Maybe I’m dense, but I don’t get the punch line: this poll has a granularity of 1 year. If you want to criticize the poll @ValueAnalyst made, it’s in another thread.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla Full Self Driving Availability Prediction
- Thread starter D3xDt3Reaction
- Start date
ValueAnalyst
Closed
Please stop with the trolling. If you have a problem with people having opinions that are different than yours, this is not the forum for you.
I’m good if @electracity has a legit difference of opinion - the criticism just didn’t even make sense to me here.
No point in skewing a poll that’s at least somewhat serious, though. I like the first amendment, but if you’re going to be critical, at least critique the right thread.
Last edited:
There are a whole suite of applications that are falling like dominos, as a consequence of a revolution in image classification that began in 2012. It was made possible by advances in processing power and new techniques for training multi-layer (deep) neural networks.
Melanoma is one is one of those. Detection of various diseases from MRI imaging are another. The creepy way that facebook can pick people out of images is another. And the current self driving systems use the same basic techniques as well.
These DNN systems are shockingly effective, demand lots of processing power but are pretty simple. And they're completely different from what will be required for true L5 self driving. It's the difference between a 8 inch model rocket and the Saturn 5. They look kinda similar, maybe have the same roots, but aren't remotely the same.
I think media saw the rapid advances to lane following on L2 systems and assumed that L5 was just an extension. It's not. It's almost a completely different problem, and massively more complex.
You spam this whole site with your ignorant self promotion and then you criticize me for "trolling"? Make points you can support or go away.
Remind me again: When does Tesla start buying back shares?
Wouldn’t you agree that such advances have made problems more trivial that weren’t even close to trivial in 2012? If so, why couldn’t further advances make the autonomous car problem tractable in a few more years with the right application of technology, whether that is via neural networks or otherwise?
ValueAnalyst
Closed
I'm open to discussing ideas as adults, but your constant personal attacks, ad hominem, and trolling are detrimental to discussion quality.
Sorry, I'd rather discuss ideas with those who are open to different opinions instead of constant ad hominem.
So... I have to speak with the qualifier that there are likely people on this forum who know vastly more than me. And I'm happy if they chime in.
The L5 problem is that humans rely on a ton of contextual knowledge to deal with all of the unusual circumstances that crop up in an urban environment. And the only way that Waymo is dealing with that is through a ton of hand coding, rule sets, etc... That's not what people do of course. We have an enormous NN in our head that's evolved over millions of years and trained over decades to navigate changing and unpredictable environments.
The paradigm shift that made L2 possible can't simply be extended to L5. We would need many orders of magnitude more processing power.
I do believe we'll get there; but I believe it will be small incremental improvements over decades.
Thanks - classy answer. Would like to hear from those people too.
Not sure if you’ve seen https://web.stanford.edu/class/cs37...ls/KarparthyUNREASONABLY-EFFECTIVE-RNN-15.pdf. I’m a fan of Andrej Karpathy who now works at Tesla. I’m sure he’s just tinkering in that article with “toy” networks, so if he really got going I’d probably be amazed. If anyone can make an AI breakthrough he’s probably on the short list.
Last edited:
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.
MelaniainLA
Member
What hasn’t been getting enough attention here are the financial implications of Tesla releasing FSD. Tesla is heavily incentivized to do so ASAP to continue funding its operations because:
1. For those who prepaid for FSD as part of the purchase (do we know how many?), Tesla cannot realize that income on its balance sheet until it has delivered the product. Activating FSD would add a one-time bump to Tesla’s cash.
2. Activation of FSD would get many to buy Teslas and would get countless existing owners to pony up the $4000 to activate it after purchase, especially once they have had the opportunity to do a free trial. Elon has been waiting to activate the free trial of EAP once the “E” features became available (looks like that might happen soon). Sounds like a good time to start offering the trial. An FSD trial would certainly generate a lot of $4000 sales, all of which would come at a cost of ZERO. Even Model 3 buyers who wanted a “cheaper” car would be enticed to pull the trigger on a post-purchase $10,000 EAP+FSD package, which is almost 1/3 the price of the car!
1. For those who prepaid for FSD as part of the purchase (do we know how many?), Tesla cannot realize that income on its balance sheet until it has delivered the product. Activating FSD would add a one-time bump to Tesla’s cash.
2. Activation of FSD would get many to buy Teslas and would get countless existing owners to pony up the $4000 to activate it after purchase, especially once they have had the opportunity to do a free trial. Elon has been waiting to activate the free trial of EAP once the “E” features became available (looks like that might happen soon). Sounds like a good time to start offering the trial. An FSD trial would certainly generate a lot of $4000 sales, all of which would come at a cost of ZERO. Even Model 3 buyers who wanted a “cheaper” car would be enticed to pull the trigger on a post-purchase $10,000 EAP+FSD package, which is almost 1/3 the price of the car!
All those FSD payments are going to give Tesla one big class action lawsuit if they stumble. Payment for undelivered product. Technology keeps improving by leaps and bounds and here we are paying for software tech that's not coming in 3 years based on last years hardware. Stupid!
Personally I love driving, so I cannot appreciate FSD.
Personally I love driving, so I cannot appreciate FSD.
Driving policy is a hard task, that Waymo tackled well with a lot of work, programmation and simulation, and that Mobileye is pushing the formalization through is RSS paper (Responsibility Sensitive Safety).
But the best driving policy worse not much if you don't have a perfect situational awareness.
Waymo and Cruise are getting there thanks to LIDAR/vision fusion + HD map. Mobileye is pushing an even more ambitious solution with HD map (collected thanks to his EyeQ4) + 2 fully independent system pure vision on one side and pure lidar on the other side, which offers the huge advantage to allow the development of incredible ADAS system based on pure cheap vision, and level 4/5 self driving with the two systems.
Even if Tesla if cracking the HD map+vision as Mobileye is on the way to do, the question will still be the lack of real redundancy for FSD. Radar is a plus, but not precise enough to confirm or infirm the vision result : is it a strange shadow or is it something real I will hit? Is it a strange pothole shape or is it a small kid lying on the road?
Vision+radar will surely make a lot of progress but there is no chance it will be as reliable as redundant vision+radar and lidar as Molbileye is pursuing.
Don't get me wrong, I'm a Tesla fan and investor that think that among the many domains Tesla have disrupted, they are huge success but also a few mistakes. And autopilot communication for me is one of them.
As long as autopilot is level 2, its a hugely convenient system (that will get even better with end to end) but very far from a safety improvement due to the necessary imperfect automation+human dimension (the better the system work, the worse the human get on average).
Getting to level 3 (on highway for instance) not even talking about FSD is a huge leap as it imply full liability.
Tesla and Musk have shown their ability to recognize mistakes : Falcon wing doors, too much complexity on the X, automation of model 3. Perhaps it would be time to admit that the FSD, autopilot safety communication was overoptimistic, and target a more realistic true level 3 on highway (which would be already a huge achievement), and focus on pilot monitoring as autopilot will be a level 2 system for still many year perhaps even forever without new material.
Tesla should implement pilot monitoring quickly, even at the price of retrofitting S and X with internal camera.
Leading computer vision technologies for aviation, automotive, rail, medical
But the best driving policy worse not much if you don't have a perfect situational awareness.
Waymo and Cruise are getting there thanks to LIDAR/vision fusion + HD map. Mobileye is pushing an even more ambitious solution with HD map (collected thanks to his EyeQ4) + 2 fully independent system pure vision on one side and pure lidar on the other side, which offers the huge advantage to allow the development of incredible ADAS system based on pure cheap vision, and level 4/5 self driving with the two systems.
Even if Tesla if cracking the HD map+vision as Mobileye is on the way to do, the question will still be the lack of real redundancy for FSD. Radar is a plus, but not precise enough to confirm or infirm the vision result : is it a strange shadow or is it something real I will hit? Is it a strange pothole shape or is it a small kid lying on the road?
Vision+radar will surely make a lot of progress but there is no chance it will be as reliable as redundant vision+radar and lidar as Molbileye is pursuing.
Don't get me wrong, I'm a Tesla fan and investor that think that among the many domains Tesla have disrupted, they are huge success but also a few mistakes. And autopilot communication for me is one of them.
As long as autopilot is level 2, its a hugely convenient system (that will get even better with end to end) but very far from a safety improvement due to the necessary imperfect automation+human dimension (the better the system work, the worse the human get on average).
Getting to level 3 (on highway for instance) not even talking about FSD is a huge leap as it imply full liability.
Tesla and Musk have shown their ability to recognize mistakes : Falcon wing doors, too much complexity on the X, automation of model 3. Perhaps it would be time to admit that the FSD, autopilot safety communication was overoptimistic, and target a more realistic true level 3 on highway (which would be already a huge achievement), and focus on pilot monitoring as autopilot will be a level 2 system for still many year perhaps even forever without new material.
Tesla should implement pilot monitoring quickly, even at the price of retrofitting S and X with internal camera.
Leading computer vision technologies for aviation, automotive, rail, medical
That's the crux of the case against the modest hardware in the current production Tesla. How will Tesla ever do a cost benefit and decide that the cars are level 3 "eyes off the road". Tesla at that point essentially becomes liable for every accident on autopilot. How does Tesla price that cost, both financial and reputationally, into the car? Is the cost of defending Model 3 level 3 lawsuits in the current gross margins quoted?
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. Siri's usefulness as an AI has been disappointing, but apples well designed and well differentiated products still do fabulously well.
I totally concur with your analysis and Siri comparaison. Its not their main advantage against competition, and they are even lagging others (Siri, map, iMessage against google assistant, Alexa, google map, Waze, WhatsApp) but it make sense for them to develop it anyway to be independent.
As FSD with current hardware has few chance to ever happen and autopilot will be level 2 for the foreseeable future, it would greatly improve safety adding driver monitoring (not very complex to achieve and better user experience with eye on, hand off) and a smart HUD with a lot of Autopilot surrounding information, an excellent way to naturally keep your eyes on the road rather than looking at the central screen.
Current autopilot technology with pilot monitoring (and HUD) would not be the expected FSD but represent an excellent system with optimized safety level (its 100% sure that pilot monitoring and HUD improve driver attention).
One domain where driving without pilot would be a huge progress and that isn't discussed much is parking. Especially in Europe we have plenty of underground parking and it would be terrific if you could get out of the car at the entrance and let the car find an available spot and park by itself. If Tesla can crack this problem, it would be great.
MelaniainLA
Member
Disagree respectfully. Watch this video from CNET from a couple weeks ago, where they took a test drive using Mobileye's "FSD" equivalent through Jerusalem highways and surface streets. You'll notice an EERY similarity between Mobileye's FSD platform and Tesla's, down to the precise camera numbers and positions. Mobileye has managed to get it to work pretty well thus far even WITHOUT LIDAR and other fancy tricks, and without Tesla's 100,000+ AP2.0 and AP2.5 vehicles (and growing, thanks to Model 3 ramp up!) to provide data to the neural net.
An autonomous trip through Jerusalem in Intel's ride - Video
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!
It’s hard for me to bet on them or against them. On the one hand, FSD is a hard problem that hasn’t yet been publicly demonstrated well in spite of lots of smart people working for years. On the other, it’s hard to believe that even Tesla in their hubris would be so overconfident as to telegraph something that was complete vaporware. To use a term Elon likes, that would be “boneheaded”.
Yeah, from a revenue standpoint, this becomes extremely important as Tesla is currently holding these deposits in its "deferred revenue" line item on its financials which becomes a liability for them until they actually deliver. I'd imagine this is a pretty large liability that they'd likely want to turn into revenue very very soon.
