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

Please support this podcast by checking out our sponsors:

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:

SOCIAL:

[LexClips]

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.

[wertytrewqa]

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

[rmmm]

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

[brunogauthier]

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

[mattstevenson]

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.

[robertcameron]

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.

[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

[tombittikoffer]

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.

[thomasm.]

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...

[Aktenverwalter]

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

[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.

[EliWallach]

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.

[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.

[allhitstaken]

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.

[cag]

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.

[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

[KRTB]

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.

[benstrachan]

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.

[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.