No One Understands Tesla's AI.

[@farzyness]

Check out JD's channel: www.youtube.com/@AIDRIVR

[@mattsenkow6986]

"What the fuck is that - that's crazy!?" That's also how a human brain works. We're just a lot slower. You're going to witness the birth and early training of a human intelligence. You know how Elon Musk repeats important sayings to drive home their importance? You do and should do the same. AI a system modelled off a human brain (neuron) that learns in the same way we do = repetition and reward/discipline.

[@marriagepartnersministry5942]

The oil companies lobbied to get legal liability caps put on oil spills. This was done "for the public good" so oil companies would continue exploring and digging. Tesla will HAVE to lobby to get a liability cap for fsd accidents and here's why. A car crash lawsuit has the liable driver paying for actual damages witch is usually medical bills, vehicle damage, and property damage. But, if a driver violates the law and injures someone then they can get hit with PUNITIVE damages which can go into the hundreds of millions or even billions of dollars. Without "if/then" codeing then you have a control system that is not exactly and precisely known what the car will do in a given corner case. It will be pretty easy for a clever law firm to show Tesla has gross negligence which is the threshold for triggering a punitive damages award.

[@TheChr3880]

I have been investing in other companies. When I noticed FSD V.12.3, I was for the very first time "wow". my positions have now changed to 50 % to Tesla - I cant sit outside anymore. I feel lucky for my entry at 166 average price. I am long for around 10-20 years - maybe longer. I am excited for the future with FSD, Robots, energy and extremely good cars.

[@davidkendall2272]

You will not regret your decision to invest in Tesla long term. I invested at 2010 IPO and am still long!

[@GET2222]

10,557 shares since 2018. I have not sold a share and considering buying more.

[@Detroit12boi]

I can hear my wife now. Asking AI why are you parking here ?

[@viktorianas]

Underrated comment.

[@andrewsaint6581]

Then you know they've nailed it.

[@sportbikeguy9875]

"yeah, theres clearly a spot 2 vehicles closer, why pic this one?" lol

[@edstahl5651]

Thanks!

[@wlmsears]

Good comment on camera placement. "There are a lot of sports cars with really long hoods". A Mustang has a hood that is nearly two feet longer than a Model 3 hood, so the driver ends up with similar visibility to the B-pillar camera.

[@Alexander_S18]

21:00 minutes in ... lol Hans is also in the call hahaha

[@erikbrown77]

Tesla may not ever reach 1,000,000 employees, unless you count Optimus robots as employees.

[@gohansaru7821]

If the government mandates a UAW chip, they will never get close.

[@bru512]

Tesla may not ever reach 500k employees

[@bru512]

When will the total number of Tesla employees decrease?

[@bru512]

Will Tesla reach peak employment before 2030?

[@205rider8]

True!

[@rodneysmith1750]

I was fortunate enough to drive FSB Beta from Aug 2020 until about May 2021 and I am flabbergasted to see what 12.3 can do. Listening to version 11 and 12.3's driver feedback difference is amazing.

[@craighermle7727]

Did you pay for the beta software?

[@mhfs61]

40:59 @HansCNelson IMO one of the reasons that the Cybertruck doesn’t have FSD yet is the fact that it has rear wheel steering, which is quite a different driving dynamic. On top of that it has an additional front camera and variable ride heights. Seems to me very complicating issues, with insufficient time to test. I expect that all the Cybertrucks delivered are collecting data and training FSD is shadow mode.

[@cmw3737]

The direct mapping from the input data taken from human drivers in millions of Teslas to the output of the car driving is so much more suited to machine learning that all the stages in between where they were trying to turn the data into a model with occupancy network and labelling objects more to convince humans that it knows what it's doing than actually teaching it to drive.

[@davidellsworth2427]

Farzad, this is fantastic! Please remember that a human in these creep forward scenarios would actually lean forward. The vehicle can only accomplish this by rolling.

[@Mrbfgray]

If you only sample one FSD channel it should probably be AI Driver.

[@allieflounder5764]

Cyberlyft and chuck cook are actually the best

[@Mrbfgray]

@@allieflounder5764 Lol, not even close, Chuck is good but one dimensional, Cyberlyft is so full of himself and talking about himself I can barely listen to him. AI DRIVR offers in depth consideration, analysis, criticism and comparisons.

[@jeffs.9299]

For the last time the model Y is not a dam Suv it is a med size hatchback 1:14:19

[@briansilver9652]

The last time?😊 The AWD Y feels like any larger crossover on the inside and drives like a tank in the snow. It's sporty and highy utilitarian. It's also a vehicle. That's S.U.V. It's just semantics. Have a great day.

[@RX-8GT]

@@briansilver9652exactly 👍

[@jmattoxriskpro]

I see the guys who know the important stuff to discuss, semantics, have arrived.

[@davidkendall2272]

Fascinating to hear two former Tesla employees talking excitedly about how Tesla innovation and vertical integration continues to bring Tesla to the forefront as the obvious leader in FSD. V-12 is epoch in my view as someone still waiting for OTA. Having observed so many FSD drivers reaction to 12.3, I am beyond excited to get OTA. I am planning trip from WA to NM in May and plan to have my MS drive us there. Tesla please give me the OTA. Uber driver CYBRLFT recently demonstrated how effective 12.3 was in correcting mistakes made while driving. He repeated the portion where he had to intervene to show that 12.3 learned from the mistake and drove correctly every time he repeated this portion of drive. Very impressive demonstration IMHO.

[@davidellsworth2427]

Farzad, this S, X curbing could be simply those models recogning their own girth/height. My neighbor is 6'9". He always ducks his head when coming through the front door.

[@BongoWongoOG]

10 mins in and already hooked

[@bernardradcliffe6240]

The mark of a good driver is the degree of how safe your passenger feels. When FSD is perfected the passenger will feel 100% safe 100% of the rides. And after 1yr of being driven by FSD at this level the passenger will unlikely ever want to drive again. As most people if asked would love a private chauffer on call 24hr. Think about it. Your own driver on call 24hrs.

[@gregorstoutbelly5598]

Women have had that for a long time ... They call them husbands/boyfriends.

[@Techtalk2030]

Robotics, A.I, self driving, vr/ar and new discoveries in drone tech and longevity are going to change everything forever. Exciting times ahead.

[@Throwthatassback737]

Universal food and housing is here earlier than expected, thanks to AI.

[@StaticFreq]

A believer! 😊

[@leonari]

⁠@@Throwthatassback737housing? Please explain how AI will build housing in cramped city centers

[@Throwthatassback737]

Mass-producing 3D printed modular homes are the only solution to the housing problem. Elon talked about this a while back. @@leonari

[@merlinmidday4644]

​@@leonari I'd say it'll be by design. Smarter design than we Fail by, now!

[@audience2]

A single camera in b-pillar can be compensated for by moving the vehicle slightly. The NN will implicitly do parallax distance calculations between camera frames.

[@CFLDumpsters]

Such a cool time to be alive! Imagine, self driving semi trucks carrying goods, that humanoid robots loaded, moving to its destination with a human monitor going along for the ride, that is then unloaded by humanoid robots. Planes, ships, trains that all can do the same etc, businesses would explode, no longer being limited by work force costs, HR issues etc. I think we will see it in our lifetime.

[@pshattuck1]

The disruption that FSD is going to bring to the Trucking Industry is going to be huge!

[@hwillia204]

Waymo is nowhere close to TESLA FSD! Waymo has people remotely controlled, though I'm seating my TWSLA. It INCREDIBLY HANDLE difficult highway and side road situations without my interfering! Let’s do a test Waymo vs Tesla and Blue cruise driving in Colorado Rockies back roads to Boulder I’m ready!

[@briansilver9652]

Farzad's comment about the 3rd gen compact being the solution makes so much sense. By being shorter they can move the side cameras forward. Existing training would work because it would err on leaving more room on the front. Those cars then can be used to collect new training data. Once enough model 2 data is collected and trained the camera placement can be moved on the rest of the models when refreshed. Chicken/egg solved.

[@kevinbecquet]

01:15:40 he isn’t joking, I’ve driven in L.A. everyday at the worst time for traffic to get to work and it is a real nightmare and if you had a version of autopilot that didn’t nag you, it would change all of our lives

[@darinl848]

what about all the people that fall asleep at the wheel? i never hear about how FSD will save lives for those are too tired / sleepy to drive.

[@Pizza-Solutions]

I'de love to see Tesla FSD navigate Naples. That would be the ultimate test.

[@revmsj]

So I’ve been dying to know and asking everywhere if Dojo is online or even completed yet and what you guys said in this video answers me that it is in fact not yet online. Will we know for fact that it’s online when it is on line?? Will it be a noticeable step change in FSD capabilities when it’s utilized? Or will they tell us?? I’m really super curious and interested in what Dojo will bring. I also fear that Dojo will be obsolete and unimpressive by the time it’s completed…

[@RussInGA]

I think the blocked turning camera placement issues is being looked at all wrong. It is a very specific use case that could be solved (if necessary) by simple sensors with built in chips that put out an is it clear signal. This could be used with FSD or with human driver. Simply if you are easing out into an intersection and something is coming dont let the car go. This could work easily and be cheap. It would not require ANY retraining and could be added to any vehicle for a relatively low cost. Probably not necessary though.

[@OrganicStuff1]

This is my go to Tesla stock pump Channel keep it up bro I’m long

[@sakhi2ya]

Great colab with AIDRIVR🐐 !!

[@jamesmarinelli2177]

Ai learning must feel like what I do before I take a LONG trip. Years ago we could not go on google maps and ACTUALLY see real life pics of the highway the scenery and landmarks. Before a trip I go on google maps and kinda drive on google maps practically all the way looking at the pics of streets, exits, what the signs look like , what the wendys , gas stations, landmarks look like. So even though Ive NEVER been there after I see all this on google maps its as though I have been there and experienced this trip. When I actually do go on the trip I do remember most of the landmarks, what the signs, etc etc look like so this AI learning must be the same. Although now video instead of pics showing all this footage to the computer I can see how it learns! This is kind of a way of looking at and understanding this neural learning. Cept the computer would remember everything and can absorb unlimited data.

[@rlla-hz2cl]

Can’t wait for the unintentional road rage feature

[@toughlove7706]

With respect to unprotected lefts and camera location. The b pillar camera is positioned for a rear passenger leaning forward to peer left. Whereas the driver would lean forward with their eyes in line with the side view mirror or the a pillar. In winter weather with snow banks blocking both the left and right line of sight. A camera position in the front turn signals make the most sense when inches matter the most.

[@robbiero368]

I think of it this way: imagine writing code to describe a billion rules. It is necessary to make it work. No amount of human coders could ever achieve that task without a million conflicts and bugs. A neural network can construct that from data. If the size of the network is unlimited, so is the number of rules it can encode.

[@GET2222]

FARZAD: “I have to get it out of my chest” 😳 I think you mean… I have to get this OFF MY CHEST. 😂

[@dougselsam5393]

phlegm?

[@mxguy2438]

This camera position change vs existing data doesn't invalidate the existing data because only the camera view has changed, not the position of the cars. The data itself can be used to recreate the data from the new camera perspective which can then be used to train for the new camera position.

[@erictheblue7256]

If the front view left and right cameras are not in the housing of the left and right headlights, why not? With wide angle lenses they could collectively view near 90 degrees left, right plus center, and "see" several feet in front of the driver's eyes. The B pillar cameras could still be used for left and right side view watching for T-bone (side impact) threats. If not the the left/right headlights, another location could be either at the base or top of the left/right A pillars for additional height from the ground. In any of these positions the left/right front facing cameras are still in front of the driver's eyes.

[@robbiero368]

The capacity of compute a company owns should be part of the valuation calculation. No-one is talking in those terms yet.

[@geraldmarshall22]

Do Teslas have binocular forward vision (Two or more forward facing cameras)?

[@jeremiah6617]

The point about Tesla self driving being criticized involved in an accident is correct. It will need to operate safely the majority of the time at a higher percentage than a human driver. Evasive maneuvers mentioned seems like a natural progression of advancements that will be added with more learning. Farzad’s discussion videos are getting better.

[@jackofalltrades1582]

Tesla should make a video of the model s going around the nürburgring as fast as physically possible while only using fsd. Do it with no one in the vehicle and push the vehicle as hard as it can go and see how fast of a lap it will do. If that could be accomplished on fsd then there would be no questioning it's capabilities

[@FEEL-X93]

Excellent video guys !!! Very informative & well thought out !! 🎉😮❤

[@garycarson3128]

Tesla may have tons of data, but Scott Walter commented that the Tesla driving data may have a data bias - the data is biased to the driving habits of early adopters and perhaps the use of the FSD Beta by the 400k early adopters had dropped during the FSD V11.x deployment. This could be the reason Tesla is requiring their sales staff to demonstrate FSD for every new customer delivery and the one month free FSD. By expanding the adoption of FSD, Tesla might eliminate their early adopter data bias and increase the breadth of data to include more caution drivers, more female drivers, more rural and middle of the country drivers and edge cases that exist in areas of the country where there are few early adopters.

[@mxguy2438]

The increase in performance from a subsequent improvement is not necessarily 1:1, particularly when you have multiple exponential growth curves interacting. Moore's law was always just an observation.

[@Ronnieleec]

i’m very interested in how it handles weather-rain, snow, ice and fog?

[@pshattuck1]

I’ve witnessed V12’s ability to operate in moderate rain, but snowy conditions are a different story. The neural net needs to send more cautious signals to the control systems, particularly turns at intersections.

[@daverodenwald9757]

If FSD can reverse the car, then it can react to a unsafe excessive creep faster then a human could and be totally safer then the human. No camera changes needed!

[@ramacvr12]

Just got my V 12.3 downloaded , will try first thing in the morning.

[@WhodatIzz]

Great insightful chat! Thanks all!

[@stevemartin8994]

Great conversation guys!

[@TedKidd]

Wow, for me it seems more incremental. Which is saying a lot given it's a whole new stack and isn't a huge step back like each other version update started out as a step back. My feeling is it will accelerate quickly, but not that it's made a huge jump yet.

[@wasimmahmood561]

any thoughts on nvidia and their potential working with oems, to devolop fsd?

[@khuo0219]

The problem with that is a, pack of data and b partnerships rarely work

[@rr6013]

ONLY nvidia can vertcal market - “car guys” wait until it trickles down

[@perjohanohlsson]

About those very dangerous unprotected left turns, the speed limit must be reduced, which will annoy people who are in a hurry. But as more and more people get FSD, they will be less stressed anyway.

[@DanWall-dn1vb]

When Tesla comes out with a $25,000 car, it will be game over for the Dino burners. All that will be left is the time it will take for them to bleed out.

[@garywright8758]

Mad Max training... 👍

[@chrisheath2637]

Maybe the final piece of the puzzle (that's never solved! ) is humans interacting with speech, if there is a problem - "Try going round again to find a parking spot" - "Yes, I was going to try that...."

[@AIDRIVR]

I'll um, like try to get better at um, talking to the camera

[@farzyness]

You did great brother.

[@magitekarmor6875]

Perfect :D

[@jeremyPieLover]

You did totally fine

[@patricke.2181]

This was a great one to watch/listen to. Exciting times ahead. Appreciate your insights and your channel is one of the best.

[@miked9474]

Try public speaking it’s easy to judge hard to do!

[@garycarson3128]

According to Tesla, Hardware 3 will be able to handle full autonomy. This is because Tesla has optimized a small percentage the FSD software running on Hardware 3. Optimizing the software isn’t a priority for Tesla right now so they optimize the code as more compute is needed on Hardware 3.

[@pshattuck1]

Do you think Tesla will offer upgrades to HW4 for cars that currently have HW3? I’ve heard that the form factor is different and currently won’t allow such an upgrade. Similar to the inability to upgrade cars with HW1 to HW3.

[@garycarson3128]

@@pshattuck1 No, Tesla will not be upgrading Hardware 3 to 4 because of the form factor, increased power requirements and because it is not necessary. James Douma explained in a KZbin video that Hardware 3 despite it’s age is more than capable of running FSD at Level 5. It was designed way ahead of its time. Switching to end-to-end neural networks offloaded a lot of processing from the CPU and GPU and is now running on the much more efficient neural network hardware.

[@garycarson3128]

@@pshattuck1 James Duama has a video that explains why Hardware 3 is more than adequate for FSD level 4 or 5 especially after moving to end-to-end neural networks. Hardware 3 was developed way before its time and it is just now taking advantage of the processing power that it was designed for.

[@ghislainbourassa1906]

In a world where AI/robots take care of everything for all humans, humans would only have to play together and enjoy life. If you do 't need to work and make millions to enjoy life, why would you instead of spending all of your time with the people you love?

[@perjohanohlsson]

How hard would it be for Tesla to develop an "FSD kit" to sell to fit to old cars? Some sort of safety solution in order to reduce the number of accidents.

[@michaelhess4825]

The last update (non-fsd) basically made autopilot useless. It doesn't see the half of the speed limit signs it used to. All the 70's are now 30's! A. Are they making autopilot worse to push folks to buy fsd? B. Do they have no developers on autopilot anymore? C. Do they not care about rural folks? (Middle of working, not a complex environment!) This will lead to LESS enthusiasm for FSD if users with autopilot just see it getting worse and worse...

[@pshattuck1]

I’ve noticed this, too. Trash cans no longer show up on the car’s awareness display, which raises the question in my mind…is the car still seeing these obstacles?

[@johnbarber5891]

Can it handle emergency vehicles now? A fire, police or ambulance coming towards you, or coming up from behind, or crossing ahead of you at an intersection: does it know what to do now? Needs to, if it doesn't yet.

[@danw6542]

The best unprotected left turn is no unprotected left turn

[@briansilver9652]

3 rights make a left

[@joshuasellers4434]

Elon's already talked about how the cars and optimus are using the same models. These embodied AI are going to be "body" agnostic. This isn't like the old way to create ML models.

[@robant555]

~30:00, It's foolish to think that the prices of these goods and services will go down. They will stay the same while the cost of production goes down. Billionaires will continue to pocket the larger profits. Look at all of the efficiency gains we've had in the past 50 years of computing. Do we work fewer hours? Do we get paid more per hour? Not without a revolution.

[@barrydaugherty5528]

Good thought by JD, but my 2018 model 3 also hit the curb going around a sharp curve on a 2 lane road. So not just model y and x

[@davidfenwick987]

Reverse would be very helpful.

[@viktorianas]

What about a scenario when most of cars become FSD and they literally communicate between each other, the safety would be mindboggling. Once ALL cars become FSD you would not need traffic lights and congestion could be reduced dramatically.

[@mattsenkow6986]

That ending! Hype!

[@Chris-jo3nb]

there are techniques to morph HW 3 camera data to synthesize camera data from new, adjusted locations (like near the front of the car). Possibly Tesla could use this approach to at least start the training process for a future hardware platform until they have more real data for it.

[@chuinnim24]

@farziness regarding your three biggest remaining issues, here’s one. I have a path with one left followed immediately by another left and the first one has two left turn lanes. Even on 12.3 the car consistently gets in the right-most left turn lane, makes the first left and then realizes it has to get into the the left lane again. In traffic that could be difficult if there’s no gap, which would lead to blocking cars wanting to continue straight. I imagine the training will have to take into account a few turns ahead in the path and it’s apparently not doing that yet.

[@rr6013]

you are generations away from FSDpathways where TeslaFSD history stacks routes of familiarity on-board optimizing its internal trip stack down to a single model route

[@chuinnim24]

No one said anything about history, though that would be great. The problem appears to be that FSD’s not considering the next couple of turns. And now that it’s trained end-to-end rather than much (or any) heuristic code, that would mean they’d have to include either the destination or the next few upcoming turns as training inputs for every few road segments maybe so that it could learn in situations such as mine, humans tend to get to the left most lane or whatever. Based on its performance, I doubt its being trained that way currently.

[@rr6013]

@@chuinnim24 that FSD isn’t caching upcoming turns from route planner is pulling back the curtain. Now we can see realtime in-car compute is overburdened here. Tesla may deliberately have prioritized RT over pre-time compute as an FSD logic protocol - - RT is baked-in

[@chuinnim24]

I feel like you’re making quite a leap with that statement that the computer is overburdened. You could be right but I think it could be anything. Even if it is an issue with calculating upcoming turn data in real time, they may be able to precalculate for the specific route initially and store in lookup table. Who knows what the holdup is? But I think they’ll figure it out. It just doesn’t seem unsolvable to me even with current computer though I acknowledge we’re all just speculating.

[@WhodatIzz]

Just a guess, the biggest wrinkle in Cybertruck fsd is steer by wire and four wheel steer changing the required control outputs at different speeds.

[@rr6013]

nah, the second pivot pt variable is programmatic in realtime. Tesla have some choices to make with Cybertruck. For example, does it allow variable rear tracking which enables the front and rear tire tracks to follow inside each other? That allows the possibility of driving CT into trouble with predictable paths. Drivers may find CT cutting when its least expected. Size of CT is a global re-configuration dynamic for FSD as Tesla state-of-art AI. That more than anyrhing is a huge rewrite type of effort whether architectural, scale of ML

[@WhodatIzz]

@@rr6013 someone Heard from an engineer that the challenge is largely rear steer as it's super important for fsd to know where it's hips are to properly read camera movement on the sides and CTs back end moves different.

[@rr6013]

@@WhodatIzz THAT means those cameras are analyzed RT before FSD can act so CT needs parallel compute for that function

[@BatmanBoss]

Thanks guys

[@BeijingTesla]

I'm trying really hard to get on board with this FSD, but everyone raves about the improvement in the releases, but really I dont care about that, i only care about how far away we are from the end? What is missing, each release upon release does not quantify how far we have to go, some help here would be appreciated.

[@asiando]

I have a theory. When the time comes, they will retrofit the side blinkers of the cars with FSD, which currently have one camera looking backward, with other ones with a set of 2 cameras (one looking backward, and another one looking forward).

[@richardjones8395]

Yes, they can include the current cameras and add an additional camera in the same unit. The cars would be able to continue using the software trained on the current camera system until enough data has been collected to train a new system with the additional cameras. Once the new system is performing better than the current one, they could just roll that out to more and more of the fleet that have the new hardware.

[@TheNativeTwo]

I just want to point out that all they have to do is disable to steering wheel and pedals, and it’s level 5. These ASE levels don’t mean what you guys seem to think it means. I am a control systems engineer who worked on autonomous vehicles. I built a level 5 system. I think these “levels” are not well thought out. It should be based on the environment. For example, my car was level 5, ONLY ON THE TEST TRACK. It can be level 5, but only geofenced, or only on the freeway, or only on paved roads. Cuz think about it, when Tesla solves FSD and has a “level 5” system, does that mean it can drive on gravel roads? Dirt roads? The beach? The grass/lawn? Can it drive around my farm? Can it drive in the desert? Can it drive on the moon or mars?!? If not, well, it’s not level 5 then right? And nothing ever will be, because there will always be new driving environments. Or is it level 5 but only in certain environments? And if the latter is the case, then we should stop talking about what level it is because it is meaningless and we already have level 5, just constrained to specific environments.

[@bogdanstroescu1932]

It should drive anywhere where humans typically drive. Roads, parking lots, driveways. It's pretty simple

[@pshattuck1]

Excellent point!

[@TheNativeTwo]

@@bogdanstroescu1932 humans normally drive on dirt roads and beaches… off road is a thing… so it won’t be level five until it drives off road? The point I am making is engineers don’t use these levels. They are silly. We already have level 5 cars, Tesla included, but only in certain environments (and the standard actually says all environments, which is impossible). So we either need to change the standard to specify EXACTLY what environments, or just throw it out. If it can drive without a person controlling it, it’s level 5, which means all teslas are already level 5. The standard also doesn’t set safety margins or accuracy. It’s just not a good standard and not one any engineer would use. No Tesla engineer is sitting there saying “how much more percentage accurate do i need to be to get to level 5.” They are asking themselves how they can do better based on actually useful metrics.

[@MrJinske]

Correct me if I am wrong, it seems Hans does NOT have fsd, is this correct??

[@MrJinske]

Nevermind at an hour and 3 min he does not have fsd so his opinions are not valid, he has zero experience with v12, that is what I thought and his opinions are not based on experience hence no value Farzad, and Ai drive along with myself have exp with v 12

[@Alarix246]

@@MrJinskeyour conclusions are not valid as well - Hans sees and analyses many recordings of FSD, is in touch with experts, isn't blind nor braindead. You can say he doesn't have a hands on experience, yet knows things (as a mechanical engineer) that the hands on folks don't necessarily have. It is a fusion of opinions in such discussion that is net positive for all of us. His opinions are valuable.

[@mikegagnon1]

Knowing the style of driving in different places like farms vs San Francisco, the model should be trained to use current GPS coordinates?

[@markoconnell804]

1:47 you would need to keep “home groups” mentality with a limited number in each group. Each group working to keep best practice outcome paramount.

[@Ronnieleec]

Tesla and Xai were both on the Blackwell list as well.

[@brianschroeder3381]

They can just simulate the camera placement in their virtual model. Most of their training data for the training set is virtual.

[@Matzes]

How do you know most data is simulated/virtual? Like what's your source?

[@PaulHigginbothamSr]

At this point at night in the rain or daytime or nightime in a snow storm I will be driving.

[@garycarson3128]

Tesla’s FSD Beta has been at autonomy Level 3 for several years. Mercedes has been certified as Level 3 for their Drive Pilot software. but if you read what Drive Pilot can and cannot do, you will realize that FSD Beta has been doing what Drive Pilot can do and much more with less restrictions for several years now. Elon refuses to play the recertification game until Level 4 or 5 is achieved. Recertification at any level lower than 4 or 5 would severely restrict Tesla rapid pace of development of FSD, especially when Tesla updates the FSD beta monthly and sometimes weekly. By the time regulators re-certifies FSD Beta at an intermediate level, Tesla would have made dozens of revisions to that version of the FSD Beta, making the recertification worthless, because no-one would be using that 3 to 9 months old version of FSD Beta. Elon has chosen to avoid that recertification HELL, by avoiding it until the product is finished - Level 4 or 5.

[@pshattuck1]

How close are we to pursuing full government certification in the US?

[@garycarson3128]

@@pshattuck1 It is a state by state process for regulatory approval. By default autonomy is allowed unless states have laws restricting the use of autonomous vehicles. 10 states have laws permitting autonomous vehicles with restrictions, several states have laws allowing autonomous vehicles and sone states have no laws at all addressing autonomous vehicles. Today autonomous vehicles are allowed in 40 states.

[@jmf6000]

I was listening to this in the background and I thought that voice was familiar. Just clicked the video to check and of course its AI Drivr. Only ever heard the voice.

[@Mr.E.Shoppa]

The "creep" problem at dangerous intersections could be solved cheaply with a periscope solution, only the periscope would extend horizontally in front of the car, and a tiny ball on the end of a car antenna type skewer would see in all directions. if it got hit, it would not be serious, and it could prevent any creeping out accidents. It would extend and retract only when needed, and would do so instantaneously, so it would never be damaged by an oncoming car.

[@lucadellasciucca967]

Esponencial growth will blind side everybody

[@khuo0219]

Put two cameras inside the cabin near the a pillar, solved. If people aren't driving then they won't mind those cameras taking up space

[@EmergencyButtons]

With enough robotaxis they can share their camera views with each other

[@jemm92122]

38 minutes in and I completely agree that Tesla should just do what people do and avoid dangerous unprotected left turns by finding a safer route. Many times that will be just turn right and make the first possible u-turn.

[@lourdessilva6442]

Incrível muito esclarecedora conhecimento e vida nos liberta sucesso sempre

[@the_teknomads]

Good discussion. Yes, it seems the world largely continues to overlooked the levity of what Tesla is bringing to global economies, as a fore-runner in mobile compute and AI, and the implications that it's vehicles and distributed compute will have in the world. Tesla has transformed from an Electric Automotive MFG, into a pioneering Cloud AI compute entity, leveraging a highly sophisticated integration of NVIDIA compute modules within millions of sedans. Tesla vehicles host compute modules akin to those being utilized in the forefront of cutting-edge robotics-that make top news headlines daily in 2024. The strategic incorporation of always on mobile compute, not only enables Tesla vehicles to capture and relay real-time environmental data but also positions these cars as mobile nodes in a vast, distributed computing network. There is simply nothing else like it in the world at this time. Beyond this already unparalleled distinction, is the reality that the energy necessary and capital investment for the computational resources onboard these vehicles are being wholly financed and maintained by their owners. There is nothing like this decentralized approach to data processing and AI development in any competing AI company. Tesla has access to training data that extends WAY beyond traditional laboratory, or server farm paradigms. A radically innovative synergy of automotive technology, AI, and user-supported infrastructure. Tesla's position in a totally new archetype in the realm of Cloud AI computing is totally bonkers! Beyond that, once Tesla vehicles are permitted to operate fully autonomously during charge cycles, the implications extend profoundly into both technological and socio-economic realms. There is no previous president to use in order to gain comparative understanding. We are about to see Tesla transform the fleet into a truly dynamic and self-optimizing network of mobile computing resources, available 24/7. Economically, it revolutionizes the concept of vehicle ownership, turning these assets into a dynamic income generator for their owners, as vehicles will certainly offer services like ride-sharing, or possibly a parked vehicle lend's it's idle compute to data processing as a node in the network. The integration of fully autonomous Teslas into daily life and industry is going to usher in a new era of economic innovation.

[@merlinmidday4644]

These versions already need a pop up hand 👍 to acknowledge other drivers....😂

[@darwinboor1300]

FSD does not produce human behavior. FSD generates human-like output responses to pixels in that mimic human behavior. FSD is not sentient. FSD is incapable of reasoning or acting volitionally. FSD V12 is great. It clearly adds to driver/passenger safety and reduces driver fatigue. FSD has a new set of clothes. FSD V12 remains unable to pass a behind the wheel driver's exam. It still needs an intelligent supervisor. FSD is not programmed. FSD is a set of keys (output controls). With V12 any key is selected statistically based upon its training and the current pixels in. No two driving circumstance are identical. FSD can and will choose the wrong key on occasion. Trying to go straight with a left turn only green. Speeding in an active school zone or on enterring a housing development are simple examples of FSD V12 choosing the wrong key. FSD V12 is still level 2-3 autonous driving. Adding up-to-date geofencing might get FSD to level 4. (The B pillar cameras are too far back and need better distance resolution.) FSD training is complex and compute intensive. The FSD team does not get enough credit. We spend way too much time talking about FSD and way way too little time discussing the training system that trains FSD and its AIs. They are evolving to be Tesla's realworld Q*. When FSD errors are identified. New datasets are acquired. A new round of training is performed until the errors appear to have resolved. If necessary additional data is acquired and/or generated and training is repeated. Eventually, a release candidate is sellected. Testing moves into validation to ensure the errors are fixed and to screen for new errors that emerge from the most recent training. At the same time training begins to fix new errors. Validation is first performed in simulation. If the candidate is passed, it moves out to Tesla testers, then to early beta testers, and only reaches the wide pool of beta testers in stages if it passes the earlier validation stages. I vote for FSD passing a practical driver's test, fixing the error that pulled FSD freeway driving, and being able to use ALL controls including reverse.

[@70lilthumpert]

Yes. Side repeater cameras just need about 15 degrees more.

[@tacoyaky8872]

If they can just mount even a very low res wireless camera to the front (just connected to a power source), I think they can add this to their fleet fairly easily and cheaply (these already exists as wireless security cameras). It doesn't even need to be super responsive. It can have a full 1 second lag and still help verify if any objects are approaching from a blind spot in the front.

[@garycarson3128]

People accept that people aren’t perfect and humans are fallible. Humans hold machines to a higher standard. People are less forgiving or have a much higher standard for machines killing people even if humans create these machines. It is psychological harder for parents to accept a death of a child or love one by a Robotaxi than by a drunk driver especially if the Robotaxi company is a huge Trillion Dollar corporation! When Tesla Robotaxis becomes ubiquitous throughout the USA and are use for perhaps 80% of passenger miles, it is conceivable that Robotaxis reduce the death rate from 117 per day to 1 per day or 365 FSD caused deaths per year, the public will have a harder time accepting each death and people will expect much higher settlements than they currently do. The rarer traffic fatalities become, the more attention will be paid on each fatality.

[@johnfurr6060]

Man I can't wait to for the upgrade. The context clues part is killing me! I so want to use this. All these things have to be right or humans just want trust it. So jealous of you guys using this already!

[@Alarix246]

VW announced that they are the first company which has a Level 5 car and system: will be released in 2026. 🤣 Couldn't stop laughing.

[@richardjones8395]

With regards to nobody wanting to buy anything but a Tesla once FSD is solved, this is the way I think about it... The demand for Tesla vehicles will obviously go up and if it goes up dramatically, it is likely that Tesla will have to raise prices again to match supply and demand. The market will dictate what a Tesla is worth and what other 'non-Tesla' vehicles without FSD are worth. People who can't afford to pay the premium of what a Tesla is now worth will still buy vehicles from other manufacturers. When this happens though, you know that Tesla is going to put the peddle to the metal with expanding production and Gen 3 is going to be ramping up to huge numbers. Other automakers are already in a difficult position (ICE sales will be going down and they need to ramp up EVs that they lose money on), but they will also have another disadvantage in not having FSD on top of that. There will be a lot of consolidating in the industry and automakers are going to shrink from where they are now and a number of them will probably go bankrupt. If they want to survive, it is likely that automakers will have to bite the bullet and partner with Tesla (just like they are doing with NACS and the supercharger network), or perhaps partner with a Chinese company. It's going to be an interesting time and this is what happens when disruptive technologies come along...

[@trent_carter]

Great discussion

[@zytuben]

FSD has been driving better than humans since day 1. Humans can't even use their indicators.

[@Mr.E.Shoppa]

Elon is more "sandbaggy" because he has to think about Tesla stock. TSLA. Get some today while you have a chance.