efficient and mechanically effective way with purpose-built apparatus for those applications, more so than with a bipedal drivetrain. And people creating specific niche solutions for that, will always have a superior product.
Switches can be controlled and operated much more efficiently with a purpose-built apparatus for those applications, (typically wireless and remote), more so than a mechanical device on a bipedal drivetrain physically moving itself to that switch. A dedicated webcam in a fixed location will always be more resource efficient than a movable camera on bipedal or quadrupedal legs. If you need your camera to be mobile, then flying drone, tracked or wheeled drivetrains will always be more efficient.
Yes there are some few and unique terrains were wheeled and tracked vehicles might have difficulty, but even the US Army found that four-legged quadruped robots were not as efficient as small tracked vehicles for moving heavy loads of cargo. Even if they have to drive a bulldozer through terrain to clear a path for trucks, it's simpler to do that then using quadruped robots.
And with fixed locations involving heavy traffic, it will always be simpler, more resource efficient and locomotivally superior to build a road or platform or ramp and drive a trolley up some stairs, then to make a stair climbing bipedal robot.
How many science fiction films from the 50s and 60s (i.e. George Jetson, etc), show a humanoid robot vacuuming. It was historically inevitable that the Roomba was always going to be the most efficient solution. And the Roomba is a great paradigm for everything that is overhyped about the humanoid form factor as purportedly being the most efficient solution for the mechanical locomotion of an articulating robotic set of arms.
@@CyberKyle With great respect, can we agree to disagree.
>> Re: "Everything in the world is made for the bipedal form factor".
No it is not. Doorways, stairs, elevators, sidewalks, etc, can be traversed in a much more resource efficient and mechanically effective way with purpose-built apparatus specifically built for those applications, more so than with a bipedal drivetrain. And people creating specific niche solutions for that, will always have a mechanically superior product.
Factories, tools, etc, will always have available to them more resource efficient purpose-built apparatus for those applications. And people creating specific niche solutions for that, will always have a mechanically superior product.
>> Re: "You don't need to fit a new form factor to the world"...
My view is that it will always be more resource-efficient and mechanically superior to do just that, than to have a generalised bipedal movable platform, just to support and relocate a pair of articulating mechanical arms.
Switches can be controlled and operated much more efficiently with a purpose-built apparatus for those applications, (typically wireless and remote), more so than a mechanical device on a bipedal drivetrain physically moving itself to that switch.
Sure, there might be some unique terrains were wheeled and tracked vehicles might have difficulty, but even the US Army found that four-legged quadruped robots were not as efficient as small tracked vehicles for moving heavy loads of cargo. Even if they have to drive a bulldozer through terrain to clear a path for trucks, it's simpler to do that then using quadruped robots.
And with fixed locations involving heavy traffic, it will always be simpler, more resource efficient and locomotivally superior to build a road or platform or ramp and drive a trolley up some stairs, then to make a stair climbing bipedal robot.
>> Re: "The bipedal form will be the most easiest and most general to adapt to our world".
Nope, I can't see it. It has multiple pinch points that can amputate your fingers, it is unstable and can fall over an injure people, and hence virtually with most, if not all applications, a tracked or wheeled drivetrain will always be safer, mechanically more effective and a more resource efficient way of moving your articulating arms to the job site.
How many science fiction films from the 50s and 60s (i.e. George Jetson, etc), show a humanoid robot vacuuming. It was historically inevitable that the Roomba was always going to be the most efficient solution. And the Roomba is a great paradigm for everything that is overhyped about the humanoid form factor as purportedly being the most efficient solution for the mechanical locomotion of an articulating robotic set of arms.
PERSONALITY Vs MECHANICAL-GENETICS
Thank you for interviewing Ye and not taking the interview down. I didn’t like some of the things you said but please understand the plight of the black American.
On whatever projects that Lex is working on or studying, I wonder what his next objective/obstacle is that he is trying to overcome 🤔
Let's talk
Chat me up
"I think humans want to be good to each other" and then has Ben Shapiro on degrading Palestinians. crickets
Quick disclaimer: I am not an authority or an expert, nor do I claim to be. I'm just stating an opinion.
Now something you often hear said is that everything in the world has been designed for people to navigate through it with two legs. That's actually getting it all back to front.
From a planetary formation perspective, it's biological organisms that have evolved to navigate their environment, (and not vice versa). And different organisms have developed different locomotive solutions. Some have developed quadrupedal locomotion and humans have bipedal locomotion. Others of course have wings and flippers, etc. And there is nothing special about any of these organisms, or their particular method of locomotion.
And so from a human construction and engineering perspective, we've just extended the terrain and environmental paradigms that exist in the natural world, in a way that accommodates our bipedal locomotion. However that then doesn't necessarily mean that bipedal locomotion is the most superior kind there is for our environment.
That's why we've invented tracked and wheeled locomotive platforms for just about anything that we need to move and relocate itself. And when we take to the skies we don't use wings like birds do, we use propellers and jets. Just because birds use wings, it doesn't mean that we can't engineer a superior aviation solution.
And likewise, just because we use bipedal locomotion, it doesn't mean we can't engineer a superior terrestrial locomotive solution. And it doesn't mean that humans have therefore evolved the "perfect" locomotive solution, which in turn would also be the optimal solution for generic mechanical based locomotion.
Hence, this bipedal locomotion idea is just a very human-centric cognitive bias which assumes that our own biological solution is the epitome of locomotive perfection. Nothing can be further from the truth.
To give you some examples. We're currently evaluating a stair climbing motorised wheelchair (Scewo wheelchair) for someone who has mobility issues. Now in theory, you could put bipedal (or quadrupedal) legs on the wheelchair and have that platform climb up the stairs. In fact, there are prototypes of such a device. And maybe one day this idea will be perfected, just like Boston Dynamics have perfected their bipedal robot Atlas. (n.b. Not to be confused with the Tesla Optimus which is a toy by comparison). Now sure, the legs and torso of Atlas with a chair on top could also act as a stair climbing robot. And also a larger version of their "Spot" quadrupedal robot could climb stairs with a human passenger.
But these solutions will never be as mechanically effective or as resource efficient as this tracked solution with wheels, when it comes to climbing stairs. In short, this is the point! The bipedal robotic solution will never be as efficient at "CLIMBING STAIRS" as a "purpose built" stair climbing solution with tracks and wheels. The fact that bipedal robots can do other "generalised" things like dance and play hopscotch is irrelevant. A paralysed person will ALWAYS want this particular solution rather than a bipedal or quadrupedal robotic stair climbing solution.
The US Army Corps of Engineers also did a lot of research and evaluation into locomotive solutions for all the various types of terrain that they're required to navigate. And on military construction and civil works projects, they evaluated experimental quadrupedal and bipedal bots and found that there was virtually no terrain that couldn't be traversed much easier using just conventional trucks and tracked vehicles. Which in turn had load-bearing capacities greater by many orders of magnitude. Motorcycles were more efficient than a mechanical horse and even small forklifts were found to be more efficient then a mechanical mule.
In short, just because bipedal locomotion works for humans, that doesn't necessarily mean that they will also work for machines in a way that is more efficient than tracks or wheels. Bipedal robots are just a reflection of our own human-centered cognitive bias, and not a deeply analysed engineering solution.
When trying to predict the future, it's not uncommon for people to get it completely wrong. The George Jetson idea of a humanoid robot vacuuming the floor was of course proven wrong. We've developed the Roomba which in turn is a more efficient purpose-built and application-specific solution. The Scewo wheelchair is likewise a similar optimised solution for climbing stairs with a heavy payload. We're also using it to transport small loads of cargo up the stairs.
So hence, the idea of a "one stop shop, all purpose robot" will always be superseded and made redundant by individualised solutions for every type of problem there is, whether it's climbing stairs or carrying heavy engineering loads through the forest, or up a staircase, etc.
And if designing thousands of different "roomba" style solutions for the thousands of specifically different applications ends up being what's required for the creation of an optimal solution, then that's what will happen. And if it requires thousands of manufacturing ramp-ups going into mass production vs ramping up one mass-produced generalized solution, then the bipedal robot will always lose out, simply because machines will always have superior locomotive options available to them, than the bipedal mechanism of a humanoid robot.
can't do that if I know before.like this
Imagine yearly basis computation between Tesla and Boston dynamic, eventually others will follow, it will be the Olympics of robot 🤖🤣😂
Robot olympics would be cool! But I think there should be a requirement that the robots are all humanoids with skin that looks human.
Elon should just buy BD and integrate that hardware and controls expertise into the optimus program. The kinematics is a solved problem. Don't reinvent the leg.
The kinematics needed are much different than what BD has. It's like asking Elon in 2008 to buy Lamborginhi because their cars have the best performance instead of making their own super EV
@@davidsc4680
Funny enough first Roadster was based on existing car, Lotus Elise, which was one of the worst decisions Elon made by his own admission. Too many changes have been made, that it would have been simpler to start from scratch than to use existing one and modify it to your own needs.
The kinematics are child’s play. It’s the software that’s hard.
man that guest talks like i have the video on 1.5 speed
If the Tesla bot has holes for people to "explore", all of the fancy algorithms won't matter.
Hopefully people realize Andrej Karpathy quit tesla July 2022. Have to wonder why
The day is COMING ___Optimus will be available for the Greater Good.
Perhaps we see a New Reality in this decade! Game Changer___ _PERIOD_
Dopamine along the way. @mark zuckerberg
Waiting for Karpathy to go back to Tesla and get on the job of building the AI that will run these bots !
The world (desperately) needs it
And Lex, wow! You hit the nail on the head when you mentioned that jealousy for others will actually harm you! And that Twitter amplifies that and lowers the productivity of the world!! Hope Karpathy takes that that message to Elon, since now they own that platform!
Let's talk
Chat me up
I would be scared af having robot in home on in the backyard doing stuff. Would be unnerving. Safety protocols must be rigorous, like ridiculously safe and robust. I think this will be the problem in the future. Trust in the technology and the company. Because if something goes wrong we have Metalhead case in real life.
What is interesting with that is that you dont require such robust protocols from person you would hire to do the job in that same backyard who could cause allot more harm if he wanted.
Teslabot would have a occupancy network, will never crash/ conflict with Humans in same space.
Another pipe dream
Says another fool.
Me and lex used to fool around under the bleachers
Doing what, building robots?
@@michaellowe3665 lex was a sex bot
,
insightful
I will not be surprised, if Tesla bots become very common in all fields.
Sure, why not.
Vaporware is very common nowadays.
Just an other PR stunt of Musk
I’m pretty sure Elon is an A.I bot from 3022
Keep it up bro
Chat me up
‘They could drive cars’ … lol
Let's talk
Chat me up
It's "brain" can already drive pretty well and non-self driving cars will be around for awhile, so it's not far fetched.
Driving would be the happiest thing for it.
Misleading. Go watch thunderf00t.
Worst example out there.
@@Zripas why so? he is right. elon is a con-man. he lies again and again. he has a long track record of lies.
You really think thunderfoot is better qualified on the subject than Karpathy?
@@Josh92Hughes even if you ignored thunderf00t, the evidence in his videos is solid. Musk overpromises and does shit all. Just look at evidence, no need for any qualifications to get to a conclusion that elon is a fraud.
@@amirhasanbasic2193 ridiculous - look at his achievements. thousands of satellites sent to orbit, millions of cars sold, billions of miles driven autonomously, to name just a few... you call that fraud? what has thunderf00t achieved apart from smug, self-serving UA-cam videos?
Man, this was really Andrej's chance to show he's not a Tesla shill, but he failed the test.
What was false in what he said?
“This was his chance to think like it do but he doesn’t”
Tesla Bot is a joke
It has some real potential.
I was excited about Tesla till Elon went full MAGA. He deserves every hate he is getting and it's his own making.
I was going to buy Tesla but didn't. And I actively stop anyone else in my circle to buy it.
"The Celestine Prophecy"
I predict Tesla bots WILL develop smooth movements and increasingly convincing human form within 7-10 years, far surpassing Boston Dynamics's models, which are dedicated, not general-purpose. we know Tesla never sleeps, constantly works to make both incremental and breakthrough gains, and the "hive mind" enabled by the bots' feedback to the mother ship should expedite "evolution." And the whole project fits so well with Elon's others (paralyzed person controls bot with mind, via Neuralink, for example). Tesla's overall AI prowess and Dojo chip deployment will benefit from every additional user, whether vehicle, robot, or rocket guidance system.
Let's
Ok bro
Tesla bots, tesla Semis, tesla hyperloops, tesla "full self driving"... pay attention to how many projections aren't met and the goal posts keep moving... 🤔
Tesla -Con- Coin 😏
We need to stop giving infinite good will to Tesla just because they work on sexy new things that are hard to do. They have yet to make a 4 wheeled car that can drive safely on it's own from point A to point B. Let's not get ahead of ourselves in thinking they're going to do better than everyone in everything in a robotics domain that's much harder.
*Haaaaa!*
_Elon Musk ain't gonna build shit._
_Well... maybe some day... when he's an Emperor of Mars._ 🙄
lol@ "Building a car is not that different then building a robot" . So much bs...
Autonomous car*
This is so exciting to have Musk driving this effort. Perhaps I will see an android in the few years I have left. Then I will have someone to put flowers on my grave.
Elon should have acquired Boston Dynamics instead of Twitter.
I'm sure there are reasons this didn't happen, and I'm sure this is wrong, I'm just not able to see why.
Let's talk
Chat me up
Sometimes its better to start from scratch than to try adapt existing technology to your own needs. Elon learned this with his first mistake when he tried to adapt Lotus Elise chassis to fit Roadster needs.
Zrips That's the only thing I can really imagine being the case. Maybe the Boston Dynamics approach is just so radically different that spending money to aquire them would be a waste.
@@julians7268
BD likes to use lidar and robots are in general more bulky. So majority of that would need to be redone from scratch which would defeat entire point in buying that company.
If you have your own resources, your own smart people, then its better to just go with it and do it in your own way you actually want to do it. Just compare to how Tesla bot looks like and how BD bot looks like.
"We just need to retrain neural nets". How? For FSD it takes years to train them and still they are not that good as humans. But FSD learns from human drivers that are behind the wheel and correct FSD blunders. Who is going to be correcting Tesla bot blunders? It is driverless by nature
Frankly, I am disappointed by this part of discussion with Andrej. He is not talking about the elephant in the room at all.
How do you train child if you don't have direct control of him? Its self learning AI, it will learn to do things correctly in time. Currently main game stopper is processing power, mainly for the "mothership" servers which processes raw data and trains neural network, but this is being worked on, heavily.
@@Zripas Please do not compare humans to current state of neural networks. Humans have millions of years of evolution behind in their genes.
@@Zripas When you say "self-learning AI" you just show that you do not understand how NN training works. Tesla has massive infrastructure that involves humans at different stages (drivers input, labeling and classification of images), huge offline NN training farms.
And this is just to solve autonomous driving where the input is mostly 2 dimensional: rotation of the steering wheel and the vehicle speed. Humanoid robot has over 20 degrees of freedom just in mobility part, not including arms and fingers. This makes the control problem for humanoid robots at least 2 orders of magnitude more difficult than driving. And we all know that FSD has not been solved, and we do not have the date when it will be solved.
@@andreylukyanov2179
"Humans have millions of years of evolution behind in their genes."
What the hell has this todo with anything in here? Are you saying that 2 things cant be compared if they are not same age? Neural networks of AI are based on brains, we skipped quite a few million years of evolution here while creating AI... We dont need to wait another few million years for us to be able to compare those things...
"Tesla has massive infrastructure that involves humans at different stages (drivers input, labeling and classification of images), huge offline NN training farms. "
Yes... Just like you went to school and you got someone else teaching you things, yet you are still capable and you did self learning part on your own... There are multiple stages to this... Self learning is one of them. Just check videos on youtube about AI learning how to perform task over multiple generations... You can train network, or you can release AI into real/simulated world and allow it to learn on itself, or do both things.
" and we do not have the date when it will be solved."
Because there wont be one... Its gradual improvement, not on/off switch.
Robotics is based on AI understanding environment, what is what and what is where. If AI can understand this then it can perform appropriate tasks without big issues even if you have 5000 limbs. Moving limbs is easy, understanding where this limb is in relation to an object you want to interact isint.
@@Zripas "you can release AI into real/simulated world and allow it to learn on itself" :) Right, are you really sure this is how NNs learn? Learning by itself? (C) UA-cam University :)
Wow, cameras and microphones that walk around the house and record everything we say and do. Move over Zuckerberg. Elon is going to get a medal from the NSA for this one. Meanwhile if I can't block the driver facing camera and still have self driving, I'll pass.
lol@The world is "designed for human form factor"... No it's not.
Yes it is... Environment we made for us is made, well, for us. It's designed by us to fit human form factor.
I wonder if an AI lawyer could win against the gov't for FBA slavery reparations.
Um, what?
Let's talk
@@ChitFromChinola Chat me up