Andrej Karpathy Reveals Key to SUPERHUMAN Tesla FSD
ฝัง
- เผยแพร่เมื่อ 3 เม.ย. 2024
- Tesla's end-to-end Full Self Driving v12 has made HUGE strides, and is almost as good as the best human drivers now; but what if Tesla wants FSD to get 2x better than a human driver, or 10x better? At that point imitation learning from human data fails and you have to come up with another method--and Andrej Karpathy in a recent interview gave us some really great clues as to how this might be done, and how it relates to Large Language Models (LLMs).
**If you are looking to purchase a new Tesla Car, Solar roof, Solar tiles or PowerWall, just click this link to get up to $500 off! www.tesla.com/referral/john11286. Thank you!
Join this channel to get access to perks:
/ @drknowitallknows
**To become part of our Patreon team, help support the channel, and get awesome perks, check out our Patreon site here: / drknowitallknows . Thanks for your support!
Get The Elon Musk Mission (I've got two chapters in it) here:
Paperback: amzn.to/3TQXV9g
Kindle: amzn.to/3U7f7Hr!
**Want some awesome Dr. Know-it-all merch, including the AI STUDENT DRIVER Bumper Sticker? Check out our awesome Merch store: drknowitall.itemorder.com/sale
For a limited time, use the code "Knows2021" to get 20% off your entire order!
**Check out Artimatic: www.artimatic.io
**You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you!
* USA: amzn.to/39n5mPH
* Germany: amzn.to/2XbdxJi
* United Kingdom: amzn.to/3hGlzTR
* France: amzn.to/2KRAwXh
* Spain: amzn.to/3hJYYFV
**What do we use to shoot our videos?
-Sony alpha a7 III: amzn.to/3czV2XJ
--and lens: amzn.to/3aujOqE
-Feelworld portable field monitor: amzn.to/38yf2ah
-Neewer compact desk tripod: amzn.to/3l8yrUk
-Glidegear teleprompter: amzn.to/3rJeFkP
-Neewer dimmable LED lights: amzn.to/3qAg3oF
-Rode Wireless Go II Lavalier microphones: amzn.to/3eC9jUZ
-Rode NT USB+ Studio Microphone: amzn.to/3U65Q3w
-Focusrite Scarlette 2i2 audio interface: amzn.to/3l8vqDu
-Studio soundproofing tiles: amzn.to/3rFUtQU
-Sony MDR-7506 Professional Headphones: amzn.to/2OoDdBd
-Apple M1 Max Studio: amzn.to/3GfxPYY
-Apple M1 MacBook Pro: amzn.to/3wPYV1D
-Docking Station for MacBook: amzn.to/3yIhc1S
-Philips Brilliance 4K Docking Monitor: amzn.to/3xwSKAb
-Sabrent 8TB SSD drive: amzn.to/3rhSxQM
-DJI Mavic Mini Drone: amzn.to/2OnHCEw
-GoPro Hero 9 Black action camera: amzn.to/3vgVMrH
-GoPro Max 360 camera: amzn.to/3nORGYk
-Tesla phone mount: amzn.to/3U92fl9
-Suction car mount for camera: amzn.to/3tcUfRK
-Extender Rod for car mount camera: amzn.to/3wHQXsw
**Here are a few products we've found really fun and/or useful:
-NeoCharge Dryer/EV charger splitter: amzn.to/39UcKWx
-Lift pucks for your Tesla: amzn.to/3vJF3iB
-Emergency tire fill and repair kit: amzn.to/3vMkL8d
-CO2 Monitor: amzn.to/3PsQRh2
-Camping mattress for your Tesla model S/3/X/Y: amzn.to/3m7ffef
**Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or TH-cam - / @zenlee_music
Tesla Stock: TSLA
**EVANNEX
Check out the Evannex web site: evannex.com/
If you use my discount code, KnowsEVs, you get $10 off any order over $100!
**For business inquiries, please email me here: DrKnowItAllKnows@gmail.com
Twitter: / drknowitall16
Also on Twitter: @Tesla_UnPR: / tesla_un
Instagram: @drknowitallknows
**Want some outdoorsy videos? Check out Whole Nuts and Donuts: / @wholenutsanddonuts5741
Source video from AI Explained: • Why Does OpenAI Need a... - วิทยาศาสตร์และเทคโนโลยี
![ปกติรีบตรง ไม่มีทรงรีบโกย : Gameboys [#73]](http://i.ytimg.com/vi/TnHXBorXoUU/mqdefault.jpg)
My dream is a self driving rv. It would be like having a train that takes you wherever you want to go. Make dinner watch TV site see out the windows all well the rv takes you to the next destination
Hopefully not as far off as we think, I want one!
Sort of like Star Trek.
no need for a driver's seat
Also a teslabot helper please
My dream as well!
It’s pretty crazy to consider that if you bought a Tesla in late 2016 you could park in in a storage unit, and let it just sit there for 8 years, and all it would need is a HW 3 upgrade and it is capable of supervised self driving.
The necessary hardware has been built in for almost a decade.
No other automaker can claim this. No one. It’s just crazy
What really? That’s nuts!
@@trevorsoh2130yup.
@@trevorsoh2130yea. Since October 2016 every Tesla manufactured is compatible with HW3 (up until HW4 vehicles).
Tesla started development on HW3 back in early 2016, and just had to work within the power envelope of HW 2.5 (100 Watts of compute) to allow for upgrades.
Well, comma ai has been offering retrofits (also just a "hardware upgrade" to an existing car) of "supervised self driving" for years and many makes and models. They claim to support more than 250 models, some of them starting in the 2014 model year like Audi A3, and VW e-Golf, or 2015 VW Passat or 2015 Honda CR-V, and so on. In the Tesla investor or enthusiast bubble it is a narrative that only Tesla could do certain things. But most of it is either limited knowledge about the rest if the industry or plain ignorance or misleading the public.
In the end, "supervised self-driving" is "supervised driving" (SAE Level 2) and not "self-driving". Therefore, the producers/makers of these driving assistance features are not liable in any way and can always hand over to the driver - who therefore must stay alert and monitor all the driving. With Tesla's claim on "full self-driving" for also many, many years, at least since 2016/2017, they are still nowhere near this actual goal, but had (a small share of) customers pay for an illusion and promises for years. And for more years to come.
Thank You! You’ve expertly answered a lingering question for me on how FSD gets beyond being only as good as the best human driver
Only as good as the best human driver is pretty damn good. I consider myself to be in the 99.999 percentile in driving skill, and I've never caused a crash in 50 years of driving. You can't get much better than that.
@@darylfoster7944So you consider yourself one in a hundred thousand. Would you not trust FSD before it was one in a million drivers? I certainly would, because even if it would beat 10,000 it would nearly always do so, and never tire, have a medical crisis or lose its mind completely, which good drivers do on occasion.
You have a talent for explaining complex systems in a manner we can all understand! Thank you!
I think there is a bit of a fallacy here - even with just imitation FSD has the potential to be super human. ONE, it harnesses stellar driving from an enormous amount of individuals, and TWO the hardware does not suffer from human deficiencies - no drugs, no anger, no preoccupation with something else
Combined, even imitation will get FSD far beyond statistical human level driving
I agree the potential is there, but perhaps it is because FSD is not legal in the EU yet, but even something much simpler like Cruise Control isn’t working here without phantom breaking, because it is not able to understand or read traffic like humans. Left-turning cars at an intersection will cause a catastrophic break even if the turning car is not even close to you because it doesn’t understand the car will leave the intersection long before you ever get there. It isn’t even able to register if it is raining or not, so talking about superhuman abilities is a bit premature 😊 But it will get there - and I cannot wait, but right now FSD is being priced like it is a robo-taxi, while the reality is that you will need to pay superhuman attention to not get killed. It is natural to trust the FSD, and that is when your brain relaxes and gets you into a fatal accident. That is for me not worth paying $12000 for. FSD is a cool gimmick as it is now, but at $12000 you could get an Apple Vision Pro, a Macbook, a used car, and you get to keep your life. FSD at that price is a hard sell.
@@QuadHealer FSD outside the US and Canada is a joke and anything less than a transfer when buying a new Tesla would be stealing money.
That being said, all Hardware 3 buyers should feel way better now with FSD 12 which cleared a path to have above human driving skills on this hardware.
Reactions in milliseconds, eyes on back, radar, GPS in the head, real knowledge of millions of drivers. I think these are superhuman abilities.
@@Markoss007 Radar? Forgive me if this was an answer to another thread but new Tesla cars do not have such luxuries as radar, LiDAR, parking sensors, or even a rain sensor, but you may have been talking about something else. It does have superhuman abilities to slam on the brakes, which can come in handy, but it can also cause an accident if the car behind me isn’t equipped with superhuman abilities. I turn off cruise control when coming close to traffic lights to avoid accidents when oncoming traffic in the opposite lane turns left (way before I get to the intersection). It will slam on the brakes every time because it cannot understand there is no danger, and the car will be long gone when I get there. Perhaps it will improve when the FSD you have in the US becomes legal in Europe, but I seriously doubt I will buy it at $12000. That is an insane price for something that isn't a robo-taxi, and will cause dangerous situations, but I might share your enthusiasm once I experience it firsthand.
@QuadHealer Yes I know. But other FSD cars have Lidar. It was in the context of learning only on human videos. I think that is enough to be accepted as superhuman. If is polished, of course, fantom breaking, is not acceptable...
I think even when other cars don't have FSD, the car can see 360° and prevent even crashes from the back or side. With other sensors, FSD can see 300 meters even through houses. Therefore can prevent crashes on crossroads where humans can't see. We are just in the beginning. I see a future where you can talk with the car. And if there is something that the car can't decide, it will ask. Or you can just say go slower, there is a bump near the school.
AlphaGo played itself. Call it adversarial learning. Nice thing about board games is that the game can be simulated with perfect accuracy and the adversary can simply be a copy of itself. Both copies improve each other that way in simulation. With driving the sides of this adversarial constellation would be the car and the world. One could train the car model to get better and then train the world model to get harder to drive in and so on, but the problem is that one can’t simulate the world with sufficient accuracy.
You can train an infinite number of times on an infinite number of variations. The real world does not have infinite variations, ergo the real world will be contained and mastered within the training.
@@SirHargreeves In a more granular sense, the real world (even just limited to driving) does in fact have finite number of variations. Similarly, a game might have so many possible outcomes that it appears infinite to us dumb humans, but is in fact limited. But the there are many many more variations in real world driving than any game...and the problem is that in a game, regardless of the middle steps, winning is the goal. Whereas with driving, the goal isn't just to get from point A to point B. There are a ton of rules in between with regards to timing, traffic, safety, comfort etc. So those are the nuances that make it an incredibly difficult problem to solve. Possible, but very difficult.
@@noobystokNope, a creative form of simulation could make many more variations than the whole real world has, just by expanding its computational and storage capacities. The real world is nowhere near its physical limits of variation in driving situations. Think about it. Any existing real situation could be varied in one important detail. Then already you have twice the possibilities in simulation as the real world has.
Use procedural generation to create countless worlds populate it with AI agents driving around on various levels of competency and add in pedestrians and random environmental effects like rain hail storm and various other things. Simulate this to train better AI agents. The simulation needs to simulate not only competent drivers but also bad drivers because the real world has a lot of bad drivers that the AI agent needs to avoid.
I'm pretty sure that Tesla has had generated virtual simulations as part of FSD training for a long time.
They have. It was presented in one of the AI days ages ago. There was an example with virtual people jogging in the middle of a highway, I think. At least something like that.
Not exactly the same as what's being talked about here, but certainly a step in the same direction..
Agree but the massive ramp up in compute power by multiple players introduces a step change /improvement in the usefulness of simulations?
They haven't talked about it lately, but it would be a way to let several FSDs play against each other.
For example let them drive as agressivly as possible. But punish if you forced a car with right of way to react. That's too agressive.
A simulated environment would be perfect to learn how to minimize the effect if accidents, when it has to choose the least dangerous action when something do goes wrong. So, you also need some rulebreakers there to get accidents.
Real data would then be used mostly to build virtual replicas, more random than those a programmer would build.
I guess they have to use virtual world learning to get it better than good human drivers.
Everyone' talks about how important real world data is. But I belive most of the learning, if not all, can be done in a virtual world.
That would also be the way forTesla's competition to build something from less real world data.
During training, let two or more cars talk to each other, like NASCAR teams, and see what full Co-op mode could do while driving. You could have in city coordination and on highway strategies that would develop over time. (You'd also want them to not really interfere with random single cars humans/other companies' self-drivers on the road.)
@@larsnystrom6698Yes, you need rule breaking routines or even rule breaking agents to create not necessarily accidents, but what I call "pre-accident situations" that FSD would not create in driving against itself. Then FSD can learn to cope with the unexpected in huge variety, which is what trips up humans more often than not. A superhuman driver will cope not only with known types of edge cases, but with unknown cases that are not the product of human experience or imagination. An example for me was encountering roadways after the Loma Prieta earthquake that had split in half LENGTHWISE,where half the road had suddenly plummeted into a valley below. Would FSD drive on the remaining half, or decide to get the hell out of there and turn around immediately? (I'm pretty sure I didn't make the best decisions in such disorienting conditions.)
The more I delve into this, the more I regress.
Great vid, John. I've wondered about this very point myself for ages.
Well summarized, thank you.
Thank you ... very helpful perspective ...
Thanks for that. Quite fascinating!
I have watched the video for you all. The non-stop talking of 18 minutes summary is that the host is speculating that Tesla may or is doing synthetic self learning (with alpha zero as a metaphor) to improve beyond the best human driving. No more and no less.
Very smart and easy explanation. Watching your videos is joyful learning moment.
It's the difference between learing by watching vs by doing.
When you do you seek the level of optimal learning.
Had to step in and take control twice Tuesday evening on a 15 mile trip home from a meeting. It had done way better on this trip before.
Very well communicated for mere mortals 👏
Yes! When I drive, I anticipate what the cars around me are going to do. I can usually predict what a car is going to do and have worked out how I will react accordingly. I also know which areas to avoid based on heavy traffic or construction. If Tesla can incorporate this kind of behavior, it’s going to eliminate the need to ever drive again.
Very interesting topic...... chasing the 9's. This is definitely the key to AGI and functional and safe FSD. Perhaps similar content in the future as this will be the key to a revolution. Thanks Doc. 😊👍
Cool video. I think you definitely ask the right question here
The beauty of this is that when a model can train itself, it can define and operate upon its own internal semantics and procedural structures instead of being squeezed into ours.
This would be good if you can prove that it is "better". Or even prove that it works at all. If you can't understand what it does, you are in trouble.
@@tedmoss Definitely. There are ways to have a fairly good idea via verbose outputs, but ghosts in the machine are definitely an issue of concern.
We are still dabbling in AI.
Today's digital AIs have almost nothing to do with human brain function.
We should think of today's digital AIs as an entirely different and evolving lifeform that predominantly shares an ability to mimic human output.
Digital AI requires massive amounts of energy compared to the human brain. Digital AIs process 10 million times faster than human brains. Dispite this overwhelming speed advantage, children have skills that digital AIs have yet to acquire.
Our brains are analog and appear to be based upon multilevel logic. Today's AIs are digital and for the most part are limited to binary logic.
Agents brought better AI task performance through task specific training. Encorporating agents into teams substantially improved performance. Building a supervisor AI to build task specific teams to complete specific tasks with autonomous error checking and correction further improved performance. The next steps reportedly are pushing on the edge of general intelligence through massive compute and massive energy consumption.
Analog compute has the potential to reduce energy cost by multiple orders of magnitude and at the same time offer an easier path to multilevel logic.
PS V12.3.3 is substantially better than V12.3. Still making questionable routing changes just as you demonstrated.
Super interesting and I like the new editing style.
Wow that was so bonkers rate of bringing me up to speed in this area. Ty
terrific show
Fascinating
cybertruck drives past in the background 2:16
lol
Wow, good eye.
Elon really is ruling the world i guess 💀
You saw the gorilla! That's impressive!
I had a bad lane choice yesterday, with the car choosing a non-lane shoulder on a city street rather than its usual choice of the actual right hand lane.
excellent video!
Drove all over town today to two doctor appointments and a haircut. Have driven 5 segments plus one auto parking, rear in. Everything was PERFECT. No interventions!
FSD cannot drive straight into parking spot. Not perfect.
FSD cannot handle diagonal parking spots. Not perfect.
FSD cannot understand signs: DO NOT ENTER. ROAD CLOSED. Not perfect
FSD cannot avoid potholes or broken glass in the road. Not perfect
FSD often exceeds posted speed limits. Not perfect
This list goes on and on.
If your virtual arena had only Tesla FSD vehicles, would be eye-opening to see the future.
Thanks! Would like to see radar included in the Tesla safety suite. Ultimately the system should be superior to human and computer vision.
Why do you want to make it a three body problem?
Step one is imitation learning. Step two is adversarial competitive learning in a simulated environment. Step two requires more compute. Tesla has the compute now.
Question: if Tesla is in chill mode or sport mode does it drive differently in FSD?
Tesla's vast amount of training data is great for exposing the model to the huge number of situations the driving environment may provide but having all reactions based on a good human driver will have a improvement ceiling.
Alpha GO was still super human at plsying go. ZERO was much stronger yet.
Didn’t we have this same conversation with V11? I am pretty sure we will have the same discussion for V13, and probably v126. Getting so close!
Along those lines, it was *eight* *years* ago when Elon said FSD would drive 2800 miles across America, automatically. In April of 2024, that is a joke.
Incredibly insightful for the layman, thank you Dr knowitall
Regarding agents’ improved outcomes via competing in synthetic worlds: this video is excellent but it would really illuminate the kind of processes you are trying to describe if you could go into one or two actual examples of the types of improved outcomes you are thinking about, and break down some of the factors involved in the improvement. I know this might be hugely difficult because any example would likely involve a huge number of variables - but even a relatively simplified model/explanation might be very illuminating in getting down to the nitty gritty of how the synthetic world improvement of outcomes might come about. Perhaps an obvious simplified example might concern your point about setting an aim of maximising the efficient use of energy - so, for example, how different speed choices might maximise this, or improving energy efficiency via corner turning choices, or even utilising certain techniques to maximise the benefits of regenerative breaking. (I mean, in all this the models are not deciding based on human example, but simply by ‘playing out’ many options.)
Or perhaps, even better, discuss another, more challenging example (though with similarly simple level of detail) where the aim of the virtual ‘game’ in the synthetic world might be something else - such as maximising safety. I think this might be a good subject for a separate video!
One way to get better than human driving is to add to the model route data from the drivers as a function of time of day. The most efficient and safest routes are not necessarily the shortest.
We have an already set up actual version of that in the Mexican way of bus routing. (Using paper and pencil and your brain).
I was presuming they might already be doing something like this.
If I am understanding this correctly, what is needed is a massive amount of synthetic data with every conceivable situation coupled with the basic rule of the game, which in this case is first, not to crash, and secondly not to hurt anyone in or around the vehicle. Probably followed, closely, by don't hit pets or animals, or other autonomous vehicles - which we will soon see hitting the roads and byways in increasing numbers for delivery and other purposes. Let the system go on that and it will train FSD to be better and better, eventually far better than human drivers (which is a pretty low bar to begin with anyway). Probably somewhere in the rules of the game should be driving smoothly so as not to alarm the occupants.
In airline pilot training, there is something called AQP -- something like Advanced Quality Program. Instead of just training pilots on how the plane works and how to do all the standard things, AQP presents them, on simulators, with specific scenarios, generally taken from events/accidents that have occurred in real life -- and develops specific methodologies for recognizing and dealing with them. Usually these are scenarios that occur under intended or unintended instrument flying, where human perceptions cannot be relied on. Possibly presenting the driving system with simulated scenarios and then grading its reactions, you could improve things. My situation yesterday with the choice of a poorly marked shoulder lane over the correct one would be a good scenario example. I get the feeling V12 relies a little less on the map data for what it does at times, which could explain this new behavior. But I don't know.
But until there are no humans driving around, accidents will still happen to even the best drivers, human or otherwise.
It's a good thing that self driving vehicles will be better than human drivers. At least (hopefully) one day, maybe you won't have the option to disengage FSD for the way you drive as you express your views in your first drive videos that FSD is not driving like you do! Promoting speeding, lane changing, navigation, yielding at intersections, etc.😊
Although superhuman is obviously the ultimate goal, I think as long as it shares the roads with humans, it should be as predictable and human-like as possible. But I think acting superhuman will distance FSD from human-like and predictable driving, like we saw alpha zero get super human and very unpredictable, although that's admittedly part of the strategy to win there.
And probably human-like behavior actually makes the passengers feel better as long as the seats are oriented to the front at least :)
I actually think it can be superhuman in a good way (so without loosing predictability or human-like behavior) being trained with humans driving already because it can see in all directions at the same time.
John thanks.
It is very tempting to set the stage purposely and see what FSD does .. throw a cardboard box out in front, have a person walk in the middle of the street and then jump out in front of you coming up, have a dog race across the road, have a truck drop a ladder off the back … someone should purposely create some edge cases!! What happens you you start purposefully covering some of the cameras to simulate mud or purposefully do some weird hand body motions or abruptly open a car door … on and on … throw a bucket (like a splash) of water across the drivers’ windshield. Why not setup a competition to who comes up with the best crazy situation!! Challenge FSD on purpose!!! Why not put up a sign saying accident up ahead, take another route and see what FSD does .. the more ridiculous the better .. it’s time to have fun 😮😅
I’m new to fsd using the free trial. I was eager to try it. While it may be much better than V11, it sure hasn’t sold me on considering purchasing it. Today it was driving slow for unclear reasons, and then needed disengagement because it became perplexed on which lane to drive in shifting left and right and left. Then it wouldn’t yield to allow two cars to merge on the highway in front of me. It’s pretty cool when it works, but not ready for general distribution yet, in my opinion.
The cars getting on the highway are supposed to yield, not the ones already on the highway.
@@darylfoster7944 I was well behind them. I could see them much better than they could see me. Just easing off the accelerator would easily allow them to merge. at the rate we were going, I would’ve either caused one or both cars to have to brake. My way seemed more considerate, and a simpler approach, although not to all drivers. In my car I want things to operate the way I feel comfortable.
@@BobbieGWhiz ok, but they were driving incorrectly. If you were well behind them, and they were going the same speed as you, they would have merged easily. A vast majority of drivers try to enter a highway going below the speed of the traffic flow, which is bad driving.
@@darylfoster7944 They were merging and had not come up to highway speed yet, but at the rate we were all going we would have occupied the same space at the same time. I’d either have to slow down or accelerate, or they would have had to brake. My passenger is a much more aggressive driver than me and he felt the FSD was making the wrong decision too. There was very light wet snow on the ground and discretion was the better part of valor.
@@BobbieGWhizIt is illegal for merging traffic to force highway traffic to change speed or lanes. On the scales of courtesy, the burden is on merging traffic. That said, I have seen FSD show great deference in many situations for extra safety. Mods to individual driver preferences are part of the future, I feel certain.
Really interesting John. What if you took an exotic list of edge case scenarios: kids and other things running into the road, road closures, potholes, odd signage, lights or lanes, other driver aberrant behavior, weather, etc. and massively and randomly threw those events at otherwise normal driving simulations? Also has potential to erode Teslas FSD data lead / moat?
How would you know what you are doing is correct? That's why they are using real data, you will make less mistakes and have less traps to fall into. In the end, what you say is exactly what will be done to prove the system works. The real proof is the hard part.
Looking at the snow instead of the dog is exactly how models "beat" humans eventually. They correlate what we don't notice.
got ya...
Excellent show. I had wondered how Tesla could get past best human level.
Thanks. However doesn't this mean that v12 FSD is actually the first "Alpha Go"? If so then it will still be awhile before we get to the next step.
No, it won't. This is changing very fast. By actual data, Tesla is 15 times better than all other cars in safety already. (On FSD).
@@tedmoss By next step I meant that if their current FSD 12.3.x is going to be training the next version (as Alpha Go trained the next version that became superhuman) then we have just started to see with 12.3 what's coming. Elon said it will be superhuman, and from what the Dr said I can now see how they can achieve it. Previously I was wondering how it could get better than human if it was training on only human vids.... now I know.
@@tedmossBeware. Safety is only one metric for excellence, though the most important.
Interesting would be, how FSD will perform in Europe where you have sometimes very narrow streets. The problem is there that you must reverse or do other crazy staff (e.g. drive into an entrance or gateway to a garge) to get arround each other and be able drive further ...
I remember seeing Ashok (Tesla's FSD guy) say this will be really ready for the world if it can drive by itself in India! And I agree! 😅
There is no limit to what a neural net can learn provided it has enough good data and compute. So India, Italy, Egypt, South Africa and other crazy driving environments are perfectly learnable.
@@balaji-kartha That is his bias, there are plenty of crazy drivers in other parts of the world. All it takes is one. It can drive in India because it can drive anywhere.
@@michaelnurse9089 Time, it takes time. If it can drive in one place, it can drive in all. Even with different rules. We are human after all. Notice that FSD slows down all traffic. I have done experiments on this many years ago. Driving becomes safer when you slow down slightly. Other people copy you.
12.3.3 is solid. just needs an option to find parking now. but as robotaxi, it's so close. on a recent 50 mile city drive with multiple stops, it was flawless. navigation is now the weak link.
There are videos on TH-cam that show FSD tested in San Francisco, New York, and Toronto.
Lots of mistakes and human interventions.
As I see it there is no such things as a 10x better than the best human. It is not a competitive sport. The goal is to get from point a to b, without damaging the car, and violate any rules. It’s sufficient to do it just good. The bonus is that AI won’t get distracted.
In a simulator, how Do you train the car to be "feel confortable"?
I've been a tutor in Maths and Physics since 1979 and know EXACTLY the process of teaching and empowering a neural network. Nerds who have no social acumen will take a long time to learn this.
Tesla's neural network has been training for about 90 days, but FSD is still making lots of mistakes on the road. Every day.
Your opinion?
@@DerekDavis213 With humans, memory needs initially to be rooted in "cheat sheets", which are kept all times visually present to provide context and focus thinking.
@@patrickshanahan7505 And what does that have to do with the primitive simple minded FSD which continues to make mistakes every day, after 90 days of intense training?
@@DerekDavis213 that answer will cost you $3,000. I don't give my secrets away for free to the peanut gallery.
So, to go next level FSF, will the mountain of real world videos be used, or something else, like simulations? If the latter, then competitors might catch up to Tesla, if they have enuf compute, which can simply be purchased (tho for a lot of $).
If you like going off road in your 4x4 having fun Injoing nature or just working on a homestead or ranch. Will or is Cyber Truck Full Self Driving System Work Off Road without getting stuck or scratching your truck up from Holders, trees, bushes, etc. Would about going through a steam, creek or river?
Just like a human has to learn how to drive off road, even when they already know how to drive on asphalt, it needs to learn that as well - so more training.
Thanks. A big question answered. Go Alpha Go Zero, Go Mu Zero, Go FSD.
I had a chance to try out V12.3.3 last night and this morning in slush and snow...big fail, maybe drives like a California driver ;-). It was worse than V11, so I guess it needs more training in snow. (note that it could see the line/lanes based on visualization, it had a tendency to go over the line on the irght and onto the shoulder where there was the most slush and snow) - So hopefully Tesla will succeed to go beyond imitation and FSD becomes 10x better than humans!
Careful making comparisons on to little data, the car has to be calibrated.
@@tedmoss yeah, I always wondered about that. Does it recalibrate after each version(major maybe)? It does “feel” like it’s gotten better in good conditions in the few days I had it, I hope it’s only that and I just need to drive it more
It has been pointed out that the visualizations are generated after the fact of all decisions and do not form input to the net. They are a backend driver display only. So it could be putting lines in where it imagines they could have been under the snow.
Driving on snow, slush and ice is extremely challenging. After more than 40 years of practice I still don't consider myself expert at it. (And please don't come north without snow / all weather tires at least!)
@@user-gv4cx7vz8t A guy in Canada reviewed FSD here on TH-cam.
With light snow on the ground there were about *10* *interventions* on a short trip. FSD did very poorly.
I would love too but I'm still waiting for my v12 update.🙄 I have v11.1. AI creating its own language can be dangerous, we may not be able to decipher it.
I think AI will take FSD to great advancement in the next few years for all the reasons you have discussed. Karparthy is very interesting to listen to.
Alpha Go is simple compared to driving tho
Simple for you; difficult for me.
@@tedmoss I know what you mean, it's a difficult game but with practice you would improve! At the speed of AI you'd be a Grandmaster like Lee Seedol within weeks
@@davec2211 Where's my Neuralink!
Why did Karparthy leave his job at Tesla, if he is the AI expert?
There are no experts in AI. They all have bad misconceptions. None of them have solved many of the most important problems yet. They haven't internalized The Bitter Lesson. I believe Karpathy held Tesla back.
What’s good driving really though? Besides not having an accident, what you consider good I may consider to be bad. The game Go however is easy to measure, there is always one winner and one loser. There is no winner in FSD. If two version got to the destination, who won?
What I hate about FSD is its caution when it doesn’t need caution. Especially with the latest versions speed limits. Or at a stop sign always letting the other cars go even though it’s my turn. I get why it’s cautious and it should be. But that’s not how I drive or want to be driven. Some people will prefer it or not care. Maybe they can delineate the “style” a little more. Make the “assertive” setting actually be assertive.
I want to see how well the core driving AI could chew up a rally course or race track, not having to worry about pedestrians/traffic and othe road normative features just to see how good its basic intuition for pedal to the metal is...
Compared to the amount of driving data currently used, how much more data will be collected on a month of free fsd?
About 4 times as much in a month, maybe more. 400,000 users vs 4-5 million.
@@tedmoss Tesla will never reveal the *true* number of people who did the 30 day free trial. It goes against their interests.
Is level 4/5 FSD a matter of "when" not "if"?
Since 4/5 has *never* been done before, the answer to question is IF.
Observe the continuous nature of human consciousness. The AI needs to become a sentient being so we can begin to share our world together. Maybe not a perfect union, but closer than we were before.
I think Andrej meant RLHF shouldn't be called reinforcement learning, because the human feedback isn't fit for the way LLMs learn.
A computer beat Gary Kasparov in 1997. By now, we should have fully autonomous cars.
If future training to be super human will be in simulation, does that diminish the value of tesla's data moat?
I want to see FSD take the James Bond/Vin Diesel/Jason Stratham escape driving course and then improve on it.
I can’t wait till I can tour a city in a Tesla and Xai or Grok gives me an autonomous driving tour explaining key historical cites.
Sites too
And sights.
Regression to the mean doesn't seem like an appropriate name for the phenomenon you're talking about. Models don't regress towards the "mean", instead they regress towards something that gives them high reward.
I prefer the term "model collapse"
Very interesting. But if this the case and agents can compete and get better in a virtual world, there will be no need for millions of miles of actual driven miles. Then Tesla's advantage/moat could evaporate and Nvidia or someone else might come up with autonomous driving.
I would appreciate some help here. This is not a complaint. I’m actually calling for help. Whenever I use this latest version of FSD I have to disengage after a minute or two because the car is going way too slow, often below the speed limit as other cars are going 15 to 20% over the speed limit, which is about right.
reboot your car and try again.
Try matching the traffic speed and then engage FSD. See if it slows down or maintains speed.
Felt like if ppl figured out some rules (mainly safety wise, not "legal" wise, coupled with efficiency, time and energy, as a reward) just like Go, then FSD can learn driving by itself. The hard part will be a deep understanding of the "rules" put there as there are no clear rules like Go. On the other hand, I can't believe "safety" is not the TOP measure for what Tesla is currently doing. Therefore one can be almost certain that FSD is very safe, even if it still looks awkward sometimes.
There are no clear rules to driving? There most certainly are, it's just that a certain percentage of humans don't follow them. Driving is really pretty simple. Stay in your lane, look before you change lanes, and don't tailgate. Follow those rules, and crashes virtually disappear.
@@darylfoster7944Except that "virtually" means "not," almost literally so. How do those rules help when an airplane lands on the road in front of you?
@@user-gv4cx7vz8t you aren't serious? Plane in the road, that's your unsolvable edge case? If a plane landed in front of me, the odds of which are one in a billion, I would slam on my brakes, which is what FSD would do.
What you or the AI would do there is not my point. Following your rules yet not avoiding all accidents is the point. It's impossible to be prepared for all contingencies, so rules are insufficient to "solve" driving. The best hope is highly adaptive intelligence, faster reflexes, and pattern recognition to replace almost all rules in the actual, unpredictable event.
@@user-gv4cx7vz8t and my point is that all contingencies will never be 100% unavoidable, due to the laws of physics. No matter how smart you are, there are going to be situations that just happen. At least until all cars have FSD, which will be quite a while. If there are a few car crashes a year, so what?
As theses LLM's learn on their own accord by cycling through their own data, they can never forget there carrying precious cargo!
The very approach to the problem in the first place is to avoid a circular answer, you have to go out in the world to find the answer, not in a closed system. That's why the internet is so important to this exercise. Still you have to be careful, it becomes harder and harder to check your solution.
You have to have some novel or rather unpredictable source(s) of real variation in your simulated RL models. For instance, with solid physical world models and the ability to predict or imagine potential future video frames/experiences dynamically based on recent past and currently incoming frames, people and cars can plan alternative actions ahead of time to deal with potential real events. For reinforcement learning, both AI "players" can replay the the same situations over and over again by random selection among alternative futures and each other's behaviors.
Likewise, one superpower of the AI is to think and act much faster than human. We learn to play chess (or drive) by playing each other and improve slowly over a lifetime. AlphaZero plays many lifetimes of games with similar opponents very quickly. For a skill like driving, with 360 degree perception and much higher bandwidth frame rates and planning/reaction times, the car should outperform humans just like athletes and others highly skilled outperform the semi-trained and untrained (or predators catch prey). Thus the AI car should be able to recognize and avoid, out-maneuver/escape, or minimize hazardous or dangerous situations in ways not even possible for human drivers.
❤
I don't need human language to be adapted to become more efficient in conveying concepts from one computer to another - I would not mind if LLM developed human language to become more nuanced in expressing concepts and emotions.
Why couldn't cars that are using the full self driving in the real world be used to compete against each other?
Human needs to communicate with FSD while driving.
I would simulate that some 'high' driver wants to jam a Tesla and fsd trains to 'feel' safety limits and avoids collision
I love listening to real experts! Regression to the mean is a very mis-understood concept that must be controlled to achieve meaningful results, not just statistically significant. I have had a question about how "rules" are applied - every State, and country has their own set of traffic rules. How are these applied when the dataset is collected so widely?
Regression to the mean might be fundamental problem for our education system. And the "woke" issue. When everyone needs to think the same way and do things the same way then creativity and innovation will begin to disappear. We end up as humans being more like non advanced machines and software clones of each other. F*ck that. I want my own individuality.
@@19951998kc Being smart enough to know when to be different.
@@19951998kc "woke" doesn't mean everyone needs to think the same way and do things the same way. Ideally, a "woke" argument on some social injustice issue should change your mind that you should think and act differently.
John, going beyond what Andrej eluded to and with human learning in mind it occurs that Humans don't learn how to tell a dog from a Wolfe in isolation! The human learns concepts in stages for example a human might learn that the animal in the picture could be more likely a Wolfe if in a snowy background too however this would be separate from the task of learning how to recognise a dog or Wolfe in a picture... It's a question of assigning ontology and at what stage of the task that ontology is assigned.... I believe that we will end up with cascading models with varying ontological assignments at the boarders of these models. In this way the model designer can direct the importance and weighting of the various ontological inputs to each successive model. Tesla I think are already doing this, maybe they haven't yet realised the power of this approach!
Oh, yes they have. They even said so.
A few things. First, I've seen my son-in-law playing Grand Theft Auto. And was amazed at the lifelike city scapes and driving conditions. And I was wondering if something like that could be used to train an AI. But, of course, it would be very human intensive to create those scenarios. We can play with concepts like: _what would a self driving car be like that was orders of magnitude better than a human?_ But that's like mathematicians discussing what a 100 dimension cube would look like. Does it make any sense in the real world? I think the first thing we want is safety. We want a car that is *never* in an _at fault_ accident. And, apart from extreme circumstances, *never* breaks the law. That's all. I believe there are, not me, but humans who've done that. Also, it will be a long time, if ever, that we have purely self-driving cars on the road. And until then, do we want self driving cars to behave in non-human-like ways? I think not.
Watching this video and then thinking about what's coming in the future, when will AI develop its own language that we won't understand, so that it can learn faster and faster. Terrifying.
As a person with no driving liesence i hope they Can make it happen and make the model 2 with no steering wheel and sell it world wide as a robo taxi.
If everyone drove as well as the best driver, wouldn't all accidents be avoided? Do we need anything better than that?
I got a letter of rejection from Tesla for using my driving input!!
Change your avatar.
In addition to virtiual simulations, maybe they need to put a bunch of FSD Teslas on a formula 1 course and get them to race. Maybe throw in some random obstacles and bots pushing baby caridges. Who knows what kind of skill levels might emerge. Just the fact that the car might develop skills with traction control, differential throttle settings and and braking to individual wheels, etcetera. Access to fine controls human drivers can't access might create superior driving skills.
It's true that current paradigms don't quite get it but for driving, an alert able human is able to drive flawlessly so there isn't as obvious an opportunity there for great AI except for emergency driving. You could imagine AI doing completely outrageous things to traverse a distance at maximum speed and that would require super human understanding. Also if the rules of driving were changed to only allow AI driving it could be stepped up dramatically. Too dramatic for most. It would be like a rally ride.
Actually, imitation could work because the human pattern is hidden in the language.
The difference between humans and AI is we have a pre-programmed brain that already has patterns built in to help sort through the "noise" whereas the computer is starting out with a clean slate, Tesla's approach will have the machine thinking like a human as opposed to making a completely alien machine which is what Andrew's approach will produce. Lots of data will make up for the lack of a framework to sort out the "noise". This will be the difference that will make AI controllable.
I'm sort of favoring the alien machine as superior, with all of Asimov's laws in effect. Being perfect means never having to say you're sorry.
It would be fun to put FSD through a driving test. Invite a qualified driving test examiner to ride along with you. In the UK there is an advanced driving test to assess your ability to drive more defensively, anticipate hazards, and handle different road conditions compared to the standard driving test.
I do that myself. It passed.
@@tedmossPlease elaborate. You made us so curious!
Is there a limitation in LLM itself? Even if GPT x moves away from imitation modell and achieve insanely good writing skills. LLM is just LLM? It wont understand any logic and it wont make up new things for us? Genuine question btw.
These driving neural nets aren't LLMs. The second "L" stands for language, and FSD cannot even read non-standard road signs, I believe. Also, it's impossible now to have a conversation with FSD. Scenes are processed visually, "from photons," as Musk says.
On your broad question, LLM's should be specfically trained with logic and math to acquire these skills, just as humans are. We cannot depend only on natural language inprovements to instill proficiency in other special disciplines.
The audio is slightly out of sync and it's wiggin' me out lol
I don't think this analogy is necessarily applicable to driving, other than in some very unique scenarios.
If you let self driving develop "its own psychology", I imagine that you may be foregoing the comfort of the passenger which humans will prioritize. The AI agent may forego that to achieve higher safety scores, but may lead to lower trust from humans. Comfort and trust from humans is paramount right now.
Tesla should just have thousands of vehicles driving in a simulation - similar to how agents were trained in mazes. Give it the rules but get some agents to drive randomly, drop random objects, etc... I believe this is 99% determined by computing power, which would become less of an issue in coming months and years. BTW, as for optimize for energy, Tesla should add an FSD profile optimized for energy consumption. They would be able to reduce consumption by 10-25% easily - driving a bit slower on highway, less throttle, utilize reiterative 99.9% of the time by breaking earlier and reduce likelihood of emergency braking (eg, optimize for right turns, keep more distance, etc).
"Reiterative" should be regenerative braking.
👏🏻👏🏻👏🏻👏🏻👏🏻✌🏻
Interesting. Was wondering why not train on just syntethic data, if this is the case Nvidia would probably have a huge advantage over Tesla. But based on the video, you need actual data to avoid reverting back to the mean. All synthetic data would just produce crazy results
Correct, a circular solution will get you nowhere.
Tesla is buying Nvidia GPU systems to implement DOJO, in addition to the in-house processor.