When will full self-driving be solved? | Andrej Karpathy and Lex Fridman

Lex Fridman Podcast full episode: • Andrej Karpathy: Tesla...
Please support this podcast by checking out our sponsors:
- Eight Sleep: www.eightsleep.com/lex to get special savings
- BetterHelp: betterhelp.com/lex to get 10% off
- Fundrise: fundrise.com/lex
- Athletic Greens: athleticgreens.com/lex 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/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/mZXMdWBvgrKjmJI Lex Fridman podcast channel: kzbin.info 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.

[@cheponis]

Exacty why is he 'legendary'? Seems to me, he's woefully undereducated and certainly under-experienced. He is responsible for for having CAMERA-ONLY software released into the wild, which has resulted in dozens of unnecessary DEATHS. He is, in fact, an accomplice to murder. And you have him on your show and you don't grill him about the negligent murders he has committed? Shame, Lex. For SHAME !!!!! ('76 6-3 here)

[@wertytrewqa]

has anyone else ever noticed that this guy works on autonomous driving, and his name is car-path-ee

[@Michael-il5wd]

nice observation

[@johnseifert6927]

Crentist. Sounds a lot like dentist.

[@nehorlavazapalka]

It's from a mountain range I can see right now. It's not pronounced like that.

[@pdorism]

Yeah Tolkien was such a hack

[@nbarbettini]

Nominative determinism says hi

[@rmmm6725]

I’m not sure if I would really miss driving. If they made cars with beds in them I’d be ok. Sleep every-time I go on the road. Imagine just punch in the address and go to sleep. Amazing

[@NewCastleIndiana]

And keep it going. You don’t need to go 30 miles an hour or more. Think about working from home being your solar, powered van, life, snail, autonomy thing… And going to sleep while you drive to the next thing on the world in the earth. To see. Few years or decades time you should be able to see the entire place.

[@mattstevenson5849]

I have worked in automation for 20 years. Robotic arms with vision systems struggle to differentiate between parts and shadows. Shit, picking parts out of a bin robustly is not well developed. Skynet is a long way away.

[@jimj2683]

How do you explain Tesla FSD beta then?

[@mattstevenson5849]

@@jimj2683 explain what about it?

[@wintonsam496]

Not really, I can fix all those issues. I only ask for a small loan of $100 mil.

[@quintrankid8045]

@@wintonsam496 Come up with a small demo of, say, a 10% improvement to state-of-the-art and you won't be able to beat off the people who will throw money at you.

[@Lalalalalelo]

It's obvious matt doesn't work at Tesla

[@brunogauthier1]

Meanwhile, Tesla is promising Level 5 ready for next year since 2017.

[@kingiam9271]

True

[@Anonymous_Lee19]

Never trust Elon Musk's promises on deadlines and timings...

[@kingiam9271]

@@Anonymous_Lee19 never trust Elon period. You think you investing in Tesla, find out your actually buying twitter 😒

[@tombittikoffer412]

I'm increasingly shocked that my self driving experience in 2017 was so good. At night, 7 or so miles in the city with good lanes marked and stop signs and even going on and getting off highway for a moment went FLAWLESSLY. I feel perhaps my experience was very unique and not typical.

[@cheponis]

wow, it worked, once. Isn't that amazing!!!!!

[@jeffgorchynski]

@@cheponis so you're choosing to ignore the millions of miles driven daily without intervention?

[@cheponis]

@@jeffgorchynski are you actually serious? It doesn't matter if you drive a bazillion miles until the one day in which you smash into a barrier and kill yourself and your passengers. Or worse, bystanders. Are you even aware of what you are suggesting?

[@kingiam9271]

@@jeffgorchynski we ain't talking about cruise control goofy

[@pogo1140]

@@jeffgorchynski how many of those were after Tesla took off the radar?

[@robertcameron1738]

When I compare AI-based self driving to human driving, I see TWO major deficiencies in the AI-based solution: the FIRST is capturing local knowledge of specific trouble spots. Human drivers, when going through a trouble spot will learn and remember the local conditions and apply that remembered specific knowledge the next time they traverse that trouble spot. Whereas AI systems only apply their general rules in repeat transits through the trouble spot and essentially re-encounter it with no local knowledge each time. If cars could learn local knowledge that could help. SECOND: bringing in V2V and V2X car-to-car and car-to-everything communication systems to resolve local conflicts and uncertainties. People do this with hand-waves or shouting or whatever. Cars need to be able to do this as well.

[@LarryPanozzo]

If you listen to Andrej at 0.75X speed, he sounds normal.

[@itsnotallrainbowsandunicor1505]

ROFL, That means we need to operate at 1.25x our current speed to keep up with him.

[@fredrik3685]

Ha ha, you are right.

[@ThrowFence]

@@itsnotallrainbowsandunicor1505 actually that means we would need to operate at 1.33x current speed to keep up

[@cag66]

I don't know if autosteer uses different technology than FSD but my model Y can't drive more than a few KM before it needs an intervention.

[@Aktenverwalter]

Google's AI still doesn't understand me when I talk to Google assistant. Their AI predicts wrong what I wanna type. The KZbin algorithm predicts stuff of which 95% I'm not interested in. So far, AI seems pretty over hyped to me.

[@Venefica]

Right? KZbin thought I'd want watch this video instead of just skipping right to criticizing the premise without actually watching it. That actually doesn't do Lex Fridman any favors because while comments on a video are helpful if they promote interaction, downvotes and non-watches are not a KZbin Creators favorite thing. The worst algorithms on KZbin govern identifying videos under copyright, because they have no actual way of distinguishing samples in songs from the original material, covers of songs, and so on. That means more erroneous copyright claims that can overwhelm people simply trying to create good content, as many talking heads on KZbin have pointed out.

[@cornevanzyl5880]

One thing to keep in mind is that they have to do these things at scale, not just for you but for millions of people many times a day. There are tools that can achieve what you want, but doing it at such a large scale becomes very expensive en computationally demanding. So they resort to basic algorithms that run fast and cheap

[@Buggabones]

YT algorithm is programmed to show you new stuff along side stuff youre interested in. 9.5/10 videos on your front page are def not random videos youre not interested in.. Stop the exaggerating.

[@sebastianreyes8025]

Are you doing this on a Google pixel 6 or 7?

[@Swizzenator]

Sorry but Siri is one dumb bee otch. Most questions and she'll go to the web. But she can set an alarm.

[@rickjones2912]

If self driving systems were based on models where pay as you go vehicles were most prominent then it would likely be with us far sooner. Set routes much like subway lines could service densely populated areas and cost of operation could be spread out amongst users which could generate huge profits for whoever operates the service.

[@musa2732]

This

[@OliverOttoman]

You could also just expand and revamp existing subway lines and save even more money. But i guess it's more fun to dump money into unproven technology for rich people to look cool!

[@ryanleemartin7758]

That's a good idea but it's not self driving.

[@chrispeacock1257]

Then we could save money by putting it on rails and using cars with loads of seats

[@rickjones2912]

@@chrispeacock1257 far less capital upfront wouldn’t you think? Adjusting routes for city expansion could be easily accommodated as cities grow out of infrastructure which becomes dated and inefficient. So much of rail car infrastructure is lost in greedy politicians pockets lined by preferred contractors. No where hide if autonomous vehicles filled the void.

[@SiimKoger]

With all the small little details that driving needs I don't see fully self-driving cars to be let loose during the next decade. 98% self driving? Sure. 100%? No way. Unexpected roadwork, high snow/rain/fog, situations where it's even hard for humans to understand what is going on. Understanding the difference between a child, adult, a group of people, teenagers - predicting how different people might act by the road. I'm sure cars are better drivers on the highway but small townsquarey areas where road looks like everything else and you have to open the window to talk to people? Don't think so.

[@Tranzero]

Is this comment based on knowledge about the current capabilities of fsd beta 10.69.3?

[@zajec11]

Is this comment based on any knowledge of Tesla self driving?

[@bobthebuilderhecanbuildit]

self driving has already been “solved” in my eyes (2 years engineer at Waymo) the difficulty is image, marketing and deploying it in more cities (waymo has to map cities in-depth) the software is already there

[@kufo1000]

Mapping cities is not the end game. mapping cities works short term because the city doesn't change much day to day. the second the city gets a brand new skyscraper and changes the roads around it the mapping does not work in that area anymore. the way waymo is doing things is idiotic in the long term. my Credentials- nine years of coding experience, bachelors in computer science.

[@gregexplorer8901]

He didnt dare to make prediction 😮

[@goodthings5772]

Who is taking interview or whom?

[@EliWallach72]

I don't believe they are even close to being "over the hump" yet. They still need fairly pristine and normal driving conditions in order to function properly. I recall watching a driver testing his tesla on an overpass where you could clearly see that cars had stopped ahead beyond the bend, but they were mostly hidden by the concrete wall. He let the car go but it was moving fast and not slowing at all so he had to take over. They are nowhere near that kind of reasoning. But, what do I know. I do believe however that we will have decently good self-driving before we have commercially viable fusion reactors, which I don't expect until late this century, well beyond my personal timeline.

[@IonorRea]

The question is if FSD can be solved with the current configuration and hardware, personally I don think so, and here is why: Vision of cameras right next to the glass similarly to vision through prescribed eye glasses get far more affected even by few droplets in front of the camera than when you looking at windshield half meter away from driver's position and hardware is not capable enough to compensate for smugged image in a way satellite pictures can be reprocessed to predict detail when smudged shape get reproduced to most likely jet fighter limited visual data point towards through machine learning trained on low resolution images of known objects, that kind of processing power in real time at least 30 frames per second is simply not there in any Tesla current car only on supercomputers and due to placement of cameras next to glass while lacking thermal cmaera to complement lack of processing power and bad sensor placement in bad weather means that government unlikely ever approve Tesla cars in current hardware configuration for full self driving without driver constantly watching over car not matter how many new Tesla FSD software updates will come. For cameras to have decent enough visibility even during rain they would need to have comparable visibility to a driver which would be easiest to do by placing cameras between front seats (perhaps on a retractable arm inside the car activating only when necessary) or into the co-driver seat (which would make it unusable during autopilot). Alternatively and likely the most realistic solution, just put a thermal camera which cannot be blinded by incoming light sources so easily unless having a high thermal signature apart from not being affected by rain or fog very much even when mounted on the windshield or directly above it like LIDAR on current Lucid cars and enforce by law new road signs and road surface markings that are visible on thermal cameras which can't see "colors" (different visible wavelengths reflection and emitting properties) but only see differences in materials if they have sufficiently different thermal properties in order for the thermal camera to by reasonable bad weather replacement for visible wavelengths cameras mounted on the windshield that become useless during heavy rain to a point where reading street signs become unreliable to bet human life on it...

[@MrVTeta]

I too was fooled that FSD was very near. Nowadays I think a personal car without a driving wheel is at least 20 years away. There will be no revolutionary leap, forget about that.

[@itsnotallrainbowsandunicor1505]

With more people being interested in robotics, automation and autonomous vehicles, it may happen sooner than 20 years.

[@mrt2895]

20 years is still nothing

[@mearino]

I think constrained environments are the key. For example, highway driving in good weather is largely solved as long as traffic is predictable and no animals jump out. If you control the weather and animals, driving becomes so much simpler eg in a tunnel. Boring tunnel. I think tunnels combined with FSD have a very good chance of working together.

[@mubbywub]

The problem they’re trying to solve is environment independent, which is the right way to do it. Tesla can’t change roads, but they can change their cars.

[@mearino]

@@mubbywub I just don't see them being able to solve all the variables right now. I think we need to add electronic signals to roads and signals that cars can read (like big QR codes hah hah), dedicated lanes and other ways to control the environment for at least some aspects of the drive to be properly autonomous.

[@Nadzap]

I see you've fallen squarely into Musk's aura of delusion. How can anyone seriously believe in the tunnel idea?

[@howtomanagetech]

@@Nadzap Have you looked into the tunnels in operation in Las Vegas? What aspects strike you as delusional? They have succeeded in a number of important ways. The tunnels can be drilled at a smaller diameter because they don't have to exhaust fumes. This saves money. The boring machines have been electrified, and are being steadily roboticized to increase tunneling speed. The cars within the tunnel are driven by humans as a fire regulation requirement. I'm doubtful FSD could handle the passengers pickup and take-off but I can think of ways they could mechanize that if the cars couldn't be programmed specifically. As for travelling within the tunnels, it was demonstrated that the cars are more than capable of travelling at high speed by following the two painted lines in the tunnel. Its definitely early days but I would say the tunnels show promise, and are anything but delusional. Where does your research reveal otherwise?

[@Nadzap]

@@howtomanagetech Have *you* seen the Boring Company tunnels in Las Vegas? It’s an absolute joke and nothing like what they were initially contracted to build. There are countless articles about why it is a slap in the face to public transport and a fundamentally flawed idea. Even with the improvements in cost you mentioned, building out an entire single-car based infrastructure with this “technology” in America would cost trillions of dollars. And that does not account for the many unanswered questions regarding the safety, viability, and maintainability of these tunnels. If the Vegas project is anything to go by, those questions are likely to go unanswered for a long time.

[@darktower74]

I think one of the things that will bring about fully autonomous driving will be ubiquity. When all vehicles on the road at least have the function available. Basically, even if a vehicle can perform flawlessly on a test track or optimal conditions, it won't necessarily mean...poof... now we're all driving autonomously. If half the cars on the road still have the chaos factor of humans behind the wheel, we will find a way to fuck up the perception of autonomous driving. That lowered perception will hinder full legislation. If everyone HAS cars that can do it, I think it will become a thing.

[@phaandorpertwee6981]

there will be a moment like when I saw the first Falcon 9 Booster landing flawlessly which literally brought me to tears - and from that moment on FSD will just be a normal thing.

[@HistoricalGaming]

I dint think it will be solved until roads have regulations enforced that help keep the cars aware if conditions. Too many confusing lines in the road and lack of lines at times are too confusing

[@craigkdillon]

I have not seen or heard of the liability issue being addressed. Accidents will happen. How will liability be handled? 1. People are killed in another in an accident who is liable? Seems to me the manufacturer of the self driving car must be held liable. Today, the driver is held liable. Some even go to jail for negligent homicide. Future victims will sue manufacturers for compensation. How will manufacturers deal with that? 2. A self driving car relies on properly working sensors. Sensors can go bad. Will sensors work when covered in mud or slush, which you get a lot here in the Midwest. Also, can the sensors sense street lanes and boundaries in snowy and rainy conditions?? What about in low light conditions -- street lamps not working, or not present?? What about country roads?? Dirt roads?? 3. Can self driving cars deal with deer or other animals in the road? For me, the biggest question is liability --- if my car is involved in an accident, while I am a mere passenger, I don't want to be held at fault, just because I am the owner. Elon Musk should be held liable and pay the money for this pointless invention.

[@untouchable360x]

"Maybe that is a question you should ask Chy-na." DJT

[@JohnSmith-ut5th]

Wouldn't that be "Jy-na"?

[@benstrachan3418]

Regarding self-driving: I don't understand why they don't take the easier option. Popular Non-Highway routes are created on existing roads with colour-coded markings. Need to meet certain criteria from Tesla but local government implement. Don't stop long term goal but make short term progress.

[@milescoleman910]

Humans haven’t mastered driving yet. We crash everyday with bad decisions and stupidity and bad eyesight etc. don’t know if we will ever think they are good enough.

[@sicaris415]

Majority of people have never master anything in their life, even after 10k hours. Too complacent

[@movingsocks579]

My Tesla often tries to throw me into oncoming traffic in the Full Self Driving Beta. It does well most of the time..though.

[@auricschiller]

No big deal right?

[@movingsocks579]

@@auricschiller yeah I've learned to be ready to disengage on any turns that involve waiting for incoming traffic to clear. In ideal conditions it works flawlessly.

[@googlyboobly51]

just seeing if you are you paying attention lowly human 🖥💻🤣🖥💻

[@bossman4401]

As long as it only try to trow you into oncoming traffic sometimes, thats not too bad right ?

[@kingiam9271]

@@movingsocks579 sounds like cruise control that cost thousands lol

[@theburnetts]

I’ve watched videos of Waymo in Phoenix and that fleet of vehicles seems to be fully self driving without any real issues. What am I missing? Are the conditions in Phoenix not varied enough to really be FSD?

[@patrickoneil2973]

Waymo pre-maps the area using HD lidar scans. The reason why people don't consider that "solving" autonomy is because that approach doesn't scale. If the map changes they have to remap the area. Waymos don't work unless they are in their tiny HD mapped area. Teslas work everywhere but not at the reliability that Waymo is at. True autonomy would be Waymo reliability but anywhere.

[@theburnetts]

@@patrickoneil2973 gotcha. Thanks. I figured it must be something like that. Appreciate the answer.

[@mithos56]

@@patrickoneil2973 That was may question as well, thanks for answering it!

[@DurangoC]

It’s a playpen, everything completely mapped to the centimeter. Has nothing to do with real world self-driving.

[@mho0]

@@DurangoC If it drives itself why it is not self-driving?

[@tschulz]

My new Tesla Model Y does not even understand the very bright LED speed limit signs on the German autobahn. Not to speak of the interim speed limits at construction sites. Identifying a traffic sign is a straight forward classification task. If even that doesn’t work, how would you trust the autonomous driver to make logical assumptions on the behavior of other traffic ( pedestrians, especially children, cars,…)? Speaking with Daniel Kahnemann, I guess we must solve both fast and slow brain functions. the state of either is dismal as of November 2022.

[@lapin46]

the signs in Germany are plainly not coded properly and trained in the software. It has nothing to do with system or software capabilities. All the focus is on the US for now unfortunately. They could at least implement proper sign recognition.

[@monsenrm]

I hate to say it but I think it might be 5 or more year away. I have FSD and it is about 5% done.

[@logofthelex2668]

I want to see a computer navigate the narrow back streets of a coastal town or even Boston.

[@hooshisar2435]

Here are the hectic streets of San Francisco. This was about 4 months ago as well. The system has gotten mildly better since then. kzbin.info/www/bejne/nKjHpptooLedaa8

[@brentwgraham]

Tesla AP doing San Francisco is pretty good.

[@EugeneHaroldKrab]

You will lmao. Sooner than you realize

[@user-qc7dr3cd3q]

Or white-out snow and ice conditions on rural Great Lake's snow belt roads.

[@logofthelex2668]

@@hooshisar2435 make those streets two way (instead of one way) and you get an idea of what a truly old fishing town is like. That's true for both new and old England.

[@user-kr4jl6en7r]

I wish he would have just answered the question. Just give an estimate. We know it's not 100% certain, but if anyone is in a position to give an educated prediction it's Andrej

[@mho0]

That should tell something about how soon he expects it to come. Probably not 'next year'...

[@IonorRea]

The question is if FSD can be solved with the current configuration and hardware, personally I don think so, and here is why: Vision of cameras right next to the glass similarly to vision through prescribed eye glasses get far more affected even by few droplets in front of the camera than when you looking at windshield half meter away from driver's position and hardware is not capable enough to compensate for smugged image in a way satellite pictures can be reprocessed to predict detail when smudged shape get reproduced to most likely jet fighter limited visual data point towards through machine learning trained on low resolution images of known objects, that kind of processing power in real time at least 30 frames per second is simply not there in any Tesla current car only on supercomputers and due to placement of cameras next to glass while lacking thermal cmaera to complement lack of processing power and bad sensor placement in bad weather means that government unlikely ever approve Tesla cars in current hardware configuration for full self driving without driver constantly watching over car not matter how many new Tesla FSD software updates will come. For cameras to have decent enough visibility even during rain they would need to have comparable visibility to a driver which would be easiest to do by placing cameras between front seats (perhaps on a retractable arm inside the car activating only when necessary) or into the co-driver seat (which would make it unusable during autopilot). Alternatively and likely the most realistic solution, just put a thermal camera which cannot be blinded by incoming light sources so easily unless having a high thermal signature apart from not being affected by rain or fog very much even when mounted on the windshield or directly above it like LIDAR on current Lucid cars and enforce by law new road signs and road surface markings that are visible on thermal cameras which can't see "colors" (different visible wavelengths reflection and emitting properties) but only see differences in materials if they have sufficiently different thermal properties in order for the thermal camera to by reasonable bad weather replacement for visible wavelengths cameras mounted on the windshield that become useless during heavy rain to a point where reading street signs become unreliable to bet human life on it...

[@FractalPrism.]

"FSD is basically solved, we're only waiting on california regulatory approval, we expect to roll out FSD by EoY 2016" -musk in 2016 here we are in 2022 and their so called FSD price has gone from $5k to $15 and still does not work.

[@CDX7]

What a laughably lazy analysis lol

[@viksusername]

It's solved indeed but not by Tesla...

[@DurangoC]

Actually it does work. Mostly. Watch the recent videos on KZbin - they’re very close.

[@Michael-il5wd]

everyone is a critic. FSD is literally the hardest problem humanity has ever attempted. What role do you play in this?

[@reginald7214]

😆 🤣

[@mayetesla]

Maye the force

[@BekTesfa]

Thank you for asking this question. He believes the problem of FSD is tractable by expert intuition. Clearly this thing will determine the fate of Tesla.

[@jeffsteyn7174]

What's the use asking the guy that failed at it.

[@b21-soaring]

"When" is the KEY question, and I've asked it of over 400 people since 2015. Instead of "What do you think is the timeline..." and getting an essay for a non-answer, ask "In what year will you be able to buy or hail an autonomous car, get in it at home, take a nap, and wake up at work?" (hint: the answer is a number). In 2016 the common answer to me was 2021, and since then the answer has been going backward in time faster than real time. Actually the common answer in 2016 was "I think Tesla can do it already, but the legal and insurance framework isn't in place yet". Obviously there is zero chance any Telsa currently manufactured will have full self driving, but the industry persists in giving waffly non-answers to the "when" question to keep that marketing myth alive.

[@rolisreefranch]

Q: When will full self-driving be solved? A: About 5 minutes before the government bans human driving

[@waffles23456]

Thats too real

[@techfan6129]

Actually once self driving is fully solved, it should be safe for an 8 year old to drive. The system could intervene enough to keep them safe and from being a road nuisance.

[@DurangoC]

It’s insane that we allow humans to drive. 50k dead per year in the US alone.

[@rumfordc]

@@techfan6129 well yea that is what "fully solved" means...

[@charlesalexanderable]

So.. he describes the poor state of things in 2017 yet Musk promised it by the end of the year that year...

[@CapnSnackbeard]

When will Elon Musk predict it will be solved? I mean the next time he gets it wrong, not all the other ones

[@Venefica]

I think it's time someone replaced the 8087 in Elon, it seems to be spitting out unforced errors. Either that or he needs a clean install of his OS with some updated code. Maybe both. Beep boop.

[@mho0]

We all know that: "next year"

[@VinnyTheory]

Finally a new video about the future and automation!

[@nicholasshook7513]

When we redo all the roads

[@charleschen4766]

This is a solvable problem and it is just a matter of time. We all know what computers can do and can improve the performance exponentially. The computer can beat the best Chess player and the best Go player. There is no reason to not believe that it can perform better than a human driver.

[@StefanNorberg]

An open world problem cannot be compared to Go or protein folding. 99% of researchers will tell you computer vision is not ready for safety critical systems, not even in radiology. The S-curve is real and every decimal of reliability gets harder and harder. Camera only autonomy for personally owned vehicles will likely not happen before 2030, and I don't think it will happen in my lifetime.

[@teflonheadbuddha]

It will need a new infrastructure. Adding AI to an analog system is working backwards.

[@thomasm.7058]

I think we don't need self-driving at all. I use for longer distances, mostly the train, and all the other drives, it will get too complicated and at the end controlled too. What most bother me is, when they think it's far enough developed, they will remove the self steering in the cars. And you can't even freely choose the destination without interference. It's all about control... Be carefully what you wish for...

[@nycgags]

So the answer to the click bait title, is there is no answer, today.

[@allhitstaken6200]

This is where he is wrong. 99.9% of the self driving problem is what they call “edge cases”. But he and they think edge cases are only the 0.1% of the problem. So FSD is actually 999 times harder then he thinks. That’s why FSD has been, and will continue to be, delayed and delayed and delayed.

[@Nick-kh6vf]

Yeah dude, you certainly know more about AI than one of the world's leading experts! You should take over his old job man. Genius!

[@cag66]

Yeah, It feels like the difficulty of FSD is not uniformly distributed over the driving task. It basically really easy most of the time and extremely hard in a small number of situations. If you can't solve those small number of situations and they occur with any sort of regularity then the technology is useless. I feel like tesla are still trying to solve the easy part of FSD and haven't even addressed that important hard parts.

[@Joker1531993]

you hit percentages like a expert, why you not solve it if you know how hard is it, dudeee.

[@michaelfletcher940]

Lex set us back with his interview with kanye

[@J.D.Vision]

I'm legally blind, and can't wait till A.I. is advanced enough for this dream to become a reality.

[@OcaOca]

This guy taught millennials how to rubix cube as badmephisto on youtube, now he’s leading self driving wow

[@pinochio8778]

Until we can code consciousness....

[@bossman4401]

Most people seem to think only tesla do self driving, when in fact there are cars like Mercedes that does it even better right now. As a Tesla and Mercedes driver I have used both.

[@techfan6129]

Really? Mercedes has a system that drives on city streets, just like my Tesla did tonight?

[@Finnspin_unicycles]

@@techfan6129 That is in large part because other manufacturers don't put a beta systems on the market. They geofence it to the regions where it is known to work extremely well. That's also why Mercedes is currently the only manufacturer (as far as I know) who assumes liability for their system. Don't get me wrong, Teslas system is very impressive in cities, but I'm 99.99% sure that if most Auto manufacturers had a system at that level, they would not roll it out - still too many mistakes.

[@techfan6129]

@@Finnspin_unicycles And since Mercedes does not have have a City Street’s capability (regardless if your opinion on the reason is true or not), then the two systems are not comparable. I.e., your comparison was invalid and suggesting you somehow magically know Mercedes COULD do it, but is too conservative, does not change the fact that the two systems are not comparable (even if I grant your magic power to know Mercedes’ thought processes).

[@rumfordc]

​@@techfan6129 even little kids can drive on city streets once and get away with it. OP said other companies "do self driving" and he is correct. you are the one trying to draw inappropriate comparisons.

[@techfan6129]

@@rumfordc What? He claimed it would be smart to ban humans because they would be less safe than the software. I’m claiming there’s no need to ban human driving, because the software would probably be capable enough to intervene when necessary to make the human driving safe. Like having a copilot that can instantly assume control. Therefore even children could safely drive.

[@Attorneysean]

Need smart roads with sensors.

[@rumfordc]

we have those already. they are called trains.

[@TimUckun]

Elon is too busy driving twitter into the ground to work on self driving cars.

[@ednauseum3060]

If you ask me, edge cases like fog should all be put in a bucket, where the vehicle pulls off the road & stops, declaring "I can't do this". And the MAIN cases are solved, while the edge cases are saved for later upgrades. But then, I think logically, and use common sense.

[@Assywalker]

You have no clue, what you're talking about. FSD struggles with completely different things than fog. The fog he mentioned was figurative speech and a video game analogy.

[@quintrankid8045]

@@Assywalker Actually Fog of War is an expression that predates video games by around 70 years. According to the intertubes.

[@ednauseum3060]

@@Assywalker It's a comment made by Lex, not Andrej. (paraphrase) "You're on a mountain, in the fog..." Lex says this +/- a minute from Andrej's "fog of war" statement. Yes, "fog of war" is different from being in the fog on a mountain.

[@Assywalker]

@@ednauseum3060 Sure, but they both do not talk about the actual act of driving with that phrase. It's about the progress und development strategy being hard to evaluate on a problem this complex and new. No one said, that the car can't deal with fog on the street. Andrej and others have talked about the core issue at other occasions and it's quite simple: A driver of a car has to consistently do inputs, that don't lead to an accident, especially in an urban environment. When you misinterpret things for even a second, you can crash the car. And every encounter is unique to some extend, so you never really know if the system can handle it correctly or not.

[@Assywalker]

@@quintrankid8045 Interesting :) According to wiki it come from a Prussian book from 1832 (in German) or its translation from 1873 (in English) by a military analyst called "Carl von Clausewitz".

[@chrisbarry9345]

The issue is that if EVERYONE was self driven it could work fine. The cars could "talk" to each other to keep things safe

[@Finnspin_unicycles]

I would like to be allowed to walk on public roads please. Don't just think about cars when you think about traffic.

[@rumfordc]

wrong. there will always be variables such as animals, potholes, weather, and vehicle degradation

[@freedom_aint_free]

I think that when self driving will be totally solved, the computer science field will have advanced so much that in the process we will have solved way more important problems that will see self driving as toy project (specially the neuro science behind how humans actually drive, as even kids can drive a car after some training and yet the biggest brains in the world with billions of dollars so far were not able to crack that secret, what means that the treasure behind it will be epic), although the economics implications will be far reaching and irreversible, will be like going from steam engines to internal combustion.

[@Aaron25thinfantry]

It's gonna get expensive

[@SevenDeMagnus]

Hi, full autonomy should just be a standard not a goal (like be perfect for God is perfect in catholicism but one will never be perfect, it's a great standard though). It might just be Level 4.99 at max, which may take a few years until the end of time (which could anytime but it's best to put at trillions of years from now), A.I. does not have a mind (an intagible that can't be in code) but Level 4.99 is a good tool but will still be babysat by us humans but carry on, got for Level 5, fully autonomous so we can just chill while the cars and Optimus earn for us in the background. :-) God bless.

[@vimalnelson4378]

It is not possible . The hardware is not ready for it and software to do it… I can understand what Elon is doing.. making an effort and moving forward to achieve something.

[@Ant3_14]

Current 10.69.3 looks pretty good to me. Needs to be verified in edge cases but there was video it doing 100 miles without intervention.

[@K1RTB]

No Tesla on the road today will ever be Level 5. Tesla already has your money. They have no incentive to take on all this liability that comes with real full self-driving.

[@koontekinte0]

from within the autonomous driving industry: I call bullshit on tesla. level 3 is coming soon (not to tesla, though, I believe the first level 3 car on the streets will be some Chinese auto maker), but level 4-5 are so far away it is useless to even think about it. let us put a floor on this timeline - at least a decade before level 4 (assuming capital flows to the industry are kept at some level that enables employing ~30000 engineers in the different companies). I have seen what autonomous vehicles are capable of and I can tell you that if not geofenced, they amount to nothing. yes, they work perfectly 75% of the time, but that is not nearly enough to be considered safe. the lack of accountability, the lack of regulations, the sheer apathy of these companies to human lives will ruin the industry.

[@u-d-0-n]

FSD beta has been improving fast this past year, with all the new teslas hitting the road the speed in which FSD will be solved will improve at a more significant rate. I don't think the growth is going to be linear. The past 5 years has gotten it to this point, this year the growth was rapid, but the next couple of years we may see some truly amazing things from FSD.

[@anubizz3]

Yeah right I ask undecided if Tesla can detect pothole, he said next version. 3 years later still next version.

[@vankiv3627]

Progress has plateaued, my friend!! You don’t have good memory of where it was last year.

[@kingiam9271]

Tesla been the same crap for years now. It can fart when you honk the horn tho

[@u-d-0-n]

@@vickykennel7414 angry karen spotted please call animal services

[@jeremylink7701]

If he cant in anyways give a timeline even from intuition shows they are nowhere near solving it.

[@RunForPeace-hk1cu]

It will get solved ... eventually ... but no one should be paying $15k for a feature that doesn't work as claimed .... it will not do FSD in the lifetime of the current Tesla cars. It's a goFundMe.

[@mho0]

True, but it is everybody's own choice to invest that $15k or not (I consider quite horrible one but everybody can do their own decisions). It is just quit fascinating that some people do ... maybe they just have too much money :)

[@RunForPeace-hk1cu]

@@mho0 not true . Tesla is misleading consumers by calling it FSD. There’s nothing self driving about it. The car holds no legal liability. The driver does. The proper terminology is “Drivers Assist” it gives the consumer a false sense of security and it should be illegal

[@mho0]

@@RunForPeace-hk1cu Basically agree with you...

[@fredturk6447]

The interviewer talks too much. Please let Andrej talk he’s who we want to hear from!

[@sungam69]

FSD means it safely drives your kid to Kindergarten while you mow the lawn.

[@mahadevovnl]

Full self-driving isn't going to happen because we don't have the computing power to process the data. Computers can't be creative problem solves like humans can be, not for a long time to come.

[@georgekyria2619]

May I gently point out, is it not a transparent delusion to say "no one has created self driving", if Waymo/Cruise etc have for at least one year been charging the public cab fares to drive them around in a fully autonomous . . . completely driverless vehicle . . . . known as a Robotaxi ! In Phoenix and San Francisco . . . . and Austin etc by the end of 2022 !!

[@DurangoC]

That is not self-driving, it’s a playpen, like a train on a track. A train that cannot leave its track can drive itself around in a circle. It can even charge for it. But that’s not self-driving.

[@georgekyria2619]

@@DurangoC You are plainly wrong, like a flat Earther, and like a flat Earther, facts dont matter, have a nice life.

[@brandodurham]

Self driving is 💯 not going to happen with just the crap cameras Tesla uses as hardware, or the current CPU’s. It doesn’t even matter how good ai is or isn’t. Even with perfect ai, and perfect weather, the latency is not good enough without much more powerful processing

[@thomaskong5]

That was the longest none answer ever

[@dd-ow6pe]

they will solve it .. just matter of time.

[@Dan.50]

It's a jobs program. It ain't never happening.

[@TheRitualChannel]

I hope they get human drivers out of the equation and see how that works.

[@rumfordc]

uhh thats exactly what they tried and it _didn't_ work which is why this conversation is happening!

[@ChipMIK]

Believing FSD ever will be on anything near the current hardware thats in a Tesla today is like going 70 years back and believing in cars would be nuclear driven in the 1960´s yet people still fall for the BS and not realizing its their own dream they pay for instead of a driveable car with propetary sensors and automations that actually works.

[@victordegrande1628]

Is Tesla even the leader in this technology? Musk makes lots of predictions that are always too optimistic, but Waymo has driven far more mileage.

[@vanrozay8871]

I doubt Elon would have discarded radar if he had ever driven into a sudden extremely heavy rain shower (always short-lasting) as I have 3 times in a long life. Under such pounding, I could not see the front of my car's hood, no less vehicles ahead of or behind me. Each time, slowing to a crawl, I edged toward the shoulder, luckily never hitting or being hit. But I'm certain that NO driving system using only cameras could have handled such conditions better than I did. Meanwhile, a radar would have shown me vehicles near me, making it much safer. Someone should make Elon aware of this issuie, which may never have been considered as a factor in FSD capability; why NOT have a car with BETTER than human "vision?"

[@DurangoC]

Because competing information between radar and vision sensors creates lethal confusion. The problem you’re describing will be solved when cars talk to each other - not that complicated.

[@mho0]

@@DurangoC there will be only lethal confusion if sw decided that. You could probably do a sw which does not make lethal confusion out of that, right? Anyway real reason is cost saving...

[@chunksloth]

I think radar and maybe even LIDAR are inevitable on Teslas because they DO offer superhuman capabilities. However, I think Tesla is correct to do only vision right now. The problem is so large that it is better to solve just one part, vision, before solving the rest. Once vision is solved then adding the other sensors and training using vision as a guide should be relatively easy.

[@vanrozay8871]

@@chunksloth Thanks for thoughtful reply. You are right, of course, but I fear Elon may have closed his mind to non-visual sensors. Maybe it will take a horrible fatal failure of FSD to shake him up. And if he were true to his ideal of always improving everything, why not employ sensors that assure BETTER "vision" than human?

[@rajaravivarmar]

Andrej "Karpathy" "கார் பத்தி" பேசிட்டு இருக்கார்!

[@pencilsandlight1318]

Does anyone really want self driving cars? Many people enjoy driving!

[@MMATopGs]

I think never

[@Swizzenator]

Tesla should be testing in Asian countries.

[@boosthit811]

So much fud in this comment section

[@adelinaquijano1083]

I know but I'm not own the car

[@structuralengineering3406]

ask the leader who just quit the game? lol

[@tomjensen618]

Elon predicted 2017?? No smart money on that one! Going from LA to New York without assistance is far in the future.

[@XAP113]

lost lost boys ,, Robert Bosch GmbH are on track

[@ziqueez]

No one has built autonomy? You're making Waymo and Cruise with their autonomous deployments sad.

[@Venefica]

Full Self-Driving is never meaningful going to happen outside of being a hyped-up marketing term. The people promoting that term are being dishonest with who they promote it to, and perhaps themselves. That's "when". The best way to ensure a vehicle is consistently guided to the location it's meant to go is simply to put it on rails. If you want the ease of not having to drive, your most realistic option is simply to put money into freely-available public transportation.

[@kuruman1]

People are unpredictably stupid. If all cars could be autonomous at the same time it probably wouldn’t be so hard, but developing autonomous driving that accounts for idiots is infinitely more difficult.

[@FlorinArjocu]

Ask mr. Musk, it will be level 5 ready in 2.17-2018-2019-2020-2021-2022 and so on 😀

[@teddysthaiadventure2534]

Its coming on all Teslas in 2016 according to elon

[@FlipBoxStudio]

He’s already admitted that he was wrong and that solving true FSD is harder than he originally assumed. He’s also changed much of his language in terms of what their FSD system will be capable of when fully released, which is that it still require a driver to take over until the software is perfected and regulators allow them to operate without a driver. Keep up buddy.

[@teddysthaiadventure2534]

@@FlipBoxStudio Yet is still selling "full self drive". There is plenty of vids showing him saying they already have it working and he will release it in a few months. Every year for the last 6

[@kingiam9271]

Nvidia will solve automation before Tesla does

[@liamvogel9151]

How come

[@kingiam9271]

@@liamvogel9151 they are running simulations to be able to indetify real world scenerios. They don't need real world data when they can simulate it. They will be providing the software to all car manufacturers in the next 2 years

[@liamvogel9151]

@@kingiam9271 Tesla does that too. I think real world data is needed as well. The world is too unpredictable and inconsistent

[@kingiam9271]

@@liamvogel9151 exactly. The real world is too unpredictable, simulating every scenerio, no matter how relevant, is necessary for FSD. Tesla would need data from every car on the road to gather to understand every possible scenerio. Nvidia simulates it. No one does simulation better than Nvidia. They are the best in the world. Time will tell

[@kingiam9271]

@@liamvogel9151 plus Nvidia will be providing this software to a number of car manufacturers, not just one manufacturer like Tesla

[@baqirhusain5652]

I love him so much..his smile his laugh....i am a fan ..in the his croud of millions

[@OliverOttoman]

Man hate all this focus on self driving cars. Invest less in highways more in buses and trains and general public transit, both within and between dense areas. Self driving cars are cool but we already have the technology to go on long trips and not have to drive it's called taking a train. America just hasn't put money into them like most other developed countries have. Self driving cars are addressing a need that only exists because we are stupidly dependent on cars.

[@ziggs123]

That technology will be in place for every vehicle obviously. The next one will be Tesla Semi, which is the most important one imo.

[@MovingUp7]

If the goal is 100% autonomous driving, yes we are a long ways off. But if you have driven a Tesla with FSD or even better FSD beta, you know that it very much can do most of the heavy lifting. I work in my car now while the Tesla drives to my job sites.

[@franciscochavez1709]

We will go to the moon in 5 years, using materials that have not yet been invented. vs this...

[@Bootman899]

Andrej speaking pattern keeps throwing me off. I've looked at my playback speed like 10x to make sure I don't have it on 2x speed. He speeds up his words as he starts speaking lol...great clip, thanks Lex!

[@karlInSanDiego]

Sorry , but this insider is just not convincing. He's essentially admitting that his confidence in pulling this off has not shifted. Yet all the others in the space recognize that they vastly underestimated the complexity of the problem. I suppose you couldn't continue to work there for years without that kind of unfaltering hubris, but it's not convincing. It actually makes me less confident of his ability to do this. Because he's not capable of seeing it as a worse solution than a human driver. He is convinced, so machine, some day will be better, leaving no room for the possibility that all of his work is for naught. It'll always be inferior to humans, especially when it's camera only.

[@1paris1942]

I read an article very recently that was critiquing self driving cars , and one of the engineers interviewed for the article basically said, "We don't know how to program intuition because we don't what it is." That really hit me.

[@DurangoC]

You’re uninformed. It’s already safer than a human driver and getting better every hour.

[@1paris1942]

@@DurangoC I'm not. And it's not.

[@karlInSanDiego]

@@DurangoC Elon convince you of that or are you not aware of how incapable FSD is?

[@technicalmaster4054]

Saying machines will always be inferior to humans in driving is like people in 90s saying humans will always be better at chess than a computer or even people in 2010s saying that machines will never draw as good as a human artist. Self driving and AI in general has already improved so much over the last couple of years that it is inevitable that machines will one day drive on a level humans can't even imagine.

[@npc1386]

Won't be fixed for at least 50 years.

[@josha4128]

He talks like an AI trying to sound human.