Yann Lecun on Tesla Optimus and humanoid robots | Yann LeCun and Lex Fridman

Lex Fridman Podcast full episode: • Yann Lecun: Meta AI, O...
Please support this podcast by checking out our sponsors:
- HiddenLayer: hiddenlayer.com/lex
- LMNT: drinkLMNT.com/lex to get free sample pack
- Shopify: shopify.com/lex to get $1 per month trial
- AG1: drinkag1.com/lex to get 1 month supply of fish oil
GUEST BIO:
Yann LeCun is the Chief AI Scientist at Meta, professor at NYU, Turing Award winner, and one of the most influential researchers in the history of AI.
PODCAST INFO:
Podcast website: lexfridman.com/podcast
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com/feed/podcast/
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

[@LexClips]

Full podcast episode: kzbin.info/www/bejne/a6WUp4eCipymapI Lex Fridman podcast channel: kzbin.info Guest bio: Yann LeCun is the Chief AI Scientist at Meta, professor at NYU, Turing Award winner, and one of the most influential researchers in the history of AI.

[@JCIIIBULLET]

hello friends

[@Anx3lx]

Kinda wild that a few days after this conversation Figure 1 was demonstrated doing some (very basic) household chores

[@AgentMeech]

I run winding machines for Panasonic- we supply tesla with their batteries- Optimus will be able to do my job with zero re-configuration of our machines. Quality technicians will be replaced, machine operators will be replaced- there will still be need for maintenance technicians- but I foresee full-autonomous manufacturing plants before these units are shipped to homes

[@brandon_youtube]

I don't see that within the next 15 years still. This is very early and the robots today and in a few years still need refining. Maybe we're in the Apple Lisa stage of robots and AI

[@user-hq8lb1yv9k]

Some early adopters will have them in homes, to deliver the specific training data for home application. As with FSD, it will be a work in progress, which is the only viable solution at the moment in my opinion.

[@user-hq8lb1yv9k]

@@brandon_youtube You are way off with your 15 years time horizon, just watch the progress of humanoid robotics within the last two years, it is mind boggling.

[@mellowphantpi9168]

@@brandon_youtube This is a delusional timeline.

[@user-io4sr7vg1v]

...and then there will be new things to do

[@ReflectionOcean]

By YouSum Live 00:00:07 Future of robots in physical reality. 00:01:02 Emergence of robotics industry in the next decade. 00:01:18 Challenges in developing autonomous robots. 00:02:37 Importance of AI progress for robotics advancement. 00:03:00 Humanoid robots aiding in factory automation. 00:03:17 Complexity of automating household tasks. 00:04:47 Direct interaction with AI through robots. 00:05:07 Training world models for innovative progress. 00:06:04 Planning actions with learned world models. 00:07:12 Unsolved hierarchical planning in AI. By YouSum Live

[@jamesatotago]

Like many tech entrepreneurs, Yann judges the world by the standards of how he lives and what he would want from a robotic companion. Imagine an 80 year-old person who just wants to remain in their family home. There are many such people. They are happy to live simply and just want help with basic tasks - tidying up, opening jars and such. We are not that far of humanoid robots that can meet those requirements, maybe 2. - 3 years.

[@shutinalley]

A decade is soon. Hell, i sneeze and 20 years goes by.

[@Mikolaj_Kapusta]

That's a long lasting sneeze.

[@shutinalley]

@@Mikolaj_Kapusta It gets shorter the older you get.

[@capitanalegria]

an era of so many brilliant minds working on ai: lecunn, hinton, goodfellow, bengios, fei-fei li, sutskever, etc, etc.

[@peterepstein4175]

What percentage of full autonomy is required for Optimus vs. for EVs? Might 95% of fully autonomous be enough for Optimus even though 99.99% might be required for FSD in EVs?

[@Michael-il5wd]

Probably greater than 99%. Imagine you made a mistake 1/20 times., that's a mistake every 20 seconds.

[@user-hq8lb1yv9k]

I think he is underestimating the potential usefulness of humanoid robots. It must not be completely humanlike to be useful. For example, if you need to put all ingredients which the robot needs for coking on the kitchen table, place a pot and a pan at a certain place of the stove, to facilitate robotic cooking, it would be already a big help, even if the robot can cook only 30 different recipes. I think we are already very close to that, as some students demonstrated with their university project.

[@lureup9973]

I don’t hear much about what the percentage of fallibility for personal robots that lead to injury or death is acceptable and how do you ever test the almost infinite scenarios that might lead to that kind of failure?

[@stevej.7926]

That’s going to be a fascinating realm of jurisprudence, philosophy, and legislation.

[@user-hq8lb1yv9k]

Not a big problem, it is very easy to limit the motion of a robot when humans come too close. A car is a death machine, knifes are very dangerous, a highway with high and fast traffic is certain death, if you try to cross it.

[@Jokerwolf666]

There will be some Industries where the only humans are the ones that need to do maintenance on the robots you're going to see next to none in dangerous Industries like Steel Etc

[@Xtn1Insecticide]

Loading a dishwasher is one thing, washing the dishes in a sink is another

[@Interzone1]

It’s simple, redesign the kitchen. Then introduce the robot. The same with other environments currently only specified for humans. Dual use environments. Why does a robot need to walk in the street? Upload itself to another robot at a new location, then go to work. Robots will be like telephone boxes - on an as needed basis. Why have one robot to look after an old lady when you can have three or four autonomous devices instead. The hyper focus on humanoid robots as well as human-like AI is a misnomer - AI will look like and think like AI not necessarily like a ‘human’ and nor should it need to in many many use cases.

[@romanzelgatas]

Robots and ai actually is a way to explore humans relationships with ourselves, cold metal will remain cold metal

[@mraarone]

Nobody is mentioning the reason why a parent would need a robot such as having a baby and the safety of a robot around the baby… who will ‘try that’ and not have CPS called on them when they show it off (the robot cleaning dishes… not the baby… maybe the baby).

[@RollandMcGriggs]

Once AI gets even more advanced, I think there will be robot run child daycare. As insane as it may sound. Some people may be single parents or have other valid reasons for needed help whether it's from human run or AI run daycare. As for safety, there no way most people are going to go "Wow the Nannytron version 1 just came out. The first of its kind! I'm going to immediately throw my kids at it!" Some people don't even like simple robots that can carry food and drink in restaurants. Most people are going to wait for the kinks to get worked out but I do believe there's going to be rich tech bros who get these daycare robot prototypes day 1 and it could likely if not will likely spark controversy just as self driving cars have.

[@TheShinorochi]

It so easy for them right now to distinguish between baby and dishes

[@ForceOfChaos1776]

7:52 the hierarchy is interger based? Maybe with primes and merssenn primes. Butchered word but. The formula isn’t easily printed with my iphone

[@Iplayforfood88]

"Vision of the future and the robots!" - Lex

[@gmarie701]

The thing that Yann LeCun does not really seem to grasp yet, is the overall iterative approach Tesla is taking in creating and developing the AI kernel to be used in the Optimus humanoid robot. In his case, this void is not due to a lack of innate intelligence, but maybe to a lack of awareness of the current state of and trajectory of Tesla's progress in creating an AI that can solve for the creation of a Whole World model very rapidly by early participation in that world, and the resulting heuristic cycling that is to occur as the robot's daily participation observations are cycled back through the algorithms being used to produce and cumulatively refine the AI's problem solving capacities in that world. Optimus will appear as a viable commercial and or domestic option far more rapidly than what these guys are thinking. And that will happen by early 2026.

[@jimj2683]

You have been tricked by the marketing from Elon Musk.

[@Clinkety]

I suspect that prediction of yours is way off the mark. I think creating a functional humanoid robot that can perform day-to-day tasks effectively is even more difficult than solving full self-driving. Coupled with the facts that it is very challenging to scale humanoid robots -- severely limiting the amount of training data -- and that humanoid robots are computationally limited, I suspect there need to be several breakthroughs in AI before it is actually possible.

[@BongoWongoOG]

I’m with you. Commercially available in 2026 for sure. I believe that the breakthrough needed is for bots to be learning from few ( or no) examples. They need to reason through a problem set, using existing knowledge bases, and apply those learnings to the new sequence of actions necessary to carry out a verbally (text based) requested task. This has been demonstrated in simple environments, complexity will need compute.

[@user-hq8lb1yv9k]

I agree, but it will be later than early 2026, but quite sure within this decade.

[@Gnaritas42]

@@BongoWongoOG figure robotics is already doing that, the coffee pot demo was from the model training itself watching a human for 10 hours.

[@xxxyyyzzzzzzzzzz]

Imagine the end line of near bright future with robodogs from Boston.

[@ForceOfChaos1776]

3:32 ayo… 😳

[@nightbeezer]

I was quite surprised at how such an important AI professor could read reality so inversely. It seems to stem from the reflex of thinking "They need us." Contrary to what he said, AI cannot create an accurate world model without humanoid robots because there is no data source created in synchronization with its image, voice, and sense of touch. Without humanoid robots, there is also no place where you can monetize that data. The companies working on humanoid robots are dealing with something far more important than everything you've been working on at Meta so far.

[@Gnaritas42]

Nonsense. Embodiment doesn't require a humanoid shape.

[@user-tx9zg5mz5p]

It would be fun to test out if the tesla bot is bulletproof 😂

[@tripleog3567]

Yann LeCon is underestimating Tesla

[@mervstash3692]

Nope

[@user-hq8lb1yv9k]

At the end of this decade we will have for sure millions of humanoid robots deployed. There are much more jobs with simple repetitive tasks than a few million in the developped world. He is vastly underestimating ai. It will be in some areas much superior than humans and in a few inferior, but that will not be hindering its deployment. Chat GPT-4 is already significantly superior in the aspect of constructing a virtual reality that is behaving similar to the real world - see the videos generated only with text prompts.

[@davidperry5326]

I believe that the impact the industrial robot and AI will have on the value of human labor will happen long before the common man will have access to owning one. These robots build by corporations will be built for corporations and labor long before they reach the general public. An unemployed public.

[@ichoudhury007]

Robotics will reach closer to its potential that we envision in general after Neumorphic Computing resolves some of its challenges.

[@thedanielkarim]

What could go wrong…

[@avgjoe5969]

Its too early to address when robots can navigate the world as a helper or nurse. The real benefit in the next 5-10 years will be how well they can do work in the limited world of factories (much easier to navigate). If you can replace 75% of the line workers in 50% of the world's factories and warehouses... then you can sell all the robots you can produce and benefit yourself as well directly (in Tesla's case reducing the cost of their products and having robots building robots). I wouldn't sell to people who could afford $20k robots until I saturated the market for $50k robots with factory skill sets. Individuals are more likely to sue you and giving the robots tasks that require going to the market and interacting with thieves, children, pets etc... too much down side, too little up side as the companies will (in general) pay more for a less sophisticated robot.

[@michelrouleau5755]

Instead of simply selling their robots, Tesla should probably look at a financial Model for possibly leasing/renting them for let's say $15 per hour (maintenance included), for 16 hours a day (2 shifts of 8 hours per day) in manufacturing environments.... 1,000,000 robots X $15 per hour X 16 hours per day X 365 days per year -> $87.6 Billion dollars (!!!) of Top Line Revenue "per year". Compare this hypothetical revenue to simply selling the robots for $20,000/$25,000 per robot + 10%-20% yearly maintenance contracts.... Now, let's multiply this 1,000,000 robot number, by 10..... by 20..... by 50.... The Total Revenue numbers go crazy really fast !!! Food for Thought..... Regards to you All. :-)

[@Demonica1313]

I suggest they test the AI in a virtual home if possible before putting it into a hardware bot.

[@ForceOfChaos1776]

“Aight looks clean type ai” vs “I wanna see you eat off it” ?!? A catalog of all the dishes in your house would be cool 😏

[@hardyvonwinterstein5445]

So in a few years I print my robot at home and download some apps to make him mow the lawn, clean up the shed, repair the roof, dig some trenches and so on. Great. But it will make my wife unemployed. And then...

[@briandoe5746]

Clean unloading the dishes and cooking can all done by mobile alpha. I thought this guy kept up with what was going on.

[@ForceOfChaos1776]

Nice ✂️ lol fr

[@Jai_Lopez]

You a funny dude lex lol 😅

[@richardjohnson1387]

Can AI think?

[@djcardwell]

Yann reminds me of a robot that Google programmed with all sorts of bias built into him (them).

[@AK-np2fh]

Yann gets dissed by OpenAI every few weeks. He says something is impossible in the next 5 years, and openAI does it next week :)

[@peterepstein4175]

Although 100s of millions of robots are coming, why must they be humanoid? Can't AI be embedded into non-humanoid robots?

[@riveteye93]

it can be, but both our infrastructure and our monkey brains make humanoid shapes somewhat more efficient than they should be in a vacuum. Wheels are a must though.

[@steve.k4735]

It can but we built the physical world that currently exists around the human form.

[@darylfoster7944]

The world was designed for humans, so robots can best navigate the world in human form.

[@algorusty]

Humanoid is the most versatile shape for Earth

[@raziele92]

I think it's more of a psychological issue than anything else. The machines in The Matrix started producing non humanoid robots because of utility. They used insects, arachnids, and things like squids as models later on in their existence. Humans, of course, had built them to look humanoid.

[@jacksouthern7929]

With respect your guest seems 5 to 10 years behind from current robots. Learning from video to neuro-net systems. Boston Dynamics level…

[@darylfoster7944]

This guy isn't familiar with FSD v12.

[@flazerflint]

??

[@piuniverse9306]

AI generated images are not that interesting i.e, I can't connect with them emotionally, which begs the question, how will AI cooked food taste?

[@navskygupta5311]

Do you think that commercially available intelligent robots capable of performing like human maids and cooks will be widely available when we lie on our death beds alone and miserable (say by year 2044)? You can ponder over it for over a week before replying

[@chriskramer9311]

I think humakind's obsession with robotics is a futile distraction. I dont think humans need help with dishes or garbage. Its a project with no point. THOUSANDS of human compute hours so that we can glance at our maid-bot-2000 and say "neat".

[@phishingtube]

how lazy are we going to get before the world burns

[@JCIIIBULLET]

🐰silly rabbits, robots are so last year.. - U.S. Government

[@user-tg8be3uw3o]

Is there such thing as a gay-trans gendered a.i. or robot/ics system yet? Forward 50yrs later 2084 I was born a woman but I'm really a man 😮...

[@darylfoster7944]

Garbage in, garbage out

[@bengsynthmusic]

Why would there be such? Robots would be retrofitted with appropriate modules.

[@restonthewind]

I love this guy. After so many hype-rich videos on humanoid robots, he tells it like it is.

[@tacitozetticci9308]

I've heard from David Shapiro that he isn't credible. I still don't know much about him.

[@restonthewind]

@@tacitozetticci9308 Is David Shapiro credible? He says AGI in seven months, so we'll know in seven months.

[@daniellivingstone7759]

He is wrong. World models are being built rapidly. He is a classic skeptic.

[@restonthewind]

​@@daniellivingstone7759 He says that world models can't possibly be complete enough. We don't even have an FSD model handling sufficient edge cases, and the public road network is orders of magnitude less complex than the world a humanoid bot must navigate. That's just a fact. Skepticism of online hype is very appropriate. Optimus is not on the cusp of becoming Mr. Data. It can't possibly have been trained to do what humans generally do in the way that FSD is trained to drive as humans drive. If you know how Tesla accumulated all of this training data, enlighten us. I'll believe it when I see it, and you shouldn't believe it until you see it.

[@tacitozetticci9308]

​@@restonthewind it looks pretty accurate, it wasn't the only prediction he made. Of course, very optimistic but still

[@Adak2030]

dude, who is this guy.? Tesla's very close on FSD, maybe a couple years out. Tesla Bot progress from Bumblebee to Gen 2 in less than 2 years is nothing short of astounding. At this rate of evolution Tesla Bot should be working in factories within 2 years and manufacturing is not even biggest use case. 90 million Americans looking at dying in horrible nursing homes could be given good end of life with a Tesla Bot, aka not end up in a nursing home... I don't even specialize in tech and I can read the tea leaves.

[@FractalPrism.]

musk in 2016 "we have fsd, its ready, we're only waiting on regulatory approval to roll it out, the cars already have all the hardware they need" vs real world tesla - its level 2 it is not level 5

[@davidpacheco5501]

Lol, this is one of the most respected AI researchers in the world. Who are you? I've been following self driving cars for 14 years - Tesla is far away from FSD. Look up the Dunning Krueger effect. Applies to you pretty well.

[@TheHeavenman88]

Dont say « who is this guy » lol . The dude invented some of the algorithms used in deep learning . His opinion is definitely worth hearing out

[@RollandMcGriggs]

​@@davidpacheco5501 you can be respected in a field and also not be 100% always correct in that said field. And I don't think you get to pull the "who are you" card as if you too are not a nobody just like me and most of these comments. Reading articles online for 14 years doesn't actually make you that informed about how the inside is actually going. Your just "informed" about what is allowed to be public. I think people like you truly do provide a great example of the Dunning Krueger effect because you like pretending to be smart and informed just because you've been "following" a topic for 14 years. Consider Moore's law and how technological advances can be exponential. Things are going to pick up speed, and when you're wrong you will give no apology to the people you've tried to correct based on your feelings and not the facts. You reek of emotional immaturity and pseudo-intellectualism.