FSD video completely fake?

[Matias]

Isn’t it obvious, that now 2.5 years (Sic!) later we can say, that for all practical purposes the video was completely fake.

Even if Tesla years from now will develop functional FSD, it doesn’t change the fact, that the video per se was completely fake.

 

[electronblue]

some anonymous guy on reddit is not exactly what I would call a credible source.

History has shown anonymous Tesla sources being quite often useful. But be that as it may it is still one potential source on the validity of that video that is supported by what is known of Tesla’s actual (lack of) progress — and there really aren’t many such sources.

Musk for example clearly is not a reliable source for AP2 progress info... so what else do we have?

 

[mongo]

Isn’t it obvious, that now 2.5 years (Sic!) later we can say, that for all practical purposes the video was completely fake.

Even if Tesla years from now will develop functional FSD, it doesn’t change the fact, that the video per se was completely fake.

The self driving video was on a blog post titled from Oct 2016:
All Tesla Cars Being Produced Now Have Full Self-Driving Hardware

Notice how it doesn't say software. This was in the time period immediately after the AP1 MobilEye split and Tesla needed to show that, even though the feature set was currently reduced, the cars were capable of performing the tasks.

The last paragraph is practically all disclaimer:

Before activating the features enabled by the new hardware, we will further calibrate the system using millions of miles of real-world driving to ensure significant improvements to safety and convenience. While this is occurring, Teslas with new hardware will temporarily lack certain features currently available on Teslas with first-generation Autopilot hardware, including some standard safety features such as automatic emergency braking, collision warning, lane holding and active cruise control. As these features are robustly validated we will enable them over the air, together with a rapidly expanding set of entirely new features. As always, our over-the-air software updates will keep customers at the forefront of technology and continue to make every Tesla, including those equipped with first-generation Autopilot and earlier cars, more capable over time.

Which echos the original AP blog post of October, 2014

The launch of Dual Motor Model S coincides with the introduction of a standard hardware package that will enable autopilot functionality. Every single Model S now rolling out of the factory includes a forward radar, 12 long range ultrasonic sensors positioned to sense 16 feet around the car in every direction at all speeds, a forward looking camera, and a high precision, digitally controlled electric assist braking system.

Building on this hardware with future software releases, we will deliver a range of active safety features, using digital control of motors, brakes, and steering to avoid collisions from the front, sides, or from leaving the road. (Please note this hardware is not available as a retrofit.)

Model S will be able to steer to stay within a lane, change lanes with the simple tap of a turn signal, and manage speed by reading road signs and using active, traffic aware cruise control. It will take several months for all Autopilot features to be completed and uploaded to the cars.

Our goal with the introduction of this new hardware and software is not to enable driverless cars, which are still years away from becoming a reality. Our system is called Autopilot because it’s similar to systems that pilots use to increase comfort and safety when conditions are clear. Tesla’s Autopilot is a way to relieve drivers of the most boring and potentially dangerous aspects of road travel – but the driver is still responsible for, and ultimately in control of, the car.

The Autopilot hardware opens up some exciting long term possibilities. Imagine having your car check your calendar in the morning (a feature introduced in Software v6.0), calculate travel time to your first appointment based on real time traffic data, automatically open the garage door with Homelink, carefully back out of a tight garage, and pull up to your door ready for your commute. Of course, it could also warm or cool your car to your preferences and select your favorite morning news stream.

The introduction of this hardware is just the first step for Autopilot in Model S. We will continue to develop new capabilities and deliver them through over-the-air software updates, keeping our customers at the forefront of driving technology.

October 2018:

Every single Model S, Model X and Model 3 ever built will receive new features and improvements from Software Version 9.0. This is just the latest reminder that when you own a Tesla, you own a car that will only get better over time.

(it looks likes Tesla prefers even yeared Octobers for AP updates....)

 

[diplomat33]

History has shown anonymous Tesla sources being quite often useful. But be that as it may it is still one potential source on the validity of that video that is supported by what is known of Tesla’s actual (lack of) progress — and there really aren’t many such sources.

Musk for example clearly is not a reliable source for AP2 progress info... so what else do we have?

True. At the end of the day, all we really have is what we actually get in our cars in an update. Everything else, especially Musk's quotes, are future telling which are subject to change.

In terms of the 2016 video, I do think the reddit may be partially right. I do think Tesla probably did some hard coding to make the car do what it does in the video. I say this because we know for a fact from a talk Karpathy did, that at the time of the 2016 video, Tesla was using a so-called "software 1.0" approach to FSD, ie they were trying to hard code everything. So Tesla probably used AP1 software + hard coding to make the car do what it does in the video. I am speculating but Tesla probably thought back then that they could just hard code FSD and it would be "easy". Hence, why they started selling FSD and claiming it just needed regulatory approval. Of course, that method was woefully flawed which is why Tesla radically changed their approach and is now using machine learning to do FSD which is bearing much better fruit.

 

[JeffK]

Why is there still debate on this? As far as the videos themselves, they are 100% real as far as the hardware did drive that route on its own. It was a hardware (mostly sensor) demonstration. It was never a demonstration of the software and that same software would never have worked on alternate routes besides the two it was trained on. The route itself was painstakingly trained upon hence the number of disengagements reported and the number of test vehicles making this exact same trip.

We also know that some of the rules were hardcoded in. We know this specifically since Sterling Anderson had mentioned it in several talks. Remember that Andrej Karpathy didn't join Tesla until the middle of 2017 after Chris Lattner messed stuff up royally.

In sum, there's nothing fake about the hardware demonstration, but it was just that, a demo. Remember Elon said:

Demoware is easy; production software is hard... It’s easy to do a cool demo, it’s hard to put something out. Especially software that’s going to work on millions of different roads all around the world in a wide range of circumstances—in winter, in summer, in rain, in dust—there’s a world of difference there.

 

[lunitiks]

after Chris Lattner messed stuff up royally

Please expand

 

[JeffK]

Please expand

After the release of hardware 2.0, Sterling, who was actually qualified for that role since he has a history in assisted driving, failed to deliver software parity with AP 1.0 by the end of 2016. Chris, who might be qualified to lead but definitely not qualified for that role, then came in forced a major rewrite of the NN architecture and also failed to delivery parity in 6 months. He was then ousted for not producing results and generally wasting everyone's time. In addition, later Tesla was subjected to the lawsuit about failing to deliver software parity in a timely manner forcing them to issue partial refunds. Andrej, who actually is qualified for that role, comes in, has the team do yet another rewrite to finally achieve parity with AP 1.0.

I feel that if Andrej would have come on board sooner they could have saved six months or, in the least, moved the timeline forward.

 

[Bladerskb]

After the release of hardware 2.0, Sterling, who was actually qualified for that role since he has a history in assisted driving, failed to deliver software parity with AP 1.0 by the end of 2016. Chris, who might be qualified to lead but definitely not qualified for that role, then came in forced a major rewrite of the NN architecture and also failed to delivery parity in 6 months. He was then ousted for not producing results and generally wasting everyone's time. In addition, later Tesla was subjected to the lawsuit about failing to deliver software parity in a timely manner forcing them to issue partial refunds. Andrej, who actually is qualified for that role, comes in, has the team do yet another rewrite to finally achieve parity with AP 1.0.

I feel that if Andrej would have come on board sooner they could have saved six months or, in the least, moved the timeline forward.

What kind of revisonary history is this? Sterling didn't fail to deliver parity by end of 2016, it was impossible and not only that but to deliver EAP aswell that elon was claiming. The amount of deflection from obvious lies that elon says is amazing. Sterling didn't fail. Elon lied plain and simple. Lastly it took Andrej 12 months to achieve anything resembling parity in performance with AP1 and this is with all the leg work, firmware and software framework that had already been done before he took over. yet elon wanted EAP by end of 2016, not even just AP1 parity.

 

[Matias]

Funny thing is, i’ve been under the impression, that U.S. is the promised land of class action lawsuits and “see you in court”, yet Tesla customers just bend over...

 

[electronblue]

Interestingly for all his prowess Andrej Karpathy hasn’t been able to reach AP1 parity in speed-sign recognition and rain sensing. MobilEye has done the former for over a decade and dedicated sensors have done the latter even longer.

The rain sensing in AP2 is close to useless and of course speed-signs are not sensed at all.

 

[verygreen]

The rain sensing in AP2 is close to useless and of course speed-signs are not sensed at all.

rain sensing works ok during daylight for me. Dedicated sensors use a different idea so they are not really comparable.

 

[electronblue]

rain sensing works ok during daylight for me. Dedicated sensors use a different idea so they are not really comparable.

Of course because AP2 lacks an active light source for rain-sensing. Why did they have to shoot themselves in the foot over this too?

 

[mongo]

Interestingly for all his prowess Andrej Karpathy hasn’t been able to reach AP1 parity in speed-sign recognition and rain sensing.

AP0 and AP1 both used part #1011582-00-C for rain sensing, so I'm not sure what you are getting at...

 

[electronblue]

AP0 and AP1 both used part #1011582-00-C for rain sensing, so I'm not sure what you are getting at...

Of course? As I said, dedicated sensors have done this for years. My point is Karpathy has not been able to match their performance.

 

[JeffK]

elon wanted EAP by end of 2016, not even just AP1 parity.

No, from the autopilot conference call:

So we expect to reach feature parity following field validation of hardware 2.0 probably in December (2016) so maybe two or three months from now. So for two or three months is actually hardware 1.0 car will be better than a hardware 2.0 car and then approximately every two or three months thereafter is when we expect to release significant improvement in autonomous capability.

 

[electronblue]

@JeffK The EAP Design Studio text did fool a lot of people believing EAP would be ready by December 2016 though.

 

[iTech]

Interestingly for all his prowess Andrej Karpathy hasn’t been able to reach AP1 parity in speed-sign recognition and rain sensing. MobilEye has done the former for over a decade and dedicated sensors have done the latter even longer.

The rain sensing in AP2 is close to useless and of course speed-signs are not sensed at all.

ME has a patent on sign recognition, so Tesla has to get around that.

And of course, the CV based rain detection will never work in the dark.

 

[mongo]

Of course? As I said, dedicated sensors have done this for years. My point is Karpathy has not been able to match their performance.

Right, but that was a feature on the car at the same time as AP1, not a feature of AP1. That Karpathy has any rain sense working is feature beyond AP1 unless AP1 had a method to indicate to the car that it needed the windshield cleaned.

 

[JeffK]

@JeffK The EAP Design Studio text did fool a lot of people believing EAP would be ready by December 2016 though.

It certainly did, yes, I don't know if anyone expected that kind of delay... I mean even in this timeframe George Hotz demoed his own version (obviously not to the same standard, but goes to show you can crank out something even remotely similar in a short timeframe especially when you already have the expertise.)

When asked about how much of the data from AP 1.0 they could reuse for 2.0 Sterling said nearly all of it.

 

[electronblue]

Right, but that was a feature on the car at the same time as AP1, not a feature of AP1. That Karpathy has any rain sense working is feature beyond AP1 unless AP1 had a method to indicate to the car that it needed the windshield cleaned.

Of course but that’s splitting hairs. Tesla decided to solve rain-sensing differently between AP1 and AP2 cars and Karpathy (amonst others) was tasked to solve this for AP2+ cars. So far they have pretty much failed in my books.

I am of course not blaming Karpathy for the decision to leave out a rain-sensor. He was not there then.