Tesla Full Self Driving Availability Prediction

[victorplus]

I bet anything that Tesla already has test vehicles out there running FSD that can meet or exceed what Mobileye has been showing off with their system in this video, but that they're simply waiting to "flip the switch" to release it to all cars with FSD once it's ready for prime time. I was losing faith in Tesla's ability to get FSD to work with current hardware until I saw this. After all, as humans we have 2 eyes, pointing in one direction (typically straight) with no radar or ultrasonic sensors, and we generally go just fine. Why can't 8 cameras all around + forward long range radar + 12 ultrasonic sensors do the same or better with excellent software/processing behind them? I'm optimistic now![/QUOTE]

Mobileye is the most advanced on the way to vision only self driving, and Tesla is taking the same road (even if there HD mapping is less clear and less advanced, they should be able to get there). But if you follow closely Mobileye development they are aiming for an independent Lidar only self driving working on their EyeQ5 to get the redundancy required for a true level4/5.

Vision only can work very well, but is it well enough to cover all the very complex edge cases required to get to a high reliability level, its less than certain. Human don't have 8 camera and radar but they can understand context. A simple shadow can be very tricky even for the best NN vision system, so the system is not sure if there is something or not.

 

[Barklikeadog]

I bet anything that Tesla already has test vehicles out there running FSD that can meet or exceed what Mobileye has been showing off with their system in this video, but that they're simply waiting to "flip the switch" to release it to all cars with FSD once it's ready for prime time.
.

Lol thanks for that it is a dreary day here in portugal

 

[Phrixotrichus]

I bet anything that Tesla already has test vehicles out there running FSD that can meet or exceed what Mobileye has been showing off with their system in this video, but that they're simply waiting to "flip the switch" to release it to all cars with FSD once it's ready for prime time.

If you´ve been following the reports and demos of similar systems in Audi/Merc/Nissan etc.then I´d even dare to say that ALL of them have systems comparable in capability in testing right now.
However I´m also willing to bet that none of those will actually be released to the public as a true lvl 4/5 system in the near future. Demos with a fixed thorroughly tested route are one thing, but true everyday use without having to worry about being sued for errors is a completely different chapter.

Contrary to the classic manufacturers Tesla has been known to release beta software/features so maybe they`ll just release some 90% reliable lvl4/5 system in the next 2-3 years and label it as level 2 until they can guarantee it`s actually always safe to use it some years later.

 

[Matias]

Yep. I know it well. It's one of the classics of current machine learning and the results are fascinating. And it's examples like this that the media has latched onto to leap to the conclusion that the AI overlord has almost arrived.

Under the hood, what Karpathy did here was pretty simple. And, while it puts out results that look almost like a real document, it will never be capable of producing something truly coherent. To do that requires deep knowledge of subject matter, and that's a much more complex problem.

People have done some funny things working from Karpathy's RNN. One produced surprisingly good bible verse. Another one took the entire text of Deepak Chopra's work, and then auto-generated more Chopra-like text. This latter example produced word salads almost indistinguishable from Chopra's actual words. But that says more about Chopra's "wisdom" than it does about the wisdom of the RNN.

Some trivial things cannot be done with NN’s. E.g. Mobileye’s founder Amnon Shashua told in one of his lectures, that no one has managed to multiply large numbers with NNs. So you can train NN to recognize hand written numbers, but if you want to multiply those numbers, you have to use traditional digital logic.

Recognizing hand written numbers has been difficult for computers until the advent of NN. On the other hand multiplying numbers has been done with computers since 1930’s.

My point is, that you have to combine different techniques. NN are not the silver bullet for everything.

 

[sandpiper]

Some trivial things cannot be done with NN’s. E.g. Mobileye’s founder Amnon Shashua told in one of his lectures, that no one has managed to multiply large numbers with NNs. So you can train NN to recognize hand written numbers, but if you want to multiply those numbers, you have to use traditional digital logic.

Recognizing hand written numbers has been difficult for computers until the advent of NN. On the other hand multiplying numbers has been done with computers since 1930’s.

My point is, that you have to combine different techniques. NN are not the silver bullet for everything.

Exactly. A 100 year old mechanical calculator can do a much better job of multiplying than the world's largest biological NNs.

 

[victorplus]

Contrary to the classic manufacturers Tesla has been known to release beta software/features so maybe they`ll just release some 90% reliable lvl4/5 system in the next 2-3 years and label it as level 2 until they can guarantee it`s actually always safe to use it some years late

Agreed. Autopilot will be level 2 for still a very long time. Even if the system can get much better avoiding the obvious deadly collision with road obstacle and stop objects, I'm not sure at all people will be confortable sleeping in their car on the highway or in the city trusting 100% the system especially without the added redundancy of an independent Lidar system.

Aviation automation is highly valuable for safety analysis of self driving systems :
- automation is saving life reducing human error.
- but automation can lead to accident of their own
- human can save the day when automation fail or the situation is too tricky
- automatic and human cooperation is difficult and out of loop lack of situational awareness of what the machine is doing is deadly even for pilots who much are more trained than drivers (AF447, autopilot problem leading to stall...)
- Full redundancy is key (and even with 3 redundant system you can fail due to a common cause as pitot freezing on AF447)

Even with level4/level5 people will not sleep in their vehicle and will keep monitoring especially if you don't have the best of the best of systems (Vision only), except perhaps at low traffic jam speed where accident consequences are less serious.

Even with the best systems, accidents will happen due to very complex edge cases or hardware failure, accident that would not have happen to a basic human driver.

As it mathematically sure that independent vision + independent lidar is superior to only vision, only the first one has a chance to be authorized without a driver as Tesla as been touting. Nobody will take the risk of car driving itself for pure convenience ( summon) killing somebody, when the accident would have been avoided with superior hardware.

Again, being an investor and fan I don't want to bash Tesla, but the Summon everywhere FSD vision by construction (other constructors companies choose to use better hardware) is not guaranteed at all to happen. Autopilot will be incredibly close to self driving but will still need monitoring (level2) except perhaps at low speed and parking.

As safety is a genuine Tesla concern I don't see how they will avoid adding pilot monitoring to their system. Would be better to recognize it before waiting for the authorities to impose it.

With pilot monitoring and smart HUD to maximize pilot awareness, Tesla would be able to offer close to self diving convenience with genuine best safety level.

 

[D3xDt3Reaction]

Exactly. A 100 year old mechanical calculator can do a much better job of multiplying than the world's largest biological NNs.

Sounds plausible, but what precludes Tesla from using a hybrid of NNs and traditional programming, and how do we know they haven’t already done it?

 

[sandpiper]

Sounds plausible, but what precludes Tesla from using a hybrid of NNs and traditional programming, and how do we know they haven’t already done it?

That is exactly what they're presently doing. As are Waymo and others. The problem is that there's been no revolution in this traditional programming bit. I think it's going to take a long time to develop a really reliable system that will drive anywhere under all circumstances.

 

[Matias]

My guess is that current cars end up level 3ish. Tesla may never publish firmware with terms where "eyes off the road" is acceptable, but current cars eventually will do a really good job of self driving on the highway in good weather.

But I don't think this version of the future hurts Tesla.

What about the fact, that Tesla has sold FSD to (soon) hundreds of thousands cars? Tesla has to reimburse customers who paid FSD, if it can’t deliver it.

 

[Matias]

Sounds plausible, but what precludes Tesla from using a hybrid of NNs and traditional programming, and how do we know they haven’t already done it?

They will most certainly do a hybrid solution.

 

[D3xDt3Reaction]

That is exactly what they're presently doing. As are Waymo and others. The problem is that there's been no revolution in this traditional programming bit. I think it's going to take a long time to develop a really reliable system that will drive anywhere under all circumstances.

At least no public revolution. But if Tesla really was so far from that revolution, why put so many chips on the FSD number by advertising it is relatively imminent when you don’t have to?

 

[Matias]

At least no public revolution. But if Tesla really was so far from that revolution, why put so many chips on the FSD number by advertising it is relatively imminent when you don’t have to?

It is just Elon being Elon. Tesla’s culture is to jump from the airplane with sewing material and needle and hope you have enough time to make a parachute before hitting the ground.

 

[sandpiper]

At least no public revolution. But if Tesla really was so far from that revolution, why put so many chips on the FSD number by advertising it is relatively imminent when you don’t have to?

My opinion is that Elon got ahead of himself, as he's prone to do. nd now they're stuck with it. The T&Cs when you buy FSD are pretty loose; "you'll get it someday we think". I'm no lawyer, but I suspect that Tesla will be mostly in the clear. I expect them to roll out something better than what we have, call it FSD, and let the chips fall where they may.

There's been no "public" revolution, but if anybody were privately ahead of the pack, it would leak out. There's too much interchange of people between the companies doing that sort of work. For sure we're going to see many iterations of improved partial auto-driving well before any form of full self driving. And even that's been slow in coming. There are no systems on the market that are better than what I have in my 2014 AP1. With 4 years we've seen no appreciable improvement, and so I'm having a hard time seeing that extrapolate to FSD within the next decade.

 

[sandpiper]

It is just Elon being Elon. Tesla’s culture is to jump from the airplane with sewing material and needle and hope you have enough time to make a parachute before hitting the ground.

That's an awesome analogy. I'm going to have to remember that!!!

 

[D3xDt3Reaction]

It is just Elon being Elon. Tesla’s culture is to jump from the airplane with sewing material and needle and hope you have enough time to make a parachute before hitting the ground.

My opinion is that Elon got ahead of himself, as he's prone to do. nd now they're stuck with it. The T&Cs when you buy FSD are pretty loose; "you'll get it someday we think". I'm no lawyer, but I suspect that Tesla will be mostly in the clear. I expect them to roll out something better than what we have, call it FSD, and let the chips fall where they may.

There's been no "public" revolution, but if anybody were privately ahead of the pack, it would leak out. There's too much interchange of people between the companies doing that sort of work. For sure we're going to see many iterations of improved partial auto-driving well before any form of full self driving. And even that's been slow in coming. There are no systems on the market that are better than what I have in my 2014 AP1. With 4 years we've seen no appreciable improvement, and so I'm having a hard time seeing that extrapolate to FSD within the next decade.

You guys are all right and I respect your opinions. Whether you are right remains to be seen, but well said. Parachute analogy was great.

 

[MelaniainLA]

What about the fact, that Tesla has sold FSD to (soon) hundreds of thousands cars? Tesla has to reimburse customers who paid FSD, if it can’t deliver it.

It’s not just a matter of reimbursing the FSD cost. What about all these costs related to FSD purchase:

• Sales tax
• Annual registration fee (property tax) tied to purchase price
• The fact that so many bought the cars to begin with on the promise that they’ll ultimately have FSD and “come with FSD capable hardware”?

If you refund just the $3000 FSD price you’re not accounting for all of the above. And if customers argued that they felt duped into buying the car on the promise it would eventually have FSD, they can claim fraud. Finally, all owners could make a claim based on the lower resale value of the car because it doesn’t actually have FSD capability.

I’m suspecting none of this will matter. They’ll figure out the software, update hardware as needed, and get things to work. Time will tell!

 

[D3xDt3Reaction]

This poll has now had 100 votes, and the weighted average implies around 2022 for FSD depending on how you treat the "2025 or later" bucket. We've had some pretty well-read individuals who have indicated we're still far from a FSD solution. Would anyone from the "2020 or earlier" camp like to provide their rationale?

2018: 2
2019: 13
2020: 21
2021: 12
2022: 18
2023: 3
2024: 5
2025-: 26

 

[EinSV]

This poll has now had 100 votes, and the weighted average implies around 2022 for FSD depending on how you treat the "2025 or later" bucket. We've had some pretty well-read individuals who have indicated we're still far from a FSD solution. Would anyone from the "2020 or earlier" camp like to provide their rationale?

2018: 2
2019: 13
2020: 21
2021: 12
2022: 18
2023: 3
2024: 5
2025-: 26

I don't have a specific date in mind -- at this point it seems very difficult to predict with any accuracy.

I did run across an article I thought had some interesting nuggets suggesting that people may tend to radically overestimate how long it will take AI to solve problems. The classic example is the game Go:

In 1997, IBM’s Deep Blue beat world chess champion Gary Kasparov, the first time an AI technology was able to outperform a world expert in a highly complicated endeavor. It was even more impressive when you consider they were using 1997 computational power. In 1997, my computer could barely connect to the internet; long waits of agonizing beeps and buzzes made it clear the computer was struggling under the weight of the task.

Even in the wake of Deep Blue’s literally game-changing victory, most experts remained unconvinced. Piet Hut, an astrophysicist at the Institute for Advanced Study in New Jersey, told the New York Times in 1997 that it would still be another hundred years before a computer beat a human at Go.

Admittedly, the ancient game of Go is infinitely more complicated than chess. Even in 2014, the common consensus was that an AI victory in Go was still decades away. The reigning world champion, Lee Sedol, gloated in an article for Wired, “There is chess in the Western world, but Go is incomparably more subtle and intellectual.”

Then AlphaGo, Google’s AI platform, defeated him a mere two years later. How’s that for subtlety?​

Why are AI predictions so terrible?

The author cites a number of reasons so many people got this so wrong, including "human bias" (we like to think we're really awesome) and "humans are really bad at understanding exponential growth" -- we tend to think of improvements coming in a linear fashion.

As @MelaniainLA says, all of the hardware necessary for self-driving appears to be in place (with the possible exception of the chip upgrade Tesla has said it will provide if necessary).

With Karpathy and his team refining the neural nets and the fleet collecting and analyzing reams of data, improvements could come a lot faster than many people expect, like they did with AlphaGo.

 

[Barklikeadog]

I don't have a specific date in mind -- at this point it seems very difficult to predict with any accuracy.

I did run across an article I thought had some interesting nuggets suggesting that people may tend to radically overestimate how long it will take AI to solve problems. The classic example is the game Go:

In 1997, IBM’s Deep Blue beat world chess champion Gary Kasparov, the first time an AI technology was able to outperform a world expert in a highly complicated endeavor. It was even more impressive when you consider they were using 1997 computational power. In 1997, my computer could barely connect to the internet; long waits of agonizing beeps and buzzes made it clear the computer was struggling under the weight of the task.

Even in the wake of Deep Blue’s literally game-changing victory, most experts remained unconvinced. Piet Hut, an astrophysicist at the Institute for Advanced Study in New Jersey, told the New York Times in 1997 that it would still be another hundred years before a computer beat a human at Go.

Admittedly, the ancient game of Go is infinitely more complicated than chess. Even in 2014, the common consensus was that an AI victory in Go was still decades away. The reigning world champion, Lee Sedol, gloated in an article for Wired, “There is chess in the Western world, but Go is incomparably more subtle and intellectual.”

Then AlphaGo, Google’s AI platform, defeated him a mere two years later. How’s that for subtlety?​

Why are AI predictions so terrible?

The author cites a number of reasons so many people got this so wrong, including "human bias" (we like to think we're really awesome) and "humans are really bad at understanding exponential growth" -- we tend to think of improvements coming in a linear fashion.

As @MelaniainLA says, all of the hardware necessary for self-driving appears to be in place (with the possible exception of the chip upgrade Tesla has said it will provide if necessary).

With Karpathy and his team refining the neural nets and the fleet collecting and analyzing reams of data, improvements could come a lot faster than many people expect, like they did with AlphaGo.

Lets just shorten this. Make it simple. At what point will the cars be able to make a reliable and safe illegal left turn? 3 yrs 5 yrs 10 yrs?

 

[Julie's MS]

If Porsche is planning to launch Level 4 in 2019. Lets hope Tesla is ready before them

Porsche's all-electric rival to the Tesla Model S is scheduled to go on sale in 2019 and will spawn multiple variants

Blume also confirmed that Porsche plans for the Taycan to have Level 4 autonomous driving technology (self-driving in nearly all situations, with driver attention not required), but denied that it would allow fully autonomous driving over longer distances. “There are situations in traffic jams where you will be able to read a newspaper, but our customers take pleasure from driving and this will remain,” he said.

Porsche Taycan name confirmed for production version of Mission E | Autocar