FSD Fact or Fiction? See a refund in the future?

[LoneWolfO6]

https://apple.news/AFVVA-nLLTY2ddLRBSXYG3w

Hope they can get the human factor incorporated into the FSD tech. Similar to learning English for the first time, English being the hardest language to learn due to all the slang. The FSD just can’t get the human factor down to fully move to level 5 autonomous driving. Seen this happen in the past selling a product that cant be delivered.

“Elizabeth Holmes dropped out of Stanford University at 19 to start blood-testing startup Theranos, and grew the company to a valuation of $9 billion. But it all came crashing down when the shortcomings and inaccuracies of the company's technology were exposed, and Theranos and Holmes were charged with "massive fraud."Aug 11, 2020 Business Insider.

 

[OneLapper]

Good grief dude.

Did you read this? Did you see this line?

I quote; " [emphasis added and hat tip to Gordon Johnson for the lead]"

Gordon Johnson! Probably the worst TSLA short out there. Annual returns of -7% ytd. Ranked one the worst investors to give your money to. Source: Tip Ranks Gordon Johnson

I'm sure Gordon PAYS people to write this stuff. Who on earth would take a "lead" from Gordon Johnson???

 

[Crissa]

Also, apple links get redirected based upon your platform, and they can expire, so it's best to link directly to the source.

I can't comment on the article, because I can't read it.

-Crissa

 

[Throwcomputer]

Anyone else notice the absurdity of promoting FSD as a life saving feature to reduce human error, but then implementing a driver grading system and only opening up the FSD beta to the safest drivers??

Seems to me they should open it up to the worst drivers first if they want to prove their claims of drastic reduction in human error. hahah.

 

[TyPope]

Anyone else notice the absurdity of promoting FSD as a life saving feature to reduce human error, but then implementing a driver grading system and only opening up the FSD beta to the safest drivers??

Seems to me they should open it up to the worst drivers first if they want to prove their claims of drastic reduction in human error. hahah.

The safest drivers pay attention so Tesla wants people who pay attention to help improve the system. That way, when they DO release it to everyone, including those people who don't pay attention, the car will be able to figure out sticky situations. I suppose it's the same reason that the Air Force uses test pilots to try out new equipment instead of new pilots with limited experience.

 

[Crissa]

Anyone else notice the absurdity of promoting FSD as a life saving feature to reduce human error, but then implementing a driver grading system and only opening up the FSD beta to the safest drivers??

Seems to me they should open it up to the worst drivers first if they want to prove their claims of drastic reduction in human error. hahah.

The Beta is about teaching FSD to be safe, so of course they recruited the most motivated and skilled drivers they could select.

Unlike paid test drivers, these drivers passed a test. They control when and how long they use the test software - they can turn it off at any time, They're also invested in the routes; they get to choose where it goes, and use it while they're going from place to place as they normally would.

That means Beta drivers are fresh and ready to take over, compared to safety drivers for other companies that don't get to chose when to drive, can't just stop, don't choose the routes, may not even know the route, and hence (like the uber case) won't be fresh and ready to take over.

FSD when perfected would be best to give to the worst drivers. But only when they're no longer allowed to override it. How many accidents are blamed on Autopilot turn out to be someone ignoring or overriding its warnings?

-Crissa

 

[Throwcomputer]

The safest drivers pay attention so Tesla wants people who pay attention to help improve the system. That way, when they DO release it to everyone, including those people who don't pay attention, the car will be able to figure out sticky situations. I suppose it's the same reason that the Air Force uses test pilots to try out new equipment instead of new pilots with limited experience.

Multiple problems with this belief..

They currently have to deal with error correction even with only the highest rated drivers included... otherwise you have bad data entering the model. And when its in its public production release format, it will never be locked permanent, but will always be evolving, so they have to account for error correction of bad data from bad drivers anyways at pretty much all stages of this software development. They need this bad data though.. to teach the model what not to do. So to limit the user base to only the best, simply on this idea of "they need the best drivers to teach the model" actually slows down the model improvements by having less "what not to do" examples to learn from. So they already do currently have some system of saying "don't add this mistake to the model", whether it be manual oversight by a person at Tesla viewing data and marking good/bad, or by implementing the same grading system to their data acceptance check into the model by saying something like "users with sub 80% driver rating are disallowed from adding raw data from this insance when this instance deviates from the average "acceptable" response in similar situations."

Now if they currently already do still use camera/driving data from bad drivers who are not allowed in the beta program, to provide this "what not to do" data, the argument against now allowing those same bad drivers into the beta is that regardless of driver rating (bad or good), each driver has a 1:1 chance of an accident when using FSD beta, because the error proneness is removed from the driver and placed onto the software. Driver skill no longer plays a role. They will never be able to remove the human error proneness of either the "good" or the "bad" driver from actually intervening when something goes wrong. In fact, I think it would make a good driver slightly less good because they are slightly less attentive or ready to respond instantly.

Additionally, by limiting the FSD Beta to only the best driver ratings, you lose out on the real world statistics of comparing the bottom 10% of driver rating users "errors/crashes" percentage prior to FSD Beta usage with their statistical improvements during usage of FSD Beta. This is actually where they want to be.. proving the FSD software is actually saving lives and removing the human error factor at rates they claim it to.

Limiting FSD Beta serves no real purpose from a software design or a proof of concept perspective. I realize they are being forced to do it legally from federal oversight. I was just trying to point out the futility of this restriction.

PS. This is coming from a potential customer with zero interest and or faith in FSD. Simply looking at it from a logical perspective, the driver rating requirement is pointless. I'm arguing for larger beta user base with worse driver rating because it only makes more sense to give it a better chance of learning correctly.

 

[TyPope]

Multiple problems with this belief..

They currently have to deal with error correction even with only the highest rated drivers included... otherwise you have bad data entering the model. And when its in its public production release format, it will never be locked permanent, but will always be evolving, so they have to account for error correction of bad data from bad drivers anyways at pretty much all stages of this software development. They need this bad data though.. to teach the model what not to do. So to limit the user base to only the best, simply on this idea of "they need the best drivers to teach the model" actually slows down the model improvements by having less "what not to do" examples to learn from. So they already do currently have some system of saying "don't add this mistake to the model", whether it be manual oversight by a person at Tesla viewing data and marking good/bad, or by implementing the same grading system to their data acceptance check into the model by saying something like "users with sub 80% driver rating are disallowed from adding raw data from this insance when this instance deviates from the average "acceptable" response in similar situations."

Now if they currently already do still use camera/driving data from bad drivers who are not allowed in the beta program, to provide this "what not to do" data, the argument against now allowing those same bad drivers into the beta is that regardless of driver rating (bad or good), each driver has a 1:1 chance of an accident when using FSD beta, because the error proneness is removed from the driver and placed onto the software. Driver skill no longer plays a role.

Additionally, by limiting the FSD Beta to only the best driver ratings, you lose out on the real world statistics of comparing the bottom 10% of driver rating users "errors/crashes" percentage prior to FSD Beta usage with their statistical improvements during usage of FSD Beta. This is actually where they want to be.. proving the FSD software is actually saving lives and removing the human error factor at rates they claim it to.

Limiting FSD Beta serves no real purpose from a software design or a proof of concept perspective. I realize they are being forced to do it legally from federal oversight. I was just trying to point out the futility of this restriction.

You are looking at this wrong. I didn't say to only use good input. I said only use good (attentive, actually) drivers. Tesla is building autopilot based on situations the car "sees". Good drivers and bad drivers see the same things. The bad drivers are less likely to take over should autopilot make a mistake.

Imagine if all your test drivers didn't pay attention and you ended up with a bunch of wrecked cars that were all crashed on the first few days of your trial. That doesn't exactly help further the autopilot narrative. Plus, those wrecked cars can no longer provide input.

On the other side, your attentive driver stops the impact from the accident from happening, presses the record button, and continues driving and providing input. His input from an "almost" accident is just as valuable as an inattentive driver's accident... but it's more sustainable.

 

[Throwcomputer]

You are looking at this wrong. I didn't say to only use good input. I said only use good (attentive, actually) drivers. Tesla is building autopilot based on situations the car "sees". Good drivers and bad drivers see the same things. The bad drivers are less likely to take over should autopilot make a mistake.

Imagine if all your test drivers didn't pay attention and you ended up with a bunch of wrecked cars that were all crashed on the first few days of your trial. That doesn't exactly help further the autopilot narrative. Plus, those wrecked cars can no longer provide input.

On the other side, your attentive driver stops the impact from the accident from happening, presses the record button, and continues driving and providing input. His input from an "almost" accident is just as valuable as an inattentive driver's accident... but it's more sustainable.

I am also making the argument that good drivers aren't any more or less likely to respond in a mistake, but that good drivers are actually being encouraged to be "slightly" less good at responding in those instances, because they are not actually driving, hence depending on the FSD in some way that affects response time and or attention. Maybe not "as bad", but definitely slightly less willing and able simply from not actually driving at the moment. There is lag between recognition and response when you are observing a vehicle doing things you are riding in, vs actually driving at the moment.

 

[Crissa]

They need this bad data though.. to teach the model what not to do.

They do have this data: All cars with Autopilot, opted into their feedback system let a stack run in shadowing mode. And then they conglomerate this data back at home (and my ultra detailed maps for FSD to 'remember' stuff.)

This is how they rate Autopilot, too: It's always running in the background, trying to guess what will happen next, when it's safe to turn on. And so Tesla collects that data so it can have the same sample data from people who drive around and never turn Autopilot on (but Autopilot was ready) vs people who did turn Autopilot on. That way they're not counting miles where Autopilot couldn't be active in their comparisons.

-Crissa

 

[Throwcomputer]

They do have this data: All cars with Autopilot, opted into their feedback system let FSD run in shadowing mode. And then they conglomerate this data back at home (and my ultra detailed maps for FSD to 'remember' stuff.)

This is how they rate Autopilot, too: It's always running in the background, trying to guess what will happen next, when it's safe to turn on. And so Tesla collects that data so it can have the same sample data from people who drive around and never turn Autopilot on (but Autopilot was ready) vs people who did turn Autopilot on. That way they're not counting miles where Autopilot couldn't be active in their comparisons.

-Crissa

Now if they currently already do still use camera/driving data from bad drivers who are not allowed in the beta program, to provide this "what not to do" data, the argument against now allowing those same bad drivers into the beta is that regardless of driver rating (bad or good), each driver has a 1:1 chance of an accident when using FSD beta, because the error proneness is removed from the driver and placed onto the software. Driver skill no longer plays a role. They will never be able to remove the human error proneness of either the "good" or the "bad" driver from actually intervening when something goes wrong. In fact, I think it would make a good driver slightly less good because they are slightly less attentive or ready to respond instantly.

 

[CyberGus]

Per Elon on the earnings call, the beta will be open to all FSD owners in the next month.

 

[JBee]

Yay! Now I just need a car that runs it....

 

[Crissa]

Per Elon on the earnings call, the beta will be open to all FSD owners in the next month.

A month or so.

-Crissa

 

[CyberGus]

A month or so.

-Crissa

It’s Elon, the “or so” was implied ?