Elon: Tesla FSD is its own worst enemy
Вставка
- Опубліковано 23 чер 2024
- Go to drinkag1.com/drknowitall to get a year supply of vitamin d3k2, and 5 extra travel packs of AG1 with your first purchase!
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!
**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!
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 YEAR OF EMBODIED AI Shirt? 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
Thanks to AG1 for sponsoring today's video!
**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 UA-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 - Розваги
Go to drinkag1.com/drknowitall to get a year supply of vitamin d3k2, and 5 extra travel packs of AG1 with your first purchase!
Did you give your own comment a heart? Narcissism much?
@@wreckinball11You OK?
They may not be getting interventions in California, but they sure haven't eliminated them in Colorado. The car consistently takes an exit at 75 mph into a exit with do not enter signs in a gate blocking an unused rest stop. On a 5 Mile stretch of interstate, I typically have to intervene four times. While not all of them are safety related, they are not convenience. The car is actually breaking the law. If Tesla wants data they should allow us to better annotate the times when the car is having problems without always having to yank it out of FSD. And then they actually need to use this data as I've been reporting these problems for over a year.
Similar experience
I estimate a disengagement every 10 miles in Denver
Yeah, it sometimes has trouble staying on the road here in NC. Especially rural roads.
Remember that v12.4 doesn’t use neural nets on hwy yet.
Remember the Tesla test drivers are using a version 1 or 2 versions ahead of the one that's deployed, so it could be the case that there are very few interventions
They need to look at different areas of driving. I can give them plenty of interventions. There are plenty of things to fix that don't require interventions. Examples:
1. Hard braking at some intersections and situations.
2. Navigate Rail Road tracks at correct speeds. (Please never accelerate on the tracks)
3. Don't accelerate when entering curves during take offs.
4. Don't drive on the fog line on 2 lane roads. (AKA stay in the middle of the lane.)
5. Begin slowing for turns earlier. Navigation knows the distance, use it to reduce brake use. I never use the bake but FDS does for situations that could easily avoided.
6. Fix Speed limits. It is unacceptable for the speed limit to change from 55 to 25 for no reason. Very dangerous too.
A big weakness FSD has is the inability to read signs, as in “road closed” or passing an “ Exit Speed 35MPH” sign when on a freeway and my Tesla suddenly reduces my freeway speed to 35.
I find it missing regular speed limit signs on the interstate. I also found it apparently reading hwy 50 sign and setting speed limit to 50. That’s not cool in a 35 mph zone.
It will fun to watch FSD deal with winter here. Because it becomes "whose lane is it and am I on the road" game. Parking lots are worse.
It'll be also interesting to see what happens when the cameras are "blinded". I've had that happen a couple of times simply from the mist coming off of the tires of passing cars.
Looking for chauffeur mode so my flute of champagne stays steady while I spread my Grey poopon and eating caviar.
Come have the FSD team drive in Boston I'm sure they'll have a lot more interventions to work with. Lol
or Charlotte
And remember, FSD is "overtrained" for California. I heard on FutureAza that FSD V12.3.x seems to fail on Colorado roads. Also, in roads around me in Alabama, I get too many false "collision" warnings (often due to shadows, I think, or in some cases because of the way the road immediately ahead is hidden by the rise and fall and twist of the road, sometimes due to the oncoming pickup truck wafting over into my lane -- it's Alabama, that is how "Bubba" drives, there is usually plenty of room left in my lane for me (or FSD) to fit my Model Y into it without risking any damage -- sometimes I don't even notice mild incursions until reviewing the video later ..)
The MUSTATTEERS want to hype it and Musk needs to as well. He's just a bit behind schedule and effectively is betting Telsa's future on FSD, It'll be interesting to see what is actually delivered with the roboteslas. I have a feeling that more than "Tesla Vision" is going to be involved,
I don’t know where they aren’t getting interventions, but mine tries to go through red lights and stops in the middle of intersections in San Francisco.
FSD will be better than almost all drivers in most situations but dummer than almost all drivers in very rare cases
And when do you think that will occur?
Was nice to meet you in Michigan John thank you for stopping by the tent. See you next year buddy the Highlander.😊
Hardware that turns up and down power consumption as needed is the norm for mobile and most PCs today.
I was going to mention that. The device drivers and operating system handles that. Nothing to do with the FSD software.
Secondly, FSD needs to develop local knowledge of the roadwork in which the owner drives. FSD makes the same hesitations and same in decisions at the same intersections and same stretches of roads, as though it is driving it for the first time ever, when in fact in my car, it’s driven it over a dozen or more times and had it developed local knowledge, FSD would become much smoother. For example, I’ve got a weeping willow tree that grew in the summer and branches extended to obscure or block a particular stop sign, so now FSD rapidly decelerates at that same intersection in the summertime because the stop sign is obscured by branches, when in the winter time it’s very smooth at that intersection. All it needs to know is that there is a stop sign at that intersection and it has known that for at least a year or more, this is what I mean about it must develop local knowledge, it must have some memory of local conditions and not just approach them as though it’s never seen them before. This is exactly what human drivers do.
Elon needs to send his professional drivers to Louisville, KY, because FSD performs poorly here. Interventions every 30 seconds on surface streets, and that’s being generous. IMO based on my 4 months of experience with FSD, it is 5 years away from being ready to be monetized.
MO also gets interventions.
When you analyze what 1 intervention per year means to Tesla, you have to multiply by the fleet size and you get 16000 interventions per day...
That's all unrealised statistics.
Great work as always 😎 One more thing that Elon mentioned which I thought was profound, was that the only way to solve FSD is, to have millions of cars with the system running out on the road. That would mean Tesla will be the only company to be able to achieve this for many years.
At 12.3.6, there were plenty of interventions. Not constant, but the dithering on lane choice, and improper stopping position on a certain tough intersection there was an intervention every single time. Plus, my driveway entrance is an epic hill climb, and I don’t think the software will ever decide that this is an acceptable pitch (32 degrees) 😂 Note that a human has no trouble with this. I’ll try again in a few months, since I’m not currently subscribing.
According to the Federal Highway Administration (FHWA), the average American drove 14,263 miles
With the latest version I received a week or two ago, LOTS of intervention need was reintroduced - but it makes left turns at signals better. Overall, a big step back.
It’s amazing to see how far Tesla has come.
I recently drove a friend’s model Y. He turned on the FSD. It worked beautifully, but at the same time it scared me because I’m so used to having control of a car rather than the car having control of itself. Overall, I liked it and if I buy a Tesla or lease it, the FSD will probably require me to practice using it on local streets when there’s not a lot of traffic. Ditto for the regenerative braking. Oh and yes, driving a sedan that goes faster than a gas V8 bows my mind. That’s probably why insurance is so expensive for EVs. 😊
The insurance for EV's is based on potential cost of repair and liability, it will come down when the insurance companies get a handle on it. Why do you think Tesla is in the insurance business, sort of? Tesla is also working on reducing the cost of repair of their cars.
Thanks John, very informative
Place it in India, say, where there is cows, motocycles, tricycles, people, cars, and trucks on teh road, so many distractions, at max speed, and make it work there.
Sounds trivial programming in clean environments , well defined roads, and no
complex traffic!
Run three, five or 7 versions and get them to mark each other. If the majority say proceed then proceed and markdown the one that was marked wrong.
You raise an interesting point that you disengaged for personal preference. Maybe shadow mode could be used for custom driving profiles so the car drives the way you like it.
I suspect that many interventions are due to the driver’s personal preferences, not safety. For example, my S on FSD loves to go bombing around curving mountain roads much faster than I would drive. I know the car can safely go that fast, but it’s too nerve wracking for me!
As someone who frequently drives in rural areas I’ll be interested to see how FSD deals with gravel roads in winter when the straight line winds create snow drifts and bone dry stretches of road.
It’s interesting to read how inconsistent FSD is depending on location. I was using 12.3.6 in Reno, NV and it was flawless. Odd that it can perform so well in Reno yet suck in Colorado.
Different cars.
@@tedmoss What? You can't be serious. FSD is supposed to be a universal product, right?
With around 400,000 FSD users on the roads you would think that Tesla gets many interventions per minute. Why does Elon claim that they do not have enough interventions to grade the different versions?
Dojo is doing exactly that improving itself in simulation. But there is a lot of ways to drive from A to B. Alpha 0 has no guideline, just one purpose. Shadow mode is an engine that has to find the best set of guidelines for FSD to grow. Different! You don't see google come out with any FSD do you?
Interventions are not rare in my everyday 2024 Tesla Model Y FSD driving experience. On two-lane highways, caution signs are ignored warning of town speed limits ahead, therefore speeding is common and slow to correct when clearly posted. Speedup is fast before 55 mph signs actually take effect.
Speed traps are common entering towns so intervention is enforced by police. This happens town after town, it never learns, update after update.
There are many narrow roads in my neighborhood. Narrow roads with big deep ditches, and sometimes with the inside corner of turns eroded out of existence. When FSD turns too tight it doesn't scuff a wheel, it tries to tear it off. Also the entrance of my driveway is quite wide, but narrows a little as it bridges over a deep culvert. A few times FSD has made the turn without making me think it is going to drive off into the culvert, but in general I don't even let it try. We don't have a service center in my state and I don't want to have to have Perl towed to Atlanta to get her repaired should Ego try to take a dive off the driveway. (ha ha, Full Self Diving: I'd prefer to keep the 'r' in the word. )
I guess I've used all my interventions until 2032.
Yeah, that proclamation might have been more than a bit optimistic, but that's sales for ya!
Tesla needs to do testing in northern BC. interventions are very numerous. Turned off my wife. Will definitely still get FSD on our CT, but not our current Y. Loved hearing idea of free FSD trials when new versions come out… could maybe change her mind.
Doesn’t do well when no side lines/sidewalks. it needs to learn how wide a typical road is and position itself accordingly so that it isn’t driving off into the soft shoulders.
If my wife is any criterion, she will come around. I have Foundation CT on order.
It stopped showing traffic cones and trash bins in the visualization.
Tesla’s worst enemy is when a Know It All asks Elon for a two hour interview rather than a real question.
Well, you can't blame the previous 9s for the difficulty of the current 9 :) and there is more to it
No collision or near misses is good. 7:53
Very good
Has it been suggested for Tesla to race on a track ,multiple cars in FSD ,at very high speeds to acquire more interventure possibilities?
Not real world.
Thanks John for sharing. Yes, FSD is really getting good but I still maintain that they could give it help with a better route planner. The planner must 'guide' FSD by showing it the perfect route lane wise so that, yes still use all its fsd magic to move through the traffic but always gravitate toward the perfect lane (guide line). Gravitation towards this guide line should then increase as it gets closer towards a change on direction. This guide line should also give information like speed limits, school zones and the like. FSD needs a buddy but can still be the best system out there.
As far as I can see, it already does that. Google Earth needs an upgrade.
As much as I love Elon and Tesla, they are far, far away from FSD becoming areality, at least in the direction that they are pursuing it in.
- It doesn’t have the ability to read and understand discrete text that it has never seen before. I would hate for that one intervention per year to be that it missed the “bridge out” sign because it couldn’t read and understand the text because of hadn’t seen it before.
- It doesn’t have a values system. Research the brains Reticular Activation System (RAS) to understand this more. By not being able to recognize and evaluate the difference between a small child and a dog, it can only rely on what it’s seen in the past to make a decision as to which one to hit.
- It doesn’t have knowledge of physics outside of the car so can’t predict when something heading towards it that it’s never seen before might hit it.
These are all things which are learned by the human independently and then come together as a complete decision-making system. FSD simply doesn’t have that ability at this point. The brain is not a flat, linear networking model. It’s made up of multiple, layered systems which have evolved over thousands of generations to work together.
I’m sure they will get better and better and maybe even address all of the above items. I only hope he doesn’t bet the farm on FSD at this point because one death caused by this system could possibly get it shut down by the government. Just my 2 cents.
Yep! And there’s an old problem in neuroscience and other fields tackling the reasons behind why humans need to be conscious or self aware at all. Why can’t we be mindless automaton taking in data and producing output in sounds and movements, all without being consciously aware. Being so far from the answer, I think it is tremendously arrogant for Elon to be so optimistic about what AI can truly achieve in its current approach. It is all very exciting and hyper impressive to a surreal level, especially compared to the days back when I was studying software engineering. But we best take a lesson from the Wizard of Oz and the man behind the curtain.
@@jackl593 I agree, but it could be argued that Elon‘s extreme optimism eventually seems to always lead to leading edge solutions. So I wouldn’t necessarily call it arrogant but rather ignorant of the amount of time that it takes to do something that’s never been done before. I was also in software engineering for many years before I eventually left and would love to be involved in this development phase of FSD but what I think they really need, as you pointed out, is someone who deeply understands neuroscience and psychology in order to bring this to reality. We shall see…
Sounds like Tesla needs more cars in the fleet helping them develop FSD faster. Why not just do another free FSD trial for three months this time and extend it if necessary to finish their product. I guarantee that they’ll get the needed interventions to build their software if everyone had and used FSD.
Here in Ft Lauderdale I sometimes go two weeks between safety interventions. This is why shadow mode is the way to go because you can never have enough test drivers before you release every model.
Just ask Microsoft or IBM, they let their users solve the problems after the testers have their shot at it.. I drive in Phoenix with no interventions.
Elon claiming there is 10,000 miles between interventions ........... geez, he is so FOS.
How in the world would you know? You don't have the information.
@@tedmoss Because I see the videos where the latest Tesla runs straight through red lights at a railway crossing. I see vids where there is 4 or 5 interventions. Weird that this should only happen once per year per driver, yet even the pumpers cant publish a months worth of driving with zero interventions. And THATS how I know kid.
Lewis V: This comment is just a bit of sillyness of mine to counter your very annoying AG1 advertising, Apparently Tesla hardware V is using a new form of energy that actually has intelligence. At approx timestamp 4:31, you say that in more complicated traffic situations, hardware V responds "by using more mental energy - some conscious energy". I suggest that this is very significant and could be better highlighted, no?
And thank you for the very good work that you do! Also, in another video, Miss Information's contribution to the discussion behind the decision whether or not to buy the Cybertruck was much appreciated. There in that discussion, you also mentioned the Cyberlander truck-camper. That was very interesting as I hadn't heard the Cyberlander being talked about lately, and I'm encouraged to learn that it seems to still be very much a serious and active project. Thanks for that.
At some point, even if FSD never makes a mistake, it'll need interventions to handle other drivers' mistakes ... or incredibly unlikely events like a meteor crashing into the street ahead of you.
You are thinking the car couldn't tell there was an obstruction in the road, this is unlikely. It avoids potholes.
Tesla fanboy believes the dancing robot is has ALREADY WON THE RACE 😂
FSD will train drivers to be safer. Omar Qazi has stated he has toned down his driving after using Auto Pilot and FSD. FSD will also have a similar effect on new drivers when they first start to drive solo, as most of their passenger miles while growing up will be in FSD driven cars. We all know families who have poor and dangerous driving habits, such as following too closely, that have been passed down from parents to children. FSD will help break that cycle. I remember one side of my family who had a shared habit of steering around corners by making multiple 'mid-course' corrections, rather than steering smoothly. Very strange. I probably have my own bad habits.....
I think Elon is great, but that’s BS about interventions. I find FSD pretty useless in CA on city streets. Perhaps these aren’t considered fair interventions? On city streets, FSD drives too fast over dips, drives into giant pot holes, Cues too far behind the limit line. To slow to react on turns pissing off people behind me, and no turn signals sometimes. Just the short list.
Great on highways
The test drivers are using later, unreleased versions and mostly operating in the San Francisco bay area. It's not BS, just not applicable outside that zone. He wasn't saying it because he expected you to see the same results. He was saying it because his test drivers are not seeing interventions and it's taking a lot of time to test the versions. Maybe they should do testing in all the major cities.
You didn't say what version, 12.3.6 does not do those things with a calibrated car. (At least not in the past month or so in my car). (I admit my car is slightly out of calibration compared to ideal, but it is close enough).
@@billweberx My ver. 12.3.6 works well, applicable anywhere. I don't put much miles on it though, it is hard to draw conclusions from a small sample.
@@tedmosssorry, 12.3.6
Different experience, different area I guess
I own a 2020 MY with HW3 and a 2023 MY with HDW4; these cars drive very differently. I don't no where these cars are that have few interventions but both of mine still make serious errors i.e. turning incorrectly into a turning lane, indecision with the steering wheel shifting left and right, following to closely to car in front, pulling halfway into an intersection and then getting caught with a red light. HW4 is much better than HW3. But where are the no intervention drives?
4:00
I suspect it is already built into the design.
Perhaps I am oversimplifying things, but it seems like the scaling would happen automatically.
Assume the AI input is an MPEG stream or equivalent. Given their preprocessing to create a vector map - this could be how they do it...
If a 30-frame-per-second frame represents very few changes, the tokens passed to the AI are greatly reduced.
It could also operate in such a way that it only needs to use the inference engine when the car predicts an object intersection or when the car will move from proper lane placement - to a minimum of 10 inference checks per second... example.
Elon already said they were running simulations. There isn't enough compute to run two programs at the same time in the car.
I have an idea of how Tesla could help train for the edge cases. What they could do is close down a road, street, neighbourhood, etc. like a film shoot would do. They would then stage special scenarios using their own drivers in Teslas and other vehicles. I feel that this would be a great way to train for emergency vehicles approaching and other occasional situations.
Emergency responders say Teslas are no worse than any other cars, having been one myself, I tend to agree with them. (I'm not a car yet, so I was referring to the latter, not the former).
and then repeat that on any street, in any town, in any state....
In theory yes, but it would fall under the naive approach. As someone pointed out edge cases are for all practical purposes, infinite. And the world continues to change day by day. Humans can understand, neural nets just do matrix multiplication. Well that’s a simplification but it’s basically glorified linear algebra plus brute forcing data and computing power. But we are falling for the illusion. A very impressive illusion which has many benefits, but true zero intervention FSD will never happen with the current approach. Also, for systems to be trustworthy they still need to be explainable and testable. Just because it’s neural net based does not magically change that. Policymakers and lawmakers do not understand the technology deeply enough to make informed decisions, and the public are mostly riding on the hype train now. It’s all super Impressive for sure and we should be excited, but not ride on the hype train forever
700W is a lot of power just for computing in a car.
That's max, not continuous.
That's what Elon said!
They should release FSD Supervised in Europe. I'm sure they'll get tons of interventions.
First you handle the conditions you know, then you take on the unknown.
Correlate this with your assertion that there would be personal versions of self driving for each driver.
Really every car would be different, they are you know.
I don't know how they are training all these different hardwares and then putting them all together so that they are all just as good.
Right now, all training is for Hardware 3 (HW3). Cars with HW4 run an emulator that is pretending to be HW3. Later this year Elon says HW4 will get its own, distinct, native version of FSD. Tesla uses basically the same video clips to train each version, adding clips to try to solve issues that didn't come out the way they expected.
Its not hardware, its different software.
Why do his expectations not meet reality???
That's simple, his reality is two versions ahead of yours and his car is calibrated better.
Rhapsodizing about adaptive processor usage is to be unaware that Apple products do that all the time.
When you wish to promote a sponsor Incorporate it within context. Do not become a robot reciting a speech. ❤🌵🕊️
Where do you put the AG1? There is no gas tank.
Now, what might you be referencing?
So I think I remember Doctor-Know-All addressing some of these issues that you guys are raising here in that the car at some point needs to have an AI operate at the car level, meaning that the car has the ability to think for itself, independent of the neural network, which is an over, arching idea of how the cars should interact with the environment, based on data being fed into the computer stack, and it’s going to be the same with the Optimus robot, because how will a neural-network work function, when there are so many of these robots working in unique, and different situations?
The other thing is, what happens or how will the chat box that we’re now hearing so much about coming to the car, will interact with the cars occupants be it the driver or passengers, because I would think the idea would be that the chat box would have information about the local area that may not be contained in the neural-network system.
I can't disagree more about what you said about FSD getting so good that you can't use human drivers to train it. That's not how it works. Humans are perfect drivers except when they make a mistake or deviate from being perfect. You don't train FSD on video's where humans are driving poorly. They only use video of when humans are driving perfectly. Going down the road in a straight line will is not going to be improved by FSD. It's always perfect. The improvements to FSD is using video of people doing difficult things correctly. We have that video. Once FSD can turn a difficult corner perfectly, it is only as good as a human doing the same thing, and can't get any better. This is not like Go, where there's a near infinite number of moves and it requires a massive calculation to determine the next best move. The hard part of FSD is getting all the edge cases trained, but there's no reason you can't use human video of doing those edge cases so that FSD is perfect at them.
There are an infinite number of edge cases.
@@tedmossthis.
Once they get to a certain point they don't need to keep re-training FSD anymore. You are going to get to a point where the interventions are so few that it's better to manually fix them than to re-train the entire model. You can get to near 0 accidents which would be a monumental achievement and then just hunt down the scenarios and adjust the model so that it doesn't encounter tha problem again.
If they were using the software you invision, yes; but they are not.
@@tedmoss I've been guessing on how it works from what Elon says, if you can provide an explanation please share.
It’s a fair bit more complicated than that. How would you define almost zero accidents? How would we standardise a way to classify FSD accidents? What about close calls? What’s considered something that needs fixing? How do you actually “fix it”? Write complex code every single time to handle so-called edge cases? Who will maintain this code base that will no doubt grow and change as the world (not just some states in the USA) changes. And so on.
In the vision clip you have shown I think Elon was hinting that one or more car maker will subscribe to FSD and he explain why that is a very good news for FSD improvements.
Obviously Ford.
Was just typing this mid video, then you proposed multiple models at the end. Delete, delete... But, not covered...
Having A and B shadow mode (or D...Z) models in the same car analyzing the same data allows direct comparison of model decisions rather than using statistical comparison by installing A in 1/2 fleet and B in the other half. Should allow much better than 2x efficiency in model evaluation.
As well, specific incident data where A and B disagree can be exactly captured and analyzed as well as becoming part of the test harness for future models.
This is becoming similar to triplicate safety systems where A & B disagree, C breaks the tie (perhaps human is 4th if still relevant). Running 3 models could not only allow more than 3x testing, but might be a further safety improvement. Although 1 of the 3 might carry the wrong weights for a situation, hopefully things aren't so homogenous that 2 of 3 make the wrong decision.
Just say it, the more FSD learns, the slower it learns… This is the law of diminishing returns. It’s a simple concept. The question is how to be creative to keep up the learning rate at a productive clip.
This shows that computers are not all that smart, about the same as people I would guess.
There has to be additional criteria to FSD beyond "safety critical". One area of concern is determining speed limits. Does it pick up speed limit signs for the parallel access road and apply it to the main lane? (Answer: yes, acceptable? No). Does it realize that the speed limit changes when exiting a construction area that does not have a resume speed or equivalent sign? (Answer no.) Does it change speed limit determination for no apparent reason. (Answer yes). Is it a "safety critical" issue? Maybe not, but driving at 55 mph in a 70 mph zone can be.
Another issue. How close to a concrete barrier does it go? Maybe not too close for FSD, but still too close for driver/passenger comfort.
In my area there is a LOT of road construction. In these areas, FSD is NOT ready.
I have an important question about Optimus that no-one has asked yet. Tesla cars have millions of cars driving around to collect training data, other than virtual bots, what data sets are they going to use to train Optimus on?
Their will be a catalog of training for the bots, a lucrative money maker. You pay for the software to do a job.
When are they going to run FSD in the Formula-E?
What would be the point?
From experience, I see the vision he is seeing, but the timeframes and power consumption will be more than he states. Dynamic power consumption…gee, i get it.
Elon is very aware of the problem of power consumption, as am I. being a former grid operator. You should know that the real reason robots didn't take off sooner is the battery and power use. Elon saw the possibility of using a power efficient system that his team created for more than one use.
Here is an idea to run by Elon. Train a version of FSD to race at a track. Then have driverless cars race. This would be a wild promo for FSD and seems very feasible to me with current tech.
Yes, but not related to the real world, which is where FSD needs to be solved.
The real world I has no clue about FSD. The point is not to solve with this but to promote. Also important
More difficult terrain like Roads of India and Africa may have more edge cases. That may be the answer. What do you think Doc?
Not Doc here just a BS, the roads in Africa and Asia for the most part are all edge cases.
What does AG1 have to do with FSD?
You drink AG1 so you can keep up with FSD.
Back before I retired whenever we developed a new piec eof software we first tries to see if it worked the we tried to see if we could break it. Knowing how the code worked helped us to do that. So rather than just have bad things happen by accident maybe they can develop an AI that learn how to break things and then that it is trained to do that so that then the code can correct for those things. It is like one has a good guy AI and a bad guy adversary AI that keeps trying to break the good guy AI and each using machine learning the get better at their respective tasks.
For example the bad guy AI (BGAI) might notice that the Good Guy AI (GGAI) had a bit of a problem with a lane change in a run and then set out to change the initial condition to make that problem worse, like adding more traffic, changing the rod a bit, changing the visibility etc. to see what makes it worse and the the GGAI tries to learn form that and make it better.
Thus one might need what one might call inherent goodness measures like how much did the car accelerate and decelerate or deviate from being centered in the lane and such, thes enot be problems but portents of perhaps potential problems. Thus the BGAI is trained to find stressors.
Sounds like Tesla needs fewer FSD subscriptions not more. In order to run in shadow mode and compare to human driving FSD must be off or nonexistent.
It is there, watching you.
@@tedmoss Sorry, Ted. My Model X has no cabin camera.
Does it do shadow mode while not in fsd?
If it didn't why would they put it in the car?
@@tedmoss so why all the fuss about not enough interruptions?
You're misunderstanding the tech.
AI systems always outperform the experts they're trained on. The AI that learned from chess experts still turned out better than them. You can train an LLM for accurate code completion on a 100% error rate dataset (where every example contains a mistake).
The car needs to eventually learn from its own experience, like: Here are two almost identical situations, in one I almost crashed, in the other I kept a wide distance. Learn the second.
👍👍
Hate to say it but Elon still seems to be clueless about FSD. Or he's lying. Such cringe to hear him talk about it not having interventions
Seems like it should be relatively straight forward for FSD to rate both itself and human drivers in shadow mode and this rating should be much better than just assuming the human driver is best (at any stage of development). My hunch would be that even a very poor FSD will be better than human almost all of the time since it will not suffer from distraction or fatigue. Problem is when FSD gets it wrong it is likely to be badly wrong! Humans will be more likely to have close shaves, FSD to have really serious accidents. FSD problem is prediction, but after the fact it should be possible to evaluate how close actual events came to disaster. Humans would be very bad at that, FSD should be flawlessly attentive and objective.
FSD has not shown itself to have many accidents, in fact it is quite the opposite. In fact again, it is very cautious.
@@tedmoss I did not mean to suggest accidents with FSD would be frequent, just that it is likely to eliminate minor accidents altogether so those that remain, even if exceedingly rare, will be major ones. On the other hand it is in a position to 'shadow mode' itself: what if I'd reclassified that bridge as an artic just a little later, or not at all, how would I have reacted and what would have been the outcome? It does not have to rely on learning from disengagements or cases where shadow mode diverges from the actions of a human driver.
Shape up on the road HUMANS! 🤨🫵🏽🫡
We have been trying to get them to do that for 130 years now.
No. It. Definitely. Is. Not. There. Yet….sorry…
If other car makers subscribe to FSD tesla could reach its goal much faster as few more millions of car will be collecting data for tesla to process them.
Probably not necessary but would alleviate the tremendous cost involved.
Assume the expert, Tesla-employed human is better whenever they differ. There's more to improving FSD than reducing interventions. Humans will be uncomfortable when the car doesn't drive in a way that is at least "superhuman," which is to say human-like but safer, rather than simply "supersafe" and "supernotdangerous." I think the best humans, when not in trouble, provide an adequate model.
I don't consider overriding lane choice or speed to be an intervention. Like airplane autopilot, you should have some input into the decision process.
The human, expert or not is not a "good" driver, just ask the insurance companies.
All laptops and phones already control their power usage - this is nothing new. So, regarding Telsa, this isn't technology to get excited about. Oh yeah...and likewise AG1 is pure marketing BS. - nothing for anyone to get excited about.
Tesla's Auto-crash program is a liability to everyone on the road. Elon is making Tesla into the Tyrell Corporation from the "Blade Runner " movies.
Yet the crash data seems to suggest otherwise
I now believe FSD was NEVER intended for use by ANY Tesla that is now in existence.
It was designed for the RoboTaxi or CyberCab - whatever that is.
and elon knows this
he needs billions of dollars and billions of miles driven to make FSD work
Eventually, FSD will work, but I, as an owner of it, will get NO benefit.
I can't go to sleep while FSD drives
I can't watch a movie
I can't fiddle with my iPhone
I can't do anything a non Tesla owner can't do
there is NO difference since the law applies to EVERY driver on the road
no special treatment that allows me to watch a movie in the moving car with FSD and my neighbor can't
So FSD will find a home in the RoboTaxi
and hopefully, I will get my $10,000 back with interest and a lot of penalties
Hundreds of new laws will be passed regarding the RoboTaxi, and police and courts will be instructed on how to handle a driverless car.
Just get in the back seat and let someone else supervise the FSD, your problem is solved. Or you could wait for the robo-taxi and let that drive you where you want to go far cheaper than owning a car. Just look at the money you will save!
@@tedmoss Why not just walk to the bus stop - that's even cheaper.
Or ride a bike
or just skateboard
Robotaxi is going to have to compete with taxis, Lyft, Uber, etc
What makes anyone think Robotaxi will be cheaper?
Those of us that have FSD need to stop turning it off just before we get to a place it might have trouble.
We all do it to save inconvenience, but that doesn't give Tesla what they need. We all need to let FSD struggle and fail.... yes it can be hard to do.
It seems to me that Tesla's paid test drivers are only concentrating their drives in some areas. That doesn't help after awhile as most cities and areas have unique road configurations.
Do we know how much data that Shadow mode gives them versus an FSD intervention?
I’m not paying to test a product out.
@@singed8853 No worries. Those of us who did are making a bet that once it is perfected that Tesla will raise the price of it by a lot. We might be wrong but we might be right. Either way we'll have a full time chauffeur.
@@nickmcconnell1291Chauffeurs do the driving for you and dont require you to oversee their driving.
@@TheSpartan3669 You obviously don't know the chauffeurs I know.
I never do that.
John, I thought of a cool way to use teslas Ai tech. If all teslas could keep and eye on existing amber alerts and notifications the driver if it sees something. No criminal could get away with a crime. And it won’t violate and privacy cause it’s all public information. This is something an update could do and bring value by saving many lives
I'm not sure that people would like having their cars used to surveille their surroundings.
@@TheSpartan3669 it can be optional like a toggle switch. I would think if people had the opportunity to help find criminals especially I. Their areas they would opt in
@@TheSpartan3669 the car already surveillance its surroundings
@@chillguy2758 Yes, but that data currently doesn't get sent anywhere.
@@TheSpartan3669 it doesn’t have to get ent anywhere. It can just notify the driver if it sees something. And the driver will have to report it or not
Well, I can tell you he's (Elon) full of crap. I have way more interventions on FSD 12 than I did on 11.
Recalibrate.
Elon in 2019 - "You're dumb to buy any car except Tesla because in the next 12 months there will be over 1 million Tesla Robo-Taxis on the roads making their owners over $30k a year.
Fanboys in 2024 trying to explain how Elon isn't a liar, he's just "too good" for his own good.
It's not lying when you make a bad prediction. Musk believes what he says will happen. He's bad at predictions.
Optimistic, visionary yes, but liar no. If I told you something that I knew to be false, that would be a lie. Elon believes or he would not do.
If these interventions are becoming so rare, I might have to be careful what I'm saying when I'm reporting them 😂
As long as FSD is better than every human that is enough, but preferably much better.
That is hardly necessary for success.
@@tedmoss absolutely necessary. Humans are terrible drivers so it's not a too hard to reach goal.
A video game could enable training 100 times the driving videos. Do I wish for a FSD system that has gotten expertise from high challenge situations being constantly thrown at it or one that doesn't see realistic and difficult situations except rare random not typical scenarios? VIDEO TRAINED FOR ME!!!!!
You are not in the real world.
Let's help the doc with a like and comment!
Click bait 🙄
Which thing? I see a lot.
I foresee the day when people will be scared to drive manually.
Ok. After AG1 ad have to unsubscribe 🤦🏼♂️ can’t watch these videos. just like Steven McRyan’s bs 💩
To be fair, I think the same, but you really have to try it before coming to such extreme conclusions.
AG1 Scam
Hope lies eternal on the human breast.